Extraction of Garlic in the North China Plain Using Multi-Feature Combinations from Active and Passive Time Series Data

Abstract

Garlic constitutes a significant small-scale agricultural commodity in China. A key factor influencing garlic prices is the planted area, which can be accurately and efficiently determined using remote sensing technology. However, the spectral characteristics of garlic and winter wheat are easily confused, and the widespread intercropping of these crops in the study area exacerbates this issue, leading to significant challenges in remote sensing image analysis. Additionally, remote sensing data are often affected by weather conditions, spatial resolution, and revisit frequency, which can result in delayed and inaccurate area extraction. In this study, historical data were utilized to restore Sentinel-2 remote sensing images, aimed at mitigating cloud and rain interference. Feature combinations were devised, incorporating two vegetation indices into a comprehensive time series, along with Sentinel-1 synthetic aperture radar (SAR) time series and other temporal datasets. Multiple classification combinations were employed to extract garlic within the study area, and the accuracy of the classification results was systematically analyzed. First, we used passive satellite imagery to extract winter crops (garlic, winter wheat, and others) with high accuracy. Second, we identified garlic by applying various combinations of time series features derived from both active and passive remote sensing data. Third, we evaluated the classification outcomes of various feature combinations to generate an optimal garlic cultivation distribution map for each region. Fourth, we developed a garlic fragmentation index to assess the impact of landscape fragmentation on garlic extraction accuracy. The findings reveal that: (1) Better results in garlic extraction can be achieved using active–passive time series remote sensing. The performance of the classification model can be further enhanced by incorporating short-wave infrared bands or spliced time series data into the classification features. (2) Examination of garlic cultivation fragmentation using the garlic fragmentation index aids in elucidating variations in accuracy across the study area’s six counties. (3) Comparative analysis with validation samples demonstrated superior garlic extraction outcomes from the six primary garlic-producing counties of the North China Plain in 2021, achieving an overall precision exceeding 90%. This study offers a practical exploration of target crop identification using multi-source remote sensing data in mixed cropping areas. The methodology presented here demonstrates the potential for efficient, cost-effective, and accurate garlic classification, which is crucial for improving garlic production management and optimizing agricultural practices. Moreover, this approach holds promise for broader applications, such as nationwide garlic mapping.

1. Introduction

As the global population expands and dietary patterns evolve, the cultivation area and production of garlic have steadily increased [1]. China consistently dominates in garlic production, exportation, and consumption, positioning garlic as a critical crop within the nation’s economy [2]. In recent years, garlic production has exhibited a steady increase, underscoring the significance of promptly accessing information regarding its planted area and distribution to inform assessments of garlic prices and food security. Conventional garlic planting area surveys rely on field sampling techniques, which are characterized by high costs, prolonged sampling periods, and challenges in providing timely information on garlic cultivation [3]. The maturation of the remote sensing industry offers a solution, enabling rapid, large-scale, and cost-effective monitoring of garlic cultivation through satellite remote sensing data [4,5,6]. In China’s primary garlic-producing regions, mixed cropping of garlic with other crops is prevalent, resulting in numerous mixed pixels in low spatial resolution remote sensing imagery. This phenomenon poses a challenge to accurately identifying garlic cultivation areas. Therefore, the utilization of high spatial resolution imagery emerges as the preferred approach for precisely delineating the spatial distribution of garlic within these key production zones [7,8,9]. Furthermore, the predominant mixed cropping of garlic with winter wheat and rape in the study area yields similar spectral signatures during specific stages of vegetation growth, leading to the manifestation of “foreign body in the same spectrum” in multispectral images [10]. Consequently, discriminating between garlic, rape, and winter wheat solely based on multispectral images at a particular stage of the vegetation growth cycle becomes inherently challenging, significantly impeding garlic identification accuracy.

The duration and diversity of time series data for agricultural crops are directly correlated with the complexity of ground vegetation types [11,12,13,14]. Leveraging remote sensing imagery provided by satellites to form time series data enables the rapid and precise extraction of crop types. Time series imagery has found widespread application in classifying both single and multiple crop types [15,16,17,18], predominantly by constructing the NDVI (normalized difference vegetation index) time series from remotely sensed optical imagery to serve as an indicator of crop phenology [19,20,21,22,23]. However, time series data typically encompass a significant period during which environmental factors are complex and variable, and the crop state undergoes changes. Therefore, relying solely on NDVI as an indicator may not be universally applicable across all environments. To establish a more robust basis for classification, scholars have explored alternative indices to mitigate environmental influences and thereby enhance classification accuracy. For instance, the Enhanced Vegetation Index (EVI) [24] and Modified SAVI (Modified Soil Adjusted Vegetation Index, MSAVI) [25] have been investigated. Through the integration of various indices to construct new time series data, it becomes feasible to more accurately reflect crop growth dynamics.

Although significant progress has been made in vegetation classification using time series data from either active or passive remote sensing images, both modalities possess distinct advantages and limitations. Obtaining high spatial and temporal resolution time series data exclusively from one type of remote sensing image remains challenging. Optical imagery, often utilized for constructing time series curves over the crop growth period, is susceptible to cloud cover and shadowing, diminishing its usability [26,27]. On the other hand, SAR is sensitive to environmental factors such as rain, snow, and atmospheric humidity, potentially compromising image quality under adverse weather conditions. Chen et al. demonstrated the potential of combining active and passive remote sensing for garlic research by extracting garlic using both stepwise and simultaneous approaches, yielding favorable results [28].

While prior classification studies have primarily focused on enhancing classification accuracy [29,30], insufficient attention has been devoted to exploring the factors influencing classification accuracy. However, investigating these factors may offer novel insights for improving classification accuracy. There is no doubt that ground crop cultivation characterized by good regularity and uniformity is advantageous for remote sensing image classification. However, due to the unique circumstances in our country, it is often challenging to achieve this ideal situation. There are differences in the cultivation patterns between garlic and winter wheat in the study area, with garlic cultivation being less regular and more dispersed. To accurately describe and understand the impact of garlic dispersion on classification accuracy, further investigation into the degree of garlic dispersion is warranted.

