Monitoring of Farmland Abandonment Based on Google Earth Engine and Interpretable Machine Learning

Abstract

In recent years, China’s hilly and mountainous areas have faced widespread farmland abandonment. However, research on farmland abandonment and its driving mechanisms in hilly and mountainous regions is limited. This study proposes a transferable methodological framework that integrates Landsat data, Google Earth Engine, a time sliding-window algorithm, and the interpretable XGBoost–Shapley Additive explanation (SHAP) model. The time sliding-window algorithm is used to robustly detect long-term land cover changes across the entire study period. The SHAP quantifies the contributions of key drivers to farmland abandonment, providing transparent insights into the driving mechanisms. Applying this framework, we systematically analyzed the spatiotemporal evolution patterns and driving factors of farmland abandonment in Ji’an City, a typical city located in the hilly and mountainous areas of southern China and ultimately developed a farmland abandonment probability distribution map. The findings demonstrate the following. (1) Methodological validation showed that the random forest classifier achieved a mean overall accuracy (OA) of 91.05% (Kappa = 0.88) and the abandonment maps achieved OA of 91.58% (Kappa = 0.83). (2) Spatiotemporal analysis revealed that farmland area increased by 13.26% over 1990–2023, evolving through three stages: fluctuation (1990–2005), growth (2006–2015), and stability (2016–2023). The abandonment rate showed a long-term decreasing trend, peaking in 1998, whereas the abandoned area reached its minimum in 2007. From a spatial perspective, abandonment was more pronounced in mountainous and hilly regions of the study areas. (3) The XGBoost–SHAP model (R² > 0.85) identified key driving factors, including the potential crop yield, soil properties, mean annual precipitation, population density, and terrain features. By offering an interpretable and transferable monitoring framework, this study not only advances farmland abandonment research in complex terrains but also provides concrete policy implications. The results can guide targeted protection of high-risk abandonment zones, promote sustainable land-use planning, and support adaptive agricultural policies in hilly and mountainous regions.

1. Introduction

Farmland abandonment has become an increasingly severe issue in global land-use change, particularly in the hilly and mountainous southern China [1,2,3]. Compared to the more favorable conditions in plains, farmlands in hilly and mountainous regions face numerous environmental challenges, such as steep terrain, inferior soil quality, and intense erosion. Furthermore, factors such as fragmented land parcels, low mechanization levels, and high costs associated with transportation and information exchange significantly hinder large-scale agricultural operations and technological innovation [4,5,6,7,8,9,10]. Additionally, with economic transformation, laborers in hilly and mountainous regions have shifted to non-agricultural industries in search of higher-paying jobs, further exacerbating the farmland abandonment problem. Since the 1980s, this phenomenon has gradually spread across most regions in China, with more severe cases occurring in hilly and mountainous areas [11,12,13,14]. Farmland abandonment not only leads to the waste of land resources but also poses a threat to regional socioeconomic stability. Therefore, an in-depth exploration of methods for identifying farmland abandonment and understanding its driving mechanisms in China’s hilly and mountainous regions is crucial. This is essential for effectively curbing the trend of abandonment and promoting the rational use of land resources, while also supporting national policies and playing a key role in promoting rural revitalization.

Currently, the identification and routine monitoring of farmland abandonment face significant challenges. Traditional methods such as surveys and remote sensing are the primary means of obtaining spatiotemporal distribution data on abandoned farmland [15,16,17]. Although traditional surveys can accurately capture the underlying mechanisms of farmland abandonment, they suffer from issues such as long time consumption, low efficiency, and limitations in sample size and technical methods due to farmers’ subjectivity, the sensitivity of the topic, and other constraints [18,19,20]. In contrast, remote sensing, with its advantages of long temporal coverage and broad spatial scale, has become the primary tool for investigating farmland abandonment. Remote-sensing-based methods for identifying abandoned farmland include spectral feature analysis, time-series data analysis, and land-use/land-cover-based approaches. Spectral feature analysis relies on the band differences in a single image or a small number of images to distinguish land types, but in hilly and mountainous areas, it is easily affected by mixed pixels of terrain and land features, making it difficult to distinguish between short-term fallow land and long-term fallow land [21,22]. Although time-series methods can capture vegetation growth cycles and long-term changes, in hilly and foggy areas with cloudy and rainy weather, it is difficult to construct a complete and continuous time-series due to the interference of clouds and shadows, and most of them are based on exponential changes, which are difficult to directly associate with driving factors such as terrain and socioeconomic factors [23,24,25,26]. The land-use/land-cover-based approach addresses the problem from a macro perspective by analyzing land-use types and cover conditions. It integrates factors such as surrounding land-use information, topography, and landforms to make comprehensive judgments, while fully considering the spatial correlation and integrity of land-use. This approach complements the shortcomings of spectral feature and time-series data analysis and demonstrates effective identification of abandoned farmland in complex geographical environments [27].

