Abstract: Cropland abandonment is a widespread global land-use change issue that directly endangers food security and the sustainable utilization of land resources. Nevertheless, the long-term evolutionary patterns, regional disparities and driving mechanisms of cropland abandonment across China remain insufficiently understood systematically. This study addresses that gap by constructing a spatially explicit dataset of abandoned cropland across China for 1995–2023 and investigating spatiotemporal dynamics, spatial clustering, and driving mechanisms at national and agricultural sub-regional scales.Abandoned cropland was identified by integrating two complementary land-cover datasets: the China Land Cover Dataset (CLCD) at 30 m annual resolution, and the China Multi-period Land Use Remote Sensing Monitoring Dataset (CNLUCC) used as a cropland mask to constrain detection domain and reduce misclassification. A temporal consistency correction was applied to CLCD time series via a five-year sliding window to eliminate spurious land-cover transitions caused by classification errors or cloud contamination. A pixel-level land-use trajectory coding framework for the 1990–2023 sequence was adopted to identify cropland continuously converted to non-land for two or more consecutive years, the abandonment threshold matching China’s high-intensity farming conditions. We quantified abandonment for seven target years and classified plots into short-term (≤10 years), medium-term (10–20 years), and long-term (>20 years) groups by abandonment duration, which serves as a critical basis for differentiated reclamation planning. Based on 13 multi-dimensional variables covering terrain, climate, soil, population and economy, an interpretable XGBoost-SHAP framework was built to quantify factor importance and directional effects in 2000, 2010 and 2023, revealing temporal shifts of abandonment drivers nationwide and in sub-regions.The results showed abandoned cropland expanded from 75,900 km² in 1995 to 182,200 km² in 2023, with the abandonment rate rising from 4.33% to 10.35% in a fluctuating upward trend: abandonment expanded rapidly during 2000–2010, slowed after 2010, and slightly declined between 2020 and 2023 yet stayed at a high level. Obvious spatial clustering emerged: high-abandonment hotspots clustered along the Hu Huanyong Line, including the Loess Plateau, Inner Mongolia and Great Wall Fringe, and mountainous Southwest China, while low-abandonment cold spots concentrated in major grain-producing plains. Southern agricultural regions were dominated by short-term abandonment, whereas northern plateau and agro-pastoral transition zones held more medium and long-term abandoned land.Nationwide, topographic indicators and annual temperature served as dominant drivers, with natural constraints stronger than socioeconomic factors. Agricultural divisions presented divergent driving patterns: the arid Northwest and Southeast were climate-controlled with opposite temperature effects; the Loess Plateau, Inner Mongolia fringe, Huang-Huai-Hai and Middle-lower Yangtze zones were restricted by stable terrain and hydrothermal conditions; Northeast, Southwest and South Yangtze hilly areas were jointly shaped by natural conditions and rising population outflow. The Qinghai-Tibet Plateau was uniquely governed by elevation with nonlinear altitudinal effects, and regional economic development gained growing explanatory power. From 2000 to 2023, population density and economic level carried rising explanatory weight in most zones, reflecting stronger coupling between natural endowments, urbanization and labor migration.This research innovatively couples long-time-series trajectory detection with interpretable machine learning to distinguish abandonment duration gradients and compare zone-specific driving dynamics over nearly 20 years. It provides empirical evidence for differentiated abandoned cropland governance and hierarchical reclamation strategies.