In recent years, the application of machine learning methods in remote sensing classification has seen rapid development. Among these methods, random forest (RF), support vector machine, extreme gradient boosting, decision tree, and deep learning are commonly employed. Random forest, in particular, has demonstrated excellent classification performance in remote sensing classification and has garnered widespread recognition from previous researchers [31,32].

While there exists a growing body of literature and practical experience in land cover classification using time series image sets [33,34], the focus of interest in this study lies in the identification and extraction of individual crops based on time series data. Unlike most past studies, the study area under consideration includes the presence of a same-season crop (winter wheat) with a phenological period very similar to the crop of interest, with mixing of the two being common. Distinguishing between these two types of crops using only optical time series data, strongly correlated with phenological characteristics, poses a significant challenge. Moreover, the complexity and diversity of vegetation types in the study area make it prone to the occurrence of the same spectral signature in remote sensing images. Consequently, achieving the desired extraction accuracy by using remote sensing images at a specific moment to differentiate crops with similar phenological characteristics is difficult [35,36]. Additionally, traditional classification methods based on time series data have limitations and are particularly challenging in distinguishing between different crops with similar phenological characteristics.

In this paper, we present a study aimed at extracting garlic in the study area through time series data, with a focus on the cultivation area and distribution of garlic, without making detailed distinctions between crops other than garlic. Situated in the eastern part of China, the study area predominantly features garlic and winter wheat as the main winter crops. The aims of this study were to address the following three research questions: (1) whether the superposition or splicing of temporal information can help improve the classification accuracy of garlic; (2) whether the superposition of optical data and Sentinel-1 SAR data can help identify garlic accurately; (3) whether the constructed garlic fragmentation index can effectively express the fragmentation of garlic distribution in different regions and help analyze the reasons for the differences in garlic extraction accuracy in different regions.

2. Data and Methodology

2.1. Study Area

The study area is situated within the North China Plain, spanning from 114° E to 121° E and from 32° N to 40° N. Characterized by low terrain, this region is conducive to the growth of agricultural crops, rendering it one of the primary grain-producing areas in China. Six primary garlic-producing regions were selected for this study: Jinxiang County, Laiwu District, Lanling County, Pizhou City, Qixian County, and Zhongmou County. These areas span a geographical range from 113.5° E to 118.5° E and 33.7° N to 36.6° N, covering a total area of 8597 km². These regions experience a temperate monsoon continental climate and typically engage in a two-season crop rotation system. The predominant crops cultivated include garlic, winter wheat, vegetables, and peanuts. Among these, garlic cultivation encompasses 228,500 ha, representing 26.58% of the total area. Garlic planting commences in early October each year, with growth accelerating in April of the following year as temperatures rise, culminating in harvesting in June. Conversely, winter wheat is sown in mid-October, exhibiting slightly slower growth than garlic, and harvested in June as well. Mixed cropping of the two is widespread in the study area, with most areas featuring a blend of the two crops in smaller units. Given the study’s focus on garlic distribution and acreage, other ground cover types are not meticulously distinguished. Vegetation other than garlic (such as winter wheat, vegetables, peanuts, peppers, or other crops) is uniformly classified as “other”. The fertility period of garlic in the study area is detailed in Figure 1. The distribution of the study area is shown in Figure 2.

2.2. Data

2.2.1. Field Survey Data

Between 27 January 2023 and 3 February 2023, researchers conducted a week-long survey in the study area to comprehend garlic and winter wheat cultivation practices. The survey route commenced in Tai’an, continuing to Jinxiang, then to Qixian and Zhongmu, followed by Pizhou and Lanling, and concluding in Laiwu. Before departure, points of interest within each county of the study area were selected based on satellite remote sensing imagery and time series data, with the sampling route planned according to the distribution of these points. The counties in the study area are major garlic-producing regions in China, where garlic and winter wheat cover nearly all cultivated areas, occupying vast expanses of the landscape, with minimal field weeds. Small patches of vegetables and other crops adjacent to garlic fields present challenges to classification accuracy. Trees are mainly concentrated near mountains, towns, villages, and along the roads between fields. Vegetables and other vegetation are predominantly found around villages, with much of it grown in greenhouses.

During the survey, samples of various vegetation types were collected, with a particular focus on garlic and winter wheat. Given the vast expanse of the study area and the intricacy of its terrain, achieving uniform sample collection was challenging. Consequently, in addition to field sample points, visual samples were also selected. To ensure the accuracy of visual samples, high-resolution remote sensing images were initially employed to classify the sample categories, followed by integration with Sentinel-2 NDVI time series remote sensing images for final categorization. Subsequently, the accuracy of the data was verified via meticulous comparison of all samples with high-resolution remote sensing images. Ultimately, the samples utilized in this study were compiled from ground-truthing samples and visual samples.

2.2.2. Synthetic Aperture Radar Data

The Sentinel-1 (S1) data are obtained from the “COPERNICUS/S1_GRD” remote sensing image collection on Google Earth Engine (GEE). Within this study, the Sentinel-1 radar image is directly accessed via the GEE platform, having undergone official preprocessing procedures, including updating orbital metadata, removal of GRD boundary noise, elimination of thermal noise, radiometric calibration, and ortho-correction. The data mode employed for this study is the IW mode, which is well-suited for remote sensing studies of land surfaces, particularly with VV + VH polarization.