The issue of farmland abandonment in the hilly and mountainous area of southern China has long been a focal point for research. However, due to the complexity of the topography and the limitations of available data sources, research on farmland abandonment identification and its driving mechanisms in these areas remains relatively scarce [28]. Existing datasets, such as the China Land Cover Dataset (CLCD), China Land Use/Cover Dataset, and China 30 m Resolution Cropland Dataset (CACD), are all national-scale datasets. These research findings reveal discrepancies in both the quantity and spatial distribution of farmland, which also deviate from the statistical data reported by agricultural authorities [29]. Consequently, current research focused on the southern hilly and mountainous areas suffers from two major gaps. First, a lack of refined monitoring methods: There is a critical need for approaches capable of accurately and robustly identifying highly fragmented, dynamically changing abandoned farmlands across complex, high-relief terrains. Second, limited interpretability: While machine learning models are increasingly used to identify drivers, most are “black-box” models that fail to provide transparent and quantitative explanations of how specific driving factors influence the abandonment process. Therefore, this study takes Ji’an City, Jiangxi Province, as a typical case study area, aiming to employ refined identification methods that integrate remote sensing with ground surveys to accurately detect farmland abandonment. Furthermore, it explores the underlying driving mechanisms of this phenomenon. A comprehensive understanding of these mechanisms will not only provide scientific decision support for formulating targeted cropland protection and abandonment control policies, but also ultimately contribute to the establishment of a future abandonment risk early-warning system, regional food security, and the national Rural Revitalization Strategy.

This study proposes core hypotheses: (1) Farmland abandonment exhibits significant spatiotemporal heterogeneity; and (2) farmland abandonment is driven by multidimensional factors, including both natural geographical elements (digital elevation model (DEM), slope, etc.) and socioeconomic factors (population density (PD), gross domestic product (GDP), etc.). Building on this, the research objectives are determined as follows: (1) To map spatiotemporal farmland abandonment patterns from 1990 to 2023. (2) To utilize the interpretable model to quantitatively analyze the relative contributions and mechanisms of various natural and socioeconomic driving factors on the abandonment process.

This study establishes a robust and transferable methodological framework, integrating open-access Landsat time-series data, a long-term time-series change detection algorithm, and an interpretable XGBoost–SHAP model. Leveraging the global availability of both the data and the integrated methodologies, this framework can be readily deployed in other hilly and mountainous regions facing similar land-use challenges, underscoring its high transferability. The refined spatial maps and quantitative SHAP-based insights into key drivers empower local authorities to precisely target and prioritize high-risk abandoned parcels for effective cropland protection and tailored compensation mechanisms. Furthermore, this research facilitates the establishment of a dynamic early-warning system for abandoned land, providing critical support for designing precise cropland compensation and protection mechanisms, and ultimately contributes to the implementation of the Rural Revitalization Strategy by promoting the synergy between land-use management and sustainable socioeconomic development. More broadly, the framework can serve as a reference model for farmland abandonment monitoring in other mid-latitude regions worldwide.

2. Materials and Methods

2.1. Study Area and Data Sources

2.1.1. Overview of the Study Area

Ji’an City (113°46′–115°56′ E, 25°58′32″–27°50′ N) lies in the central part of Jiangxi Province, covering an area of 25,300 km². The area is characterized by a mountainous–hilly–basin landscape, with elevations of 200–800 m and complex, fragmented topography typical of southern low mountainous and hilly regions (Figure 1). The climate is a subtropical monsoon, with synchronized rainfall and heat providing favorable conditions for agriculture, though topographical variation leads to pronounced spatial heterogeneity. According to the national land survey, farmland totals 396,313 ha (15.79% of the land area). However, fragmented terrain, low mechanization, and labor migration have made farmland abandonment a pressing issue. As a representative region of southern hilly and mountainous areas, Ji’an provides an ideal case for exploring the mechanisms driving farmland abandonment.

2.1.2. Data and Preprocessing

The primary source of remote sensing data was the Landsat satellite image series accessible on the Google Earth Engine. Owing to the long study period, three Landsat datasets (Landsat-5, -7, and -8) were compiled to provide usable images. Natural environmental data included soil type, DEM, and slope. The slope data is derived from the DEM. Socioeconomic indicators included population density, GDP, and other factors.

Table 1. Detailed information about the data used in this study.

