Land Use/Land Cover Remote Sensing Classification in Complex Subtropical Karst Environments: Challenges, Methodological Review, and Research Frontiers

Abstract

Land use/land cover (LULC) data serve as a critical information source for understanding the complex interactions between human activities and global environmental change. The subtropical karst region, characterized by fragmented terrain, spectral confusion, topographic shadowing, and frequent cloud cover, represents one of the most challenging natural scenes for remote sensing classification. This study reviews the evolution of multi-source data acquisition (optical, SAR, LiDAR, UAV) and preprocessing strategies tailored for subtropical regions. It evaluates the applicability and limitations of various methodological frameworks, ranging from traditional approaches and GEOBIA to machine learning and deep learning. The importance of uncertainty modeling and robust accuracy assessment systems is emphasized. The study identifies four major bottlenecks: scarcity of high-quality samples, lack of scale awareness, poor model generalization, and insufficient integration of geoscientific knowledge. It suggests that future breakthroughs lie in developing remote sensing intelligent models that are driven by few samples, integrate multi-modal data, and possess strong geoscientific interpretability. The findings provide a theoretical reference for LULC information extraction and ecological monitoring in heterogeneous geomorphic regions.

1. Introduction

LULC data serve as a critical source of information for analyzing the complex interactions between human activities and global environmental change. They play an essential role in a wide range of applications, including climate change studies, agricultural monitoring, water resource management, natural disaster assessment, and land change evaluation [1,2]. High-confidence LULC data are crucial for the sustainable use of natural resources and for addressing climate change challenges [3,4], and improving their reliability is key to enhancing the accuracy of decision-making processes [5]. However, in heterogeneous Earth surface systems, generating high-accuracy LULC data remains a significant challenge. Accurate LULC remote sensing mapping has long been a research hotspot. Over the past few decades, the global scientific community has paid considerable attention to the remote sensing mapping of global land cover [6,7,8]. Yet, due to differences in data sources, classification schemes, and classification methods across various products, their suitability and accuracy in specific applications remain uncertain. In heterogeneous Earth surface systems, high-accuracy LULC (land use/land cover) data still face numerous challenges, and LULC remote sensing mapping has remained a prominent research focus. This is particularly true in karst regions, which exhibit unique geomorphological structures and ecological fragility, presenting complexities distinct from other areas. The highly fragmented terrain, poor soil conditions, and spatiotemporal heterogeneity of karst landscapes not only result in severe spectral confusion and difficulties in land cover interpretation from remote sensing imagery, but also make the dynamic changes in land use and cover more sensitive to regional ecosystem services and human activities.

Karst geomorphology is a special landform type formed by the long-term dissolution, erosion, and deposition of carbonate rocks [9]. It is widely distributed, covering about 15% of the global land surface. In China, karst areas account for one-third of the national territory. The karst region of southern China, in particular, is characterized by highly complex terrain and diverse habitats, making it a globally significant biodiversity conservation area. Meanwhile, the weathering process of carbonate rocks plays a crucial role in the global carbon cycle and climate regulation [10]. According to the results of the Fourth National Survey on Karst Rocky Desertification (KRD) conducted by the National Forestry and Grassland Administration of China, by 2021, the area of rocky desertification land in China reached 7.2232 million hectares, accounting for 14.92% of the total karst area. The expansion of rocky desertification has been comprehensively curbed, and its severity has continued to decline [11]. However, the ecosystems in these regions are inherently fragile, and rocky desertification has become one of the most prominent ecological problems, manifested in vegetation degradation, soil loss, and the decline of ecological functions, which in turn threaten food security, water resource security, and regional sustainable development. As a core representation of the relationship between humans and nature, conducting high-accuracy LULC remote sensing classification and dynamic monitoring in complex karst environments is of great significance for degradation assessment, quantifying restoration effectiveness, and supporting ecological management decision-making.

Fine-scale LULC monitoring and classification in karst regions are closely related to the global environmental governance framework, including the United Nations Sustainable Development Goals (SDGs) and the United Nations Convention to Combat Desertification (UNCCD). Among these, SDG 15 (Life on Land) emphasizes halting biodiversity loss, combating desertification, and restoring degraded land, while the UNCCD advocates achieving land degradation neutrality. Accurate LULC remote sensing mapping provides essential data support for assessing the status and trends of karst ecosystems and plays a vital role in global biodiversity conservation, sustainable land management, and climate change mitigation.

Remote sensing technology, with its unique advantages of wide coverage, rapid acquisition, and repeatable observations, has become the preferred means for obtaining large-scale LULC information [12,13]. Although existing studies have achieved substantial progress in LULC classification at the global scale, research focusing on the complex environments of karst regions remains insufficient. In particular, there has been a lack of systematic and targeted exploration on how to address challenges such as topographic shadowing, the mosaic pattern of vegetation and exposed rock, and image acquisition and interpretation under persistently cloudy conditions. These limitations not only restrict the applicability of remote sensing classification methods in karst areas but also directly affect the scientific support for land degradation monitoring, rocky desertification control, and regional ecological restoration projects. When remote sensing technologies are applied to the highly complex natural settings of subtropical karst in southern China, they encounter more severe scientific and technical challenges than in other regions, which are mainly concentrated in the following aspects (Figure 1):

(1)

At a macro scale, karst landforms are characterized by rugged terrain units such as peak clusters, depressions, and gorges, resulting in a highly fragmented surface. Land cover patches are typically small in area, irregular in shape, and have fuzzy boundaries, forming a complex landscape with a tightly interwoven mosaic of vegetation, soil, and exposed bedrock.

(2)

The phenomena of same object, different spectra and different objects, similar spectra are especially prominent in karst areas. For example, weathered carbonate rock outcrops and urban impervious surfaces such as concrete rooftops and roads exhibit very similar high reflectance in the visible and near-infrared bands, making them extremely difficult to distinguish in high-resolution imagery. Similarly, seasonally dry croplands and sparsely vegetated grasslands or shrubs often share highly similar spectral features. Even the same tree species can display drastically different spectral responses on sun-facing versus shaded slopes due to differences in illumination intensity and angle.

(3)

The dramatic topographic relief in karst regions leads to widespread terrain shadows, especially on steep mountain slopes. These shaded areas receive very weak electromagnetic signals, resulting in significant loss of land surface information and becoming a major source of classification errors.

(4)

Subtropical karst regions (such as Southwest China) are frequently shrouded in clouds and fog, and their humid, rainy climate severely limits the availability of high-quality optical remote sensing imagery. This makes it difficult to construct complete and continuous time-series datasets [14].

The four major challenges of fragmented surfaces, spectral confusion, terrain effects, and data limitations are intricately interwoven and mutually coupled in karst regions, collectively forming a significant barrier to accurate LULC classification using remote sensing in such complex natural settings. As a result, developing a high-precision LULC remote sensing classification framework tailored to the unique geomorphology of karst landscapes has become a critical issue for both ecological management and scientific decision-making. Although substantial progress has been made in the remote sensing classification of land cover in karst areas, existing studies tend to focus either on specific regions (e.g., individual watersheds or protected areas) [15,16] or on a single technical approach (e.g., deep learning alone) [17,18,19]. There remains a lack of systematic reviews addressing the common scientific challenges and methodological evolution in karst remote sensing classification. Research on emerging technologies such as UAVs and LiDAR, as well as critical issues like scale effects and uncertainty, also requires further development [20,21]. Especially in the current context of rapid advancements in artificial intelligence, a key unresolved challenge lies in how to develop intelligent remote sensing models that effectively integrate geoscientific knowledge, thereby improving model robustness and interpretability in highly heterogeneous environments.

Strengthening LULC classification research in karst regions is not only a breakthrough at the academic level but also of profound practical significance. Starting from the core challenges faced in LULC remote sensing classification, this study conducted a systematic literature search using databases such as the Web of Science Core Collection, Scopus, CNKI, and Google Scholar. The search combined both English and Chinese queries, including (“karst” AND “land use/land cover” AND “remote sensing”), (“LULC classification” AND “complex terrain” AND “karst”), (“karst spatial heterogeneity”), (“Remote sensing uncertainty”), (“Land use Land cover accuracy”), and (“Multi-scale Land cover Land use”). The search period was limited to publications before 30 June 2025. Inclusion criteria were as follows: (1) the study area involves karst or other complex geomorphic environments; (2) the research focuses on land cover/land use remote sensing classification methods or applications; and (3) the literature is published in peer-reviewed journals, conference proceedings, or book chapters. Through systematically reviewing the multi-source data support systems, typical technical pathways, evaluation frameworks, and uncertainty modeling strategies of LULC remote sensing classification in karst regions, this study focuses on critical bottlenecks such as samples, models, and scales. It further explores future directions of LULC classification in complex karst environments, aiming to provide methodological support for karst remote sensing classification as well as insights for LULC classification research in heterogeneous geomorphic regions.

2. Analysis of Karst Complex Environments and Challenges in LULC Remote Sensing Classification

2.1. Evolution of LULC Remote Sensing Classification Research in Karst Regions

The development of land cover remote sensing classification techniques in karst regions is closely linked to advances in remote sensing data acquisition, computer science, and the deepening of geographic knowledge. The evolution pathway clearly illustrates the progression from processing single-pixel spectral information, to integrating multi-dimensional features, and ultimately to leveraging deep learning models for feature extraction (

Table 1. Evolution of land cover remote sensing classification in karst regions.

Development Stage	Core Method/Model	Primary Data Sources	Key Advantages and Features	Limitations and Challenges
Pre–2000s	① Supervised/unsupervised classification: Maximum Likelihood Method, ISODATA, K-Means clustering; ② Linear Spectral Mixture Model.	Medium and low-resolution satellite imagery: Landsat MSS/TM, SPOT.	① Beginning to use remote sensing data for large-scale LULC mapping; ② Suitable for monitoring regional-scale macro-level LULC changes.	Pixel-based classification often produces “salt-and-pepper” noise, ignores spatial and textural context, and struggles to handle the fragmented and heterogeneous landscapes of karst regions.
2000s–2010s	① Traditional Machine Learning: SVM, RF; ② OBIA: Multi-scale segmentation combined with spectral, shape, texture, and contextual features for classification; ③ Polarimetric SAR Target Decomposition: Wishart-H/α and Cloude-Pottier decomposition.	① Medium- to high-resolution imagery: Landsat ETM+/OLI, ASTER; ② Fully polarimetric SAR data: RADARSAT-2, ALOS PALSAR; ③ Auxiliary data: DEM	① Significant accuracy improvement; ② Overcomes surface fragmentation; ③ Cloud penetration: SAR data can operate all-day and under all-weather conditions.	① The classification performance is highly dependent on the selection and design of manual features. ② Parameter tuning is complex. ③ Shadows in complex terrain and the spectral similarity of different land cover types remain challenging issues.
2010s–Present	① Deep Learning: Convolutional Neural Networks (CNN), U-Net, ResUNet, hybrid models (e.g., CNN-SVM); ② Multi-source Data Fusion: Integration of data from optical, LiDAR, SAR, and UAV platforms; ③ Phenological Feature Analysis.	① High-resolution imagery: Sentinel-2, Gaofen (GF) series, Planet; ② UAV imagery: Centimeter-level visible, multispectral, and hyperspectral data; ③ LiDAR: High-precision point clouds and derived products such as DEM, DSM, and CHM.	Deep learning enhances classification accuracy in complex scenarios by integrating LiDAR/UAV structural and phenological features, effectively mitigating spectral confusion and reducing manual intervention.	Deep learning relies on high-quality samples, which are difficult to obtain in karst regions; models have limited generalization; training requires high computational power; and multi-source data registration and fusion face technical challenges.

).

