Identification, Mechanism and Countermeasures of Cropland Abandonment in Northeast Guangdong Province

Abstract

Cropland serves as the most vital resource for agricultural production, while its security is primarily threatened by abandonment. Northeast Guangdong Province features a fragmented terrain and faces a significant issue of farmland abandonment. It is crucial to analyze the phenomenon of cropland abandonment to safeguard food security. However, due to limitations in data sources and attribution methods, previous studies struggled to comprehensively characterize the driving mechanisms of abandoned land. Using data from Sentinel time series remote-sensing images, we employed the land use change trajectory method to map cropland abandonment in Jiaoling County from 2019 to 2023. Furthermore, we proposed a novel analytical framework to quantify the influence pathways and interaction effects driving cropland abandonment. The results indicated that: (1) The overall accuracy of the abandoned land extraction was 79.6%. During the study period, the abandonment rate in Jiaoling County showed a trend of a “gradual rise followed by a sharp decline”, and the abandoned area reached its maximum in 2021. The abandonment phenomenon in the southeastern rural areas was serious and stubborn. (2) The slope has the greatest explanatory power for abandonment, followed by the total cultivated area, aggregation index of cropland, and distance to road. Each driving factor has a threshold effect. (3) Topography, location, and agriculture driving factors directly or indirectly affect the abandonment rate, with direct influences of 0.247, 0.255, and −0.256, respectively. The research findings offer valuable scientific guidance for managing abandoned land and deepen our understanding of its formation mechanisms.

1. Introduction

The world is currently experiencing rapid population growth and extreme climate changes, exacerbating the tension between the limited land resources and a rising food demand [1]. Global population suffering from hunger increased by 122 million between 2019 and 2022, especially in developing countries [2]. Cropland abandonment refers to land that has ceased to be cultivated due to a combination of natural, economic, and political factors [3]. Given the intensive nature of China’s agricultural system, land left uncultivated for over a year is typically regarded as abandoned [4]. Research indicated that developed regions, including Europe, the United States, Australia, and Japan were the regions where cropland abandonment was the most widespread [5]. Furthermore, within the countries and regions, the distribution of abandoned cropland is uneven. For instance, in the United States, abandoned cropland was primarily concentrated east of the Mississippi River [6,7]. Meanwhile, in China, it was mainly distributed in the southern regions with fragmented terrain [8]. Cropland abandonment not only poses a serious threat to food security but also has significant ecological impacts. For example, Fischer argued that abandoned cropland could threaten natural habitats [9]. Besides, Romero-Calcerrad and Rollins found that the weeds growing on abandoned farmland significantly increase the risk of wildfires [10,11]. However, other researchers discovered that abandoned cropland could enhance soil carbon storage [12], promote vegetation community recovery [13], and improve ecosystem services [14].

Despite the urgency and importance of monitoring cropland, obtaining high-quality data on the distribution of abandoned land remains a key research challenge. Traditional methods for monitoring abandoned cropland usually use a combination of field surveys and questionnaires [15]. While this approach can yield relatively accurate data, it is labor-intensive and struggles to capture changes in spatial patterns of abandonment [16,17]. In recent years, advancements in remote sensing technology have provided strong support for low-cost, high-precision, and large-scale monitoring of agricultural land [18,19]. Among various remote sensing platforms, the MODIS and Landsat series satellites, with their short revisit cycle but insufficient spatial resolution, are considered suitable for mapping large-scale or mesoscale cropland abandonment [1,20,21,22]. Recently, the launch of the Sentinel satellite series has opened new possibilities for extracting cropland abandonment data. Sentinel-1, equipped with a synthetic aperture radar (SAR), can effectively reduce interference from clouds and fog in the study area [23], providing robust support for land classification in cloudy and rainy regions [24]. Sentinel-2, with its multispectral imager (MSI), offers rich spectral information and a temporal resolution of up to 5 days, enabling precise capture of surface cover changes. In the hilly and mountainous regions of southern China approximately 50% of the areas are consistently affected by cloud and rain interference in remote sensing images [25]. The integration of Sentinel-1 SAR and Sentinel-2 MSI data can significantly enhance the accuracy of land cover classification in such areas [26].

