Detecting Long-Term Spatiotemporal Dynamics of Urban Green Spaces with Training Sample Migration Method

Abstract

Urban green spaces (UGSs) are critical for landscape, ecological, and climate studies. However, the generation of long-term annual UGSs maps is often constrained by the lack of sufficient, high-quality training samples for training classifiers. In this study, we introduce an automatic training sample migration method based on visually interpreted reference data and long-term Landsat imagery, implemented on the Google Earth Engine (GEE) platform, to produce annual UGSs maps for Tianjin from 1984 to 2022. Migrating training samples to each year significantly improved classification performance, especially for UGSs and water bodies. UGSs coverage in sample areas increased from 5% to 38%, resulting in more reliable trend detection. Our spatiotemporal analysis revealed that green coverage in the study area reached up to 40%, dominated by tree cover that is significantly underestimated in existing global and regional land cover products. Distinct temporal patterns emerged between the old built-up area (OBUA) and new built-up area (NBUA). Early UGS decline was largely driven by NBUAs, while post-2007 greening involved both OBUAs and NBUAs, as captured by classification maps and vegetation indices. Our study proposes a scalable and practical framework for long-term land cover mapping in rapidly urbanizing regions, with enhanced potential as higher-resolution data becomes increasingly accessible.

1. Introduction

Urban green spaces (UGSs), or vegetated areas within urban landscapes [1], are vital for climate-resilience, addressing ecological and environmental challenges related to climate change and urbanization [2,3,4]. Extensive research has explored the ecosystem services provided by UGSs, including carbon sequestration [5,6], urban heat island mitigation [7], air pollutant removal [8], and the promotion of public physical and mental health [9], as well as factors influencing these services and issues of environmental equity [10,11]. Clearly, acquiring the spatial distribution of green spaces is a foundational and essential step for advancing UGSs-related research.

In recent decades, rapid urbanization and greening policies have significantly altered the extent, spatial distribution, vegetation species, and landscape patterns of UGSs in cities undergoing urbanization [10,11]. Based on changes in remote sensing vegetation indices, previous studies in China have detected long-term shifts in UGSs at both a national [12,13] and city scale [14]. Despite the effectiveness of vegetation indices, land cover classification remain essential for either providing grid-based land cover type required by process models [2,15], or for enabling the exploration of UGS accessibility [16] and landscape patterns [17], etc. However, existing research often relies on a bi-temporal [18] or, at best, a coarsely multi-temporal classification map [19,20] to analyze the long-term changes in urban land cover. These low-frequency maps may fail to capture the rapid and complex land cover transitions typical of urban environments [21,22]. Due to the limited availability of urban categorical maps spanning multiple decades at annual frequency, most UGSs-related studies rely on global land cover datasets [2,15]. However, these datasets often have a minimum mapping unit that is too coarse to accurately capture land cover variations in urban environments [16]. Consequently, land cover datasets with adequate temporal resolution covering an extended time frame are indispensable to accurately monitor the long-term trends of UGSs and to understand their effects on urban ecological research and planning.

High temporal-frequency analysis of urban areas has gained prominence as a re-search frontier since 2008, when Landsat data became freely available [23]. Although high-resolution imagery can theoretically capture finer urban details [24], their application in long-term UGSs monitoring remains constrained by limited temporal coverage and high costs. In contrast, Landsat data—with its moderate spatial resolution (30 m), four-decade archive, and free accessibility—has become fundamental for analyzing urbanization’s im-pact on landscape dynamics [21,22]. Cloud-based platforms like the Google Earth Engine (GEE) and Earth Observation (EO) Browser [25] have revolutionized geospatial analysis by providing open access to multi-source satellite data, advanced classification algorithms, and computational power. However, the availability of representative training samples is still the pre-requirement for land cover classification [26,27] and access is a challenge that is particularly pronounced in highly heterogeneous urban environments.