Remote sensing technology has evolved from early pixel-based classification to the current stage of deep learning and knowledge integration. However, land cover classification in karst regions still faces multiple challenges (Table 1):

(1) Highly fragmented surfaces and spatial heterogeneity remain fundamental obstacles. The interlaced distribution of karst landforms-such as peak clusters, depressions, caves, and KRD landscapes-results in fragmented land use patches with blurred boundaries. Complex terrain leads to prominent mixed pixel issues. For instance, spectral confusion between exposed bedrock and sparse vegetation persists, with sub-pixel decomposition techniques still showing a misclassification rate of 10–15% [22]. (2) Vertical ecological gradients are poorly modeled. Existing classification systems fail to effectively distinguish ecotonal zones such as the KRD–cropland–shrubland transitions, with classification accuracies often below 60% [23]. (3) Temporal monitoring is limited. Due to frequent cloud cover and satellite revisit constraints, some mountainous areas experience up to 40% data loss in annual time series observations [24]. (4) Acquiring high-quality training and validation samples is both costly and difficult, and machine learning models exhibit poor generalizability. For example, deep learning models trained on Guizhou data show a 12–18% drop in accuracy when transferred to karst areas in Guangxi [25]. (5) Weak coupling with ecological effect analysis. Current studies often stop at land cover transition statistics and lack in-depth quantitative assessments of impacts such as carbon sequestration capacity or hydrological cycles [26]. (6) Incomplete data-sharing mechanisms. Standardizing multi-source, heterogeneous data consumes over 70% of research time [27]. These challenges are deeply intertwined, constituting the core technical bottlenecks and urgent scientific problems to be addressed in current research.

2.2. Impact of Fragmented Surface and Spatial Heterogeneity on Classification Accuracy

As one of the most distinctive and ecologically fragile geomorphic units in the Earth’s surface system [28], karst landforms are shaped by their unique material composition, hydrogeological structures, and surface morphology, which together create distinctive responses in remote sensing imagery. Compared with other geomorphic types, karst terrain exhibits extreme complexity and irregularity at both macro and micro scales. These characteristics exert profound and multifaceted influences on remote sensing imaging processes, information extraction, and classification accuracy (

Table 2. Impacts of karst terrain complexity on remote sensing imaging and analysis.

Impact Dimension	Performance	Challenges to LULC Remote Sensing Classification
Geometry influence	Image distortion (perspective contraction, overlap)	Increases the difficulty of accurate geometric and orthorectification, affecting the precision of patch localization.
	Radar image effects (overlay, foreshadowing, shadowing)	Reduces the usability of SAR data in mountainous areas, with significant information loss.
Radiation effects	Strong terrain shadows	Severe loss of land cover information in shadowed areas is a major source of misclassification and omission errors.
	Spectral variation caused by slope aspect (same object, different spectrum)	Severely distorts the intrinsic spectral characteristics of land cover features, reducing classifier performance.
Spatial pattern impact	Surface patches are extremely fragmented and have blurred boundaries	The mixed pixel problem is extremely severe, rendering traditional pixel-based classification methods ineffective.
	High landscape pattern index (patch density, shape index)	Poses challenges to the definition of classification units (pixel vs. object) and the selection of segmentation scales.
	High intra-pixel heterogeneity	Increases the complexity and uncertainty of sub-pixel decomposition (spectral mixture analysis).

). While they provide potential sources of valuable information for remote sensing interpretation, they also pose significant challenges to conventional remote sensing technologies and methods.

At the macro scale, particularly in subtropical humid climatic zones such as Southwest China, typical karst regions are characterized by complex geomorphic assemblages including peak-cluster–depression systems, peak-forest–plains, gorges, and sinkholes. The most prominent features of these landscapes are intense terrain dissection, dramatic elevation differences, and steep slopes. Representative landforms include:

Peak-cluster–depression systems (Figure 1): These consist of undulating peak clusters (or peak forests) interspersed with enclosed or semi-enclosed depressions, forming the dominant surface morphology of the region [29]. Mountain slopes are often extremely steep, with gradients reaching 60–90 degrees in some areas, while the depression floors remain relatively flat. Such drastic terrain variations directly cause severe geometric distortions in remote sensing imagery. For optical imagery, especially from side-looking sensors, pronounced effects such as foreshortening and layover often occur, distorting the true geometric shapes and spatial relationships of surface features, and significantly increasing the difficulty of precise geometric correction [15]. For Synthetic Aperture Radar (SAR) imagery, the impact is even more severe in steep and fragmented terrain. Strong layover, foreshortening, and radar shadow effects are nearly unavoidable, substantially compromising the usability of SAR imagery in geomorphologically complex areas [30].

Sinkholes and Cave Systems (Figure 2): Sinkholes, as a type of unique negative terrain, are characterized by steep and deep depressions with extremely poor lighting conditions, remaining in shadow for most of the year [31]. Optical remote sensing sensors struggle to capture effective spectral information from within these features. Additionally, the extensive development of subsurface systems such as caves and underground rivers creates blind spots that traditional remote sensing technologies cannot directly detect. These hidden underground structures have a profound impact on surface hydrological processes and ecological stability. However, the lack of remote sensing information on sinkholes creates an inherent vertical gap in our understanding of karst ecosystems.

2.3. Mechanisms and Mitigation Strategies of Spectral Confusion

The multidimensional coupling of karst geological structures, topographic variations, hydrological processes, and human activities gives rise to multi-level spectral interference mechanisms. As a result, land surface features exhibit highly complex and confused spectral characteristics, prominently reflected in the phenomena of “same object, different spectra” and “different objects, similar spectra.” These have become one of the core obstacles limiting classification accuracy [32].

2.3.1. Different Objects, Similar Spectra: Spectral Confusion Among Distinct Land Cover Types

The phenomenon of different objects exhibiting similar spectral characteristics—referred to as different objects, or similar spectra—is one of the most direct causes of classification confusion and errors in remote sensing. In complex karst environments, such spectral confusion among different land cover types is mainly manifested in the following ways:

① Bare rocks and artificial impervious surfaces represent the most typical category of land cover confusion in karst remote sensing classification. Weathered carbonate rocks (e.g., limestone and dolomite) generally exhibit high reflectance in the visible to shortwave infrared bands [33]. Their spectral signatures closely resemble those of urban impervious surfaces such as cement pavements, concrete rooftops, and roads. In high spatial resolution imagery, the two are nearly indistinguishable based solely on spectral information and can only be differentiated with the help of spatial features such as texture, shape, and contextual relationships. For example, in karst regions with scattered settlements or urban-rural transition zones, accurately extracting built-up areas or assessing the extent of KRD becomes exceedingly difficult [34,35,36]. Studies have found that publicly available LULC datasets show relatively low accuracy in complex karst natural environments. A typical example is the 2020 CRLC-10 m, ESRI-10 m, and ESA-10 m LULC datasets, which reported overall accuracies of 55.40%, 55.26%, and 49.40%, respectively [37]. Similarly, the GlobeLand30, CCI-LC, and CGLS-LC datasets have demonstrated generally low accuracy in southern China’s karst regions, with values ranging from about 40% to 52%. Moreover, the overall accuracy of LULC data in karst areas (36.9%) is significantly lower than that in non-karst areas (51.2%) [38]. This highlights the growing demand for high-accuracy LULC datasets in such regions.

② Seasonal dry cropland vs. sparse vegetation/shrubland: During the dry season or early cultivation stages, croplands often have exposed, dry soil surfaces. Their spectral characteristics closely resemble those of sparsely vegetated grasslands, shrublands, or dormant vegetation [39,40,41]. Both exhibit low vegetation indices (e.g., NDVI), with soil background signals dominating the reflectance. This similarity makes it extremely challenging to accurately delineate cropland boundaries and monitor vegetation recovery dynamics.

③ Shadows vs. water bodies: In canyons or on shaded slopes of steep terrain, pixels within terrain shadows have very low digital number (DN) values due to weak electromagnetic signal reception [42]. The low-reflectance characteristics of shaded areas in the near-infrared band resemble those of clear water bodies, leading classifiers to frequently misidentify terrain shadows as rivers, lakes, or reservoirs [43], resulting in significant overestimation of water body extent.

2.3.2. Same Object, Different Spectra: Spectral Variability Within the Same Land Cover Type

The phenomenon of same object, different spectra, refers to the significant variation in spectral characteristics exhibited by the same type of land cover under different environmental conditions, which is especially prominent in complex karst terrains. Topographic factors are the primary drivers of this variability. Even for the same vegetation type, differences in slope aspect (e.g., sun-facing vs. shaded slopes) lead to variations in solar radiation intensity and observation geometry [44], resulting in pronounced brightness differences. Sun-facing slopes appear much brighter, while shaded slopes are noticeably darker-sometimes to the extent that spectral differences within a single aspect exceed those between different vegetation types, such as grasslands and forests-severely disrupting land cover identification [45,46]. In addition, moisture stress is a key contributor to spectral variability in karst regions. Due to shallow soils and fractured bedrock, vegetation is highly susceptible to drought. This leads to reduced leaf water content, weakened red-edge effects, and increased shortwave infrared reflectance. Nutrient deficiencies and variations in plant health further exacerbate spectral inconsistency [47].

These factors collectively result in highly inconsistent spectral responses in remote sensing imagery, even for the same species or land cover type. This significantly affects classification accuracy and necessitates the use of strategies such as topographic correction, phenological modeling, or multi-source data fusion to mitigate these effects.

2.3.3. Spectral Gradient Characteristics of Different KRD Levels

Surface types with different levels of rocky desertification are unique land cover categories in karst regions. KRD is not a single land cover type but rather a continuous process of land degradation. Internationally, it is commonly classified into different levels such as potential, slight, moderate, severe, and extremely severe, with the core criteria being vegetation cover and bedrock exposure rate [48,49]. This continuous process is reflected in remote sensing spectral space as a gradient feature.

From a remote sensing perspective, KRD landscapes can be conceptualized as a mixed system composed of three primary components: vegetation (V), soil (S), and rock (R). As the degree of desertification intensifies, the spectral response of affected areas systematically changes at the pixel or regional scale, including: (1) A decrease in the vegetation component, reflected by reduced reflectance in the near-infrared band, increased reflectance in the red band, and a downward trend in vegetation indices such as NDVI and EVI; (2) An increase in the rock component, shown as overall higher reflectance in both the visible and shortwave infrared bands, leading to higher image brightness; (3) A complex pattern in the soil component: in the early stages of degradation, soil erosion results in exposed soil and a stronger soil spectral signal; however, as degradation worsens and soil is completely eroded away, the spectral contribution from soil diminishes.

The spectral gradient characteristics of different levels of karst rocky desertification determine that there are no clear spectral boundaries for distinguishing the grades in remote sensing. Traditional “hard classification” (that is, a discrete classification method which assumes that each pixel belongs exclusively to a single land cover type) is difficult to apply accurately in this context for separating different levels. Therefore, spectral mixture analysis (SMA) based on sub-pixel decomposition, along with continuous variable inversion of vegetation cover and bedrock exposure rates, is considered a more suitable methodological approach for monitoring KRD dynamics [50,51]. Determining pure endmembers that can accurately represent the complex rock, soil, and vegetation types in karst regions remains a highly challenging task.

2.3.4. Seasonal Variability and Its Interference with Remote Sensing Data, and Improvement Methods

The karst region is strongly influenced by the monsoon climate, which significantly affects water distribution and vegetation dynamics. Precipitation and vegetation cover show pronounced seasonal contrasts, introducing seasonal noise into remote sensing data. Yang et al. reported that approximately 75% of the annual precipitation in the Guizhou karst area occurs between June and October, reflecting the strong seasonality of rainfall [52]. Hou et al. found a clear positive correlation between NDVI and precipitation during the growing season [53]. Such strong temporal fluctuations in phenology and moisture lead to substantial seasonal differences in the spectral characteristics of the same land cover type, thereby affecting the stability and accuracy of remote sensing classification. Currently, multi-temporal data fusion and phenological feature extraction strategies are commonly used to stabilize land cover spectral representation. For example, Lei et al. employed SPOT-6 multispectral and Sentinel-1 SAR imagery, using dynamic time warping (DTW) to construct multi-scale temporal similarity indices. This effectively integrated multi-source temporal features and improved classification stability [54]. In addition, to address the impact of high cloud coverage, Zhang et al., proposed the flexible spatiotemporal data fusion (FSDAF) model, which integrates Landsat and MODIS data to build continuous time series. This method successfully extracted fire scars in humid karst regions, demonstrating that multi-temporal fusion approaches can overcome seasonal and cloud-related interference [55]. SAR data can maintain spatiotemporal continuity under cloudy or rainy conditions, enhancing the ability to capture vegetation and moisture variations. Meanwhile, topographic correction and illumination normalization can further reduce spectral shifts caused by seasonal illumination differences, thereby improving the robustness of classification models.