Feature Variable	Feature Abbreviation	Spatial Resolution	Data Source
Cation exchange capacity	CEC	90 m	https://data.tpdc.ac.cn/ accessed on 12 March 2025.
Clay	CLAY	90 m
Mean annual precipitation	MAP	1000 m
Mean annual press	MAPS	1000 m
Mean annual temperature	MAT	1000 m
Mean annual wind	MAW	1000 m
Mean annual minimum temperature	LMAT	1000 m
Mean annual maximum temperature	MMAT	1000 m
Nighttime light data	NL	1000 m
Potential evapotranspiration	PET	1000 m
Sand	SAND	90 m
Soil bulk density	BD	90 m
Soil conductivity	CF	90 m
Soil organic carbon	SOC	90 m
Soil pH	pH	90 m
Soil texture classification	TEXCS	90 m
Soil thickness	THICKNISS	90 m
Soil total nitrogen	TN	90 m
Soil total phosphorus	TP	90 m
Soil total potassium	TK	90 m
Aridity index	AI	1000 m	https://www.geodata.cn/ accessed on 12 March 2025.
Gross primary productivity	GPP	1000 m
Net ecosystem production	NEP	1000 m
Net primary productivity	NPP	1000 m
Relative humidity	RH	1000 m
Digital elevation model	DEM	30 m	https://www.gscloud.cn/ accessed on 12 March 2025.
Distance to nearest road	DNR	-	https://www.openstreetmap.org/ accessed on 12 March 2025.
Normalized difference vegetation index	NDVI	30 m	http://www.gis5g.com/ accessed on 12 March 2025.
GDP	GDP	1000 m	https://doi.org/10.1038/s41597-022-01322-5
Population density	PD	1000 m	https://landscan.ornl.gov/ accessed on 12 March 2025.
Potential crop yield	PCY	1000 m	http://www.resdc.cn/ accessed on 12 March 2025.
Soil conservation	SC	1000 m	https://www.scidb.cn/ accessed on 12 March 2025.
Tree cover	TC	1000 m	https://zenodo.org/records/ accessed on 12 March 2025.

offers detailed specifics regarding these data.

2.2. The Framework of the Study

The progression of the study was divided into four phases (Figure 2): (1) Data preprocessing of Landsat, digital elevation model (DEM), and other environmental variables. (2) Construction of multi-year unchanged sample points for different land types, extraction of feature information from these points, and application of the random forest classification model to obtain land-use/land-cover classification results from 1990 to 2023. (3) The distribution of farmland abandonment was identified by applying a sliding-window method. (4) The XGBoost–SHAP model was employed to analyze the driving factors behind farmland abandonment and assess the associated probability.

2.2.1. Definition of Farmland Abandonment

Different studies have provided varying definitions of farmland abandonment, with the primary difference being the timeframe used to define abandonment. The Food and Agriculture Organization defined farmland abandonment in 1995 as “arable land that has not been used for agricultural production or other agricultural purposes for at least five years” [30]. Alternatively, the 2011 International Symposium on Land Reclamation and Land Reserves characterized abandoned farmland as cultivable land lying unused for two or more years, cultivable land abandoned for an undefined period, or cultivable land damaged by poor management [31]. The two-year criterion has been widely adopted in numerous studies focusing on farmland abandonment in China [32,33], enhancing the comparability of our findings within the national research community. Based on field surveys and previous research in the Chinese context, this study defines farmland abandonment as cultivable land that has remained uncultivated for two or more years, during which natural vegetation has regenerated.

2.2.2. Land-Use/Land-Cover Classification Method

We conducted the land-use/land cover classification on the GEE using the random forest, categorizing the study area into six distinct classes: farmland, forest, grassland, water bodies, construction, and unused land. We constructed a feature set (Table S1) that incorporated texture, spectral, phenological, and topographic attributes, reflecting the characteristic properties of each land-use/land cover type. The sample points were randomly divided into training (70%) and testing (30%) subsets. To maximize classification accuracy, the Random Forest model was optimized by testing the number of trees in the range of 100 to 500, with increments of 50, and selecting the configuration that achieved the best performance. The Random Forest model’s high computational efficiency and robustness against noise made it ideal for processing the long-term Landsat time-series data on GEE.

Samples play a crucial role in supervised classification, and their quality directly impacts classification results. Given the large-scale and long-term nature of this study, selecting samples annually would have been both time-consuming and labor-intensive. Therefore, stable multi-year unchanged samples were selected to construct an efficient and accurate dataset, improving both the efficiency and accuracy of land-use/land-cover recognition. The screening process of stable samples is as follows: First, preliminary sample points are randomly generated. By analyzing the time-series data of their normalized difference vegetation index (NDVI), bare soil index (BSI), and enhanced vegetation index (EVI), points with unchanged spectral characteristics are selected. Then, the LandTrendr algorithm is used to parse the pixel time-series spectral curves, so as to identify and exclude disturbed patches. Finally, visual interpretation and verification are conducted by combining high-resolution Google Earth images and Tianditu. Based on the above methods, we randomly picked a total of 899 sample points from the study area (Figure 1) to create our final dataset, including 150 for farmland, 150 for forest land, 150 for grassland, 150 for water bodies, 150 for construction land, and 149 for unused land.

2.2.3. Method for Determining Abandoned Farmland

Farmland abandonment was determined using a three-year time sliding-window algorithm designed to robustly distinguish persistent abandonment from temporary fallow across the long-term time-series (1990–2023, Figure S2). This method requires that land-use conversions persist for two consecutive years, effectively filtering out single-year misclassifications caused by climatic anomalies, data noise, or short-term fallowing. For this process, land-use types were reclassified into three categories: farmland, construction land and water bodies, and the third land-use type (includes forest, grassland, and other unused land). The algorithm defines the first year served as the baseline year, with the subsequent year designated as the test year. When a pixel was classified as farmland during the baseline year and remained in this classification in the test year, it was defined as continuous farmland. If the pixel transitioned to the third land-use type in the test year and remained in that category the following year, it was defined as abandoned. This process was applied iteratively from 1990 to 2022, with the baseline updated annually to generate a continuous abandonment dataset. Figure S2 provides a schematic overview of the classification workflow.

