Abstract: Soil salinization constitutes a critical environmental bottleneck that severely restricts sustainable agricultural development and jeopardizes ecological security in arid oasis regions. However, the effective long-term monitoring and precise governance of this phenomenon in large-scale oases are often hindered by the trade-offs between spatial and temporal resolution in available remote sensing imagery, the high computational costs of processing large-scale datasets, and an insufficient understanding of the complex driving mechanisms behind salt accumulation. To address these challenges, this study establishes a comprehensive technical framework applied to the Yarkand River Basin Oasis in Xinjiang, China. The research innovatively incorporates the compute unified device architecture-enhanced spatial and temporal adaptive reflectance fusion model (Cu-ESTARFM), a GPU-accelerated algorithm designed to fuse Landsat and MODIS imagery. This approach effectively addresses the spatiotemporal discontinuity caused by sensor revisit cycles and cloud contamination, generating a high-quality, continuous spatiotemporal dataset. Furthermore, to enhance the physical interpretability of the model, the XGBoost-SHAP (shapley additive explanations) framework was utilized to quantitatively screen and identify the most sensitive environmental covariates from a multidimensional pool of topographic, climatic, and land-use variables. Based on the fused imagery and optimized variables, five distinct machine learning models—random forest (RF), extreme gradient boosting (XGBoost), convolutional neural network (CNN), back propagation neural network (BP), and support vector machine (SVM)—were constructed and systematically compared to reconstruct the spatiotemporal evolution of soil salinity during the critical spring sowing period (April) from 2019 to 2024. The results demonstrate that the Cu-ESTARFM algorithm significantly improved computational efficiency, successfully solving the problem of data scarcity for long-term, high-precision monitoring in large-scale areas. Among the evaluated models, the random forest (RF) model exhibited the highest performance, achieving a coefficient of determination (R²) of 0.65 and a root mean square error (RMSE) of 0.16 %, confirming the superiority of integrating spatiotemporal fusion with multi-source environmental data. Spatiotemporally, the soil salinity content in the study area displayed a clear increasing trend from the upstream alluvial fans to the downstream plains, characterized by a distinct "low in the southwest and high in the northeast" distribution pattern. Land use analysis indicated that salinity in bare lands experienced a dramatic increase of 86 % to 157 %, whereas farmland salinity remained relatively stable due to consistent irrigation and management practices, despite an overall intensification of salinization in the downstream regions. The attribution analysis revealed that wind-assisted evaporation and vegetation degradation show a significant correlation with the process of surface salt accumulation, particularly in the dry spring season where high wind speeds and temperatures synergistically accelerate evaporation in the absence of effective leaching. Consequently, a "partitioned policy + edge blocking" management strategy is proposed. This includes implementing a dual control strategy for water rights and groundwater levels in upstream areas to cut off salt sources, and strengthening salt drainage and bio-chemical reclamation measures in downstream areas. Crucially, the study advocates for constructing an ecological resilience barrier within the "farmland-grassland-bare land" ecotone. By adopting intensive management for marginal farmlands and deploying three-dimensional sand control measures, the vicious cycle of "land degradation inducing salt surface accumulation" can be effectively intercepted. This research provides a replicable paradigm for dynamic salinity monitoring and offers a scientific basis for precision governance in complex arid ecosystems.