Cropland abandonment poses a significant threat to global food security, especially in ecologically fragile arid regions, where it compromises both agricultural sustainability and ecosystem integrity. Despite its importance, the spatial distribution of abandoned croplands across large geographic areas remains poorly understood. This study focuses on the Hexi Corridor, a representative oasis agricultural zone in northwestern China. Combining multi-dimensional Sentinel-2 features (phenological, spectral, textural, and physiographic) with feature optimisation, we employed a Random Forest-based post-classification change detection approach to map abandoned croplands from 2021 to 2023, achieving high accuracy (overall accuracy: 92.08%; Kappa: 0.84). (1) Landscape pattern indices and kernel density analysis revealed key spatial trends: abandonment rates declined from 7.05% in 2022 to 5.38% in 2023, while persistent abandonment covered 11,170 ha and recultivation reached 34,403 ha. (2) Spatial heterogeneity was pronounced, with core abandonment zones concentrated in Jiuquan (Yumen, Guazhou) and Zhangye (Shandan County), whereas Jinchang (Yongchang County) and Wuwei (Minqin County) underwent significant rehabilitation. (3) Abandoned croplands exhibited increased fragmentation, driven by a 21% rise in patch density, whereas recultivation was characterized by smaller, scattered plots (patch density > 109), indicating a shift from contiguous to fragmented cultivation. Kernel density analysis revealed an 'oasis-edge aggregation' pattern in abandonment, characterized by high-density clusters in Shandan County following remediation efforts and a distinct recultivation hotspot in Yongchang County. Our multi-feature RF approach establishes a robust framework for large-scale abandonment monitoring and delivers critical insights to support sustainable land management in arid regions.