Currently, remote sensing-based methods used for identifying cropland abandonment can be broadly classified into two categories [27]. The first is the vegetation index change-detection method, which distinguishes between active and fallow croplands by analyzing the annual phenological curves of surface cover [28]. The second is the land use change trajectory method, which tracks pixels that transition from cropland to grassland, shrubland, or forest using multi-temporal land classification results, and classifies them as abandoned land [8,28,29]. Random Forest (RF) is adept at handling high-dimensional data and complex land cover spectra and does not require extensive parameter adjustments. It has been shown to effectively depict land use in previous studies [29,30,31]. The Google Earth Engine (GEE) cloud platform, which provides a vast, free repository of Earth observation datasets and powerful cloud computing capabilities, has been widely used in various remote sensing studies [32,33,34]. Therefore, it is a feasible approach to use the GEE platform and integrate Sentinel-1 and Sentinel-2 data with an RF-based land use change trajectory method to extract abandoned cropland in southern China.

Understanding the mechanisms behind cropland abandonment in a region is crucial for providing policymakers with valuable references for developing relevant policies. However, the driving forces behind cropland abandonment vary across different regions. Urbanization and industrialization are considered the primary driving forces for cropland abandonment in developed regions such as Europe and Japan [10,35,36]. Ze et al. identified topography and climate as the main environmental drivers of cropland abandonment in the karst regions of China [37]. Traditional attribution analysis methods such as the Tobit model [38], logistic regression [39] and spatial regression model [40] can only identify the magnitude and direction of the driving factors, but cannot reveal the interactions between multiple factors. Geo Detector is considered a common analytical tool for attribution of geographical phenomena [29,30,41]. However, Geo Detector cannot determine the direction of the driving factors’ effects and requires manual discretization, which introduces additional uncertainty. Machine learning algorithms are adept at capturing complex interactions among variables. However, due to their limited interpretability, they are often considered “black box” models. The Shapley additive explanations (SHAP) method is a recent tool for interpreting machine learning models, which not only measures the contribution of driving factors but also identifies their impact directions and interaction effects [42]. Currently, the integration of machine learning models with SHAP for driving factor analysis has been widely applied in fields such as urban heat environments [43,44], air pollution [45,46], vegetation changes [47,48], and ecosystem services [49]. However, to our knowledge, there is limited research on cropland abandonment. Additionally, most current studies overlook the impact of the combined effects of multiple factors on cropland abandonment and lack a comprehensive analytical framework for understanding the driving mechanisms of cropland abandonment. Therefore, coupling machine learning models with SHAP methods and exploring the combined effects of multiple factors on the driving mechanisms of cropland abandonment holds significant potential for application.

Our study focuses on Jiaoling County to map the distribution of abandoned cropland and analyze the influencing factors of cropland abandonment. The main objectives of the study are as follows: (1) To create a high-precision map of cropland abandonment and analyze the spatiotemporal distribution characteristics of the abandoned land; (2) To explore the driving factors and influencing mechanisms of abandoned land; (3) To provide targeted recommendations for the remediation of abandoned cropland in Jiaoling County based on the research findings. By analyzing the phenomenon of abandonment, a scientific basis can be provided for the government to formulate relevant policies, which is of great significance for food security, rural development, and environmental governance.

2. Materials and Methods

2.1. Study Area

Jiaoling County is located in the southeastern part of China, in the northeastern region of Guangdong Province (Figure 1). It covers an area of approximately 960 square kilometers, with elevations ranging from 60 to 1140 m. The study area is surrounded by mountains and features a terrain that slopes from north to south, with 82,690 hectares of mountainous land and 5900 hectares of cropland, making it a typical mountainous and hilly region. The primary crops in the area include rice, vegetables, and peanuts, with the local specialty rice, “Silk Rice”, accounting for over 70% of the rice cultivation in the county. The land use types in the region are complex, and the cultivated land is highly fragmented and dispersed, making it difficult to quantitatively monitor cultivated land resources. Additionally, the region’s serious population outflow, destruction of agricultural facilities, and poor road accessibility restrict the sustainable cultivation of cultivated land. Studying cropland abandonment in this region can provide valuable insights for policy development and direction.

2.2. Daasets

2.2.1. Satellite Data

The satellite data used in this study were all obtained from GEE, including Sentinel-1 GRD and Sentinel-2 MSI data from 2017 to 2023. For 2017 and 2018, Sentinel-2 were obtained as top-of-atmosphere (TOA) reflectance products, while for 2019 to 2023, surface reflectance (SR) data were used. To standardize data quality, atmospheric correction was applied [50]. For each year’s harvest season (October 31 to December 31), images with less than 80% cloud cover were selected. Cloudy images were removed using the QA band, and median composites were created to produce cloud-free imagery covering the study area. The annual vegetation index was calculated using Sentinel-2 data, with an S-G filter applied to reduce noise in the time-series data and reconstruct the phenological curve, then maximum values and standard deviations were computed to capture key crop growth stages [51]. Sentinel-1 radar data were preprocessed, and the mean annual time series of VV and VH backscatter coefficients were extracted. Subsequently, the gray level co-occurrence matrix (GLCM) was used to compute texture metrics, including angular second moment, correlation, and inverse difference moment, for each 3 × 3 pixel window in the NDVI, VV, and VH datasets [52]. Elevation and slope characteristics were derived from DEM data. In total, 30 classification features were obtained (Table S1), and Figure S1 shows the distinguishability of these features across different land cover types.