Supervised land object classification based on machine learning algorithms (MLAs) relies on independent variables derived from remote sensing imagery, such as spectral bands, vegetation indices, and texture features, as well as dependent variables represented by labeled reference samples [28,29,30]. Despite advancements in MLAs, the quality of long-term image classification still heavily depends on the availability of adequate and representative training samples [31,32]. To fully capture urban landscape dynamics from the 1980s to the present [33], utilizing diverse remote sensing sources is inevitable due to data gaps and varying satellite service periods [21]. Yet, spectral characteristics of the same objects can vary across sensors and years, making it challenging to maintain classification consistency [21,22]. Therefore, training a single classifier based on single-date training data and applying it to every year within the study period can significantly compromise accuracy, while training a classifier for each year using corresponding reference samples is more reasonable. However, obtaining high-quality reference samples annually through visual interpretation is labor-intensive, time-consuming, and often hindered by cloud contamination [26]. Although some studies suggested deriving training samples from previous land cover maps [34], it is unsuitable for urban areas due to the coarse resolution of source products, which fail to identify fine-scale urban land objects. Recently, Huang et al. [16] proposed an automatic method for migrating reference samples across time periods by measuring the spectral similarity and distance between spectral reflectance data from Landsat 8 and Landsat 5. Their approach achieved over 90% accuracy for migrated no-change samples, and classifiers trained with these samples performed comparably to those using original 2015 samples. Similarly, Ghorbanian et al. [35] demonstrated that migrated training samples improved land cover mapping in Iran. Through reducing the reliance on manual interpretation and addressing spectral variability, the training sample migration method offers a valuable solution for long-term urban land cover mapping with high frequency.

In this study, we attempted to map the distribution of UGSs on an annual basis from 1984 to 2022 across the central urban area of Tianjin City, utilizing the GEE platform and Landsat series imagery. Firstly, we migrated visual interpreted reference samples to each year within the study period. Secondly, we trained classifiers annually using migrated samples and generated urban land cover maps. Finally, we analyzed the spatiotemporal dynamics of UGSs in the study region. To validate the reliability of our results, we compared long-term NDVI trends and global land cover datasets with our annual UGSs maps, highlighting discrepancies and their implications for UGSs-related environmental and climate change research.

2. Materials and Methods

2.1. Study Area

Tianjin is located in the northeastern region of the North China Plain (116°43′–118°04′E, 38°34′–40°15′N), adjacent to Bohai Bay. The city experiences a warm temperate semi-humid monsoon climate with distinct seasonal variations, featuring an average annual temperature of around 14 °C and a mean annual precipitation ranging from 360 to 970 mm [36]. The central urban area (Figure 1a), encircled by Tianjin’s outer ring road, serves as the city’s administrative, cultural, and commercial center. It is characterized by the highest population density and the largest urban development zone, encompassing approximately 4.2 million people (36% of the city’s population) within an area of about 371 km² (Tianjin Statistical Yearbook, 2022). Tianjin’s urban expansion has progressed from the inner core outward [37]. Districts such as Heping, Nankai, Hexi, Hongqiao, Hedong, and Hebei were well established by the 1980s and are defined as the old built-up area (OBUA) in this study (Figure 1b). In contrast, the surrounding regions, which have gradually developed with industrial, residential, and commercial zones over the past decades, are defined as the new built-up area (NBUA).

2.2. Dataset

Land cover classification was performed using multi-temporal Landsat surface reflectance collections (1984–2022), accessed via the Google Earth Engine. The core dataset comprised atmospherically corrected Level-2 products (Landsat TM/OLI, 30 m resolution) processed with LEDAPS/LaSRC algorithms and CFMASK cloud filters [38]. Additional remote sensing imagery was employed for supplementary validation, as follows: (1) 10 m resolution Sentinel-2A [39] multispectral imagery from the European Space Agency (ESA), and (2) 250 m resolution moderate resolution imaging spectroradiometer (MODIS) products obtained through NASA’s official data portals [40].

The 4695-reference training sample database was established through systematic visual interpretation of Google Earth imagery from 2015. Using Landsat imagery, we performed sample transfer based on the principle of spectral consistency, migrating the 2015 sample points to other years to serve as training data for each year. These samples were categorized per the Chinese Land Use Classification Standard (GB/T 21010-2017) [41] into the following five land cover types: impervious surface, tree, grass, crop, and water (

Table 1. Description of land cover type and corresponding number of sample points.

Primary	Secondary	Number	Description
Green space	Tree	1484	Arbor forests, afforested land, artificial young forests, and shrub
	Grass	790	Natural grassland and artificial grassland
Non-green space	Crop	613	Cropland
	Water	560	Rivers, lakes, reservoirs, channels, ponds, and other similar water features.
	Impervious surface	1248	Building land, in this study refers to land use that is categorized as neither green space, cropland, nor water bodies.