2.3. Research Methodology

2.3.1. XGBoost Model

In this study, the XGBoost model was utilized to probe the association between the selected driving factors and farmland abandonment. XGBoost is an efficient, adaptable, and scalable machine learning algorithm grounded in the gradient-boosting framework, which has seen extensive application in the realm of abandoned farmland in recent times [34,35,36,37].

Farmland abandonment is influenced by various factors, and XGBoost can quantitatively estimate the influence magnitude of each factor. In the study area, a fishing net tool was used to generate 25,965 1 × 1 km grids, and the area of abandoned farmland, farmland area, and driving factor values of each grid were extracted. When constructing the model, the abandonment rate (ratio of abandoned area to total farmland area within each grid) served as the response variable, while the driving factors acted as explanatory variables. To improve the model accuracy, we employed the random stratified sampling approach to refine the sample distribution, dynamically adapting the sample size based on the farmland abandonment rates observed across different years. The dataset was randomly split into 80% for training and 20% for testing. For the XGBoost model, the hyperparameters including the number of boosting rounds, the maximum depth of a tree, and the learning rate were tuned using a grid search approach with 3-fold cross-validation on the training set. The model was evaluated using RMSE, R², and MAE indicators to provide a comprehensive measure of its performance.

2.3.2. SHAP

Although the XGBoost algorithm has high accuracy and good flexibility, it faces the problem of insufficient interpretability, like traditional machine learning models. This “black-box” characteristic makes it difficult to comprehend and verify the model’s predictions. Moreover, due to data quality problems, it may give rise to overfitting. The SHAP method was developed to address this challenge. SHAP is an advanced interpretability method that leverages cooperative game theory and the principle of local interpretability to elucidate the impact of individual features on a model’s predictions. SHAP treats each model feature as a contributor to the prediction outcome, constructing a supplementary explanatory model for every individual prediction instance. It then computes the Shapley values, which are quantitative measures of each feature’s contribution, and offers interpretations grounded in these values [38,39,40]. SHAP substantially enhances the interpretability of the model by intuitively and quantitatively identifying the key driving factors of farmland abandonment. This transparent interpretability is particularly critical for informing subsequent policy formulation. Compared with models that rely solely on traditional accuracy metrics, the integration of SHAP provides higher reliability and scientific credibility. Moreover, it offers a mechanism-based foundation for developing effective land management intervention

2.3.3. Accuracy Evaluation Method

The overall accuracy (OA) and Kappa coefficient were calculated using a confusion matrix to validate the classification accuracy. OA represents the proportion of correctly classified pixels to the total number of pixels, with the following formula:

$$OA={\displaystyle \frac{TP+TN}{TP+FP+TN+FN}}$$

where $TP$ (true positive) refers to positive samples correctly classified by the model, $TN$ (true negative) refers to negative samples correctly classified by the model, $FP$ (false positive) refers to negative samples incorrectly classified by the model, and $FN$ (false negative) refers to positive samples incorrectly classified by the model.

In contrast to OA, the Kappa coefficient accounts for random errors in the classification results. The Kappa coefficient was calculated as follows:

$$Kappa={\displaystyle \frac{PA-PE}{1-PE}}$$

where $PA$ represents the consistency of two observations, whereas $PE$ represents the chance consistency of two observations.

Based on the spatial distribution, 210 random verification points were selected to verify the accuracy of the identification results for abandoned farmland (Figure 1). A confusion matrix for abandoned farmland was established using high-resolution imagery from Google Earth combined with field validation data. The OA and Kappa coefficients were calculated to evaluate the reliability of the abandoned farmland identification results. Compared to many existing studies, we conducted extensive field validation, effectively mitigating the limitations of relying solely on high-resolution imagery, and providing a more solid and reliable assessment of farmland abandonment.

3. Results

3.1. Annual Land-Use/Land-Cover Classification

Between 1990 and 2023, the farmland area in the study area increased by 13.26% (Figure 3), rising from 509,234.40 ha to 585,653.24 ha, with a net increase of 76,418.84 ha. The largest farmland area was observed to occur in 2021 (626,826.91 ha), whereas the lowest occurred in 1994 (454,554.44 ha). The temporal trend can be categorized into three distinct phases: a fluctuating phase (1990–2005), where the farmland area underwent several significant increases and decreases, yet maintained general stability; the expansion phase (2006–2015), during which agricultural policy adjustments and economic development contributed to an overall upward trend in farmland area; and the stabilization phase (2016–2023), during which the farmland area remained relatively stable with small fluctuations.

The spatial distribution of farmland in the study area exhibited spatial heterogeneity (Figure S1). Generally, the three counties with the largest farmland areas were the Taihe (62,890.33–83,203.18 ha), Jishui (58,427.62–75,195.73 ha), and Ji’an Counties (58,718.55–70,683.25 ha). The three counties with the highest proportions of farmland were the Jizhou District (48.73–56.97%), Xingan (30.66–35.06%), and Ji’an Counties (27.66–33.29%). Throughout the past 35 years, this spatial pattern has maintained relative stability.