2.2.2. Land Use Production

Land use data for this study were sourced from the China Multi-Period Land Use Remote Sensing Monitoring Dataset (CNLUCC), the CLCD dataset released by Professor Huang Xin’s team at Wuhan University [53], and the ESA land use data, which were used as references for selecting sample points. High-resolution historical imagery from Google Earth Pro was utilized for visual interpretation. The study involved filtering out pixels with unchanged land use types, during the research period, based on multiple land use products, and then using stratified sampling to randomly generate sample points. Visual interpretation was performed using Sentinel-2 true color images and Google Earth historical imagery for each year, removing sample points that did not match the actual land cover types, and appropriately adding sample points based on the true land cover situation each year. Details of the classification sample point selection are shown in Figure S2.

2.2.3. Driving Factors Analysis Data

Cropland abandonment is a significant land use change phenomenon influenced by various driving factors. The study categorizes these influencing factors into three types: topography factors, including elevation (ELE), slope (SLP), and aspect (ASP); location factors, including distance to road (DTR), distance to water sources (DTW), and distance to settlements (DTS); and agriculture factors, including soil type (ST), soil organic carbon content (SOC), total cultivated area (TCA), and cropland aggregation index (AI). Detailed description on these driving factors is provided in

Table 1. Driving factors data description.

Type	Variable	Unit	Resolution	Source
Topography factors	ELE	m	30 m	ASTER GDEM 30 m
	SLP	°		Calculated from ASTER GDEM 30 m
	ASP
Location factors	DTR	m		https://openmaptiles.org/ (accessed on 17 August 2024)
	DTW
	DTS
Agriculture factors	ST		1 km	https://www.resdc.cn/ (accessed on 24 August 2024)
	SOC	g/kg	250 m	https://data.isric.org/ (accessed on 24 August 2024)
	TCA	ha		Calculated by Fragstats 4.2
	AI	%

.

Higher elevations and steeper slopes are typically associated with lower soil fertility, greater erosion risks, and increased challenges for mechanized farming, all of which can negatively impact cropland productivity [54]. Aspect influences sunlight exposure and microclimatic conditions, thereby affecting crop growth. Cropland located far from roads, water sources, or settlements is more likely to be abandoned due to higher labor and transportation costs as well as limited market access [28]. Soil organic carbon content serves as an indicator of soil productivity; less productive soils require greater inputs to achieve comparable yields, increasing the likelihood of abandonment [55]. Additionally, smaller or highly fragmented cropland parcels often lack economic viability, making them more susceptible to abandonment [56].

2.3. Methods

The research process is mainly divided into three major steps (Figure 2): (1) Preprocessing of remote sensing imagery and land use products to obtain annual classification-ready images and land cover sample points. (2) Using the Random Forest classifier to generate land cover classification maps for Jiaoling County from 2017 to 2023, followed by applying the land use change trajectory method to map the distribution of cropland abandonment, and conducting accuracy assessments for abandoned land extraction by visual interpretation and field investigation. (3) Utilizing zonal statistics and spatial autocorrelation analysis to explore the spatiotemporal distribution characteristics of cropland abandonment, coupling the XGBoost with SHAP methods and the PLS-SEM model to analyze the importance, response characteristics, and influence pathways of various factors affecting cropland abandonment, ultimately explaining the causes of abandonment and providing policy recommendations for local cropland management.

2.3.1. Land Use Classification

In this study, the Random Forest (RF) classifier provided by GEE (smileRandomForest) was used for annual land use classification in Jiaoling County [57,58]. Based on the land use characteristics of the study area, the land types were divided into six categories: cropland, forest, grassland, water bodies, built-up land, and barren land. Table S2 provides a detailed introduction to the land-use classification system. 70% of the sample points were used to train the model, while 30% were used to validate the classification accuracy. After multiple experiments, it was found that when the number of decision trees exceeds 100, there was no significant improvement in classification accuracy, so the number of decision trees was set to 100. To further eliminate noise in the remote sensing data and improve classification results, the study used the RF importance evaluation tool [59] to discard the least important 20% of bands after each classification, reconstructing new band features. The classification accuracy was assessed using user accuracy (UA), producer accuracy (PA), overall accuracy (OA), and the Kappa coefficient. The formulas for these accuracy evaluation metrics were as follows:

$${U{A}_i=x}_{ii}/x_{i+}$$

(1)

$$P{A}_i=x_{ii}/x_{+i}$$

(2)

$$OA={\sum }_{i=1}^6{x}_{ii}/n$$

(3)

$$Kappa={\displaystyle \frac{\left({\sum }_{i=1}^6{x}_{ii}/n\right)-\left[{\sum }_{i=1}^6\left(x_{i+}·x_{+i}\right)/n^2\right]}{1-\left[{\sum }_{i=1}^6\left(x_{i+}·x_{+i}\right)/n^2\right]}}$$