2.4. Remote Sensing Correction Challenges of Terrain Shadow and Radiometric Distortion

Terrain shadow is one of the most prominent features in remote sensing imagery of karst regions. Shadows include not only self-shadow, caused by the obstruction of direct solar radiation by the terrain itself (e.g., hills or mountains), but also cast-shadow, where one terrain feature casts a shadow onto another (e.g., valleys or flat areas). The radiance in shadowed areas primarily originates from sky-scattered light and reflectance from adjacent surfaces. Compared to sunlit areas, the radiometric signal in shadows is significantly weaker and suffers from extremely low signal-to-noise ratios, resulting in severe information loss. This makes shadow regions a “black hole” in classification [56] and leads to several issues: (1) Severe information loss. The spectral details of ground objects in shadowed areas are greatly weakened, resulting in significantly reduced spectral differences between different objects, making effective discrimination almost impossible [57]. (2) Misclassification risks: In classification processes, shadow regions are often assigned to a separate shadow class or misclassified as low-reflectance classes such as water bodies or wetlands [58]. This not only results in the loss of LULC information within the shadowed areas but also significantly degrades the mapping accuracy and area estimation of other classes [59]. (3) Temporal variability: The extent and position of shadows vary with solar elevation and azimuth angles, which change seasonally and diurnally. This temporal variation introduces additional complexity when comparing multi-temporal imagery.

2.5. Limitations of Optical Remote Sensing Due to Weather Conditions and Temporal Discontinuity

In addition to the inherent complexity of the surface itself, atmospheric conditions and lighting variability present further challenges for remote sensing data acquisition and interpretation in karst regions. These influences are closely intertwined with terrain effects, compounding the uncertainty in data quality. The subtropical karst areas-particularly the Yunnan–Guizhou Plateau in China-fall within a humid monsoon climate zone, characterized by abundant moisture and significant orographic uplift [60], leading to persistent cloud and fog cover throughout the year. This severely restricts the availability of high-quality optical remote sensing imagery and poses a major obstacle to time-series analysis [61,62,63].

Studies have shown that in the core karst areas of southern China, the window for acquiring usable optical imagery with minimal cloud cover (less than 10%) is extremely limited-typically confined to the dry spring or autumn seasons [64,65]. These data acquisition constraints result in several critical consequences:

① Missing time-series data: The inability to build complete and regularly spaced image sequences greatly hinders studies that rely on phenological features for fine-scale classification or high-frequency dynamic monitoring. ② Difficulties in large-scale mapping: Obtaining a cloud-free, seamless mosaic image that covers the entire study area in a single time frame is often impossible, requiring data collected across multiple months or even years. This introduces spectral inconsistencies due to changes in land cover between acquisition times, undermining data uniformity. ③ Confusion between clouds and high-reflectance surfaces: Thin clouds or fog can be spectrally similar to bright features such as exposed carbonate rocks or urban structures. As a result, automated cloud detection algorithms face considerable challenges in karst areas, frequently resulting in false detections or omissions.

In addition, the scattering of solar radiation by atmospheric molecules and aerosols reduces the radiative transfer contrast between the Earth’s surface and the sensor, resulting in images that appear hazy or washed out [66]. Multiple reflections and scatterings among steep terrain surfaces can cause spectral spillover from high-reflectance features (such as bare rocks) to adjacent low-reflectance features (such as vegetation), contaminating the spectral purity of target pixels. Although atmospheric correction models can partially mitigate these effects, achieving high correction accuracy in areas with rugged terrain remains a significant challenge.

2.6. Remote Sensing Identification Features and Challenges of KRD Land Cover Types

Although the classification standards for different levels of KRD [67] are based on field-surveyed vegetation coverage and bedrock exposure rates, in remote sensing imagery these levels are reflected through a combination of surface components, texture, and spatial pattern characteristics (

Table 3. Remote sensing identification characteristics of different KRD levels.

KRD Level	Potential KRD	Light KRD	Moderate KRD	Severe/Extreme KRD
Features	Surface vegetation coverage is relatively high (typically >70%), and bedrock exposure rate is very low (<10%).	Vegetation coverage begins to decline (50–70%), and bedrock exposure rate increases (10–30%).	Vegetation degradation is evident, with coverage relatively low (30–50%), and bedrock exposure rate significantly increased (30–50%).	Vegetation coverage is extremely low (<30%), with extensive bedrock exposure (>50%).
Image characteristics	Vegetation exhibits spectral characteristics similar to those of healthy ecosystems, but with more complex texture or more sensitive responses in vegetation indices under drought stress.	Spectral characteristics show a mixture of vegetation, soil, and rock components, with increasingly complex surface textures.	Spectral signals from rock and soil begin to dominate, displaying grayish-white or yellow-brown tones and forming fragmented patches.	Characterized by large areas of bright tones, with spectral features resembling pure rock; vegetation signals are very weak, and the landscape appears predominantly white.
Real pictures
RS images

).

The complexity of karst terrain, surface fragmentation, spectral confusion, and climatic variability pose significant obstacles to the application of remote sensing technologies. These challenges also form the foundation and prerequisites for developing targeted remote sensing classification methods and data processing techniques for land cover in karst natural environments. The difficulties in remote sensing-based identification of KRD include:

(1) The core of KRD monitoring lies in accurately estimating the proportions of vegetation, soil, and rock components. However, the accuracy of these estimates is highly dependent on the spatial resolution of remote sensing imagery. Land cover aggregation patterns vary at different spatial scales, leading to different levels of mixed pixel effects [68]. Spectral mixture analysis is extremely sensitive to the selection of endmembers, and the high spectral variability of land surface features in karst regions makes it difficult to obtain pure and representative endmembers [69], resulting in considerable uncertainty in decomposition outcomes.

(2) Field-based classification standards for KRD (e.g., 30% bedrock exposure) are defined by clear thresholds, but in remote sensing imagery, the spectral characteristics corresponding to these thresholds are often ambiguous and gradual. Establishing a quantitative relationship between field survey standards and remotely observable features is a key bottleneck to achieving automated, high-accuracy mapping.

(3) Spectral confusion between bare rock and impervious surfaces, as well as between drylands and sparse vegetation, directly affects the identification of core indicators of KRD, such as bedrock exposure and vegetation degradation [70,71,72].

(4) Difficulties in the simultaneous monitoring of “rock” and “desertification.” “Rock” mainly refers to the physical properties of exposed bedrock (spectral reflectance), which can be relatively easily identified by remote sensing. In contrast, “desertification” involves declines in land productivity, soil moisture, and soil quality—ecological functions that usually require indirect estimation through vegetation indices (NDVI, EVI), soil moisture inversion models, or integrated ecological quality assessment systems. This separation of “rock” and “desertification” in remote sensing responses is particularly typical in karst landscapes. In arid and semi-arid sandy regions, however, desertification processes are often characterized by sustained decreases in vegetation cover, expansion of bare land or sand dunes, and rising land surface temperatures, which provide clearer remote sensing indicators and more distinct temporal trajectories. Therefore, the difficulty of remote sensing monitoring of karst rocky desertification arises not only from its complex eco-geomorphic coupling, but also from the indirect and weakly responsive nature of its ecological degradation signals, which urgently requires multi-source data integration and ecological process modeling for support.

3. Technical Pathways and Applicability Assessment of LULC Classification in Karst Regions

3.1. Multi-Source Remote Sensing Data Systems

The selection and optimization of remote sensing-based land cover classification methods directly determine the accuracy and usability of LULC mapping products. Given the numerous challenges in classifying land cover in complex karst environments, choosing appropriate data sources and applying targeted preprocessing techniques are essential prerequisites and critical steps for improving classification accuracy. In recent years, with the advancement of remote sensing observation technologies, ranging from traditional optical and microwave sensors to cutting-edge LiDAR and UAV-based remote sensing, an Air–Space–Ground Integrated Observation System (ASG-IOS) has been established. By integrating multi-platform, multi-sensor, and multi-scale data sources, this system provides more comprehensive observational support and a stronger data foundation for improving the accuracy of LULC remote sensing classification in complex karst environments.

3.1.1. Comparison of Optical Remote Sensing Data Sources for Applicability

Optical remote sensing, which captures the reflectance and emission of electromagnetic radiation from land surface features, provides rich spectral information and remains the most widely used and fundamental data source for land cover classification [73,74]. Based on spatial resolution, optical data can be broadly categorized into medium-resolution and high-resolution imagery (

Table 4. Applicability comparison of typical optical remote sensing data for LULC classification.

Data Type	Represents Satellite/Sensor	Spatial Resolution	Band Range	Core Advantages	Main Challenges	Typical Application Scenarios
Medium Resolution	Landsat series (Landsat 8/9 OLI, TM/ETM+) [75] (NASA, Washington, DC, USA)	30 m	0.45–2.35 μm (visible light-shortwave infrared)	Longest global time series (>40 years), wide coverage, and freely accessible.	Severe mixed-pixel problem, making fragmented land cover types difficult to distinguish.	Large-scale LUCC monitoring and dynamic assessment of rocky desertification
	Sentinel-2 MSI [76] (ESA, Paris, France/Airbus Defence and Space, Toulouse, France)	10–20 m	0.443–2.19 μm (including red edge)	Fast revisit (5 days), with red-edge bands sensitive to vegetation health.	Significant cloud and fog interference, leading to missing time-series data in mountainous areas.	Vegetation classification, ecological monitoring, and stress diagnosis
	Gaofen series (GF-1/6 WFV) [77] (CAST, Beijing, China)	16 m	0.45–0.89 μm (including red edge)	Dense observations in China, with extensive historical archive data.	Medium resolution, with persistent mixed-pixel issues.	Regional-scale LUCC monitoring and agricultural monitoring
High-Resolution	WorldView-3/4, GeoEye-1, Pleiades [78] (Maxar Technologies, Westminster, CO, USA/Toulouse, France)	0.3–2 m	0.04–1.0 μm (including near-infrared, partially including SWIR)	Rich detail, suitable for object-based classification.	High cost, limited coverage, and strong shadow effects.	Extraction of urban built-up areas and identification of small patches of rocky desertification
	Gaofen-2 (GF-2) [79] (CAST, Beijing, China)	0.8–4 m	0.45–0.89 μm (visible–near-infrared)	Well-suited for fine-scale applications.	Tasking and cost constraints, with very large data volumes.	Fine-scale LULC mapping

).

Medium-resolution remote sensing data serve as the foundational resource for large-scale dynamic monitoring in karst regions [80,81]. The Landsat series, as the world’s longest-running remote sensing program, offers 30 m spatial resolution and a 16-day revisit cycle. It has been extensively applied in studies of LULC change, vegetation restoration, and KRD dynamics in karst areas [82,83,84]. Notably, its shortwave infrared (SWIR) bands are sensitive to rock and soil moisture, providing crucial support for monitoring KRD. However, its resolution is still inadequate for highly fragmented landscapes, and the mixed pixel problem remains significant. Sentinel-2, with 10–20 m multispectral resolution and a 5-day revisit period, includes red-edge bands sensitive to vegetation chlorophyll content, health, and nitrogen levels [85], enabling more refined vegetation classification and stress detection-an advantage particularly relevant in karst environments. The Gaofen (GF) series, such as GF-1 and GF-6 with WFV (Wide Field of View) sensors, offers 16 m resolution and red-edge capability, further enhancing continuous monitoring and classification potential in karst regions [86,87,88].