2.2.3. Remote Sensing Image Data

Utilizing GEE as a remote sensing big data processing platform enables direct access to Sentinel-2 and Sentinel-1 remote sensing images, circumventing the laborious storage and processing requirements on local computers, thanks to its cloud computing capability [37,38]. The Sentinel-2 (S2) remote sensing image utilized in this study is sourced from the Sentinel-2 image dataset on the GEE platform with product class 2A. This dataset undergoes ortho-correction and atmospheric apparent reflectivity processing by the GEE official, and the imagery retrieved from the cloud platform corresponds to the “COPERNICUS/S2_SR_HARMONIZED” product, representing surface reflectance data. In this investigation, five bands of Sentinel-2 remote sensing satellite data are employed, namely, red, green, blue, near-infrared, and short-wave infrared, with a ground resolution of 20 m for the short-wave infrared band and 10 m for all other bands.

2.3. Methodology

The technical methodology of the study, illustrated in Figure 3, encompasses the following steps: (1) conducting a field survey in conjunction with Sentinel-2 NDVI curves to characterize the growth dynamics of garlic and winter wheat; (2) utilizing Sentinel-2 NDVI time series data to extract crops during the winter season within the study area employing random forests; (3) identifying a specific time point within the vegetation growth cycle and plotting time series curves using two different vegetation indices before and after the identified time point, subsequently stitching them together to generate comprehensive crop time series data; (4) formulating diverse combinations of classification features based on time series data from primary passive remote sensing bands or indices, employing random forests to reclassify winter crops using several different feature combinations, and validating the results and accuracy; (5) conducting a comparative analysis of the classification accuracy and outcomes for each combination of classification features and determining the optimal combination of features; and (6) utilizing the constructed garlic fragmentation index to assess the fragmentation of the classification results in each region. These steps represent a systematic approach to enhance the accuracy and reliability of winter crop classification, with a particular focus on garlic, within the study area.

2.3.1. Sentinel-2 Time Series Curve Composition

The Sentinel-2 NDVI curves time series were obtained from October 2021 to June 2022 using the GEE Sentinel-2 dataset and combined with field survey data to accurately understand the climatic characteristics of garlic, winter wheat, and other crops. The Sentinel-2 NDVI was calculated using the following formula:

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}={\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}-{\mathsf{\rho }}_{\mathrm{R}\mathrm{E}\mathrm{D}}}{{\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}+{\mathsf{\rho }}_{\mathrm{R}\mathrm{E}\mathrm{D}}}}$$

(1)

where ${\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}$ is the reflectance in the near-infrared band and ${\mathsf{\rho }}_{\mathrm{R}\mathrm{E}\mathrm{D}}$ is the reflectance in the red band.

Figure 4 shows that October corresponds to the sowing stage, while June marks the harvest period for both garlic and winter wheat, which are planted and harvested nearly simultaneously. Notably, the NDVI patterns of garlic and winter wheat exhibit similarity, with considerable overlap between the upper and lower boundaries of the two crops, posing challenges in their differentiation solely based on NDVI values from a single day. It is noteworthy that the overall changes in winter wheat were smoother relative to garlic, consistently displaying higher NDVI values than garlic throughout most of the growing season. Additionally, although both crops reach their peaks at nearly the same time, garlic demonstrates more rapid fluctuations in its rise and fall. These distinctive variations aid in facilitating a more accurate differentiation between the two crops.

Due to various factors such as weather conditions, obtaining high-quality time series data for all image elements can be challenging. To maximize data quality, this study employs a two-step optimization process for time series data. Firstly, monthly median values of each image element are used to generate time series data. Subsequently, a method utilizing historical data is employed to correct data anomalies within the vegetation growth period, thereby enhancing the data quality. This method identifies anomalies in each image value through harmonic regression analysis and replaces them with normal historical data, thereby improving the reliability of the time series data.

The method for splicing the time series data involves using MSAVI or SAVI to calculate the time series data for the study area during periods with low NDVI values. For periods with higher NDVI values, NDVI, GNDVI, or EVI are used instead. For instance, MSAVI time series data from October 2021 to January 2022 are combined with EVI time series data from February 2022 to June 2022 to create complete time series data, referred to as double index time series data (DITS). A time node between January and March was selected for splicing the DITS data based on NDVI time series data and field surveys. Several potential time nodes were tested to determine the optimal one, and through comparison of classification results using different time nodes, January was identified as the most suitable time node for DITS.

The formulas for MSAVI, SAVI, EVI, and GNDVI for Sentinel-2 are provided below, respectively:

$$\mathrm{M}\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}={\displaystyle \frac{1}{2}}\left[2\ast {\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}+1-\sqrt{{\left(2\ast {\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}+1\right)}^2-{8\ast \left({\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}-{\mathsf{\rho }}_{\mathrm{R}\mathrm{E}\mathrm{D}}\right)}^2}\right]$$

(2)

$$\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}={\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}-{\mathsf{\rho }}_{\mathrm{R}\mathrm{E}\mathrm{D}}}{{\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}+{\mathsf{\rho }}_{\mathrm{R}\mathrm{E}\mathrm{D}}+\mathrm{L}}}\left(1+\mathrm{L}\right)$$

(3)

$$\mathrm{E}\mathrm{V}\mathrm{I}={\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}-{\mathsf{\rho }}_{\mathrm{R}\mathrm{E}\mathrm{D}}}{{\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}+6{\mathsf{\rho }}_{\mathrm{R}\mathrm{E}\mathrm{D}}-7.5{\mathsf{\rho }}_{\mathrm{B}\mathrm{L}\mathrm{U}\mathrm{E}}+1}}$$

(4)

$$\mathrm{G}\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}={\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}-{\mathsf{\rho }}_{\mathrm{G}\mathrm{R}\mathrm{E}\mathrm{E}\mathrm{N}}}{{\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}+{\mathsf{\rho }}_{\mathrm{G}\mathrm{R}\mathrm{E}\mathrm{E}\mathrm{N}}}}$$

(5)

where ${\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}$ is the reflectance in the near infrared band, ${\mathsf{\rho }}_{\mathrm{R}\mathrm{E}\mathrm{D}}$ is the reflectance in the red band, ${\mathsf{\rho }}_{\mathrm{G}\mathrm{R}\mathrm{E}\mathrm{E}\mathrm{N}}$ is the reflectance in the green band, L is the soil conditioning coefficient, and ${\mathsf{\rho }}_{\mathrm{B}\mathrm{L}\mathrm{U}\mathrm{E}}$ is the reflectance in the blue band.