(4)

where $n$ represents the total number of samples, $i$ refers to the land cover type, $x_{ii}$ represents the number of correctly classified samples on the diagonal of the confusion matrix, $x_{i+}$ and $x_{+i}$ represent the total number of observations in row $i$ and column $i$, respectively.

2.3.2. Identification and Quantification of Cropland Abandonment

(1) This study proposed a cropland abandonment extraction method based on land use change trajectories, which identified abandoned cropland by tracking the changes in cropland pixels (Figure 3). Importantly, cropland abandonment is defined in this study as land that has not been cultivated for two consecutive years and is covered by natural vegetation. Based on this definition, the land use classification results were first reclassified: cropland pixels were reclassified as 1, grassland and forest pixels as 0, and water bodies, barren land, and built-up land as 2. To further eliminate the effects of fallow land and classification errors, pixels that are converted from cropland to grassland or forest and remain in this state for two years are considered abandoned. This study constructs a detection window with a time span of three years and a step length of one year. If a pixel is classified as cropland (1) in year $t$, and classified as grassland or forest (0) in years $t+1$ and $t+2$, the cropland is considered abandoned. The final year of conversion to grassland or forest ($t+2$) is marked as the year of abandonment. The window slides to the right starting from 2017 to obtain the abandoned land identification results from 2019 to 2023.

(2) To evaluate the accuracy of the cropland abandonment mapping, recall rate was used as the evaluation metric. For the cropland abandonment mapping from 2019 to 2022, 50 random points were generated annually within the abandoned cropland areas, and accuracy was assessed through a visual interpretation of Sentinel-2 and Google high-resolution historical imagery. For the accuracy assessment of the 2023 cropland abandonment extraction, a combination of human–machine interactive interpretation and field survey was applied. First, 40 sample points were randomly generated within the abandoned areas, followed by field investigations conducted in October 2023 in several towns within Jiaoling County, where 10 actual cropland abandonment points were collected. This resulted in a total of 50 sample points for 2023. The formula for calculating the recall rate is as follows:

$$Recall=TP/\left(TP+FN\right)$$

(5)

where $Recall$ represents the recall rate, $TP$ is the number of correctly identified abandoned cropland samples, $FN$ is the number of samples incorrectly identified as abandoned cropland.

(3) The abandonment rate (AR) reflects the extent of cropland abandonment in a region, and analyzing its changes can reveal the scale evolution characteristics of abandoned cropland. The AR was calculated using the following formula:

$$C_i=\left(F_i/S_{i-2}\right)\times 100\%$$

(6)

where $C_i$ represents AR in year $i$, $F_i$ is the area of abandoned cropland in year $i$, $S_{i-2}$ is the cropland area in year $i-2$.

(4) The proportion of abandoned land reflects its spatial pattern. We generated vector polygon data for each plot through vectorization and classified the plots based on their size. Subsequently, we calculated the proportion of abandoned areas for different plot categories:

$$P_k={\sum }_{i\in k}{A}_i/{\sum }_{i=1}^N{A}_i$$

(7)

where $P_k$ represents the proportion of abandoned land under area category $k$, ${\sum }_{i\in k}{A}_i$ is the sum of the areas of all plots belonging to category $k$, ${\sum }_{i=1}^N{A}_i$ is the total area of all abandoned land.

2.3.3. Spatial Agglomeration Characteristics of Cropland Abandonment

The Moran index was used to analyze the spatial heterogeneity of AR. We calculated AR at the village level and used two indicators, Global Moran’s I and the Local Indicators of Spatial Association (LISA), to assess the spatiotemporal evolution of AR. The Global Moran’s I, was used to investigate whether, and to what extent, AR is spatially clustered, and the LISA was used to identify the location of specific spatial patterns in AR by calculating the local Moran index [60]. The calculation formula are as follows:

$$Global Moran’s I={\displaystyle \frac{N}{W}}·{\displaystyle \frac{{\sum }_i{\sum }_j{w}_{ij}\left(x_i-\overline{x}\right)\left(x_j-\overline{x}\right)}{{\sum }_i{\left(x_i-\overline{x}\right)}^2}}$$

(8)

$${Local Moran’s I}_i={\displaystyle \frac{\left(x_i-\overline{x}\right)}{S^2}}{\sum }_j{w}_{ij}\left(x_j-\overline{x}\right)$$

(9)

where $W$ is the sum of all weights in the spatial weight matrix, $N$ is the total number of observations, $S^2$ is the population variance of the observations, $x_i$ and $x_j$ are the observations at locations $i$ and $j$, $\overline{x}$ is the average of all observations, and $w_{ij}$ is the spatial weight between locations $i$ and $j$.