High-resolution remote sensing data are powerful tools for fine-scale LULC mapping [89,90]. High-resolution imagery such as WorldView, GeoEye, Pleiades, Jilin-1, and Gaofen-2 provides rich information on land surface details, textures, and spatial structures, serving as important data sources for fine-scale LULC mapping, object-based analysis, and model validation. In karst regions, high-resolution imagery can effectively identify micro-patches such as stone outcrops and small farmland plots, and improve classification accuracy through spatial features such as texture, shape, and shadows [91]. At the same time, high-precision imagery can also serve as ground truth data to support accuracy validation of medium- and low-resolution classification results. However, while higher spatial resolution enhances detail representation, it also intensifies intra-class spectral heterogeneity, increasing uncertainty in pixel-level classification [92,93]. The high cost of high-resolution imagery limits its widespread application in large-scale and long-term studies. Moreover, the swath width of a single scene is typically only several to several tens of kilometers, for example, WorldView-3 has a swath width of 13.2 km. Data acquisition is further constrained by tasking and weather conditions, which limits its continuous use in regional-scale and long-term studies. Therefore, although high-resolution imagery shows significant advantages in fine-scale land surface mapping, its application still requires a trade-off between data cost, acquisition timeliness, and processing strategies.

3.1.2. The Critical Complementary Role of SAR in Karst Environments

SAR generates images by transmitting microwaves to the Earth’s surface and receiving the backscattered signals [94] (

Table 5. Comparison of the applicability of SAR data in LULC classification.

Band Type	Representative Sensor	Spatial Resolution	Core Advantages	Main Challenges	Typical Application Scenarios
C-band (~5.6 cm) [103]	Sentinel-1 (ESA, Toulouse, France), RADARSAT-2 (CSA, Richmond, BC, Canada)	3–30 m	All-weather and all-day observation, with sensitivity to soil moisture and surface roughness.	Limited penetration, with layover and geometric distortions in mountainous areas.	LULC monitoring in cloud-prone areas, as well as agricultural and wetland classification.
L-band (~23.5 cm) [104]	ALOS PALSAR/PALSAR-2 (JAXA, Tokyo & Kamakura, Japan)	10–100 m	Strong vegetation penetration, suitable for forest structure and biomass estimation.	Restricted temporal acquisition and very large data volumes.	Forest/grassland classification and rocky desertification monitoring.
X-band (~3 cm) [105]	TerraSAR-X (DLR, Friedrichshafen, Germany), COSMO-SkyMed (ASI, Rome & Turin, Italy)	1–10 m	High resolution, suitable for detailed mapping of urban areas and infrastructure.	Weak penetration, with limited sensitivity to soil and vegetation.	Identification of urban built-up areas and classification of urban–rural transitional zones.
Multi-polarization/interferometric SAR [106]	Sentinel-1 (InSAR/ESA, Toulouse, France, ALOS PALSAR (PolSAR) JAXA, Tokyo & Kamakura, Japan)	5–30 m	Provision of scattering mechanisms and deformation information, supporting surface deformation monitoring and vegetation structure discrimination.	Complex algorithms requiring high computational resources.	Forest type discrimination, terrain change monitoring, and rocky desertification monitoring.

). Its core advantages are as follows: (1) All-weather, all-day imaging. The microwaves used by SAR can penetrate clouds, fog, haze, and a certain amount of vegetation. Its ability to acquire images under all-weather and all-day conditions makes it a perfect complement to address the problem of missing optical imagery in karst regions [95]. (2) Sensitivity to surface structure and roughness. The intensity of SAR backscattering signals mainly depends on the geometry of the surface (slope, aspect), roughness, and dielectric constant (primarily influenced by moisture). It is more sensitive to the physical structure of surface features than to their chemical composition [96]. For example, forest stands, shrublands, and grasslands with very different structural characteristics, even when showing similar spectral features in optical imagery, exhibit markedly different backscattering intensities in SAR imagery due to roughness differences (typically forest > shrubland > grassland) [97]. In addition, high-resolution optical images can also differentiate forests, shrubs, and grasslands using texture features such as canopy density, leaf size, and texture roughness [98,99,100], indicating that SAR and optical data are complementary in representing structural information. (3) Sensitivity to moisture. The dielectric constant of surface features is closely related to their water content. SAR signals are highly sensitive to soil moisture and vegetation water content [101,102], which makes SAR suitable for monitoring drought conditions and vegetation water stress in karst regions.

SAR data have irreplaceable advantages in LULC classification in complex karst environments [107]. Incorporating SAR backscatter intensity and texture features together with optical spectral bands and vegetation indices as input features for classifiers can significantly improve classification accuracy [108], particularly in distinguishing different vegetation structural types and filling in missing information in cloud- and shadow-affected areas of optical imagery [109]. However, in mountainous karst regions with strong topographic relief, SAR imaging is prone to geometric distortions, including layover, foreshortening, and radar shadow, which cause distortions in the spatial position and morphology of surface features, reducing classification accuracy and even resulting in the loss of information in certain areas [110]. SAR signals are highly sensitive to surface roughness and soil moisture; therefore, in subtropical karst regions where precipitation is concentrated in the rainy season and soil moisture content fluctuates drastically, the backscatter characteristics of surface features vary significantly with seasonal and weather conditions, affecting the consistency of multi-temporal data. Previous studies have shown that in environments with pronounced moisture variation, SAR backscatter intensity and phase exhibit strong responses to groundwater levels and soil moisture content [111]. Wagner et al. (1999) further noted that the seasonal variation in soil moisture strongly influences SAR backscatter, directly affecting the stability and consistency of remote sensing classification algorithms [112].

3.1.3. Light Detection and Ranging (LiDAR) Data

LiDAR uses laser pulses to measure return time in order to obtain precise three-dimensional coordinates of the Earth’s surface, providing a revolutionary tool for the high-accuracy characterization of complex topography and vegetation structures in karst regions [113] (Table 6). Its core advantages are reflected in two main aspects:

Ground surface characterization through canopy penetration. LiDAR can penetrate vegetation canopies to capture ground point clouds, generating high-resolution Digital Elevation Models (DEM) and Digital Surface Models (DSM). These datasets provide high-quality inputs for terrain correction, hydrological analysis, slope and aspect calculation, and geomorphological feature extraction [114,115], significantly improving the accuracy of terrain modeling in karst landscapes.

Acquisition of three-dimensional vegetation structure. By analyzing multiple return signals (e.g., first and last returns), LiDAR can accurately retrieve key three-dimensional structural parameters such as canopy height, vertical structural density, leaf area index (LAI), and biomass [116]. Among these, canopy height is the core indicator for distinguishing trees, shrubs, and herbaceous vegetation. This is particularly useful in karst regions where vertical structural differences among vegetation communities are pronounced, thereby enhancing classification models’ ability to discriminate between vegetation types [117] and compensating for the limitations of traditional spectral remote sensing in vegetation type identification.

Table 6. Comparison of the applicability of LiDAR data in LULC classification.

Data Type	Representative Platform/Sensor	Spatial Resolution	Advantages	Limitation	Typical Applications
Satellite LiDAR [118]	ICESat-2 ATLAS, GEDI (GSFC, Greenbelt, MD, USA)	10–25 m	Good global-scale consistency, suitable for large-scale vegetation monitoring.	Sparse spatial sampling, unable to generate continuous coverage.	Biomass estimation and classification of forest vertical structures
UAV-LiDAR [119]	RIEGL miniVUX (RIEG, Horn, Austria), DJI Zenmuse L1 (DJI, Shenzhen, Guangdong, China)	0.02–0.5 m (point cloud at centimeter level)	High flexibility, capable of accurately obtaining three-dimensional terrain and vegetation structure, making it suitable for detailed local classification.	Limited by flight altitude and endurance, high cost with restricted data coverage.	DEM/DSM extraction, urban built-up area classification, and karst micro-landform identification
Terrestrial LiDAR (TLS) [120]	RIEGL VZ Series (RIEG, Horn, Austria), Leica ScanStation (RIEG, Horn, Austria, Heerbrugg, St. Gallen, Switzerland)	Millimeter to centimeter level	Extremely high accuracy, suitable for validation and sample construction.	Slow data acquisition, with limited scene coverage.	Vegetation vertical structure measurement and classification sample validation

Table 6. Comparison of the applicability of LiDAR data in LULC classification.

Data Type	Representative Platform/Sensor	Spatial Resolution	Advantages	Limitation	Typical Applications
Satellite LiDAR [118]	ICESat-2 ATLAS, GEDI (GSFC, Greenbelt, MD, USA)	10–25 m	Good global-scale consistency, suitable for large-scale vegetation monitoring.	Sparse spatial sampling, unable to generate continuous coverage.	Biomass estimation and classification of forest vertical structures
UAV-LiDAR [119]	RIEGL miniVUX (RIEG, Horn, Austria), DJI Zenmuse L1 (DJI, Shenzhen, Guangdong, China)	0.02–0.5 m (point cloud at centimeter level)	High flexibility, capable of accurately obtaining three-dimensional terrain and vegetation structure, making it suitable for detailed local classification.	Limited by flight altitude and endurance, high cost with restricted data coverage.	DEM/DSM extraction, urban built-up area classification, and karst micro-landform identification
Terrestrial LiDAR (TLS) [120]	RIEGL VZ Series (RIEG, Horn, Austria), Leica ScanStation (RIEG, Horn, Austria, Heerbrugg, St. Gallen, Switzerland)	Millimeter to centimeter level	Extremely high accuracy, suitable for validation and sample construction.	Slow data acquisition, with limited scene coverage.	Vegetation vertical structure measurement and classification sample validation

Although airborne LiDAR is relatively expensive and typically used for small-scale scientific studies or key area surveys, its value in land cover classification is substantial. First, it serves as an auxiliary data source, with LiDAR-derived terrain factors and vegetation structure parameters integrated into classification models alongside optical or SAR data [121]. Second, LiDAR point clouds can be directly used for 3D classification based on their geometric and reflectance attributes, enabling effective differentiation of ground surfaces, buildings, and multi-layer vegetation types [122], thus expanding the spatial dimensionality and accuracy potential of remote sensing-based classification.

3.1.4. UAV-Based Multi-Source Remote Sensing

As an emerging low-altitude remote sensing platform, unmanned aerial vehicle remote sensing (UAV-RS) offers ultra-high spatial resolution and flexible data acquisition capabilities, providing critical support for refined LULC classification in karst regions [123] (

Table 7. Comparison of the applicability of UAV multi-source remote sensing data in LULC classification.

Data Type	Representative Sensor/Platform	Spatial Resolution	Advantages	Limitation	Typical Applications
Visible light RGB imagery [125]	DJI Phantom and Mavic series (DJI, Shenzhen, China)	2–10 cm	Low cost and easy acquisition, suitable for rapid mapping.	Limited spectral information, easily affected by illumination.	Extraction of urban built-up areas and detailed classification of farmland
Multispectral imagery [126]	Parrot Sequoia (Parrot Drones, Paris, France), MicaSense RedEdge (MicaSense, Seattle, WA, USA)	5–20 cm	Provides vegetation indices (NDVI, EVI, etc.), suitable for agricultural and ecological monitoring.	Limited number of bands, with complex radiometric calibration.	Crop growth monitoring and forest cover classification
Hyperspectral imagery [127]	Cubert UHD (Cubert GmbH, Ulm, Germany), Headwall Nano-Hyperspec (Headwall Photonics, Bolton, MA, USA)	5–30 cm	High spectral resolution, capable of distinguishing spectrally similar but different objects.	Large data volume and complex processing, with high equipment cost.	Detailed vegetation type differentiation and mineral identification
Thermal infrared imagery [128]	FLIR Vue(Teledyne FLIR, Wilsonville, OR, USA), Workswell WIRIS (Workswell, Prague, Czech Republic)	10–50 cm	Provides land surface temperature information, suitable for water body and wetland monitoring.	Relatively low spatial resolution, highly influenced by climate.	Urban heat island studies and soil moisture inversion
UAV-LiDAR [128]	DJI Zenmuse L1 (DJI, Shenzhen, China), RIEGL miniVUX (RIEGL, Horn, Austria)	2–10 cm	High-accuracy three-dimensional point clouds, supporting terrain and structural analysis.	High cost and demanding data processing requirements.	Forest vertical structure analysis and classification of urban–rural built-up areas

). UAVs can capture centimeter-level high-resolution imagery, allowing clear identification of microtopographic features, individual vegetation structures, and rock fractures [124], effectively addressing the mixed-pixel problem common in traditional remote sensing. Moreover, UAVs are not constrained by satellite orbits and can fly on demand to meet the needs of high-frequency and time-sensitive data collection.

