A spatiotemporal dataset of cropland abandonment in the Guangxi karst mountain area

Abstract: This dataset characterizes the spatiotemporal dynamics of cropland abandonment in the Guangxi Karst Mountain Area (GKMA) from 2001 to 2015. It is derived from multi-source remote sensing and socioeconomic data, integrating MODIS NDVI time series, phenological metrics, and climatic and topographic data. The methodology employs a land use trajectory-based approach, utilizing a Random Forest classifier to generate annual 250m resolution land use maps, which are then analyzed to identify pixels with a farming cessation period of two years or more4. The dataset includes the spatiotemporal distribution of the cropland abandonment rate (CAR), spatial cluster analysis (LISA), and an assessment of driving factors using the Geographical Detector Model (GDM). Key findings indicate that 8.25% of the total land area experienced abandonment since 2001, with the CAR peaking during the 2006-2010 sub-period at 34.12%. Environmental factors, particularly mean annual temperature (MAT) and elevation (DEM), were found to be the dominant drivers of abandonment patterns. This dataset provides fundamental support for regional land use management, food security analysis, and ecological restoration policy in fragile karst environments. Key words: cropland abandonment; MODIS time series; phenology; Random Forest; land use trajectory; karst mountainous area; Geographical detector model 1. Data sources and processing The dataset relies on two categories of inputs: basic data and auxiliary data. Basic data includes MODIS Terra multitemporal products (MOD13, MOD44W) from 2001-2015, providing 16-day NDVI time series at 250m resolution, and SRTM 90m Digital Elevation Model (DEM) data. Auxiliary data comprises annual climatic data (precipitation, temperature) interpolated from surface stations, 1:1M soil and vegetation type maps, and socioeconomic statistics (road density, income gap, rural population) aggregated at the county level. Data processing involved a trajectory-based mapping approach. First, annual 250m land use maps (2001-2015) were generated using a Random Forest (RF) classifier. The RF model was trained on 37 features, including 23 smoothed NDVI values, 11 phenology metrics derived via TIMESAT, and a water mask. Training samples were manually interpreted from high-resolution Google Earth imagery19. Second, cropland abandonment (ACL) was defined as any cropland pixel experiencing a cessation of agricultural activity (transition to grassland or forest) for at least two consecutive years. Finally, the Geographical Detector Model (GDM) was used to quantify the explanatory power (q-statistic) of environmental and socioeconomic drivers on the resulting cropland abandonment rate (CAR). 2. Data results The processing yields three core data components: historical land use classification accuracies, spatiotemporal abandonment patterns, and driver analysis. The annual land use classifications from 2001 to 2015 achieved a mean overall accuracy of approximately 76%. Table 1 presents the classification accuracy for key years, showing stable producer's and user's accuracies for the cropland class (71-75%), though grassland mapping proved more challenging (63-74%). Table 2 details the spatiotemporal characteristics of the cropland abandonment rate (CAR) across three sub-periods. The overall CAR exhibited a trend of increasing first and then decreasing, peaking at 34.12% in the 2006-2010 period before declining to 21.14% in 2011-2015. Spatially, abandonment was highly clustered (Global Moran's I = 0.429), with persistent hotspots (High-High clusters) concentrated in the northern mountainous counties (e.g., HC) and the western river valleys (e.g., YRVB). Table 3 summarizes the explanatory power (q-statistic) of the determinants. Environmental factors exerted significantly more influence than socioeconomic factors. Topographic and climatic variables were the strongest predictors, with mean annual temperature (MAT) and elevation (DEM) showing the highest average q-statistics (0.140 and 0.111, respectively). Table 1 Producer's (PA), user's (UA), and overall accuracies of the land use maps. Land use types 2005 PA 2005 UA 2010 PA 2010 UA 2015 PA 2015 UA Cropland 0.73 0.72 0.73 0.71 0.74 0.75 Forests 0.85 0.77 0.85 0.76 0.87 0.75 Grassland 0.69 0.70 0.67 0.72 0.63 0.74 Water 0.99 0.97 0.99 0.99 0.84 0.92 Construction land 0.56 0.72 0.56 0.70 0.83 0.81 Overall accuracy 0.76 0.75 0.77 Table 2 Core characteristics of spatiotemporal cropland abandonment rate (CAR). Sub-period CAR (% of total cropland) Global Moran's I Key hotspot regions 2001–2005 10.35% 0.339 Initial clusters in YRVB and NJC 2006–2010 34.12% 0.505 Significant expansion; hotspots intensify in YRVB, HC, NJC 2011–2015 21.14% 0.445 Contraction of hotspots, but persistence in mountainous north Note: YRVB = Youjiang river valley of Baise; HC = Huanjiang County; NJC = Napo and Jingxi Counties; LLR = Laibin-Liuzhou region. Table 3 Mean explanatory power (q-statistic) of determinants on CAR (2001-2015). Factor category Determinant Mean q-statistic Effect direction Environmental Elevation 0.140 + Slope 0.137 + Geomorphy type 0.131 + Mean annual total precipitation 0.121 - Mean annual temperature 0.111 - Vegetation type 0.075 - Farming condition 0.073 + Socioeconomic Change of road density 0.071 + Change of income gap 0.047 + Change of rural population 0.045 - Note: '+' indicates a positive correlation with CAR; '-' indicates a negative correlation. 3. Data description and application This dataset provides a large-scale estimation of cropland abandonment dynamics in a karst region, but its limitations must be acknowledged. The 250m MODIS resolution, while suitable for regional trends, may generalize fine-scale abandonment patterns and is susceptible to mixed-pixel effects. The two-year cessation threshold used to define abandonment is a necessary simplification and may not capture all forms of temporary fallowing. Furthermore, the dataset identifies abandonment but does not explicitly differentiate between autonomous, socioeconomically driven abandonment and policy-induced abandonment, such as land enrolled in the "Grain for Green" program. Despite these limitations, the data has significant application value. It can be used to identify spatial hotspots of abandonment (e.g., HC, NJC) requiring targeted policy intervention for land management. The driver analysis (Table 3) supports the development of policies that account for environmental constraints (like elevation and climate). This dataset serves as a crucial baseline for modeling future land use trajectories, assessing impacts on food security, and planning ecological restoration strategies within China's vulnerable mountainous areas.