When $Global Moran’s I=0$, it indicates that the spatial distribution is random, when $Global Moran’s I>0$, it indicates that similar values are spatially clustered, and when $Global Moran’s I<0$, it indicates that similar values are spatially dispersed. Through the calculation of the local Moran’s index, spatial units can be divided into five spatial correlation modes, namely “H-H” promotion area, “H-L” radiation area, “L-H” transition area, “L-L” low-level area, and “N-S” insignificant area.

2.3.4. Analysis of Cropland Abandonment Driving Mechanisms

In this framework, XGBoost is used to integrate various driving factors to predict AR; SHAP explains the characteristic importance of each driving factor to determine the dominant factor; and the PLS-SEM model comprehensively analyzes the interaction mechanism of each dominant factor.

(1) XGBoost is an efficient gradient boosting framework that optimizes traditional GBDT (Gradient Boosting Decision Trees) and introduces regularization to prevent overfitting. It improves model prediction accuracy by combining multiple weak learners, typically decision trees, and can effectively capture nonlinear relationships between multiple factors and variables. XGBoost is known for its fast training speed and resistance to overfitting, and it is widely used in regression and classification studies [61,62]. The calculation formula is as follows:

$${\widehat{y}}_i^{\left(t\right)}=\sum _{k=1}^t{f}_k\left(x_i\right)={\widehat{y}}_i^{\left(t-1\right)}+f_t\left(x_i\right),fk\in F$$

(10)

where ${\widehat{y}}_i^{\left(t\right)}$ represents the prediction result of sample $i$ after the $t$-th iteration of the model; ${\widehat{y}}_i^{\left(t-1\right)}$ represents the prediction result of the model after the previous $\left(t-1\right)$ decision trees; $f_t\left(x_i\right)$ denotes the model of the $t$-th tree; $F$ is the set of all possible decision trees.

(2) The SHAP model can separate the independent effects of each sample on the predictor variables, addressing the limitation of poor interpretability in machine learning models [63]. It is an important tool in interpretable machine learning research [64] and is often used in conjunction with machine learning models like XGBoost to explore the thresholds of influencing factors. The SHAP value calculation formula is:

$$y_i=y_{mean}+f\left(x_{i1}\right)+f\left(x_{i2}\right)+\cdots +f\left(x_{ij}\right)$$

(11)

where $f\left(x_{ij}\right)$ represents the SHAP value of the $j$-th predictor variable for the $i$-th target variable. When $f\left(x_{ij}\right)>0$, it indicates that the feature increases the prediction value, thus having a positive effect. Conversely, when $f\left(x_{ij}\right)<0$, it means the feature decreases the prediction value, thus having a negative effect.

The construction and related analysis of the XGBoost and SHAP models utilized the open-source libraries (shapviz, tidyverse, and tidymodels) in R 4.4.1, and were completed on the RStudio platform. The parameters used for XGBoost in the study are detailed in Table S3.

(3) Partial Least Squares Structural Equation Modeling (PLS-SEM) is a comprehensive analytical model used to analyze complex causal relationships among multiple latent variables [65]. Latent variables can be represented by a set of observable variables, which are quantities that can be directly observed or measured, whereas latent variables are those that cannot be directly observed but can be constructed through theoretical or hypothetical means using one or more observable variables. Compared to traditional covariance-based structural equation models (CB-SEM), PLS-SEM reduces reliance on the normality assumption and can accurately derive variable scores through direct measurement of other variables [66]. Additionally, PLS-SEM introduces the concept of mediation effects [67] and is less susceptible to multicollinearity and missing data bias, making the results more reliable [68]. The path coefficients in the model reflect the direction and strength of relationships between variables. Choosing the PLS-SEM model allows for the analysis of interaction mechanisms among different influencing factors, helping to clarify the pathways of cropland abandonment.