UAVs also support multi-payload platforms, capable of carrying various sensors such as RGB, multispectral, hyperspectral, thermal infrared, and miniaturized LiDAR systems. This enables the acquisition of multimodal data with high spatial and temporal resolution, greatly enhancing the dimensionality and depth of remote sensing analysis [129,130]. In practice, UAVs have been widely used for ultra-fine LULC classification at watershed or plot scales, identification of KRD patches, and interpretation of vegetation communities. In addition, UAV imagery serves as a high-quality source for training and validating LULC classification models (

Table 8. Land cover sample images from a karst peak-cluster depression area.

Land Type	I	II	III
Arable Land
Weeds
Bare Soil
Woodland
Shrub
Water
Impervious Surface
Asphalt Road
Bare Rock

Note: The bare rock sample image was taken on 4 June 2023, while the other land cover sample images were taken on 20 February 2020. Land type refers to land cover types, and I, II, and III are screenshots showing different states of each land type.

), offering greater efficiency and spatial representativeness than traditional ground-based sampling [131].

In addition, UAVs play a key bridging role in the ASG-IOS. They can be used for ground validation of airborne and satellite imagery, and their high-precision observations can also be scaled up to connect and complement multi-scale remote sensing data [132]. Studies have shown that combining UAV imagery with machine learning and deep learning models has enabled high-accuracy classification in complex environments such as karst wetlands [133].

Although LiDAR and UAV platforms offer high accuracy and multi-dimensional information acquisition advantages for LULC classification in complex karst landscapes, their application in developing countries faces dual constraints of economic and technical conditions. LiDAR systems are costly to procure and operate, and their data processing requires high-performance computing platforms and skilled professionals. UAV data acquisition is restricted by flight regulations, climatic conditions (especially frequent cloudy and rainy weather in tropical and subtropical karst regions), and operational range. In developing countries or regions, high-precision airborne or ground survey data are relatively scarce, and research often relies on free medium- to high-resolution satellite data (such as the Landsat and Sentinel series). Therefore, under financial constraints, integrating open-source satellite imagery with low-cost UAV data, combined with crowdsourced ground observations, represents an optimal strategy for building sustainable and locally adaptable data acquisition solutions.

3.2. Review of Classification Methods

3.2.1. Pixel-Based Classification Methods: From Traditional Statistics to Improved Logic

Pixel-based classification methods represent a classical approach widely used in the early stages of remote sensing image interpretation. These methods are based on the fundamental assumption that each pixel is internally homogeneous and can be treated as an independent analysis unit [134,135,136]. In the early development of remote sensing research, such methods were commonly employed with medium-resolution imagery like Landsat to produce land cover maps of karst regions.

However, the highly complex and variable surface environments in karst landscapes present serious challenges to pixel-based methods, primarily including: (1) High sensitivity to spectral heterogeneity: Karst areas are prone to “same object, different spectra” and “different objects, similar spectra” phenomena. For instance, terrain-induced shadowing can cause significant spectral differences in the same vegetation type depending on slope aspect, and spectral similarities between bare rock and urban impervious surfaces often lead to misclassifications [137,138]. (2) Inability to utilize spatial context: The assumption of pixel independence prevents these methods from considering spatial continuity and neighborhood relationships, often resulting in salt-and-pepper noise and fragmented patches [139,140]. (3) Mismatch between image resolution and object scale: Karst landscapes are characterized by surface fragmentation and numerous micro-patches. The resolution of remote sensing imagery often does not align with the scale of land cover features, and a single pixel frequently contains multiple surface components. Hard classification, which assigns each pixel to a single class, leads to frequent misclassifications and omissions.

Thus, although pixel-based methods hold historical significance in remote sensing applications, they exhibit clear limitations in dealing with the complex geomorphology and ecological heterogeneity of karst regions.

3.2.2. GEOBIA: Workflow and Suitability in Karst Environments

To overcome the limitations of pixel-based approaches, Geographic Object-Based Image Analysis (GEOBIA or OBIA) has emerged as a powerful alternative. However, the application of GEOBIA in the complex environments of karst regions faces several challenges: (1) Segmentation quality is the key to the success of GEOBIA, and the results are highly sensitive to the setting of the Scale Parameter [141]. Karst surfaces exhibit multi-scale spatial heterogeneity, and the optimal segmentation scales vary significantly among different land cover types, such as large forested areas, strip-shaped rivers, and scattered croplands. LULC data are strongly affected by scale effects (

Table 9. Land cover characteristics at different grid sizes (Autumn).

Land Type	30 m	15 m	10 m	5 m	2 m	1 m
Cultivated field
Grassland
Woodland
Water
Construction land
Bare rock

). A typical example is that information on built-up land is severely lost as the scale becomes coarser [142]. (2) GEOBIA can generate a massive number of candidate features. Selecting the optimal feature combination and constructing effective classification rules—such as through decision trees or fuzzy logic—requires extensive expert knowledge and iterative experimentation. This process is often cumbersome and subjective [143,144].

3.2.3. Applicability of Machine Learning Methods in Karst Regions

The development of machine learning (ML) techniques, particularly the emergence of advanced classification algorithms, has provided more powerful and intelligent tools for LULC remote sensing classification. In recent years, deep learning (DL) technologies—centered around convolutional neural networks (CNNs)—have brought revolutionary advances to remote sensing image interpretation [145,146] (

Table 10. Applications of deep learning models in remote sensing classification of complex landforms.

Model Name	Fusion Data Types	Scenario Application/Features
U-Net [147]	RGB/Multispectral + DSM/NDVI	Pixel-level semantic segmentation with a simple architecture, suitable for small-scale vegetation classification in karst regions.
PSPNet [148]	Multispectral + DSM + NDVI	Extracts global contextual features, making it suitable for identifying scattered vegetation patches in fragmented karst landscapes.
DeepLabV3+ [149]	Sentinel-2/Gaofen data + DSM	Atrous (dilated) convolution enhances the ability to capture multi-scale land feature structures, making it well-suited for complex and undulating terrain.
FCN [150]	High-resolution imagery	The earliest end-to-end semantic segmentation framework, suitable for fundamental research and comparative experiments.
SegNet [151]	Multispectral + DSM	Encoder–decoder architecture, suitable for segmenting land features with clear boundaries (e.g., shrubs, bare rock).
ResUNet [152]	RGB + DSM + Spectral Indices	Combines U-Net with residual connection architecture, making it suitable for handling scenes with significant terrain interference.

).

Deep learning typically appears in the form of semantic segmentation, enabling end-to-end, pixel-level prediction from imagery to classification maps [153]. The core advantage of deep learning lies in its ability to perform automated, hierarchical feature learning. Unlike GEOBIA, which requires hand-crafted features, deep learning models can automatically learn features from raw images-ranging from low-level edge and texture features to high-level, abstract semantic representations [78]. These models possess strong spatial contextual modeling capabilities; for example, architectures like U-Net and PSPNet, especially when combined with auxiliary data such as digital surface models (DSMs), enable fine-scale identification of karst vegetation communities [154]. Moreover, multi-model fusion strategies-such as combining the results of single-class and multi-class CNNs-can further enhance the robustness of classification results.

Deep learning models show a clear dependence on data type, and different types of remote sensing data have significant differences in their characteristics. For example, Liu et al. applied GF-2 optical imagery with the U-Net model for LULC classification in complex surface environments. Using only spectral channels, they achieved an overall accuracy of about 86.3%, which further improved to around 93.6% when texture, contrast, and NDVI features were incorporated [155]. However, in SAR imagery with prominent speckle noise, the noise significantly affects image interpretation and classification results, and despeckling techniques are necessary to maintain classification accuracy [156]. Related studies have shown that specialized algorithms for polarimetric SAR (such as Cloude-Pottier H/α decomposition and H/A/α-Wishart classifiers) demonstrate high robustness in processing polarimetric data, particularly offering reliable performance in classifying complex surface types [157,158]. Furthermore, combining polarimetric scattering mechanism features with texture features, and then applying machine learning methods (such as MCSM+SVM or RF+SFS optimization workflows), can further improve classification accuracy and robustness [158,159]. Cross-modal fusion of SAR polarimetric decomposition features with CNN deep semantic features (e.g., using cross-attention mechanisms) can enhance SAR classification accuracy to a level comparable to optical imagery while preserving optical performance [160]. Therefore, in practical applications, algorithms should be selected or improved based on the characteristics of the data source: CNN-based methods are preferable for optical data; polarimetric decomposition combined with machine learning is suitable for SAR data; and cross-modal optical–SAR models are more appropriate for cloudy regions. In addition, statistical learning methods such as Random Forest (RF) and Support Vector Machines (SVM) demonstrate good adaptability to medium- and low-resolution optical imagery and SAR data, showing stable performance especially in situations with limited samples and high data noise.

The widespread application of deep learning in LULC remote sensing classification in karst regions still faces two core challenges. (1) The performance of deep learning models relies heavily on massive and finely annotated training samples [161]. However, obtaining large-scale, pixel-level, and accurate ground truth labels in karst regions is extremely costly and exceptionally difficult. The scarcity of samples remains the primary bottleneck restricting the application of deep learning. (2) The issue of model generalization. Since different karst regions vary in terrain, climate, vegetation, and lithology, improving the generalization and transferability of models for LULC remote sensing classification so that one model trained in one region can be effectively applied to other regions is both a research hotspot and a major challenge [90,162]. To address these problems, researchers are actively exploring techniques such as self-supervised learning, weakly supervised learning, transfer learning, and domain adaptation [163], with the aim of reducing reliance on annotated samples and building pre-trained LULC foundation models suitable for karst regions.

Different machine learning methods show significant differences between accuracy and interpretability, and decision-makers often need to balance the requirements for both. Random Forest (RF) and Support Vector Machines (SVM) can directly quantify the contribution of variables through feature importance indicators (such as the Gini index and weight vectors), which helps ecologists understand the roles of slope, NDVI, SAR backscatter, and other factors in distinguishing rocky desertification grades. However, they are limited in capturing complex nonlinear textures [164]. For example, when dealing with highly nonlinear and spatially complex high-resolution imagery, the accuracy of RF is usually lower than that of deep convolutional neural networks (deep CNNs). In contrast, deep CNNs, which rely on automatic feature extraction and end-to-end training, often achieve higher accuracy in high-resolution and multi-source data fusion scenarios. For instance, in UAV imagery experiments with a spatial resolution of 0.3 m, the overall accuracy (OA) can be increased to over 90% [165]. Nevertheless, the network structure of CNNs is highly abstract, and their decision-making logic can only be partially explained using post hoc interpretation tools such as SHAP and Grad-CAM [166,167]. This results in weaker interpretability, and CNNs also require large numbers of labeled samples and substantial computing power. Interpretable CNN frameworks have been proposed that incorporate physical constraints into the encoding stage (such as terrain correction layers and phenological time-series gating), so that the hidden parameters of the network correspond directly to ecological processes. By combining lightweight RF to perform post-classification correction on CNN outputs, the transparency of feature contribution can be improved by about 30% [168]. For ecological applications, the choice between accuracy and interpretability should be guided by the research objective. For studies requiring explicit identification of ecological driving mechanisms, models with stronger interpretability should be prioritized, whereas for tasks focusing on mapping accuracy, deep learning methods may be adopted, supplemented by visualization and interpretability techniques (such as Grad-CAM and SHAP) to assist the analysis.

3.2.4. Comparison of the Performance, Cost, and Applicability of Classification Methods

In complex karst natural environments, land cover remote sensing classification methods differ not only in classification accuracy but also show significant variations in computational cost, sample requirements, and production efficiency. Therefore, we conducted a comprehensive comparison of commonly used classification methods in LULC remote sensing studies (

Table 11. Analysis of the accuracy, cost, and applicability of commonly used land cover classification methods.

