Cultivated land resilience is often required for the regional food security in sustainable agriculture, particularly under the global climate change and ecological risks in arid and semi-arid regions of Western China. Among them, Xinjiang's oasis cultivated land has been a major production base for high-quality cotton and grain in China. However, an important ecological barrier can be found from multiple disturbances, such as drought, wind-blown sand, and urbanization, thus resulting in significant systemic vulnerability. This study aims to investigate the resilience pattern and driving mechanisms of cultivated land using ecosystem resilience theory. The cultivated land resilience was deconstructed into three capacities: resistance, adaptation, and recovery. A differentiated evaluation was constructed with 14 indicators suitable for the unique characteristics of Xinjiang’s oasis cultivated land, including irrigation dependence, ecological sensitivity, and economic underdevelopment. Resistance capacity was measured by the indicators, such as the proportion of effective irrigated area, soil thickness, soil pH, and soil moisture content. Adaptation capacity was represented by crop planting diversity index, multiple cropping index, agricultural machinery use per land area, and rural electricity consumption. Recovery capacity was assessed by per capita grain security rate, grain output rate of cultivated land, landscape fragmentation index, and per capita agricultural productive investment. Eight control variables were introduced to avoid estimation bias, including elevation, annual average rainfall, annual average temperature, pesticide use per land area, agricultural film use intensity, grain crop sowing area ratio, rural per capita disposable income, and per capita GDP. The XGBoost-SHAP model was employed to combine with kernel density estimation and spatial autocorrelation analysis. The spatiotemporal patterns of cultivated land resilience were examined at the county level in Xinjiang from 2000 to 2023. The strong predictive performance was achieved with the training set R² values over 0.92 for all resilience dimensions and test set R² values ranging from 0.785 to 0.919. SHAP analysis was conducted to quantify the contribution of each driving factor to the nonlinear marginal effects. The results indicate that: (1) The overall resilience of cultivated land showed a fluctuating upward trend during the study period. The peak of the kernel density curve was shifted to the right and the width narrowed, indicating a significant reduction in regional disparities. (2) Spatially, a “high in the north, low in the south” pattern was observed, with the high-value areas in the eastern oases (e.g., Yizhou District, Qitai County) and gradually clustering, while the extent of low-value areas shrank. The global Moran’s I index ranged between 0.182 and 0.310, indicating significant spatial aggregation. (3) Rainfall, per capita GDP, and elevation were identified as the drivers influencing cultivated land resilience, with contribution shares of 36.5%, 16.7%, and 13.9%, respectively. Increasing rainfall weakened the land resilience, due to waterlogging or soil nutrient loss, while the mid-elevation areas exhibited higher resilience. Reasonable investment in per capita GDP enhanced the cultivated land resilience, although excessive investment diminished the returns. The findings can provide scientific support to formulate the differentiated cultivated land protection in Xinjiang and similar arid and semi-arid regions worldwide. The interpretable machine learning models can also offer important theoretical and practical value in the nonlinear driving mechanisms of cultivated land resilience.