2.3.2. Sentinel-1 Time Series Curve Composition

Figure 5 illustrates the Sentinel-1 time series curves of garlic and winter wheat based on 5000 sample pixels. Figure 5a presents the time series data of VH/VV for garlic and winter wheat, while Figure 5b displays the time series data of VH and VV for garlic and winter wheat. Analysis of Figure 5b indicates that the time series data of VV for garlic and winter wheat exhibit a similar pattern of change to the time series data of VH, with more significant differences observed around the month of April, particularly in the troughs and the timing of their occurrence. Conversely, examination of Figure 5a suggests that the trend of the time series data of VH/VV for garlic and winter wheat is similar, although the difference is significant from February to May. Notably, the peak for garlic occurs around February, whereas for winter wheat, it is around April.

2.3.3. The Random Forest Classifier

In this study, we utilized the random forest algorithm for classification. This method reduces the overfitting risk often associated with individual decision trees by constructing multiple trees and aggregating their predictions, resulting in strong robustness [39,40]. The algorithm effectively handles high-dimensional data, maintaining powerful classification performance, even with a large number of features, thus mitigating the curse of dimensionality [41]. Its random selection of samples and features further enhances its resilience to noise, improving the accuracy when dealing with noisy datasets. Although random forest is typically regarded as a “black box” model, it offers insights into feature importance, providing some level of interpretability in the decision-making process.

Through experimentation, we determined that optimal results were achieved by configuring the RF classifier with 200 trees. The application of RF to crop classification has demonstrated its efficacy, yielding favorable outcomes [42,43].

2.3.4. Garlic Extraction

Initially, winter crop distribution maps (WCMs) were derived using RF classification based on Sentinel-2 NDVI time series data obtained from the GEE cloud platform (see Section 2.3.1). Subsequently, the WCM was segmented based on different feature combinations to produce garlic distribution maps (GDMs) for each region.

A total of nine distinct feature combinations were utilized to classify the WCM using RF, resulting in the generation of garlic distribution maps (GDMs) for each region. The accuracy of the classifications obtained from all feature combinations was compared to determine the optimal combination for each county and to generate the planting distribution map and area. The classification feature combinations encompass various time series data, with each combination demonstrating a certain level of similarity. Comparing the classification accuracy of different feature combinations allows for the analysis of the importance of specific information.

In this study, we designate x-y (where x = me, mg, mn, se, sn, n, e and y = sm, s, m, swir) as a feature combination, where x represents the vegetation index time series data included in the combination, and y represents other spectral time series data and SAR time series data. We categorize vegetation index time series data, spectral band data, and microwave data as the three types of data considered in this study. Typically, spectral bands comprise the four bands of red, green, blue, and shortwave infrared (multi-bands), while the three microwave time series data of VH, VV, and VH/VV are collectively referred to as SAR. In the feature combination, x comprises two letters to denote that the vegetation index time series data of the combination are derived from the DITS, while one letter signifies the time series data of a single index. Each letter in y then represents a feature, where “swir” denotes the shortwave infrared band only. For example, the feature combination labeled “me-sm” includes time series data from the DITS of MSAVI and EVI and SAR time series data, as well as time series data from the red, green, blue, and shortwave infrared bands. In summary, the classification feature combinations may encompass time series data from MSAVI, SAVI, EVI, NDVI, GNDVI, SAR, and multi-bands. The timing data included in each feature combination is depicted in

Table 1. Combination of classification characteristics.

Feature Sets	MSAVI	SAVI	EVI	GNDVI	NDVI	SAR	Multi-Bands
① me-sm	√	×	√	×	×	√	√
② mg-sm	√	×	×	√	×	√	√
③ mn-sm	√	×	×	×	√	√	√
④ sn-sm	×	√	×	×	√		√
⑤ e-sm	×	×	√	×	×	√	√
⑥ n-sm	×	×	×	×	√	√	√
⑦ n-s	×	×	×	×	√	√	×
⑧ n-	×	×	×	×	√	×	×
⑨ n-sswir	×	×	×	×	√	√	SWIR (only)

In the table, MSAVI, SAVI, EVI, GNDVI, and NDVI refer to the modified soil-adjusted vegetation index, soil-adjusted vegetation index, enhanced vegetation index, green normalized difference vegetation index, and normalized difference vegetation index, respectively. SAR represents time series data from vertical–vertical (VV) and vertical–horizontal (VH) polarizations, along with their ratio. Multi-bands refer to the time series data of the red, green, blue, and shortwave infrared bands. ‘SWIR (only)’ indicates that the feature combination n-sswir includes only the SWIR band from the four bands represented by multi-bands in the time series data. A ‘√’ indicates the presence of the corresponding categorical feature in the feature combination, while a ‘×’ denotes its absence.

. It is noteworthy that in feature combinations containing two vegetation indices, the index positioned at a later point in the table is situated in the first half of the DITS, while the other is located in the second half.

2.3.5. Accuracy Verification

The validation samples for the WCM were obtained from field-collected samples and randomly selected points. To ensure objectivity in accuracy validation and to minimize subjectivity in sample selection, the visual validation samples for the GDM were generated from random points within the study area. The methodology used to determine the attributes of these random points was consistent with the approach used for determining the attributes of visual samples in the training set (see Section 2.2.1). The details of the GDM validation points, regarded as ground truth data, are presented in

Table 2. Garlic distribution map verification points by county.

Region	Class Results	Jinxiang	Laiwu	Lanling	Pizhou	Qixian	Zhongmou
Ground Truth (Pixels)	Garlic	321	149	210	343	236	196
	Other	255	355	305	349	280	285
	All	576	504	515	692	516	481

.