).

We also extracted a land cover map of the study area from global land cover products and a regional urban thematic map, including FROM-GLC (10 m/30 m) [42,43], GlobeLand30 (30 m) [44], GLC_FCS30 (30 m) [45], Copernicus CGLC (100 m) [46], MCD12Q1 (500 m) [47], MODIS Land Cover (500 m) [48], and CLUD-Urban [19], for checking their performance in mapping UGSs in the study area. The normalized difference vegetation index (NDVI) was calculated using surface reflectance data from the red (620–670 nm) and near-infrared (NIR, 841–876 nm) bands of MODIS and Landsat satellite imagery. Annual maximum NDVI values, generated through a temporal compositing approach to mitigate cloud contamination and seasonal variability, were employed to quantify UGS dynamics during the study period [49].

2.3. Methods

Figure 2 below shows the workflow of this study, starting from the acquisition of satellite remote sensing datasets (in Section 2.2), to training sample migration (Section 2.3.1), to the feature index being added to assist the supervision classification (Section 2.3.2) and accuracy assessment (Section 2.3.3), and then finally to the use of the landscape pattern index to analyze the spatiotemporal characteristics of green space change (Section 2.3.4). Notably, except for the labeled sample points in specific reference years, all other datasets used in our workflow are globally available and freely accessible through the GEE platform, which makes our method easily adaptable for research in other cities and regions.

2.3.1. Training Sample Migration

To generate annual training datasets (70% for training and 30% for validation), we implemented a sample migration approach to transfer 4695 sample points from the reference year (2015) to other years (target years). Stable samples were determined through visual interpretation by comparing reference and target year imagery from three satellites (Landsat 5, Landsat 8, and Sentinel-2A). The spectral angle distance (SAD) and Euclidean distance (ED) [50] were calculated to determine the rational thresholds for spectral changes (

Table 2. Threshold information for sample points with unchanged land class across different sensors in 2010, 2020, and 2022 corresponding to Landsat 5, Landsat 8, and Sentinel-2A sensors, respectively.

	Land Cover	Impervious Surface	Tree	Grass	Crop	Water	Threshold
2010	ED_mean	0.002	0.004	0.117	0.009	0.018	0.2
	SAD_mean	0.960	0.998	0.984	0.991	0.964	0.9
2020	ED_mean	0.001	0.276	0.156	0.356	0.026	0.4
	SAD_mean	0.965	0.896	0.971	0.666	0.978	0.6
2022	ED_mean	0.028	0.094	0.075	0.029	0.026	0.1
	SAD_mean	0.998	0.972	0.901	0.987	0.967	0.9

), with the formulas as follows.

SAD is utilized to measure the directional change in angle between two temporal vectors. If the spectral angles between the reference year and the target year are identical, SAD is 1. The formula is as follows:

$$\theta ={\mathit{cos}}^{-1}{\displaystyle \frac{{\sum }_{i=1}^N{X}_{i\left(t1\right)}{Y}_{i\left(t2\right)}}{\sqrt{{\sum }_{i=1}^N{\left(X_{i\left(t1\right)}\right)}^2{\sum }_{i=1}^N{\left(Y_{i\left(t2\right)}\right)}^2}}}$$

(1)

$$SAD=\mathit{cos}\left(\theta \right)$$

(2)

ED represents the Euclidean distance between the reference spectra and target spectra. If the reference spectra are identical to the target spectra, ED is 0. The formula is as follows:

$$ED=\sqrt{{\sum }_{i=1}^N{(X_{i\left(t1\right)}-Y_{i\left(t2\right)})}^2}$$

(3)

where θ is the spectral angle, X_i(t1) is the reference spectra collected at time t1 for the training sample pixel, Y_i(t2) is the target spectra to be measured at time t2, and the variable i is the spectral band ranging from 1 to the total number of bands N.

The spectral values of samples in Landsat images in the reference and target years were extracted and compared with thresholds to identify unchanged land use points for sample migration. The number of available sample points for each year consistently exceeded 4000 (Figure 3), which is used to train the classifier annually.

