Abstract: Cropland abandonment is an important type of agricultural land-use change that reduces the effective area of cultivated land and may affect regional food security. Accurate large-scale monitoring remains challenging because abandoned cropland can exhibit spectral and phenological characteristics similar to grassland, woodland, and temporarily uncultivated fields, while reliable training samples for abandonment mapping are often limited. This study developed a time-series remote sensing framework for cropland abandonment mapping across different cropping systems by integrating phenology-sensitive features (PSFs), land-cover sample augmentation, rule-based pseudo-sample generation, and random forest classification. Two typical agricultural regions in Shanxi Province, China, were selected: the single-cropping region in northern Shanxi and the double-cropping region in southern Shanxi. Sentinel-2 imagery from 2019 to 2024 was processed through cloud masking using Cloud Score+, monthly median compositing, temporal gap filling, and spectral-index calculation to construct standardized annual time-series datasets. Based on intra-annual NDVI trajectories of cropland, woodland, and grassland, PSFs were derived from NDVI changes between key phenological months to enhance differences in seasonal vegetation dynamics. A three-year sliding-window criterion was adopted to identify abandonment: pixels classified as cropland in the first year that subsequently remained in non-cropping states, mainly woodland, grassland, or low-cover vegetation, for two consecutive years were considered abandoned, while policy-driven rotational fallow, ecological conversion, temporary idling, and short-term recultivation were excluded. To alleviate the scarcity of annotated abandonment samples, multi-year land-cover samples were interpreted using the CLCD and ESRI land-cover products, NDVI time series, and high-resolution imagery. A weighted feature-vector method was first applied to augment major land-cover categories. Subsequently, pseudo-samples of abandoned cropland, stable cropland, and other land-cover trajectories were generated by temporally combining land-cover samples according to predefined transition rules. The resulting pseudo-samples were used to train random forest models, whereas independently interpreted real abandonment samples were reserved for final mapping validation. Cross-temporal experiments were further conducted by generating pseudo-samples from land-cover data in years that did not fully overlap the target identification windows, thereby evaluating the temporal transferability of the proposed approach. The proposed framework achieved high and stable mapping accuracy in both cropping regions. For the three identification windows of 2019–2021, 2020–2022, and 2021–2023, the mean overall accuracy reached 93.45% in the single-cropping region and 94.69% in the double-cropping region, with Kappa coefficients ranging from 0.885 to 0.963. F1 scores for abandoned cropland ranged from 0.878 to 0.917, indicating generally balanced omission and commission errors. PSF comparison experiments further showed that several key feature combinations improved the F1 score for abandoned cropland identification. These results demonstrate the value of capturing crop growth, harvest-related decline, and short-term phenological fluctuations rather than relying solely on NDVI magnitude. Spatiotemporal mapping from 2019 to 2024 revealed pronounced stage-specific fluctuations in both study regions. Abandoned cropland area peaked during the 2020–2022 window, reaching 15,375.46 ha in the single-cropping region and 27,550.22 ha in the double-cropping region, followed by differentiated declining trends. At the county scale, high-abandonment areas in the single-cropping region exhibited alternating expansion, contraction, and local re-aggregation. In contrast, abandonment in the double-cropping region showed a clearer east-to-southwest shift and formed new high-value aggregation zones. These contrasting patterns reveal distinct spatial evolution pathways of cropland abandonment under different cropping systems. Overall, integrating phenological feature characterization with sample augmentation and rule-based pseudo-sample generation provides a robust approach for long-term cropland abandonment monitoring under limited-sample conditions. The resulting spatial information can support differentiated cropland management and food-security assessment in Shanxi Province.