3. Results

3.1. Results of Land Use Classification

Results of the land use classification from 2017 to 2023 are shown in Figure 4. It can be observed that our land use products have an overall accuracy exceeding 90%, with a Kappa coefficient greater than 0.89. The UA and PA for cropland were both over 90% (Figure S3), indicating excellent accuracy. To further validate the accuracy of our land use products, we compared our results with the ESA WorldCover 10 m resolution product (Figure 5). The results showed that our product exhibited high spatial consistency with ESA’s. In summary, our land use products have high usability and can be used for subsequent cropland abandonment extraction. According to the land use results, the overall land cover in Jiaoling County is predominantly forest, accounting for over 75% of the total area. Cropland resources are mainly distributed in the narrow plains in the central region and along both sides of the Shiku River that runs through the city, showing spatial consistency with the distribution of built-up areas.

3.2. Cropland Abandonment Identification

Figure 6 shows the spatial distribution of abandoned land and Figure 7 shows the scale evolution characteristics and plot area distribution characteristics of newly abandoned land. As seen in Figure 6 and Figure 7, spatially, abandoned land in Jiaoling County was primarily distributed along both sides of the Shiku River, with an overall distribution pattern resembling an “I” shape, and there were also larger clusters of abandoned land in the southeastern part. In terms of scale changes, abandoned area and rate in Jiaoling County exhibited a “gradual increase followed by a sharp decrease” trend. From 2019 to 2021, the abandoned area increased from 390.52 hectares to 417.6 hectares, then rapidly decreased to 146.6 hectares within two years, with AR fluctuating between 2.41% and 6.63%. Regarding the area of abandoned plots, most of the abandoned land in Jiaoling County consisted of small plots under 0.1 hectares, accounting for 68% to 83% of the total abandoned area, while plots larger than 0.5 hectares were less common, comprising less than 7% of the total abandoned area.

The accuracy of abandoned land extraction was evaluated using Sentinel-2 imagery, Google high-resolution historical imagery, and field surveys (Figure 8). According to field survey images, the surface of abandoned lands is mostly covered with weeds or shrubs, consistent with our predefined extraction criteria. As shown in

Table 2. Accuracy assessment of cropland abandonment extraction.

Year	TP	FN	Recall
2019	40	10	80%
2020	40	10	80%
2021	38	12	76%
2022	42	8	84%
2023	39	11	78%

, the recall rate reveals that the overall accuracy of the abandoned land extraction method used in this study is over 76%, with an average accuracy of 79.6% over five years. The highest accuracy was achieved in 2022, reaching 84%.

3.3. Spatiotemporal Characteristics of Cropland Abandonment

Using administrative villages as the study unit, the spatial distribution patterns of AR in Jiaoling County were assessed through spatial autocorrelation analysis in ArcGIS 10.8 (Figure 9). The Global Moran’s I index for the study area from 2019 to 2023 were 0.263, 0.236, 0.332, 0.410, and 0.181, showing a trend of initially increasing and then decreasing. At the village scale, “L-L” clusters were consistently found in the southern part of the study area, particularly in rural areas with lower AR, such as central Xinpu Town. The number of these clusters first decreased, then increased, and eventually stabilized. “H-H” clusters show more pronounced spatial changes, initially distributed in the southeast, moving northeast, then spreading in a ring shape, and eventually returning to the south. The number of these clusters followed a “decrease-increase-decrease” pattern. “H-L” and “L-H” clusters were sporadically distributed in the northern and southeastern parts of the study area, respectively, with overall minimal changes in their numbers.

3.4. Driving Factors of Cropland Abandonment

3.4.1. SHAP Interpretability Analysis

The XGBoost machine learning algorithm and the SHAP interpretable machine learning model were used to analyze the driving factors of the spatial distribution pattern of cropland abandonment in Jiaoling County. Based on the actual conditions of the study area, a 250 m observation scale was chosen, with 2021 being the year with the largest abandoned land area, selected as the case for the study. To further eliminate the interference of random errors on the observation results, data cleaning was performed before input into the model to ensure that each research unit contained at least 0.06 hectares of cropland. The results showed that the R² of the fit between the predicted and actual values obtained from the XGBoost model was 0.77 (Figure S4), indicating that the selected factors can explain 77% of the spatial distribution of AR. Figure 10 presents the SHAP analysis results.