2.3.2. Supervised Classification

Remote images were initially atmospherically corrected using GEE. Subsequent processing involved the following four critical steps: image mosaicking to create seamless spatial coverage, cloud masking through bitwise computations applied to QA bands, annual median compositing of multi-temporal data to mitigate seasonal variability, and geographic clipping to focus on the study area and temporal period of interest.

The model incorporated six spectral bands (Blue: 0.45–0.52 μm; Green: 0.52–0.60 μm; Red: 0.63–0.69 μm; NIR: 0.77–0.90 μm; SWIR1: 1.55–1.75 μm; SWIR2: 2.08–2.35 μm) and six spectral indices derived from Landsat imagery, which include the normalized difference vegetation index (NDVI), normalized difference built-up index (NDBI), normalized difference water index (NDWI), modified normalized difference water index (MNDWI), enhanced vegetation index (EVI), and bare soil index (BSI), which are calculated as follows:

$$NDVI=(NIR-Red)/(NIR+Red)$$

(4)

$$NDBI=(SWIR1-NIR)/(SWIR1+NIR)$$

(5)

$$NDWI=(Green-NIR)/ (Green+NIR)$$

(6)

$$MNDWI=(Green-SWIR1)/(Green+SWIR1)$$

(7)

$$EVI=2.5(NIR-Red)/(NIR+6.0Red-7.5Blue+1)$$

(8)

BSI = [(SWIR1+Red)-(NIR+Blue)]/[(SWIR1+Red)+(NIR+Blue)]

(9)

To address spectral confusion between cropland and grassland in the study area, 18 texture features were computed via gray-level co-occurrence matrices (GLCMs) [51]. These features enhance the discriminative capacity for land cover classes with overlapping spectral signatures [52]. All spectral and textural indices were added to the preprocessed image collection to create feature variables. Three machine learning classifiers—random forest (RF), classification and regression tree (CART), and support vector machine (SVM)—were trained using the prepared training samples and feature variables. Among these, RF was selected as the primary classifier due to its ensemble approach that employs both bootstrapped samples and random feature subsets [53], which provides enhanced robustness against overfitting and noise sensitivity [54]. CART employes binary recursive partitioning to generate compact decision trees, while SVM optimizes hyperplanes via quadratic programming [55]. Classification results were validated with validation samples to compare classifier performance and select the optimal algorithm for UGSs extraction.

2.3.3. Accuracy Assessment

Classification accuracy was assessed using four key metrics: producer’s accuracy (PA), user’s accuracy (UA), overall accuracy (OA), and kappa coefficient [56]. This multi-dimensional validation ensured the reliability of LULC classifications, supporting robust analysis of spatiotemporal changes and associated environmental impacts.

2.3.4. Landscape Pattern Index

Landscape pattern indices [57,58] were quantified through Fragstats 4.2 analyses, with two core indices selected, patch area (AREA) and patch density (PD), reflecting the size distribution and fragmentation of green space patches. AREA was used to reflect the size and spatial distribution of each patch, and the units used are hectares. While PD indicates the degree of patch fragmentation and landscape heterogeneity, increasing as the degree of patch fragmentation increases, and the units used are the number of patches per 100 hectares. The calculation formulas are as follows:

$$AREA=a_{ij}\left({\displaystyle \frac{1}{\mathrm{10,000}}}\right)$$

(10)

$$PD={\displaystyle \frac{n_i}{A}}(\mathrm{10,000})(100)$$

(11)

where a_ij is area (m²) of patch ij, A is total landscape area (m²), and n_i is the number of patch class i.

Since classification errors in each year may obscure the actual temporal trend of UGSs, we quantified the uncertainty in the trend estimation by incorporating classification error into the slope analysis [58]. Specifically, we first estimated the standard error (SE) of UGSs area for each year based on a simplified assumption that the error is proportional to the annual misclassification rate (

Table A1. User’s accuracy (UA) and producer’s accuracy (PA) of land use classification results using the RF classifier from 1984 to 2022, along with overall accuracy (OA) and kappa coefficient (Kappa).