Method	Accuracy (OA/Kappa)	Training Cost (Computation and Sample Requirements)	Applicability (Data Type/Scenario)	Typical References
RF	75–88% (medium-resolution imagery)	Low (low sample requirements, fast training speed)	Optical and SAR; suitable for cases with limited samples	Belgiu and Drăguţ, 2016 [169]
SVM	76–85% (performance varies significantly across different kernel functions)	Moderate (requires parameter tuning)	Optical; performs better for classes with clear boundaries	Maxwell et al., 2018 [164]
CNN	82–90% (high-resolution optical imagery)	High (requires GPU and a large number of samples)	High-resolution optical; effective for classification of urban and vegetation areas	Zhu et al., 2017 [165]
U-Net	85–92% (very high-resolution urban or agricultural imagery)	Very high (requires large-scale samples and high memory capacity)	High-resolution optical; especially suitable for complex landscapes	Ma et al., 2019 [78]
SAR methods (H/A/α decomposition, Wishart, etc.)	70–83% (polarimetric SAR)	Moderate (requires polarimetric data and feature engineering)	Cloudy and rainy regions; suitable for karst mountainous areas	Verma et al., 2023 [170]
Ensemble methods (Expert rules + ML)	80–88% (more robust when samples are limited)	Moderate (requires domain knowledge support)	Karst mountainous areas; requires integration of geological and vegetation prior knowledge	Foody, 2020 [171]

).

In terms of classification accuracy, deep learning methods (e.g., U-Net, CNN) applied to high-resolution optical imagery generally achieve an overall accuracy (OA) of 82–92% [165], with Kappa coefficients exceeding 0.80, which is significantly higher than that of traditional machine learning methods. Their main advantage lies in the ability to automatically extract high-level features, effectively mitigating confusion caused by the “same object with different spectra” phenomenon. However, the generalization ability of deep learning outside the training area is relatively limited, and accuracy may decrease by more than 10% when applied across regions. In contrast, Random Forest (RF) and Support Vector Machine (SVM) usually maintain accuracies of 75–85% on medium- to high-resolution imagery, but they perform more robustly under conditions of limited data volume or insufficient samples [164,169]. SAR polarimetric decomposition methods can maintain accuracies of 70–83% in cloudy and foggy regions, compensating for the limitations of optical imagery, while ensemble learning methods (e.g., combining expert rules with machine learning) remain stable in cases of insufficient samples or ambiguous land cover classes, typically achieving accuracies of 80–88% [170].

In terms of cost investment, deep learning methods show the highest dependence on large volumes of high-quality training samples and demand high-performance computing resources such as GPUs, making the training process time-consuming and computationally expensive. Networks such as U-Net often encounter significant time and computational bottlenecks in large-scale mapping [78]. The computational cost of SVM is moderate, but it requires manual parameter tuning, which increases the complexity of model building. RF can converge quickly under limited sample conditions, making its training cost the lowest and establishing it as one of the most widely applied LULC remote sensing classification methods. Although SAR polarimetric decomposition methods do not require large amounts of labeled data, they place high demands on data preprocessing, feature extraction, and co-registration, resulting in considerable preparatory workload. The cost of ensemble methods depends on the fusion strategy adopted and is usually at a moderate level.

From the perspective of production efficiency and applicability, RF and ensemble learning methods are relatively simple to operate and computationally efficient, making them suitable for contexts with limited resources or the need for rapid mapping. SVM performs well in tasks with clearly defined class boundaries but is less efficient in high-dimensional, large-sample scenarios. Deep learning methods can achieve high efficiency and automation when large-scale training samples and hardware support are available, but their application threshold is relatively high, limiting their widespread adoption in routine operational production. SAR methods offer unique advantages in karst regions where optical data are severely affected by clouds and fog, making them particularly suitable for subtropical areas with frequent cloudy and rainy conditions. However, their processing workflow is complex and heavily dependent on specialized technical teams. Overall, the choice of method in complex karst environments should comprehensively consider research objectives, data availability, sample conditions, and computing resources. For instance, RF is preferable when the goal is rapid large-scale monitoring; deep learning methods are more advantageous when the target is high-resolution mapping accuracy with sufficient data and computing resources; and SAR or optical–SAR fusion methods are more suitable in regions with persistent cloud cover.

In recent years, hybrid approaches combining machine learning and expert knowledge have demonstrated clear advantages in LULC classification under complex karst conditions. Traditional models such as RF and SVM are prone to spectral confusion in highly heterogeneous environments, while the introduction of geoscientific rules can effectively constrain classification boundaries [164]. For example, in its 2020 land cover mapping, the Mekong River Commission (MRC) addressed the unique shifting cultivation systems of Southeast Asia by defining an expert rule based on a “two-year vegetation index fluctuation threshold,” which corrected misclassifications where ML models confused shifting cultivation with permanent cropland, thereby improving classification accuracy by 12% [172]. In the Jammu region of India, agricultural expert rules were incorporated into LULC classification: when NDVI values ranged from 0.3 to 0.8 but texture features did not match those of typical cropland (e.g., the presence of strip-like shadows), a rule was triggered to reclassify the area as fallow land. This hybrid strategy mitigated misclassifications caused by monsoon-related cloud contamination and improved farmland class accuracy by 8% [173]. These cases illustrate that hybrid strategies not only enhance overall accuracy but also strengthen the interpretability and stability of classification results, particularly in environments characterized by fuzzy boundaries and spectral similarity. Future research should further explore the integration of expert knowledge and deep learning, such as incorporating topographic, phenological, and ecological zoning rules to build “knowledge-driven intelligent classifiers” that unify high accuracy with strong interpretability.

In summary, there is no single optimal approach for LULC remote sensing classification in complex karst environments. Different methods involve trade-offs among accuracy, cost, and efficiency. Future research trends should focus on multi-source data fusion, few-sample intelligent models, and ensemble methods (expert rules + ML) to improve accuracy while reducing dependence on computational power and large datasets, thereby achieving a balance between scientific research and practical application.

3.3. Classification Accuracy Assessment Methods and Uncertainty Modeling

3.3.1. Principles and Strategies for Sample Collection

(1) Challenges in obtaining LULC samples in karst areas.

An independent, accurate, and representative validation sample set is the cornerstone of LULC accuracy assessment [174,175,176]. However, obtaining high-quality land cover samples in karst regions is an inherently challenging task, mainly due to the following factors: ① Poor accessibility for field plot establishment: The presence of densely packed karst hills, rugged terrain, and limited transportation infrastructure makes large-scale, systematic field surveys both costly and difficult to carry out. ② Fragmented land surface: Land cover patches are small and have irregular, interlocking boundaries. When using GPS for geolocation, it is common for the recorded point to fall within feature A, while the corresponding pixel actually contains a mixture of features A and B. This compromises the spatial representativeness of the sample point. ③ Rapid temporal dynamics: Under subtropical monsoon climates, vegetation undergoes significant seasonal variation (Figure 3), and agricultural activities are frequent (

Table 12. Frequent agricultural activities.

Time	Camera Image	Time	Camera Image	Time	Camera Image
01-01		05-01		09-03
02-01		06-02		09-28
03-01		07-01		11-01
04-01		08-02		12-01

Note: The images are screenshots from surveillance footage taken between January and December 2023, located in a karst peak-cluster depression area. This figure illustrates the crop change process in a typical basin farmland landscape across different months. It provides an intuitive reference for temporal matching of remote sensing sample collection and the identification of farming disturbances. The text in the monitoring images represents the current time information displayed by the monitoring software at the moment of the screenshot.

). Therefore, the timing of sample collection must be closely aligned with the acquisition date of the remote sensing imagery; otherwise, the temporal validity of the samples will be substantially reduced.

Relying solely on field surveys is no longer sufficient to meet the demands of large-scale, high-accuracy validation. Modern sample collection strategies are increasingly shifting toward integrated multi-source validation across air, space, and ground platforms. High-resolution imagery acquired by UAVs has become an ideal tool for capturing ground truth data of land cover types. It enables the precise delineation of object boundaries within sampling areas and serves as a valuable reference for validating satellite-based classification results. In addition, very high-resolution (VHR) commercial satellite imagery (e.g., WorldView, Pleiades) and high-definition Google Earth imagery are frequently used as auxiliary interpretation sources, though their acquisition time and spatial accuracy must be carefully assessed when applied.

(2) Review of LULC Spatial Sampling Design.

The spatial and categorical distribution of samples directly determines the objectivity and effectiveness of accuracy assessment. Each sampling design approach has its own strengths and weaknesses (

Table 13. Comparison of different land cover sampling designs applied in karst regions.

Sampling Design	Principle	Advantages	Disadvantages and Challenges	Applicability in Karst Regions
Simple Random Sampling [177]	Randomly select sample points.	Conceptually simple, easy to implement, and statistically unbiased.	May result in uneven spatial distribution of samples, and rare classes (e.g., small water bodies) may not be sampled at all.	Not recommended. It cannot ensure representativeness of all key land cover types in the highly fragmented karst landscape.
Systematic Sampling [178]	Sample at fixed spatial intervals (grid-based).	Ensures uniform spatial distribution of samples with good coverage.	If land cover patterns have periodicity similar to the sampling interval, bias may occur.	Use with caution. Suitable for analyzing broad spatial patterns, but may fail to capture the patch-level randomness of micro-landforms in karst areas.
Stratified Random Sampling [179]	Use a classified map or prior knowledge to define “strata,” and perform random sampling within each stratum either proportionally or with a fixed number of samples.	Guarantees sufficient samples for all classes (including rare ones), improving the efficiency and reliability of accuracy estimation.	Relies on the quality of the stratification basis (e.g., classification map); errors in the stratification layer may introduce bias.	Highly recommended. This is the most commonly used and scientifically sound method for LULC accuracy assessment, effectively addressing the uneven area distribution of land cover types in karst regions.
Spatially Balanced Sampling [180]	Apply specific algorithms to ensure spatial randomness while achieving uniform distribution of samples across space.	Balances randomness and spatial uniformity, offering better handling of spatial autocorrelation.	Algorithms are relatively complex.	Great potential. Particularly suitable for advanced studies involving spatial statistical analysis and uncertainty modeling, as it objectively reflects the spatial heterogeneity of karst environments.

). In karst LULC classification, stratified random sampling should be considered the standard and preferred strategy for collecting validation samples, as it ensures fair representation of all land cover types-regardless of their area-making the accuracy reports more informative and reliable.

3.3.2. Construction and Applicability of Classification Accuracy Assessment Metrics

Traditional metrics such as Overall Accuracy (OA) and the Kappa coefficient are widely used in remote sensing classification assessment, but they have inherent limitations in conveying information and can sometimes be misleading [171,181]. The confusion matrix is the fundamental tool for evaluating classification accuracy, with its rows and columns representing the correspondence between predicted and actual classes [182]. OA reflects the proportion of correctly classified samples across all classes. However, in karst regions where class distributions are highly imbalanced, a high OA can mask the misclassification risk of rare classes. The Kappa coefficient adjusts for agreement expected by chance, aiming to quantify the degree of classification improvement, but its interpretation is influenced by the number and distribution of classes [183]. User’s Accuracy (UA) indicates the proportion of pixels classified as a certain class on the map that actually belong to that class, reflecting commission errors [184]. Producer’s Accuracy (PA) represents the proportion of actual ground truth samples of a certain class that are correctly identified, reflecting omission errors [185]. The F1-score combines UA and PA (equivalent to precision and recall), making it more suitable for evaluating classification performance under class imbalance. It has become one of the most commonly used performance metrics in deep learning-based remote sensing models [186].

Table 14. Comparison of common classification accuracy assessment metrics.