All classification results underwent validation by comparing the confusion matrix with the samples, a method of accuracy validation conducted at the pixel level and commonly employed in accuracy assessment. Classification accuracy was evaluated by comparing the validation predictions with the selected validation points and conducting a consistency test.

2.3.6. Garlic Fragmentation Analysis

To comprehensively analyze the factors contributing to variations in accuracy across different regions, this study leverages previous research findings [44,45] and integrates the characteristics of the study area to evaluate garlic fragmentation using the GFI. The GFI quantifies the fragmentation of garlic cultivation areas, with higher values indicating greater fragmentation [46]. The formula is provided below:

$$\mathrm{G}\mathrm{F}\mathrm{I}=\mathrm{n}/\mathrm{A}$$

(6)

where GFI represents the garlic fragmentation index, n denotes the number of plots planted with garlic, and A is the total area planted with garlic in the sample area.

Ten square samples, each with a side length of one kilometer, were randomly selected from each county in the study area to assess the degree of garlic fragmentation. The fragmentation of the garlic crop in each county was evaluated based on the GFI of the samples. To more accurately reflect garlic fragmentation, the random area range of Laiwu and Pizhou was artificially constrained.

3. Results

3.1. Winter Crop Classification Results

In this study, NDVI time series data were utilized to extract winter crops, with previous studies confirming the reliability of the NDVI for vegetation extraction [47,48,49]. The area under crop cultivation in Pizhou is 1618.03 km², constituting 77.62% of the county’s total area, making it the largest in the study area. In contrast, Laiwu has 611.78 km² under crop cultivation, accounting for 35.18% of its total area, the smallest in the study area. Figure 6 illustrates the WCM for each county within the study area. The classification results indicate that vegetation distribution in the study area predominantly extends outward, centered on anthropogenic activity intensity. Further details regarding the WCM are provided in

Table 3. Results of accuracy validation for winter vegetation map.

Region	Winter Crops		Other		Ground Truth (Pixels)		Overall
Results	User’s Accuracy	Producer’s Accuracy	User’s Accuracy	Producer’s Accuracy	Winter Crops	Other	Overall Accuracy	Kappa
Jinxiang	99.95	99.86	99.37	99.80	3804	2979	97.85	0.961
Laiwu	98.27	97.90	97.13	92.12	1335	1332	96.06	0.921
Lanling	97.26	96.24	97.31	95.72	5361	4104	98.43	0.964
Pizhou	98.67	98.21	99.95	91.26	6487	6221	94.90	0.903
Qixian	97.62	99.28	96.26	96.12	3016	3612	94.50	0.893
Zhongmou	97.23	92.03	96.52	96.89	3648	3263	92.82	0.878

.

3.2. Extraction Accuracy

First, this study successfully extracted winter vegetation using remote sensing imagery, while all other vegetation types were categorized as “Other”. The results in Table 3 demonstrate high accuracy for winter vegetation classification models across all regions (all above 92% overall). The producer and user accuracies for winter crops were over 96% (except for Zhongmou County, which had a producer accuracy of 92%), providing a reliable foundation for further garlic classification.

Secondly, by classifying the WCM, we generated the GDM for each region. The classification models employed different combinations of classification features, and the resulting GDM classification accuracy is depicted in

Table 4. Tables of overall accuracy and Kappa coefficient for garlic extraction.

Region	Jinxiang		Laiwu		Lanling		Pizhou		Qixian		Zhongmou
Results	OA	Kappa	OA	Kappa	OA	Kappa	OA	Kappa	OA	Kappa	OA	Kappa
1.me-sr	92.50	85.18	92.72	85.82	94.16	88.31	87.32	70.92	90.95	84.04	94.72	91.07
2.mg-sr	92.12	84.47	91.28	82.79	93.57	87.11	90.70	82.39	89.39	81.32	93.83	89.57
3.mn-sr	92.12	84.47	94.54	83.41	93.58	87.15	88.26	77.84	89.55	81.56	94.27	90.28
4.sn-sr	84.24	84.79	92.54	85.37	93.55	87.11	87.81	76.98	89.74	81.86	94.49	90.65
5.e-sr	92.50	85.18	93.48	87.31	92.80	85.67	87.81	76.84	89.98	82.39	95.39	92.19
6.n-sr	92.31	84.79	94.16	82.46	93.37	86.73	88.41	78.11	89.94	82.21	94.95	91.43
7.n-r	90.80	81.70	90.41	81.47	94.55	89.00	87.81	77.00	89.17	80.95	88.16	80.13
8.n-	89.49	79.07	86.52	74.40	90.63	81.52	87.79	76.92	84.09	71.60	77.41	62.97
9.n-sswir	89.31	78.64	89.36	79.28	90.86	81.48	88.21	77.70	87.82	78.53	88.82	81.44

In the table, “OA” represents overall accuracy, with bold values indicating that this method ranks among the top three for accuracy in the corresponding region.

and

Table 5. Tables of user’s accuracy and producer’s accuracy for garlic extraction.

Region	Jinxiang		Laiwu		Lanling		Pizhou		Qixian		Zhongmou
Results	UA	PA	UA	PA	UA	PA	UA	PA	UA	PA	UA	PA
1.me-sr	91.77	97.20	87.58	94.37	94.61	91.91	95.58	86.32	92.28	91.10	94.50	96.41
2.mg-sr	91.45	96.57	84.91	93.75	92.82	92.38	94.41	93.86	90.25	90.25	92.58	96.39
3.mn-sr	91.45	96.57	85.81	93.01	93.66	91.43	93.05	90.01	91.74	89.41	93.94	96.37
4.sn-sr	91.50	87.20	87.10	95.07	93.27	92.38	93.03	89.77	90.68	90.68	94.50	96.41
5.e-sr	91.77	97.20	87.26	95.80	91.08	92.38	93.33	90.01	91.45	90.68	96.39	96.89
6.n-sr	91.50	97.20	85.06	92.91	93.24	91.91	92.77	90.64	90.72	91.10	95.43	96.41
7.n-r	89.14	97.20	82.42	95.11	96.5	91.91	92.45	89.47	90.56	89.41	86.85	94.87
8.n-	88.35	96.88	78.19	91.49	91.79	90.48	92.19	89.50	81.50	87.71	79.29	92.78
9.n-sswir	87.85	96.89	81.01	90.78	93.30	86.19	92.77	90.01	87.19	89.41	89.86	95.39

“UA” refers to user accuracy, and “PA” denotes producer accuracy, with bolded values indicating that this method ranks among the top three for accuracy in the corresponding region.