Year	UA					PA					OA	Kappa
	Impervious Surface	Tree	Grass	Crop	Water	Impervious Surface	Tree	Grass	Crop	Water
1984	0.67	0.65	0.46	0.81	0.78	0.69	0.74	0.30	0.83	0.62	0.70	0.66
1985	0.72	0.63	0.39	0.64	0.93	0.60	0.78	0.25	0.81	0.61	0.67	0.52
1986	0.68	0.60	0.52	0.80	0.88	0.70	0.71	0.25	0.84	0.67	0.68	0.54
1987	0.72	0.62	0.34	0.76	0.91	0.72	0.76	0.22	0.88	0.59	0.71	0.57
1988	0.67	0.64	0.58	0.79	0.80	0.72	0.73	0.22	0.90	0.74	0.70	0.56
1989	0.73	0.66	0.45	0.84	0.83	0.80	0.73	0.25	0.89	0.67	0.72	0.61
1990	0.78	0.62	0.40	0.80	0.86	0.72	0.80	0.22	0.85	0.74	0.72	0.61
1991	0.68	0.67	0.54	0.85	0.76	0.70	0.76	0.31	0.87	0.72	0.71	0.58
1992	0.69	0.69	0.41	0.83	0.79	0.73	0.77	0.23	0.86	0.77	0.72	0.60
1993	0.76	0.67	0.56	0.83	0.88	0.76	0.81	0.26	0.85	0.66	0.74	0.61
1994	0.74	0.62	0.80	0.83	0.78	0.76	0.75	0.25	0.87	0.60	0.70	0.57
1995	0.79	0.67	0.53	0.76	0.84	0.78	0.77	0.23	0.86	0.82	0.74	0.62
1996	0.77	0.63	0.45	0.83	0.81	0.78	0.76	0.23	0.82	0.74	0.72	0.59
1997	0.78	0.70	0.65	0.80	0.86	0.78	0.82	0.25	0.91	0.79	0.75	0.63
1998	0.77	0.68	0.34	0.84	0.81	0.81	0.77	0.22	0.89	0.79	0.74	0.63
1999	0.77	0.65	0.45	0.92	0.83	0.79	0.79	0.21	0.89	0.73	0.75	0.64
2000	0.71	0.67	0.39	0.90	0.84	0.83	0.73	0.22	0.90	0.68	0.73	0.62
2001	0.78	0.70	0.66	0.88	0.83	0.83	0.82	0.24	0.91	0.67	0.76	0.65
2002	0.71	0.63	0.39	0.83	0.80	0.72	0.75	0.22	0.83	0.72	0.70	0.57
2003	0.69	0.65	0.63	0.94	0.85	0.78	0.77	0.22	0.86	0.73	0.73	0.61
2004	0.73	0.68	0.53	0.93	0.89	0.82	0.80	0.22	0.88	0.73	0.75	0.63
2005	0.73	0.66	0.53	0.98	0.85	0.86	0.78	0.21	0.88	0.77	0.74	0.63
2006	0.72	0.72	0.45	0.92	0.79	0.89	0.72	0.23	0.90	0.76	0.75	0.64
2007	0.72	0.75	0.70	0.87	0.78	0.88	0.80	0.21	0.85	0.76	0.75	0.64
2008	0.80	0.67	0.28	0.90	0.92	0.82	0.86	0.21	0.90	0.80	0.76	0.65
2009	0.77	0.74	0.65	0.97	0.80	0.89	0.85	0.25	0.91	0.79	0.79	0.69
2010	0.77	0.75	0.57	0.88	0.83	0.89	0.83	0.23	0.91	0.82	0.78	0.68
2011	0.78	0.72	0.49	0.90	0.82	0.84	0.87	0.23	0.85	0.81	0.77	0.66
2012	0.79	0.74	0.40	0.87	0.84	0.81	0.87	0.24	0.90	0.70	0.76	0.65
2013	0.81	0.78	0.65	0.99	0.78	0.93	0.91	0.26	0.90	0.75	0.80	0.71
2014	0.84	0.76	0.58	0.97	0.85	0.88	0.90	0.28	0.93	0.79	0.81	0.72
2015	0.85	0.78	0.55	0.92	0.85	0.88	0.90	0.30	0.90	0.80	0.81	0.72
2016	0.85	0.78	0.55	0.94	0.80	0.90	0.87	0.24	0.91	0.81	0.81	0.71
2017	0.85	0.78	0.62	0.93	0.92	0.89	0.92	0.32	0.91	0.80	0.82	0.73
2018	0.84	0.79	0.70	0.94	0.84	0.91	0.88	0.41	0.90	0.74	0.82	0.73
2019	0.84	0.79	0.59	0.93	0.87	0.87	0.93	0.32	0.88	0.84	0.81	0.72
2020	0.88	0.78	0.68	0.85	0.85	0.91	0.91	0.38	0.82	0.75	0.82	0.73
2021	0.83	0.75	0.76	0.93	0.86	0.88	0.88	0.36	0.85	0.73	0.80	0.70
2022	0.83	0.77	0.58	0.94	0.78	0.86	0.90	0.31	0.83	0.78	0.79	0.69