The Random Forest model demonstrated a progressive improvement in both OA and Kappa coefficient between 1990 and 2023 (Figure 4). The OA exceeded 85% for all years, with an average of 91.05%. The Kappa coefficient consistently exceeded 0.80, with an average of 0.88. The model achieved its optimal performance in 2018 (OA = 97.57%, Kappa = 0.97). Notably, post-2010 results exhibited consistently strong performance, reflecting enhanced model stability. These findings confirm the reliability of the land-use/land-cover classifications throughout the study period.

3.2. Distribution Characteristics of Abandoned Farmland

Before analyzing the distribution characteristics, we conducted accuracy validation using 210 sample points through field surveys and Google Earth imagery (Figure 1). The results showed an OA of 91.58% (kappa = 0.83), indicating a high level of reliability. Field validation combined with drone imagery further confirmed the distribution and changes in abandoned farmland. To better illustrate farmland abandonment, only the earliest year of abandonment for each pixel was presented (Figure 5).

The temporal dynamics of farmland abandonment show phased fluctuation characteristics, but the overall trend is a downward trend (Table S2). In the early 1990s, abandonment increased rapidly, peaking in 1992 at 20,890 ha (4.27%), and this was largely driven by the outflow of rural labor at the beginning of rapid industrialization and the decline in agricultural comparative efficiency. This was followed by a sharp decline after 1995, reaching a low level in 1998, which coincided with strengthened agricultural regulation and rising emphasis on food security. Between 2000 and 2009, abandonment fluctuated moderately; although short-term rebounds occurred, the overall level remained below 2%, reflecting gradual stabilization under the dual influence of farmland consolidation and urban expansion. A modest rebound occurred during 2010–2017, a trend likely associated with accelerating rural-to-urban migration and an aging agricultural labor force. Since 2018, abandonment has stabilized at relatively low levels, with the lowest rate recorded in 2021, suggesting the positive impact of farmland protection programs, direct agricultural subsidies, and policy interventions aimed at maintaining the “red line” of cultivated land.

Spatially, farmland abandonment exhibits both heterogeneity and dynamic clustering, with significant differences in abandonment intensity among counties that evolve over time (Table S3). Overall, areas with higher abandonment rates are typically concentrated in hilly and mountainous regions. During the initial phase (1991–2000), counties such as Xiajiang, Jinggangshan, and Wan’an frequently recorded the highest abandonment rates. This pattern not only reflects the impact of topographic limitations on farming costs and mechanization but also the immediate effect of labor outflow on farmland abandonment. In the mid-phase (2001–2010), abandonment rates in most counties declined significantly, largely due to agricultural policy support and land consolidation projects, hitting a minimal level in 2007. However, spatial disparities persisted, with southern mountainous counties like Jinggangshan, Suichuan, and Wan’an still maintaining relatively higher abandonment levels. In the latest phase (2011–2022), despite a low overall abandonment rate, the phenomenon remains concentrated in the southern and western mountainous counties, including Suichuan, Yongxin, Jinggangshan, and Wan’an. In contrast, counties with better economic development or flatter terrains, such as Jizhou District, Qingyuan District, and Ji’an County, maintained low abandonment rates in most years.

This persistent spatial clustering strongly supports the conclusion that abandoned farmland is predominantly concentrated in hilly and mountainous areas, where complex topography constrains agricultural scaling and mechanization. This, in turn, drives up farming costs and lowers profitability, making these fields the most susceptible to abandonment when farmers balance labor and economic returns.

3.3. Drivers of Abandoned Farmland and Abandonment Probability

This study selected XGBoost as the detection model to analyze data from 1992, 2002, 2012, and 2022. Model validation results indicated that with R² > 0.85 (Figure 6), it demonstrated a good fit and fully reflected the reliability and stability of the model.

Analysis showed that the top five variables that influenced farmland abandonment in 1992, 2002, 2012, and 2022 differed across years (Figure 7). During the four years, the potential for agricultural production in China was a key factor influencing abandonment. This is primarily because this index comprehensively reflects the land’s natural endowment. The lower the index value, the poorer the inherent conditions of the cropland, resulting in correspondingly lower farming profits for farmers and, consequently, a higher risk of abandonment. Regarding 1992, other key variables included forest cover, soil total phosphorus, mean annual precipitation, and PD. Higher forest cover may indicate the early effects of “Grain for Green” policies or competition between forest and agricultural land. Soil total phosphorus, as a key indicator of basic fertility, directly limits crop yields when deficient. Average annual precipitation is directly related to the degree of water stress experienced by crops, which was particularly important in times of inadequate irrigation infrastructure. Lower population density directly suggests a relative scarcity of local agricultural labor, reducing the intensity of cultivation. In 2002, the key variables were soil conservation, soil organic carbon, soil conductivity, and the DEM. Soil conservation capacity reflects the soil’s stability and resistance to erosion. Poor capacity typically leads to a decline in soil quality, negatively affecting crop growth and yield, consequently increasing the probability of farmland abandonment. Soil organic carbon is a key indicator of soil fertility. Low soil organic carbon content signifies nutrient-poor soil and poor crop growth, which reduces agricultural returns and, in turn, exacerbates the risk of farmland abandonment. Soil conductivity reflects the salinity level in the soil; excessively high salinity can negatively impact crop growth, leading to abandonment. DEM represents terrain factors. Steep areas severely limit the application of mechanization, increasing the cost and labor intensity of agricultural production, making it a significant driver of abandonment. In 2012 and 2022, although four variables were common—soil organic carbon, the DEM, PD, and cation exchange capacity—their rankings differed. Cation exchange capacity (CEC) is one of the indicators of soil fertility; a higher CEC helps retain nutrients in the soil, reducing the risk of abandonment. Additionally, based on XGBoost and the top 20 environmental covariates, we constructed a probability distribution for farmland abandonment from 1991 to 2022 (Figure 7).