Metric Name	Definition	Applicable Scenario	Limitations
OA [182]	Number of correctly classified samples/Total number of samples	Overall evaluation of classification performance	Sensitive to class imbalance and may mask classification errors
Kappa [183]	Degree of improvement in agreement over random classification	Adjusts for the bias introduced by random agreement	Influenced by the number and distribution of classes, making interpretation difficult
UA [184]	Proportion of samples classified as a given class that actually belong to that class	Reflects the accuracy of interest to map users	Does not reflect omission errors
PA [185]	Proportion of actual samples of a given class that are correctly classified	Reflects the sensitivity to omission errors for map producers	Does not reflect commission errors
F1 Score [186]	Harmonic mean of UA (precision) and PA (recall)	Suitable for comprehensive evaluation under class imbalance conditions	Dependent on the accuracy of UA and PA

presents commonly used classification evaluation methods.

The disagreement decomposition method proposed by Pontius et al. offers a more insightful perspective for understanding errors in LULC remote sensing classification [187]. This method divides the total disagreement between the classification map and the reference map (i.e., 1–Overall Accuracy, OA) into two independent and more interpretable components: (1) Quantity Disagreement (QD): Refers to the extent to which the total area (i.e., number of pixels) of each land cover class differs between the classification map and the reference map. For example, if a classification map overestimates forest by 150 hectares and underestimates grassland by 150 hectares, this indicates quantity disagreement. (2) Allocation Disagreement (AD): Refers to the spatial misallocation of pixels under the condition that the total area of each class matches between the classification and reference maps. For instance, both maps may contain 200 hectares of forest, but if their spatial distributions differ, this reflects allocation disagreement. This decomposition approach clearly reveals the sources of classification errors. For example, if the total disagreement is mainly driven by quantity disagreement, it may indicate a systematic bias in the classifier-such as persistent overestimation of a land cover class due to spectral confusion. If allocation disagreement dominates, it may be related to blurred land cover boundaries, mixed pixels, or insufficient spatial context modeling in the classifier.

The accuracy evaluation system is crucial in LULC remote sensing classification for karst regions. Traditional cross-validation methods are commonly used for model validation, but the choice of an appropriate scheme must be based on the characteristics of the data. Due to the uneven spatial distribution of land cover types in karst regions (such as vegetation and bare rock), standard k-fold cross-validation, while maintaining independence between training and testing sets, may unevenly allocate minority classes across folds. As a result, certain test sets may lack rare class samples, potentially leading to misleading outcomes [188]. In contrast, stratified k-fold cross-validation preserves class proportions and is particularly suitable for karst scenarios where land cover classes are highly imbalanced. It demonstrates greater stability in datasets with uneven sample distributions compared to standard k-fold CV [189,190]. Previous studies have also shown that LOOCV is appropriate for small datasets, maximizing sample utilization when data are scarce. However, it is sensitive to noise and computationally expensive [191,192]. Therefore, for complex karst environments, combining stratified k-fold CV with uncertainty modeling (such as confusion matrix analysis) can more accurately reflect classification performance.

3.3.3. Sources of Classification Uncertainty

Uncertainty in LULC remote sensing classification results is an inherent characteristic, and it is particularly pronounced in karst regions with fragmented topography and complex surface features. Systematically identifying and analyzing the sources of uncertainty is a prerequisite for improving classification reliability and model interpretability. The main sources include:

(1) Positional errors and scale effects. Karst terrain is highly rugged, and image registration errors can be significant. Even a one-pixel offset may cause misalignment of small land cover patches, thereby reducing classification accuracy. Moreover, different land cover types each have their own optimal analysis scale, and using a single resolution often introduces systematic bias. For instance, small-scale land cover types (such as built-up areas) are easily misclassified as cropland or forest when using coarse-resolution imagery [142]. The impact of image resampling on classification accuracy, especially on the precision of object boundaries, should not be overlooked. Due to the highly fragmented surface and blurred boundaries in karst landscapes, resampling may alter spatial resolution, leading to geometric distortion and information loss. Previous studies (e.g., Chen et al., 2023) have examined the impact of spatial resolution and resampling methods on classification accuracy, finding that resampled images often show decreased accuracy and an increase in mixed pixels along boundary areas [193]. Therefore, in complex karst LULC classification, resampling strategies should be chosen cautiously to reduce boundary errors and improve classification reliability. Li (2023), using multi-source data such as GF-2, SPOT-6, Sentinel-2, and Landsat-8, demonstrated that spatial resolution has a significant impact on land cover classification, and although resampling does not change scale sensitivity, it increases classification uncertainty [194]. Multi-scale effects exist in both optical and SAR data, and in the case of multi-source data fusion, inconsistent resampling may further amplify boundary errors.

(2) Semantic ambiguity. Karst ecosystems often exhibit continuous transitions between land cover types, making it difficult for discrete classification systems to capture such gradual changes accurately. For example, KRD grades are typically defined using artificial thresholds, which introduces subjectivity. Additionally, mixed pixels are common at medium and low resolutions. When hard classification methods forcibly assign multiple land cover components to a single class, semantic uncertainty is exacerbated.

(3) Model uncertainty. Different classification models, due to their distinct modeling logic, may produce varying outputs even when fed the same input data. Model selection (e.g., SVM, RF, U-Net) and parameter settings (e.g., kernel functions, number of trees) both influence the stability of classification boundaries. For instance, some studies have shown that in karst wetland classification, LightGBM outperforms RF and XGBoost [195], indicating that model choice itself is a significant source of uncertainty.

3.3.4. Methods for Quantifying and Visualizing Uncertainty

Presenting the abstract uncertainty of classification results in a spatialized form is of great significance for helping users understand and cautiously interpret remote sensing classification products. The main approaches include: ① Posterior probability maps [196,197]: Probabilistic classifiers (such as the Maximum Likelihood method or deep models using ArcGIS Pro 3.2 (Esri, Redlands, CA, USA)) output a posterior probability vector for each pixel. The maximum value in this vector can be interpreted as the classification confidence, while 1 − max(p) represents the uncertainty. This method is particularly effective in detecting uncertainty in areas with blurred boundaries or class mixing. ② Information entropy [198]: Shannon entropy is computed from the posterior probability vector to quantify the degree of probability confusion among classes for each pixel. A higher entropy value indicates greater uncertainty. Since it is class-independent, entropy serves as a comprehensive measure of uncertainty. ③ Fuzzy set theory [199]: Fuzzy classification assigns each pixel a membership value ([0, 1]) for each class, effectively capturing the “both-and” nature of mixed pixels or gradual boundaries. Uncertainty increases as membership values become more evenly distributed across classes. ④ Multi-model or multi-parameter ensemble methods [200]: These approaches assess the stability of pixel classification by analyzing the consistency among classification results produced using different models or parameter settings. Areas with high disagreement indicate zones of high uncertainty.

These methods provide quantitative tools for understanding and controlling the uncertainty and quality of remote sensing classification products.

4. Identification of Core Bottlenecks and Prospects for Intelligent Development

4.1. Core Technical Bottlenecks in Karst LULC Remote Sensing Classification Research

Karst LULC remote sensing classification has evolved from an early era of limited data-relying on medium-resolution imagery and traditional statistical inference-into the current era of remote sensing big data, characterized by multi-source data fusion and AI-driven intelligent perception. These technological advancements have undoubtedly improved classification accuracy and efficiency. However, a fundamental contradiction runs through this evolution: the inherent, multi-scale, and extreme spatial heterogeneity of karst landscapes is profoundly mismatched with the limited and relatively rigid information representation capabilities of existing remote sensing data and models. This mismatch has given rise to four major methodological and theoretical bottlenecks that are commonly faced in current research.

4.1.1. From Sample Scarcity to the Dilemma of Sample Dependence Under the Data-Driven Paradigm

Both mainstream machine learning and advanced deep learning methods rely heavily on large volumes of high-quality labeled training data [201]. However, karst regions are notoriously difficult areas for acquiring such data-rugged terrain hampers accessibility, and frequent cloud and fog severely limit the availability of long-term optical imagery, resulting in high costs and difficulties in generating precise annotations. This creates a fundamental paradox: the regions that most urgently require high-quality samples are the least able to provide them. The overreliance on supervised learning has made the current LULC classification paradigm inadequate when applied to vast, unlabeled karst regions, severely restricting the scalability and transferability of classification technologies.

4.1.2. Data-Driven Karst LULC Remote Sensing Classification Lacks Support from Geoscientific Knowledge

LULC remote sensing classification in karst regions faces amplified challenges of intra-class spectral variability and inter-class spectral similarity. Existing models-particularly deep learning approaches-exhibit the inherent limitations of “black box” architectures when addressing spectral confusion. These models distinguish land cover types by learning statistical patterns from large datasets but fail to incorporate the underlying geophysical processes that drive spectral similarities. For example, a model may learn that forest on shaded slopes appears darker than that on sunlit slopes, yet it lacks an understanding that this pattern results from the physical interaction between solar radiation and topography. This decoupling between geoscientific mechanisms and data-driven modeling leads to several critical issues: ① Sharp performance decline when the model encounters lighting conditions or land cover combinations not present in the training data. ② Lack of interpretability and scientific rigor in model decisions, which poses serious risks when the outputs are used to inform major decisions in ecological restoration or resource planning. ③ Failure to integrate decades of accumulated geoscientific, ecological, and geomorphological knowledge about karst system evolution into the modeling process, missing opportunities to constrain and guide learning in a more meaningful and robust way.

4.1.3. Lack of a Theoretical Framework for Deep Multimodal Information Fusion Adapted to Complex Karst Environments

Multi-source data fusion is widely recognized as an effective means to overcome the limitations of single-sensor information [202]. However, current fusion strategies mostly remain at a superficial level-typically feature stacking-where features extracted from optical, SAR, LiDAR, and other sources are simply concatenated and input into classifiers. This approach fails to achieve deep interaction and joint reasoning across modalities [203,204]. A unified theoretical framework is lacking to address the following fundamental questions: ① How can spectral (chemical) attributes from optical data, surface roughness and moisture characteristics from SAR, and vertical structure information from LiDAR be meaningfully integrated within the model? ② How can the model dynamically adjust the weights of different modalities during fusion, to account for variations in land cover types and regional characteristics within karst areas?

4.1.4. Uncertainty Modeling Is a Key Technical Bottleneck in LULC Classification for Karst Regions

Different land cover types have distinct optimal analysis scales, yet most existing mainstream methods rely on fixed analytical paradigms to address multi-scale effects [205,206]. For example, OBIA depends on subjectively defined, fixed segmentation scales [207], while deep learning methods are constrained by static receptive fields. This “one-size-fits-all” approach often leads to information loss or misaggregation when applied at mismatched scales [208]. The root of the problem lies in the general lack of adaptive scale awareness in current models-they struggle to dynamically adjust the analysis granularity based on local surface complexity [209]. Scale effects remain a core challenge in remote sensing analysis of complex karst environments. Although technologies such as deep learning offer great potential and are increasingly favored for their end-to-end modeling capabilities, their limitations-including data hunger, poor generalizability, and opaque decision processes-necessitate caution when applied to highly heterogeneous landscapes like karst, where sample acquisition is difficult and scientific interpretability is essential.

4.2. Future Research Directions and Outlook

4.2.1. Data Level: Integration of Knowledge Graph Structures and Full-Time-Series Coordinated Observation

In the study of remote sensing information extraction in complex karst surface environments, data fusion is advancing toward a new stage characterized by the integration of knowledge-graph-like structures and coordinated full time-series observation. On one hand, by integrating hyperspectral data for material composition (spectral), LiDAR for vertical structural information, and SAR for texture and moisture features, a comprehensive multidimensional feature representation of land cover can be constructed. With the aid of deep learning architectures such as multimodal Transformers and cross-attention mechanisms, end-to-end fusion of multi-source remote sensing data becomes feasible, potentially enhancing classification accuracy to the level of vegetation communities or even dominant species-allowing fine-scale differentiation of morphologically similar but ecologically distinct land cover types such as natural vs. planted forests or bare rock vs. urban surfaces.