). Then, we adopted a Monte Carlo simulation approach for each iteration, and we randomly sampled one value from a normal distribution defined by the UGSs area and the estimated SE for each year and combined these to form a synthetic time series. We repeated this process 1000 times. For each simulated time series, we calculated the linear trend of UGSs change over time. Finally, we summarized the distribution of the resulting slopes to derive the uncertainty bounds of the overall trend (Figure A1).

3. Results

3.1. Mapping UGSs with Training Sample Migration

We evaluated the classification accuracy of various machine learning algorithms. RF algorithm outperformed others (

Table 3. User’s accuracy (UA) and producer’s accuracy (PA) of three different classifiers smile random forest (RF), smile cart (CART), and libsvm (SVM) on GEE for different land classes, as well as overall accuracy (OA) and kappa coefficient (Kappa).

Classifier	UA					PA					OA	Kappa
	Impervious Surface	Tree	Grass	Crop	Water	Impervious Surface	Tree	Grass	Crop	Water
RF	0.85	0.78	0.55	0.92	0.85	0.88	0.90	0.30	0.90	0.80	0.81	0.72
CART	0.60	0.59	0.41	0.86	0.80	0.66	0.69	0.15	0.93	0.77	0.65	0.63
SVM	0.58	0.52	0.26	0.82	0.38	0.65	0.57	0.12	0.81	0.31	0.56	0.58

), achieving the highest accuracy with an overall accuracy (OA) consistently ranging between 0.7 and 0.8 across annual assessments (Table A1). Figure 4 further illustrates that Landsat-based classification, despite exhibiting some mosaic effects, can capture surface feature details at a level comparable to those obtained from Sentinel-2A in the example region. The estimated areas of different land cover types derived from Landsat data were consistent with those obtained from Sentinel-2A (Figure 5).

To assess the effectiveness of the sample migration method, we compared classification results with and without its application. From the perspective of spatial information extraction, the approach without sample migration struggled to accurately differentiate between land cover types and was inclined to misclassify both green space and water body to impervious area (Figure 6). In the sample area, maps without migrated samples underestimated green coverage by 87%. Because of improved spatial details, the temporal trend was also reasonable (Figure 7). For example, without migrated samples, UGSs decreased to less than 10 km² in 2010, which is impossible, as can be seen by referring to the Statistic Yearbook (Tianjin Statistical Yearbook, 2022) [59] and a previous study [18].

3.2. Spatial Distribution of UGSs

As Figure 5 shows, we found that the study area encompassed 160 km² of UGSs in 1984, with urban green coverage up to 48%. By 2022, the area of UGSs had decreased by 18.75%, among which tree cover declined to 111.78 km². Meanwhile, grassland areas expanded nearly threefold. In 1984, 60% of UGS was in the NBUA, but UGS distribution had become nearly equal between the two sub-regions. Trees were always the predominant type of UGSs, which was approximately 111.78 km² in 2022 and 27.9% larger than grass area. UGSs have also become more fragmented during the study period, which is reflected by patch density decreasing from 8 to 12 (Figure 8). Patch density in OBUAs was higher than in NBUAs, a pattern that was visually evident in spatial maps (Figure 9). Besides UGSs, the area of the water body significantly reduced from 25.24 km² to 12.61 km², with a large swath of water southwest of the study area disappearing.