The research results show that the areas with high farmland abandonment rates increased over time, mainly concentrated in the southeast of Taihe County, the northeast of Jishui County, the northwest of Wan’an County, and the southwest of Yongxin County (Figure 8). From the perspective of different years, the overall abandonment rate was higher in 1992, with areas of high abandonment rate concentrated primarily in Xiajiang and Taihe Counties. In 2002, the areas with higher abandonment probabilities decreased, with the higher rate areas concentrated primarily in Suichuan and Yongfeng Counties. By 2012, areas with higher abandonment rates expanded, with a larger area of abandonment in the western regions. By 2022, areas with higher abandonment rates had increased further, with relatively higher rates in Anfu and Yongxin Counties.

Based on the analysis of abandonment probabilities across different years, regions exhibiting a persistently high probability of abandonment (>50%) are defined as priority intervention areas, signaling deep-seated and recurring farmland abandonment challenges. Utilizing the spatiotemporal trajectory of abandonment, the contiguous western regions, particularly the northwest of Wan’an County and the southwest of Yongxin County, are designated as the highest-priority intervention areas. These areas require the urgent implementation of targeted policy measures to curb farmland abandonment and safeguard regional agricultural sustainability. Recommended interventions include developing agricultural infrastructure tailored to hilly terrain, undertaking soil improvement efforts to gradually enhance fertility, and simultaneously incentivizing the cultivation of high-value crops to boost farmers’ economic returns, thereby mitigating the trend of abandonment and ensuring regional food security. Furthermore, based on 2022 data, Anfu County and Yongxin County are identified as recent abandonment acceleration zones where rapid monitoring and preventive interventions are necessary. These include strengthening the agricultural supply chain, optimizing market channels, and stabilizing farmers’ expected returns, which will effectively prevent the large-scale expansion of new abandonment phenomena in these areas.

4. Discussion

4.1. Spatiotemporal Distribution of Farmland Abandonment

Precise farmland identification is a critical prerequisite for assessing farmland abandonment. The analysis revealed that the farmland area in 2023 had increased by 76,418.84 ha compared to that in 1990, a trend consistent with other existing land-use datasets. For example, according to the CACD, the farmland in the study area increased by 6555.05 ha from 1990 to 2021 [28]; the CLCD showed that the farmland area in the study area increased by 26,417.25 hectares from 1990 to 2021 [29]. Based on the land-use/land-cover data, this study identified abandoned farmland using actual land-use processes. The results indicate the widespread abandonment of farmland across the study area over the 35 years, with noticeable fluctuations in both the extent and proportion of abandoned farmland. These fluctuation patterns may be closely associated with policy-driven factors, including the farmland requisition–compensation balance policy, high-standard farmland construction, land reclamation projects, and the Grain for Green Program [41,42,43].

The spatiotemporal distribution of farmland abandonment in the study area showed a significant “concentration in hilly and mountainous areas” pattern, echoing the regional differentiation in the national farmland abandonment distribution pattern [44]. From a temporal perspective, the study area’s farmland abandonment displayed distinct stages. In 1992, high abandonment was due to underdeveloped agricultural conditions, low mechanization, and high cultivation costs in mountainous areas. Limited non-agricultural employment led farmers to abandon land for economic survival [45]. By 2002, government agricultural policies, including increased investments and improved production conditions, decreased the abandonment rate [46]. From 2012 to 2022, the abandonment rate gradually increased again, primarily because of accelerated urbanization and the increasing trend of rural labor migration toward cities, which exacerbated labor shortages in rural areas [47]. Simultaneously, rising agricultural production costs, volatile agricultural product prices, and decreasing comparative agricultural benefits forced some farmers to abandon their land, leading to increased farmland abandonment.

Spatially, in the early years (1991–2005), abandoned farmland was scattered across multiple counties, primarily because of the mountainous terrain and poor cultivation conditions. From 2006 to 2015, the extent of farmland abandonment continued to expand, corresponding to a period of rapid economic development and large-scale land conversion for non-agricultural usage. From 2015 to 2022, the extent of land abandonment slightly decreased. This can be explained by the government’s farmland protection and agricultural structural adjustment, with some unsuitable farmland being returned to forest or grassland.