. The results indicate that all combinations of classification features yielded satisfactory classification results (overall accuracy consistently above 84%). Within each district, multiple feature combinations demonstrated the potential for achieving high classification accuracy. Among the various counties in the study area, Zhongmou and Laning counties attained higher accuracy levels, followed by Jinxiang and Laiwu counties, while Pizhou and Qi counties exhibited more average classification outcomes.

Based on Table 4, the Kappa of each classification feature across different regions exhibit a similar trend to the overall accuracy change, with all achieving optimal results in Laning County and lower accuracy in Pizhou and Qixian counties. In most cases, as the number of features within a combination of classification features decreases, both classification accuracy and Kappa coefficients decrease (except in Lanling County), with decreases of less than 10%.

Among the various feature combinations, the inclusion of DITS in the feature combination offers some advantages over the NDVI, a benefit reflected in most areas (except Pizhou). However, when replacing the NDVI with the EVI in the feature combinations, this advantage becomes negligible, as both the feature combinations with DITS added and E-SM obtained the best results in their respective regions, showing no significant advantage. Nevertheless, this does not discount the utility of DITS, as the feature combination with DITS added was able to achieve better producer accuracy versus user accuracy in extracting garlic in most regions.

As evidenced by the classification results of the feature combination n-m vs. n-, the addition of spectral information significantly enhances accuracy, with the Kappa improving by more than 1% and even more than 6% in most regions. Relative to N-SSWIR, N-SM contains three additional RGB bands, resulting in a relatively minor difference in accuracy across regions compared to that between N-R and N- (except for Jinxiang, where the former’s accuracy difference is less than half of the latter’s in all other regions). The shortwave infrared band contributes more significantly to improving classification accuracy compared to the RGB band.

N-sm includes more SAR time series data (VH, VV, and VH/VV) compared to n-m, leading to better accuracy in most areas. Except for a 3% decrease in the Kappa in Laning County, Laiwu and Pizhou improved by approximately 1%, and all other areas exceeded 2%. The integration of active and passive time series remote sensing data has a more pronounced effect on enhancing classification accuracy in most areas. Overall, e-sm demonstrates good stability and achieves favorable results in most regions with strong generalizability. DITS exhibits advantages in enhancing accuracy in certain areas, but different combinations of vegetation indices require further testing. Moreover, the combination of active and passive remote sensing time series data can significantly enhance accuracy and is applicable across various regions.

3.3. Garlic Extraction Results

Figure 7 illustrates the GDM classification outcomes for the study area, generated from the optimal feature combinations identified in the preceding section. According to these classification results, the total area under garlic cultivation amounts to 1845.11 km². Garlic cultivation exhibits a dual characteristic of concentration and dispersion. Concentration is evident in the predominant cultivation of garlic in specific regions of the study area, while its cultivation is sparse or absent in other areas. The dispersed nature of garlic cultivation is exemplified by its frequent intercropping with winter wheat or other crops in regions where it is cultivated extensively, rather than being grown in isolation.

Among the six counties in the study area, Pizhou has the largest garlic cultivation area, covering 542.80 km². Figure 7 indicates that garlic is cultivated across the entirety of Jinxiang and Qixian counties, whereas in other counties, garlic cultivation is more concentrated in specific areas. Additionally, other types of vegetation exhibit a more continuous spatial distribution.

3.4. Classification Contribution

Table 6. Classification contribution scale.

Ranking	Jinxiang	Laiwu	Lanling	Pizhou	Qixian	Zhongmou
1	22/05 RED	22/02 SWIR	22/02 SWIR	22/05 GNDVI	22/03 SWIR	22/03 EVI
2	22/05 GREEN	22/02 BLUE	22/02 BLUE	21/10 MSAVI	22/06 EVI	22/02 BLUE
3	22/06 RED	21/12 EVI	22/03 SWIR	22/06 GNDVI	22/04 SWIR	22/05 GREEN
4	22/04 VV	22/05 RED	22/06 NDVI	22/04 GNDVI	21/10 MSAVI	21/12 EVI
5	22/06 GREEN	22/02 GREEN	22/03 BLUE	21/11 MSAVI	22/05 SWIR	22/06 GREEN
6	22/05 BLUE	21/11 EVI	22/04 SWIR	21/12 MSAVI	22/06 SWIR	22/02 GREEN
7	22/05 EVI	22/03 SWIR	21/10 NDVI	22/02 GNDVI	22/03 EVI	22/02 SWIR
8	22/05 SWIR	22/06 EVI	22/03 GREEN	22/03 GNDVI	22/04 RED	22/06 SWIR
9	22/06 BLUE	22/06 RED	22/04 RED	22/01 MSAVI	22/04 EVI	22/03 GREEN
10	22/06 SWIR	22/02 RED	22/02 GREEN	21/10 VH/VV	22/04 GREEN	22/03 SWIR
11	22/05 VV	22/05 EVI	22/06 SWIR	21/11 VH/VV	22/06 GREEN	22/04 GREEN
12	22/04 GREEN	22/02 EVI	22/05 SWIR	22/03 VH/VV	22/01 MSAVI	22/03 BLUE
13	22/04 RED	22/04 SWIR	22/06 RED	21/12 VH/VV	22/04 BLUE	22/02 EVI
14	22/04 VH	22/03 GREEN	22/03 RED	22/04 VH/VV	22/06 BLUE	22/04 SWIR
15	22/05 RED	22/06 GREEN	22/06 BLUE	22/05 VH/VV	22/05 GREEN	22/04 EVI