In contrast to our results, global land cover products consistently overestimate impervious areas and underestimate UGSs (Figure 10). Specifically, impervious areas in global and regional products exceeded our estimates by 5% to 75%. Conversely, two global land cover products underestimated tree cover by more than 50% compared to our results, while five others, including the finer-resolution FROM-GLC10 dataset, completely missed tree distribution patterns. Although CLUD-Urban is an urban thematic map and performs better than global land cover products, it still underestimated tree cover by 60%. A similar discrepancy was observed in grassland classification, except for FROMGLC10, which overestimated grassland area.

3.3. Temporal Trend of UGS

Through long-term annual mapping, we investigated the temporal trend of UGSs in the study region (Figure 11). During research period, the UGSs area experienced a significant decline at a rate of 0.69 km² per year, while impervious surfaces expanded consistently at a rate of 1.62 km² per year. The temporal patterns of UGSs in OBUAs and NBUAs differed markedly. Changes in the NBUAs drive the overall regional trend. Specifically, shifts in UGSs and impervious surfaces in the NBUA accounted for 90% and 92% of the observed changes in study area, respectively. Notably, the trends in UGSs area exhibited a distinct shift in 2007, with the rate of decline slowing to one-fourth of the pre-2007 rate. To validate these findings, we conducted an NDVI-based analysis using Landsat and MODIS data. Despite the significant difference in spatial resolution (30 m versus 250 m), both NDVI analyses detected a trend shift around 2007, aligning with our annual mapping results. The NBUA alone dominated the regional UGSs decline in the earlier phase, while both the OBUA and NBUA contributed to the greening of the region in the later phase. Despite classification errors in each year (Table A1), we found that the decreasing trend of the NBUA remains statistically significant (Figure A1).

4. Discussion

4.1. Reliability of Long-Term Annual Maps of UGSs with Sample Migration Method

For mapping long-term spatiotemporal changes in urban land cover at an annual frequency, the primary challenge lies in obtaining reliable reference samples [22,60]. Our study demonstrated that the automated sample migration method, previously applied in global and regional land cover classification [42], is suitable for obtaining sufficient and reliable reference samples in urban areas. This method enables the identification of green spaces and water bodies more accurately (Figure 6), with the estimated green space coverage aligning with a previous study [18] and the Statistical Yearbook (Tianjin Statistical Yearbook, 2022). In addition, the trend shifts in long-term UGSs can also be captured reasonably, which is supported by NDVI analyses from different remote sensing data (Figure 12).

The sample migration method demonstrates unique advantages in long-term land cover monitoring by efficiently generating training samples across multiple temporal phases [26]. Conventional automated sample extraction approaches often rely on historical land use maps [32,34], which are limited by a coarse spatial resolution, and typically migrate samples only to years close to the reference year [42]. Our sample migration method enables temporally consistent historical mapping by transferring reference samples across years within a long-term framework. By leveraging remote sensing time series, the approach achieves reliable cross-temporal sample transfer requiring only single-phase manual interpretation, which significantly reduces manual sampling costs while maintaining and improving spatiotemporal applicability. With the accumulation of high-resolution remote sensing datasets and land cover classification maps [61], we can eliminate visual interpretation entirely, enabling fully automated sample extraction and seamless cloud-based platform implementation for large-scale urban green space monitoring. Furthermore, we validated land use change detection results through independent verification using the continuous change detection and classification (CCDC) method [62]. CCDC enables the continuous monitoring of diverse land use and land cover changes with high spatial and temporal accuracy [63]. Results showed that the highest frequency of change years (i.e., 2007) identified by CCDC aligned with the annual classification outcomes from this study (Figure 13), confirming the temporal accuracy of our findings. While CCDC excels at continuous change monitoring, its annual mapping applications for UGSs face limitations, especially at large scale, due to high computational demands from dense time-series processing and reduced reliability in cloud-prone regions from insufficient clear observations [64]. Our integrated framework combines the RF algorithm with sample migration to efficiently generate independent annual land use maps, enabling robust comparative analysis while overcoming these limitations.

4.2. Long-Term Spatiotemporal Dynamics of UGS

The trend shifts in long-term UGSs around 2007 in the central urban area of Tianjin have also been found in other cities [14,65]. The decline in the UGSs area before 2007 was primarily driven by changes in the NBUA, a phenomenon commonly observed in large cities experiencing rapid urban sprawl, where original UGSs in peri-urban areas was converted into impervious surfaces [13,18]. This shift has been largely attributed to the implementation of national and local greening policies [14,65], initializing urban ecology protection and afforestation.