Overall, the abandonment of farmland has been more pronounced in mountainous and hilly areas, where complex terrain, low mechanization levels, and high cultivation costs made farming less viable. The spatiotemporal distribution variations in farmland abandonment were closely related to local economic development, policy adjustments, and terrain conditions, offering a reasonable explanation for the observed patterns.

4.2. Driving Factors of Farmland Abandonment

Previous studies exploring the driving factors of farmland abandonment have mostly adopted traditional linear regression or correlation analysis [48,49,50]. However, these methods fail to address potential multicollinearity issues, which may compromise the reliability of model results. In contrast, XGBoost–SHAP can effectively resolve multicollinearity and has emerged as a reliable approach for studying the driving factors of abandonment [51].

The model analysis indicated that potential crop yield (PCY) was a key factor influencing farmland abandonment over four consecutive years. PCY played a crucial role in determining land-use status, particularly in hilly and mountainous areas where complex terrain and poor soil conditions limited agricultural productivity [52,53]. For example, terraced farming in mountainous regions has limited production potential, and combined with factors such as labor migration and the difficulty of mechanization, many terraces have been abandoned [54]. In contrast, plots with a higher production potential are typically located on plains or in areas rich in water resources, where fertile soil and favorable climatic conditions promote crop growth [44]. Additionally, PD has been recognized as a key variable over the years. In areas with high PD, per capita farmland resources are limited, and competition for land-use is intense. Some farmland may have been abandoned owing to its conversion to other uses. However, in areas with low PD, a shortage of labor makes it difficult to effectively cultivate and manage land, thereby escalating the likelihood of farmland abandonment [11,55,56]. Conversely, in regions with a moderate PD and a robust agricultural economy, farmers are more inclined to continue farming because agricultural production provides a stable income [57]. SOC was also identified as a key influencing factor over multiple years. When the SOC content is high, it improves the structure and stability of the soil, which helps to retain moisture and nutrients, benefiting crop growth [58]. In contrast, low levels of SOC are of low quality, restrict the growth of crops, and increase the probability of land abandonment [48]. Notably, abandonment can further reduce SOC content, thus exacerbating farmland degradation in a vicious cycle. Therefore, improving the SOC content through proper management practices is an important strategy for reducing farmland abandonment [59]. Finally, the DEM is also a critical factor influencing land-use status [60,61]. Mechanized agriculture is difficult in areas with significant topographical changes, such as mountains or steep slopes. This leads to higher production costs and immense challenges to planting, thus increasing the likelihood of abandoned farmland. In comparison, flat areas are more suitable for agricultural production, and abandonment risks are relatively low. In summary, multiple factors, including PCY, PD, SOC, and the DEM, collectively influence the risk of farmland abandonment.

4.3. Suggestions for the Management of Abandoned Farmland

In the analysis of the interaction between key driving factors (Table S4, Figure S3), the strongest interaction in 1992 was between PCY and DEM. The results indicate that low PCY exacerbates the effect of high DEM on farmland abandonment, highlighting that when land is difficult to cultivate and yields are low, the risk of abandonment increases significantly. During 2002, 2012, and 2022, the strongest interaction was observed between PCY and SOC. Higher SOC levels were found to mitigate the abandonment risk associated with low PCY. In summary, multiple factors, including PCY, potential driving factors (PD), SOC, and DEM, collectively influence the risk of farmland abandonment. Therefore, these factors must be considered holistically when developing relevant policies and management strategies, to ensure the sustainable use of farmland resources.

Drawing on an analysis of regional heterogeneity characteristics and driving mechanisms, this study proposes a differentiated governance strategy system for farmland abandonment. There is a strong synergistic effect between PCY and SOC: low SOC significantly exacerbates the risks associated with low PCY, thereby suppressing agricultural productivity. Therefore, addressing SOC issues will directly and significantly impact PCY; improving soil organic carbon content is the most effective policy measure to restore crop productivity and rapidly alleviate current farmland abandonment. In areas with low SOC, soil improvement measures should be prioritized, such as promoting organic fertilizer application, straw return, and green manure rotation to gradually enhance soil fertility [62]. In hilly and mountainous areas with complex terrain and low PCY, governments should enhance agricultural infrastructure investments, including irrigation systems and road construction, to facilitate mechanized farming and transportation [63]. In regions with significant topographic relief, alongside infrastructure development, the cultivation of specialized mountain agriculture should be encouraged [64]. Furthermore, a comprehensive monitoring system should be established to promptly identify high-risk farmland abandonment areas, enabling precise early-warning mechanisms and targeted interventions to effectively mitigate cropland abandonment [65].

The process of cropland abandonment is essentially driven by the transformation of socioeconomic structures, while environmental factors determine which lands are abandoned first [66]. Labor outflow and rising socioeconomic costs are the fundamental drivers prompting farmers to make abandonment decisions [67]. Urbanization leads to a surge in the opportunity cost of agricultural labor in areas with high PD, and farmers rationally choose to abandon land with the worst environmental conditions and the lowest marginal benefits [68]. Therefore, policymaking cannot rely solely on environmental and technical measures such as increasing SOC or building infrastructure. More crucially, it is necessary to enhance the economic competitiveness of agriculture, such as by increasing targeted subsidies and improving rural public services, to slow down population outflow [69]. At the same time, land transfers should be guided to achieve large-scale operations, thereby overcoming environmental constraints and ultimately achieving synergistic governance of environmental and socioeconomic drivers [70,71,72].