As shown in Figure 10, the formation of abandoned land was influenced by multiple factors, with agricultural and topographic factors having the greatest explanatory power, accounting for 42.1% and 31.1%, respectively. The impact of locational factors on abandoned land was smaller, with a proportion of 26.8%. Specifically, the slope (SLP) among topographic factors, the total cultivated area (TCA) and aggregation index (AI) among agricultural factors, and the distance to road (DTR) among locational factors had the greatest impact, with average SHAP values of 7.55, 3.19, 2.88, and 2.81, respectively. Soil organic carbon (SOC), soil type (ST), and aspect (ASP) had the least explanatory power for abandoned land, with average SHAP values all below 1. The interaction plots show that abandoned land is significantly influenced by the coupling of topographic and agricultural factors, with the slope having strong interactions with multiple factors. When the slope is less than 6° and TCA is large, the slope has a negative impact on AR. However, when the slope is greater than 6°, AR increases with both an increase in slope and a decrease in TCA, indicating that fragmented cropland on steep terrain poses significant negative impacts on agricultural production. TCA shows a strong positive response to AR when it is below 0.4 ha; however, above 0.4 ha it suppresses AR, with this effect stabilizing at a relatively constant level. As AI increases, AR decreases, and when it reaches a threshold of around 75%, its effect on abandoned land shifts from a positive to a negative response, indicating that cropland patches with higher connectivity have a lower likelihood of abandonment. DTR shows a slight negative response to AR when it is below 200 m, but as DTR exceeds 200 m, there is a significant positive response, which strengthens with increasing DTR, suggesting that accessibility to transportation is a key factor influencing AR.

3.4.2. Analysis of Driving Mechanism of Cropland Abandonment

After conducting the SHAP analysis, the top seven factors in terms of importance were used to construct a PLS-SEM model. Based on prior knowledge, we initially made the following assumptions: topography factors, location factors, and agriculture factors directly influence AR. Additionally, topography factors may indirectly influence AR through location or agriculture factors, and location factors may indirectly influence AR through agricultural factors. Based on these assumptions, a PLS-SEM model was constructed (Figure 11). To evaluate the model’s performance, composite reliability (CR) was used to assess the internal consistency reliability [65], with the CR generally required to be greater than 0.6. Convergent validity was assessed using the average variance extracted (AVE), where AVE is typically expected to be greater than 0.5 [65]. Table S4 reflects the reliability and validity assessments of the PLS-SEM model.

The results show that all path coefficients passed the significance test with p < 0.01, indicating that our assumptions are valid. As shown in Figure 11, the constructed PLS-SEM model explains 35.9% of the variance in AR. Among the factors, topography and location factors had a positive impact on the distribution of AR, with path coefficients of 0.247 and 0.255, respectively, while agriculture factors had a negative impact, with a path coefficient of −0.256. In addition to the three direct pathways affecting the distribution of AR, topography factors also increase the positive impact on AR through their positive effect on location factors and exert a negative effect on AR through their negative impact on agriculture factors. Similarly, location factors negatively influence AR through their negative effect on agriculture factors.

4. Discussion

4.1. The Current Situation of Cropland Abandonment in Jiaoling County

Since Jiaoling County is located in the hilly and mountainous area of Northeastern Guangdong Province, cropland is scattered and the area is disturbed by clouds and rain all year round, which brings challenges to the quantitative monitoring of local cultivated land resources. The accuracy of abandoned land identification by our proposed method was above 76%, which is higher than that in the previous studies using only Landsat or Sentinel-2 data [28,31]. Furthermore, this study verified the agreement between the extraction results and the actual field verification. However, due to the consideration of fallow land, the AR obtained may be underestimated [69,70]. Our field investigations revealed that significant young labor loss and damage to irrigation facilities may be important reasons for the abandonment in these regions. Additionally, we found that these fragmented croplands were scattered around residential areas, forming a “family farm” model at the household level, with the decision to abandon it largely depending on the local villagers’ willingness to farm [71]. In contrast, “L-L” clusters were stably distributed in the plains upstream of the Shiku River, where water sources for irrigation are abundant, and there are large-scale contiguous croplands and state-owned farms, leading to a vibrant agricultural economy and thus a lower risk of abandonment.

Additionally, most previous studies focused on the impact of a single factor on abandonment [29,30], neglecting the combined effects of multiple factors and the threshold effects of changes in influencing factors. Our study used XGBoost and SHAP to provide a detailed interpretation of the threshold effects of multiple factors driving cropland abandonment and constructed a PLS-SEM model to reveal the complex interactions between these factors. Previous research has shown that topographic conditions and cropland fragmentation are major driving forces behind land abandonment in mountainous regions of China [71], which is consistent with our study. The slope is the primary driver of increased AR, indicating that sloped croplands face a higher risk of being abandoned. Compared to flatland croplands, steep sloped lands are more susceptible to soil erosion, leading to nutrient depletion and reduced water retention capacity, which in turn diminishes agricultural productivity and makes abandonment more likely [29,72]. Our field investigations found that even when the local governments invest heavily in funding and personnel for the re-cultivation of steep sloped lands, most of these lands revert to abandonment within a few years, suggesting that cultivating crops on steep slopes is not sustainable. Increases in total cropland area and aggregation index of cropland have a significant negative impact on AR, indicating that increasing the area of cropland per unit of land and connecting fragmented croplands can significantly reduce the AR [69]. When the same amount of production factors are applied to a unit of land, regions with larger total cropland areas possess greater production potential, thereby enhancing agricultural productivity [73]. Highly concentrated cropland resources facilitate unified mechanized management and cost-effectiveness, making residents more inclined to farm. We also found that the probability of abandonment increases more rapidly with greater distance from roads, a dynamic process often overlooked in previous research. In areas further away from roads, inadequate transportation capacity slows the circulation of agricultural products, negatively affecting crop production [28]. Some studies have also pointed out that underdeveloped transportation infrastructure is one of the key factors contributing to land abandonment [74].

