Accurately identifying and comprehensively managing the health of cropland systems is crucial for maintaining national food security. In this study, a more suitable framework for evaluating the health status of cropland systems in arid areas was constructed, and a systematic diagnosis of the health status of a cropland system in Xinjiang was conducted by increasing cropland stress and extending the VOR model to the VOR-S framework. The principal driving factors and spatiotemporal heterogeneity of cropland system health were investigated by using geographic detectors and GTWR models. The results showed the following: (1) From 2001 to 2023, the health level of the cropland system in Xinjiang fluctuated and increased. The proportion of areas with higher health levels (health levels I and II) in the cropland system increased from 45.84% in 2001 to 50.80% in 2023. The overall environment of the cropland system thus improved. (2) From 2001 to 2023, in terms of stress on the cropland system in Xinjiang, the overall level of HAI (human activity intensity) exhibited an upward trend, while the overall SEI (soil erosion intensity) significantly decreased, and WEI (wind erosion intensity) remained relatively stable. (3) The explanatory power of driving factors for cropland system health is ranked by order of magnitude as follows: annual precipitation (0.641) > annual average temperature (0.630) > population density (0.619) > nighttime lighting (0.446) > slope (0.313) > altitude (0.267). In addition, the combination of climate and human activity factors plays a dominant role in the spatial differentiation of cropland system health. The research results can provide scientific reference for cropland protection policies in arid areas.

Accurate monitoring of soil moisture in arid agricultural regions is essential for improving crop production and the efficient management of water resources. This study focuses on Shihezi City in Xinjiang, China. We propose a novel method for soil moisture retrieval by integrating Sentinel-1 and Sentinel-2 remote sensing data. Dual-polarization parameters (VV + VH and VV × VH) were constructed and tested. Pearson correlation analysis showed that these polarization combinations carried the most useful information for soil moisture estimation. We then applied Shapley Additive exPlanations (SHAP) for feature selection, and a Stacking model was used to perform soil moisture inversion based on the selected features. SHAP values derived from the coupled support vector regression (SVR) and random forest regression (RFR) models were used to select an additional six key features for model construction. Building on this framework, a comparative analysis was conducted to evaluate the predictive performance of multivariate linear regression (MLR), RFR, SVR, and a Stacking model that integrates these three models. The results demonstrate that the Stacking model outperformed other approaches in soil moisture retrieval, achieving a higher R² of 0.70 compared to 0.52, 0.61, and 0.62 for MLR, RFR, and SVR, respectively. This process concluded with the use of the Stacking model to generate a county-level farmland soil moisture distribution map, which provides an objective and practical approach to guide agricultural management and the optimized allocation of water resources in arid regions.

Accurate crop classification requires effective representation of crop-growth dynamics and traceable classification decisions. This study proposes a prototype-matching framework guided by quantitative temporal and phenological features to establish explicit relationships between crop-growth representations and class-level references. Using multi-temporal Sentinel-2 imagery from an oasis agricultural region in Manas County, Xinjiang, China, four major crops were investigated: cotton, maize, spring wheat, and winter wheat. The framework integrates explicit temporal statistics, quantitative phenological descriptors, and their corresponding deep representations into a unified feature space. Class prototypes are constructed from the resulting representations, and final crop labels are assigned through explicit sample–prototype similarity or distance comparisons. Prototype-space relationships are further analysed to quantify class separability and examine potential classification ambiguity. The results demonstrate that temporal representations provide stronger class separation than phenological representations, while their integration enables continuous temporal dynamics and structured phenological characteristics to be examined within a unified prototype space. Prototype-space analysis reveals clearer separation for cotton and maize, whereas spring wheat exhibits the smallest similarity and distance margins. In the primary field-level classification experiment, the framework achieves an overall accuracy of 96.03%, a Kappa coefficient of 0.9356, a Balanced Accuracy of 86.05%, and a Macro-F1 of 86.26%. In a separate experiment under common training conditions, prototype matching achieves a Balanced Accuracy of 91.36% and a Macro-F1 of 89.53%, exceeding Transformer and random forest by 7.36 and 12.78 percentage points in Balanced Accuracy, respectively. The improvements are particularly evident for spring wheat. These findings demonstrate that the framework connects quantitative crop-growth representation with explicit prototype-based classification while providing measurable evidence for interpreting class-specific discrimination and classification ambiguity.

Fine-scale shortwave-infrared (SWIR) imagery can support detailed land-surface mapping, but submetre satellites generally acquire only blue, green, red and near-infrared (RGBN) bands. Combining these data with Sentinel-2 SWIR requires an observation model that accounts for non-nested pixel supports. We developed a support-aware inverse-allocation framework using an RGBN-conditioned prior, exact polygon-overlap weights and a continuous correction field. The framework was evaluated in four regions of Xinjiang, China, using retained 20 m B11/B12 observations as reference. In a strictly matched control, replacing centre ownership with exact area overlap reduced mean joint root-mean-square error (RMSE) by 0.00174–0.00217 in reflectance units. The reduction occurred in all 15 non-nested phases and all 16 spatial blocks of each region. Continuous correction provided a further reduction of 0.000066–0.000166. Under a matched single-observation budget, the method achieved lower mean joint RMSE than adapted area-to-point regression kriging and DSen2 baselines in all four regions. Exact area overlap retained its advantage under controlled blur, ±10 m guide-image offsets and scale ratios of 2, 3 and 6, although the advantage narrowed at the coarsest ratio. Full-scene application to 0.75 m Jilin-1 RGBN and 20 m Sentinel-2 SWIR produced observation-consistent conditional SWIR allocations, with accuracy evaluated at the controlled 20 m reference resolution.