4.4. Comparative Analysis of Driving Factors for Farmland Abandonment

This study takes Ji’an, China, as an example to investigate the driving mechanisms and subsequent trajectories of agricultural land abandonment in hilly and mountainous areas. To enhance the widespread applicability of the research, we compare our findings with relevant studies from other regions, such as Europe, Southeast Asia, and Africa, to reveal the commonalities and differences in agricultural land abandonment under different socioeconomic and natural backgrounds. The study by Fayet et al. [73] provides a systematic review of post-abandonment land trajectories in Europe, highlighting that rural-to-urban population migration leads to labor shortages in mountainous areas. Additionally, the difficulty of mechanizing farming in hilly terrain and the high production costs contribute to the widespread abandonment of agricultural land. Nguyen et al. [74] found that declining land quality and unstable agricultural income are the main reasons farmers abandon cultivation in central Vietnam. In South Africa, Sibiya et al. [75] emphasized that land degradation and institutional failure are key drivers of farmland abandonment. Despite differences in natural conditions, policy environments, and socioeconomic structures across regions, the driving factors behind agricultural abandonment show clear cross-regional commonalities, including natural environmental constraints, labor outmigration, declining agricultural profitability, and the influence of institutional and policy factors.

4.5. Limitations and Future Perspectives

Although this study has expanded and supplemented research on farmland abandonment in the study area, certain limitations and shortcomings must be acknowledged. First, regarding data acquisition, farmland data obtained from remote sensing may show discrepancies compared with actual farmland data. Although remote sensing technology offers the advantage of large-scale and rapid monitoring, factors such as resolution, spectral complexity of land cover, and data processing algorithms can affect its ability to accurately reflect actual farmland conditions. In particular, the classification uncertainty inherent in remote sensing should be noted—such as spectral confusion between cropland and other vegetation types, mixed pixels in fragmented landscapes, and temporal mismatches between image acquisition and ground conditions. Future work should strengthen the data validation processes by employing methods such as field surveys and multisource data integration to minimize these errors and improve the accuracy and reliability of data.

Secondly, policy factors play a key role in the process of farmland abandonment, but due to the inherent complexity of quantifying policy impacts, this study did not conduct an in-depth quantitative analysis of this aspect. Policy formulation and implementation often involve multiple levels, and their effects have both temporal and spatial complexities, including policy lag effects, making accurate quantification difficult. Future research should focus on developing scientifically rigorous and effective methods to fill this research gap and gain a clearer understanding of the policy-driven mechanisms of farmland abandonment.

Moreover, the factors considered in this study have certain limitations and do not encompass all potential variables associated with farmland abandonment. For example, farmland abandonment can be strongly affected by factors like the clarity of land ownership, the trends in regional industrial restructuring, the level of environmental protection, and the ecological awareness among residents. For instance, unclear land ownership may create uncertainty in land-use, ultimately leading to farmland abandonment. Additionally, a rapid shift toward non-agricultural industries may drive large-scale labor migration away from agriculture, further exacerbating farmland abandonment. These factors should be integrated into the analysis in the future to create a more comprehensive and systematic framework for understanding farmland abandonment.

Finally, farmland abandonment is a complicated process affected by various factors. This study merely examined the effects of two factors without fully exploring the interactive mechanisms among multiple factors. These factors are interrelated and interactive, collectively driving the occurrence and development of farmland abandonment. Future research should comprehensively analyze the interactions between multiple factors and their effects on farmland abandonment to enhance theoretical insights into this phenomenon.

5. Conclusions

This study investigates farmland abandonment in the study area, using a combined remote sensing and machine learning framework. Our study found that between 1990 and 2023, the total farmland area in the study region increased by a net 13.26%, undergoing three distinct phases: a fluctuation period (1990–2005), a growth period (2006–2015), and a stable period (2016–2023). Farmland area remained large and stable in Taihe, Jishui, and Ji’an counties. Farmland abandonment, however, showed a long-term decreasing trend, with the highest rate occurring in 1992, primarily concentrated in hilly and mountainous areas. The XGBoost–SHAP model analysis identified key drivers of farmland abandonment, including PCY, TC, SOC, cation exchange capacity, mean annual precipitation, PD, and DEM. The study specifically highlighted the southeastern part of Taihe County, the northeastern part of Yongfeng County, and the northwestern part of Wan’an County as high-risk zones for long-term abandonment. Additionally, abandonment rates in the northern and eastern regions have shown a significant upward trend, indicating the worsening of this issue.

These findings directly support the design of targeted measures, such as promoting small-scale mechanization in steep areas, strengthening agricultural service provisioning in labor-scarce villages, and introducing sustainable soil management practices in degraded areas. This analytical framework and its outcomes enable policymakers to move from generalized support to precision conservation, effectively addressing the root causes of farmland abandonment in critical regions.