In the table, RED, GREEN, BLUE, and SWIR represent different bands of Sentinel-2 remote sensing imagery. VV and VH refer to the horizontal–horizontal polarization and horizontal–vertical polarization of Sentinel-1, respectively. VH/VV denotes the ratio between the VH and VV polarizations.

illustrates the ranking of classification contributions from the feature combinations that yielded the best classification results for each region. Due to the extensive number of features, only the top fifteen classification contributions are provided. The table highlights the significant contribution of SWIR to classification accuracy, with multiple SWIR features included (up to 56 features when considering all classified features). This finding aligns with previous observations regarding the role of SWIR features in enhancing accuracy. Notably, SAR features are scarcely represented in the table, except for Jinxiang and Pizhou, indicating limitations in using SAR features for crop species determination.

Furthermore, the table incorporates various vegetation indices, suggesting their applicability for garlic extraction across different regions. Among these, the EVI and MSAVI are more frequently observed, indicating their broader suitability for classification. The presence of different vegetation indices in distinct regions underscores the effectiveness of DITS in enhancing garlic extraction accuracy.

3.5. Garlic Fragmentation

Table 7. Table of garlic crushing by county.

Region	GFI	UA	Example	Example Coordinate
Jinxiang	0.1175	91.77		116°23′03″ E 34°57′36″ N
Laiwu	0.8261	87.26		117°28′57″ E 34°16′40″ N
Lanling	0.1001	92.82		118°05′41″ E 34°45′36″ N
Pizhou	0.3378	94.41		117°46′38″ E 34°57′36″ N
Qixian	0.1691	92.28		114°53′37″ E 34°32′05″ N
Zhongmou	0.1180	96.39		113°57′22″ E 34°50′24″ N

GFI in the table refers to garlic fragmentation and UA represents the consumer accuracy of garlic extraction.

presents the fragmentation degree calculated based on the best classification results for each region. Overall, Lanling, Zhongmou, and Qixian exhibit lower fragmentation degrees. Although there is a certain correlation between the fragmentation index and classification accuracy, this relationship is not linear. When the difference of the fragmentation degree is more than 0.2, the area with the larger fragmentation degree can be predicted, and at this time the extracted classification accuracy will be reduced accordingly. However, when the change in the degree of fragmentation is at a relatively low level, the impact of the change in the degree of fragmentation on the classification results will be reduced.

As depicted in Table 7, regions with lower fragmentation demonstrate greater stability in both user accuracy and producer accuracy. Despite Laiwu district exhibiting higher overall accuracy, its garlic identification accuracy is lower due to a higher prevalence of other vegetation types. The validation points are more distributed across other vegetation types in this district, contributing to the inflated overall accuracy.

4. Discussion

4.1. Advantages of Data Preprocessing

The study area for this experiment encompasses a wide geographical expanse, with an east-west longitude difference of 7°, necessitating approximately three image strips from Sentinel-2 satellites to comprehensively cover the area. However, due to cloud cover effects, the intervals between usable image acquisitions from different strips may exceed ten days [50]. Optical imagery remains the primary reference for classification in this experiment, necessitating a hierarchical processing approach. Despite compiling time series data based on monthly mean values, anomalies persist in many pixel values [51]. To address this, harmonic analysis was employed to identify anomalies in the vegetation index relative to the previous two years’ data and rectify them, thereby reducing the impact of cloud or other noise artifacts and improving the reliability of the time series data. Experimental results confirm the efficacy of these methods in providing more dependable data for garlic classification.

4.2. Advantages of DITS

Previous studies on crop extraction from time series data have indeed affirmed that superior outcomes can be achieved by identifying crops using such data [52,53,54,55]. The diverse vegetation within the study area, coupled with the presence of winter wheat sharing spectral and phenological similarities with garlic, presented a formidable challenge for garlic extraction. As depicted in Figure 4, garlic and winter wheat exhibit considerable similarity in their NDVI during the phenological period, rendering it arduous to discern the two solely through remote sensing imagery on a given day during the growing season. Field surveys have indicated that the foliar index of garlic during the same period is lower than that of winter wheat, theoretically suggesting that the DITS of garlic should also be smaller than that of winter wheat. However, as evident from Figure 4, the NDVI of garlic and winter wheat closely align before January, with both exhibiting lower foliar indexes in the early growth stage, thereby complicating differentiation between the two using NDVI time series data at the onset to achieve the desired effect.

From the classification results, it is apparent that the inclusion of additional classification features enhances the accuracy of classification outcomes. This is evident from Table 4 and Table 6, where SWIR bands play a pivotal role in improving accuracy, likely attributed to the substantial difference in leaf water content between garlic and winter wheat from February to June [56,57]. Both the SAVI and MSAVI mitigate the influence of the vegetation-soil environment on the vegetation spectral characteristics in the early stages of crop growth [58,59], with the MSAVI exhibiting dynamic adaptability to changes in vegetation density and soil effects. Hence, employing the MSAVI as the first half of DITS during the early stages of crop growth could prove more beneficial in discriminating garlic from winter wheat, a hypothesis validated by the results in Table 5 and Table 6. The near-simultaneous maturity period of garlic and winter wheat, when crop density is higher and soil background reflection exerts less influence, renders NDVI values of both crops convergent, posing challenges to differentiation using the NDVI. Conversely, the EVI offers a distinct advantage in observing denser vegetation and is less susceptible to noise [60], thus utilizing the EVI as the latter half of DITS may facilitate better discrimination between garlic and winter wheat. Notably, as depicted in Table 5, utilizing time series data comprising the EVI combined with microwave and several single-band time series data yields improved results in several regions, underscoring the significance of the EVI in classification contributions, as highlighted in Table 6.