On the other hand, in light of the highly dynamic and seasonal nature of karst ecosystems, dense time-series data cubes based on high-resolution satellites like Sentinel-1/2 can be constructed and combined with spatiotemporal fusion algorithms (e.g., STARFM) to extract phenological rhythms and surface structural changes [210]. Using spatiotemporal deep learning models such as ConvLSTM and Temporal Transformers, the dynamic evolution of phenology and structure over time can be jointly analyzed to enable high-frequency monitoring, causal inference, and trend prediction of ecological processes such as KRD progression, vegetation recovery trajectories, and agricultural disturbances. This approach drives a shift in remote sensing classification from static mapping toward dynamic process perception, offering unprecedented spatiotemporal analytical capacity and intelligent support for ecological monitoring and management in karst areas.

4.2.2. Algorithm Level: Few-Shot, Highly Generalizable, and Explainable Geo-AI

Land cover remote sensing classification in karst mountainous regions is significantly affected by fragmented terrain, strong spatial heterogeneity of surface features, and the scarcity of optical data due to frequent cloud cover. These conditions lead to critical bottlenecks such as sample sensitivity to scale effects and imbalanced spatial distributions. The core challenge in optimizing the sample space for remote sensing in karst environments lies in dynamically fusing multi-scale features and achieving sample distribution balance under terrain constraints.

To address the challenges of sample scarcity and surface complexity in karst areas, we propose a multi-scale constrained sample space optimization framework for land cover remote sensing in complex karst environments (Figure 4). The specific components of the framework are as follows: ① Pre-train a remote sensing foundation model on a large set of unlabeled remote sensing images to obtain a deep understanding of karst surface features. ② Design a cross-scale attention-based hierarchical fusion module, construct a scale–sample density association model, and dynamically adjust sampling strategies to reveal the evolution of sample representativeness across multiple resolutions. ③ Integrate remote sensing indices and terrain variables to quantify the spatial correlation between sparse land cover types and topographic factors. Based on this, construct a terrain–land cover probability model. By combining a boundary-aware sample generation strategy with slope threshold constraints, synthesize terrain-consistent samples to achieve balanced distribution. ④ Apply multi-class uncertainty sampling to identify low-confidence samples in transition zones between bare rock and bare soil. Then, construct a Markov decision process (MDP) using classification accuracy as the reward function to optimize the sampling path, thereby overcoming limitations in feature expression from remote sensing imagery. ⑤ Finally, develop a meta-learning-driven multi-scale coupling framework, using MAML (Model-Agnostic Meta-Learning) to pretrain across geomorphic meta-tasks, thereby enhancing model adaptability in small-sample scenarios.

Representative approaches and methods include:

(1) Geo-AI (Geographical Artificial Intelligence). Geo-AI is a novel approach that integrates geographic knowledge with artificial intelligence techniques. Aimed at addressing the lack of scientific interpretability in “black-box” AI models, its core idea is to build explainable intelligent models embedded with geographic knowledge. Combining expert rules with machine learning can effectively guide and improve the accuracy of LULC remote sensing classification. For example, by introducing geo-ecological prior knowledge, such as the influence of slope, aspect, and elevation on vegetation distribution, or the temporal succession patterns of rocky desertification, as soft constraints embedded into the loss function, models can be guided to learn feature distributions consistent with geographic mechanisms. This reduces confusion in transitional zones and ecologically sensitive areas. Transforming expert knowledge into structured knowledge graphs and embedding them into deep neural networks has been proposed as an important direction for explainable Geo-AI [78].

(2) Active Learning. Active learning is an efficient labeling strategy in scenarios with scarce samples. In karst regions, the surface is extremely complex, field sampling is difficult, and labeling costs are high. Active learning constructs a “model–human” loop, prioritizing a small number of samples with the highest model uncertainty or richest information for manual labeling, thereby significantly improving efficiency and reducing costs [211]. Studies have shown that with 40–60% fewer samples labeled under active learning than through random sampling, equivalent or even higher classification accuracy can still be achieved [212]. Uncertainty sampling is commonly used to select the most informative samples near class boundaries [213], while boundary-aware sampling further improves accuracy in class boundary recognition. When combined with Bayesian deep networks to estimate model uncertainty, information utilization can be maximized, reducing reliance on human prior labeling and achieving “human–machine collaborative labeling” [214]. For karst regions characterized by sample scarcity and strong spectral overlap between classes, active learning has great potential to reduce field sampling costs while improving the stability and generalizability of classification results.

(3) Data augmentation to enhance model robustness for rare classes at low cost. Modern data augmentation techniques provide new solutions to improve classification performance. For example, image synthesis methods based on generative adversarial networks (GANs) can generate samples resembling real remote sensing imagery, thereby effectively expanding the training sets of minority classes [78]. In addition, hybrid augmentation strategies (such as MixUp and CutMix), which perform linear combinations of pixels or regions, can enhance model robustness near class boundaries and mitigate overfitting risks [215]. Data augmentation should therefore be considered a key technical direction for small-sample classification in karst regions, complementing active learning and meta-learning.

(4) Improving interpretability alongside accuracy. Due to the ecological complexity of karst regions, classification results need not only high accuracy but also interpretability, so that geoscientists, government agencies, and international organizations can understand and validate model outputs. Recent studies have applied feature attribution methods (such as SHAP and LIME) to provide visual explanations of the decision-making basis of deep remote sensing models, thereby revealing the roles of different spectral bands, terrain factors, and temporal features in classification decisions [216,217]. Moreover, attention-based deep networks can highlight key regions, improving transparency and controllability [218]. A Geo-AI framework enhanced by explainable AI (XAI) techniques is expected to provide more reliable solutions for classification applications in complex karst environments.

4.2.3. Application Level: Automated Monitoring Solutions for Key Ecological Issues

In response to the urgent need for monitoring ecological fragility and protection in karst regions, remote sensing and intelligent analysis technologies are driving a new phase of refined and intelligent approaches to KRD monitoring and biodiversity conservation [21,219]. However, the practical implementation of karst ecosystem management remains full of challenges. These challenges arise not only from the extreme complexity of karst itself but also from deep contradictions between technological systems and practical management needs. They can be summarized as five critical gaps in “data–model–mechanism–decision–implementation” (Figure 5). Specifically, these include scarce samples, discontinuous imagery, and difficulties in multi-source data fusion; traditional models struggling with heterogeneous, cross-scale, and weak-signal features; a lack of embedded mechanisms for lithological, hydrological, and biological processes; outputs that fail to support policy-making with dynamic updates and multi-objective balancing; and disconnection between technological systems, governance structures, and community participation. Future progress requires breakthroughs in three areas: small-sample learning, cross-scale mechanism modeling, and human–environment feedback coupling, with the goal of establishing Geo-AI as the intelligent infrastructure for karst ecological management.

Future typical application areas include: (1) Comprehensive identification and early warning of rocky desertification. By integrating multi-source remote sensing (optical, SAR, LiDAR) with geography-informed models, it is possible to dynamically identify and predict the progression of rocky desertification across its stages from “potential-mild-moderate-severe,” while combining climate, land use, and socioeconomic data to analyze driving mechanisms. This supports fine-scale spatial decision-making for ecological restoration projects. (2) Habitat mapping for biodiversity conservation. Using high-accuracy remote sensing classification and landscape analysis techniques, detailed mapping of habitats for endemic species (such as karst primates and birds) can be carried out. This enables the generation of maps of food sources, shelter, and water resources, which, combined with landscape ecological models, can assess habitat quality, fragmentation, and ecological corridor connectivity, thereby supporting national park construction and biodiversity conservation strategies. (3) Carbon stock estimation and management supported by remote sensing. By integrating remote sensing classification and vegetation dynamics data with high-accuracy carbon models and flux observation results, it is possible to estimate and simulate historical changes in carbon storage across typical LULC types such as forests, shrublands, and grasslands at regional scales. This provides essential services for nature-based solutions (NbS) under the goals of carbon peaking and carbon neutrality.

5. Conclusions

As one of the most unique and ecologically fragile units in the Earth’s surface system, karst landscapes hold irreplaceable strategic importance for regional ecological security and sustainable development, making the accurate monitoring of LULC in these areas critically important. Characterized by extreme spatial heterogeneity, surface fragmentation, complex spectral confusion among land cover types, and strong topographic effects, karst regions pose fundamental challenges to traditional remote sensing-based LULC classification techniques. This study provides a comprehensive review and analysis from the perspectives of the complex subtropical karst environmental context, the evolution and applicability of LULC classification methodologies, the identification of core obstacles, and future research directions. The main conclusions are as follows:

(1) This review clearly traces the development of LULC remote sensing classification technologies under complex karst conditions. Early pixel-based traditional methods, relying solely on spectral information, often resulted in “salt-and-pepper” noise and failed to address mixed pixels, revealing their limited applicability in karst landscapes. GEOBIA significantly improved classification accuracy and patch integrity by incorporating spatial objects, texture, and contextual features, but still depended on manually defined segmentation scales and rule sets. Machine learning methods, such as Support Vector Machines (SVM) and Random Forests (RF), demonstrated strong capabilities in handling high-dimensional data and promoted the fusion of multi-source data (e.g., optical, SAR, LiDAR), becoming mainstream techniques. As a cutting-edge approach, deep learning has shown great potential for end-to-end feature learning in complex surface interpretation. However, its heavy reliance on high-quality labeled samples and “black-box” nature limits its widespread adoption in karst areas.

(2) In subtropical karst regions, the fragmented surface, spectral confusion, pronounced topographic relief, and cloudy and rainy climatic conditions make the acquisition of remote sensing data and the improvement of classification accuracy particularly challenging. The use of multi-source and multi-modal data such as LiDAR, SAR, and UAV remote sensing can effectively address issues related to mixed pixels and spectral heterogeneity of land cover types. At the same time, advances in remote sensing instruments, such as optimizing spectral bands to enhance vegetation degradation detection, improving SAR microwave resolution to reduce topographic effects, and increasing LiDAR sensor scanning frequency and data density, will further improve data quality. Future research should promote the integration of multi-modal remote sensing data and the optimization of intelligent algorithms, while embedding geographic knowledge to enhance model interpretability and robustness. The synergistic interaction between advances in remote sensing instruments and the development of model algorithms will foster multi-source collaborative observation and improve the capacity of LULC remote sensing monitoring in karst regions.

(3) Currently, LULC remote sensing classification research in karst regions faces four major common bottlenecks: ① the environment in karst areas is extremely complex, and the scarcity of high-quality LULC samples prevents the full potential of advanced intelligent algorithms from being realized; ② deep learning models lack an understanding of geographic processes, resulting in insufficient robustness and interpretability; ③ current data fusion mostly remains at the shallow level of feature concatenation, lacking a theoretical framework for deep multi-modal information integration; ④ existing methods lack adaptive scale-awareness, making it difficult to address the multi-scale heterogeneity of karst surfaces.

(4) The key to overcoming these four common bottlenecks lies in developing a new generation of Geo-AI interpretation frameworks that deeply integrate geoscientific knowledge while offering strong generalization capabilities and interpretability. Future efforts should focus on the following priorities: ① Deepening the integration of geoscientific knowledge and artificial intelligence, by developing physically constrained and interpretable “gray-box” or “white-box” models that make the decision-making process of algorithms scientific, transparent, and traceable. ② Advancing deep fusion of multimodal data, such as exploring multimodal Transformer architectures based on attention mechanisms, to enable holistic perception of land surface features from an integrated graph-structured perspective. ③ Focusing on key issues in ecological applications, ensuring that technological innovations ultimately address real-world needs such as dynamic monitoring of KRD and biodiversity conservation. This includes developing automated solutions through coordinated innovation across data, algorithms, and applications.

At the methodological level, this paper mainly adopts a qualitative review and logical synthesis approach to systematically summarize research on land cover remote sensing classification in complex karst environments. However, since bibliometric analysis tools such as CiteSpace 6.2.R4 and VOSviewer 1.6.20 were not employed to construct quantitative knowledge maps, limitations remain in the visualization of research hotspot evolution, collaboration network structures, and interdisciplinary trends. Future research could integrate bibliometric methods with text mining techniques to conduct more systematic quantitative analyses of the knowledge structure, research frontiers, and disciplinary evolution trajectories in the field of karst LULC classification, thereby further enhancing the transparency and reproducibility of reviews.