4.2. Policy Implications

This study explored the spatial distribution characteristics and driving factors of abandoned land in Jiaoling County from 2019 to 2023, providing a scientific basis for the understanding of the mechanisms of abandonment in the area and for formulating relevant policies. Based on an analysis of the results regarding the abandoned land in the county, the following recommendations for managing abandoned land were proposed:

(1) For steep sloped croplands that are difficult to cultivate and unsustainable, it is recommended to implement the conversion of farmland to forests and grasslands as soon as possible to unlock their ecosystem service values, such as carbon sequestration and biodiversity conservation [69]. For some sloped lands that remain suitable for cultivation, land leveling and terracing should be carried out to reduce soil erosion, while introducing small-scale agricultural machinery to improve production efficiency. For smaller, less concentrated croplands, farmers should be encouraged to expedite land transfers, promoting a shift from extensive to scaled and intensive farming practices, thereby enhancing the intensive use of cropland [75].

(2) The government should prioritize the impact of agricultural accessibility on land abandonment by constructing or repairing rural roads and optimizing the rural transportation network to reduce the time and cost for farmers to transport agricultural products. Early planning of farming units with a 200 m buffer zone around the farmland should be considered to allow the farmers a more convenient access to cultivate their fields. For cropland that is difficult to access and far from markets, exploring the development of rural tourism, ecological agriculture, and other multifunctional agricultural forms can increase the added value and utilization of the land. For abandoned land near markets and with convenient transportation, the priority should be to develop efficient agriculture to improve the economic benefits of the land.

(3) Enhancing the economic benefits of agricultural production is key to addressing land abandonment. Our research found that most farmers voluntarily abandon cultivation due to low agricultural income. In response, the government can boost farmers’ willingness to cultivate by introducing modern agricultural technologies and optimizing crop structures. Additionally, the government should emphasize building digital villages, including e-commerce infrastructure, and continuously innovate ways for small farmers to engage in e-commerce, especially in the southeastern rural areas. This will help revitalize rural spaces and reutilize idle land resources [76].

4.3. Limitations and Perspectives

Our study employed a cropland abandonment extraction method based on land-use change trajectories, which can only detect newly abandoned lands during the study period but cannot capture historical abandonment. The achievement of high-resolution, long-term cropland abandonment monitoring remains an urgent challenge. Future research could use cropland boundaries provided by government departments as a baseline to exclude the interference of non-cropland pixels. Moreover, our study only selected indicators from topography, location, and agriculture factors. However, the driving factors behind cropland abandonment may be more complex, with economic factors such as the farmers’ per capita disposable income, agricultural output proportion, and willingness to cultivate also affecting cropland abandonment. Future research could integrate multi-source data to further elucidate the mechanisms behind abandonment.

5. Conclusions

This study used Sentinel time series data to extract the abandoned cropland in Jiaoling County and analyzed its spatial and temporal characteristics, as well as the influencing factors. The main conclusions are as follows: (1) From 2019 to 2023, the area and ratio of cropland abandonment in Jiaoling County exhibited a “gradual rise followed by a sharp decline” trend, with the abandonment rate ranging from 2.41% to 6.63%. There was a significant positive spatial clustering of abandoned land. (2) Slope is the primary driving factor for cropland abandonment in Jiaoling County, with total cultivated area, aggregation index of cropland, and distance to road also having a significant impact on abandonment rates. Each driving factor has a threshold effect. (3) Agricultural factors had a significant negative impact on AR, with an influence coefficient of −0.256, while topographic and locational factors had significant positive impacts, with influence coefficients of 0.247 and 0.255, respectively. An analysis of the abandonment phenomenon in depth can provide an important reference for the local government to formulate arable land protection policies. In the future, we can consider extending this research framework to the study of abandoned land at a larger scale.