4.3. Advantages of Combining Active and Passive Remote Sensing

Optical images are frequently contaminated by clouds and their shadows, while SAR images serve as a dependable data source in nearly all weather conditions [61]. Processed microwave remote sensing data exhibit a sharp response mechanism to plant structure [62]; thus, the inclusion of polarized data would enhance the accuracy of feature classification. As depicted in Figure 5, a distinct disparity exists between the VH/VV of garlic and winter wheat, with the garlic VH/VV peak occurring earlier and being of larger magnitude compared to that of winter wheat. Therefore, VH/VV, by amalgamating the characteristics of VH and VV, can streamline computations and furnish additional information to differentiate between the two, thus aiding in garlic extraction. Although the addition of SAR features enhances classification accuracy, as stated by Veloso et al., observed backscatter comprises a blend of ground backscatter perturbed by soil moisture and surface roughness and vegetation backscatter altered by the 3D structure of vegetation [63]. Consequently, SAR features may not be efficacious in discerning the class of garlic. Moreover, given that the resolution of the SWIR band of Sentinel-2 is 20 m, which is relatively low compared to other visible light bands, it may impede the discrimination between garlic and winter wheat plots on the ground in most instances. Hence, the significance of SAR features in garlic extraction is underscored in this study, while the amalgamation of active and passive remote sensing techniques can enhance classification accuracy.

4.4. The Role of the Fragmentation Index

To assess the reliability of the GFI, the index was applied to multiple regions with varying environmental conditions to determine its consistency. Results showed that the GFI remained stable across different datasets, indicating its robustness. Additionally, previous studies [46] have successfully utilized similar indices, further supporting its reliability. However, it should be noted that the GFI may be influenced by factors such as data resolution and image quality. In areas with lower resolution or significant cloud cover, the accuracy of the index may be reduced.

Combining Table 5 and Table 6, it is evident that the fragmentation index holds predictive value for extraction outcomes, with less accurate results observed in areas exhibiting greater fragmentation. Thus, analyzing ground fragmentation allows for advance prediction of extraction accuracy. Additionally, examining garlic extraction results and analyzing actual planting fragmentation can inform garlic policy formulation, thereby reducing planting costs.

4.5. Summary and Future Directions

A comparison of classification results indicates that, except in Pizhou where the accuracy of the commonly used NDVI time series data was nearly equivalent to that of other features, the overall accuracy of NDVI fell short by more than three percentage points compared to the optimal feature combinations in other regions, with the gap exceeding five percentage points in most areas. Compared to previous studies, the method proposed in this paper achieves higher crop classification accuracy, with overall improvements ranging from 1% to 3% [64,65]. The extraction of garlic crop distribution and area within the study region marks a significant advancement in winter crop classification. This is particularly important because garlic, despite having a growth cycle similar to that of winter wheat, is often overlooked in studies focused on winter crop mapping [66,67].

In summary, the integration of time series data contributes to enhanced garlic classification accuracy, albeit with variations in performance across regions. While the spliced time series data improves classification accuracy in certain areas, its efficacy is diminished in a few other regions. Integration of optical data with Sentinel-1 SAR data aids accurate garlic identification [68,69], although its impact is less pronounced in regions with better classification results. Lastly, the newly devised garlic fragmentation index facilitates analysis of extraction accuracy disparities across regions.

However, this study has certain limitations and stochastic elements, which could serve as avenues for future research. A primary limitation lies in the reliance on pre-filtered remote sensing imagery for garlic classification. Despite yielding high accuracy results, this approach may still harbor errors that permit the filtration of a small number of garlic plots. Although this study judging anomalies via harmonic regression and patching time series data in the study area using previous data can effectively increase the validity of the data, this algorithm is not directly related to the surface data of the current year, and the integration of multi-sensor remote sensing data can be considered for study in the future, such as the STARFM algorithm. Additionally, green onions in the study area exhibit phenological characteristics similar to those of garlic. Although scallions are not commonly cultivated in the study area, their presence could potentially affect the accuracy of extraction.

When applying the classification method across different regions, several limitations arise. Firstly, varying agricultural practices and crop structures in different areas can pose significant challenges for accurate classification. The spectral characteristics of crops can differ regionally, necessitating the collection of additional samples to account for these variations. Secondly, when expanding the classification to a larger area, a more extensive and evenly distributed sample set is required, which often involves supplementary visual sampling, increasing the complexity of the process. Additionally, the increased scale leads to larger datasets, and processing these on platforms such as GEE may result in limitations such as memory constraints and extended computation times. To mitigate these issues, optimizing the algorithm is essential to prevent processing delays or timeouts. The garlic classification results from the six counties demonstrate that different regions may require distinct combinations of classification features to achieve optimal results. Therefore, when applying this method to different regions, it is advisable to first test the effectiveness of various classification features before proceeding with the classification. However, there also exists a robust set of features that provide high accuracy and can be utilized when precision requirements are less stringent.

5. Conclusions

In this study, we empirically demonstrated that combining active and passive remote sensing with multispectral time series data in the North China Plain enables the acquisition of garlic distribution and area. This method proves to be both economical and efficient, yielding high-quality results. By integrating active and passive imagery and utilizing time series data of red, green, blue, and short-wave infrared, we extracted six primary garlic production areas in the North China Plain using the random forest algorithm. These extractions were validated using ground sampling points and manually selected samples at a later stage. Our study identified key features conducive to enhancing classification results, highlighting the efficacy of combining active and passive remote sensing for improved accuracy, resulting in high classification accuracy (overall accuracy greater than 90%).

Furthermore, the garlic fragmentation index aids in elucidating disparities in extraction accuracy, offering insights into variations in garlic classification accuracy across different regions. The research data and platforms utilized in this study are freely accessible, holding significant potential for extension to broader processes, such as garlic extraction nationwide. Future endeavors could leverage higher-resolution satellites to further enhance accuracy.