We further investigated the drivers of UGSs greening in the post-2007 period. The increase in new green spaces in the NBUA often occurred at the expense of original UGSs, water bodies, and farmland (Figure 14), a trend commonly observed in the expansion of major cities [18]. In contrast, the greening of the OBUA was primarily driven by urban renewal and vegetation growth. To optimize land use efficiency and accommodate high population density in urban cores, many large Chinese cities, including Tianjin, have adopted strategies such as replacing bungalows with high-rise buildings [66], freeing up space for urban ecological land (Figure 14c). Additionally, small green spaces, such as corner parks or pocket parks, have been integrated into existing construction areas. Planted trees in urban environments exhibit rapid growth and larger canopy breadths compared to natural vegetation, further supported by irrigation and elevated CO₂ levels [2,67], which collectively enhance green coverage and contribute to UGS expansion. As Figure 14d shows, NDVI tripled over a 20 year period due to significant vegetation growth in a residential area.

4.3. Implications for Global and Local UGSs-Related Research

In comparison with our products, we found that global products significantly underestimate UGSs distribution in central urban areas, with urban trees nearly absent (Figure 10). While these datasets aim to map land cover broadly, they are often used in environmental and climate change research in urban areas as basic data [33]. For example, a recent study estimated China’s urban biogenic volatile organic compound emissions using FROM-GLC10 for UGSs classification [68]. Our comparative results (Figure 10) suggest BVOC emission fluxes may be higher than their estimates, given urban trees’ high emission rates [69].

Recently generated global and regional UGSs maps [19,70] should be prioritized for future UGSs-related research. However, extant large-scale UGSs datasets are often static or temporally coarse, relying on the latest imagery due to data storage and processing challenges. Even using urban thematic maps, scale-dependent discrepancies exist (Figure 10), which is caused by urban ecosystems’ heterogeneity [8,71]. Thus, high-resolution, customized long-term annual mapping of land cover and vegetation data at global and local scale should be seen as a critical research area for addressing future urban-related scientific questions and supporting accurate urban planning and ecological management.

4.4. Limitations

Despite capturing turning points of long-term UGS changes, annual maps failed to reflect the greening trend of UGSs in OBUAs (Figure 11 and Figure 12). This discrepancy highlights that, although the sample migration method improved spatial information extracted in OBUAs, coarser spatial resolution of remote sensing images poses limitations in detecting green space change in the temporal domain [72,73]. Previous studies have proved that the 30 m spatial resolution of Landsat images cannot reveal dynamic changes in UGSs in inner cities, of which patch sizes are often smaller than 0.1 ha [17,73]. In contrast, increases in green space in NBUAs are reflected in the annual maps of UGSs, which can be attributed to larger patch sizes at the urban edge (Figure 15). Furthermore, the consistent results obtained from coarser-resolution data (500 m) imply that NDVI trends in urban areas are less affected by spatial resolution, a finding also supported by the authors of [13,74]. Therefore, it is recommended to use vegetation indices for analyzing historical UGS changes and to employ high-resolution imagery for future urban land cover classification.

5. Conclusions

In this study, we generated long-term annual urban land cover maps for Tianjin from 1984 to 2022, using the training sample migration method. The main conclusions were as follows:

(1) Training classifiers with migrated samples significantly improved the accuracy of identifying UGSs and water bodies, which improved the estimation of urban coverage by 33% in the sample area. The enhanced spatial information extracted further ensures reliable temporal trends of UGSs.

(2) Spatiotemporal analysis reveals that UGSs coverage in the study area reaches up to 40%, predominantly composed of tree cover, a feature significantly underestimated in both global land cover products and regional urban thematic maps.

(3) A distinct shift in UGSs trends occurred in 2007, captured by both classification maps and vegetation index analysis, aligning with the implementation of national greening policies. Before 2007, the NBUA alone accounted for the regional decline in UGSs, while both the NBUA and OBUA contributed to the greening of the study area in the later phase.

While demonstrated through Tianjin’s rapid urbanization case, our open-data approach ensures direct applicability to other cities, particularly for resource-constrained regions requiring efficient, annual land use analysis. This workflow will achieve better application through the accumulation of finer-resolution data, such as Sentinel-2 imagery.