Multi-Source Remote Sensing for Soil Texture Mapping in an Arid Oasis of the Central Section of the Northern Tianshan Slope

Soil texture is essential for soil surveys, cropland quality assessment, and soil and water management. However, complex landforms and spatial heterogeneity challenge prediction in arid oases. Single-source data and conventional single-pixel models may inadequately capture spatial variation. This study combined field-collected soil sample data with multi-year remote sensing observations. A predictor system centered on multi-year bare-soil composites (BSCs) integrated additional remote sensing and terrain information. Following isometric log-ratio (ILR) transformation, random forest (RF), XGBoost, and Multi-Modal Vision Transformer (MMVT) were compared across spatial windows. Models were evaluated through repeated random partitions and five-fold spatial cross-validation. Grouped permutation importance and the dissimilarity index (DI) assessed modality contributions and spatial applicability, respectively. MMVT achieved the best overall performance using a 10 × 10-pixel window (approximately 300 × 300 m). Across repeated random partitions, mean R² values were 0.586 ± 0.027, 0.519 ± 0.029, and 0.706 ± 0.026. These corresponded to sand, silt, and clay, respectively. Corresponding RMSE values were 4.695%, 3.391%, and 3.335%, and RPIQ values were 2.107, 2.152, and 2.324. Five-fold spatial cross-validation yielded mean R² values of 0.550, 0.504, and 0.657, respectively. BSCs showed relatively high normalized grouped permutation importance in all three models. Its importance values were 62.68%, 61.75%, and 50.39% for RF, XGBoost, and MMVT, respectively. Mapping showed higher sand contents along the northern desert margin and greater clay enrichment in the western oasis plain. Approximately 85% of the valid mapped area fell within the area of applicability (AOA). Higher DI values occurred mainly in some urban built-up areas, low-mountain river valleys, and hilly transition zones. The findings suggest that combining multi-year bare-soil composites with spatial context at appropriate scales can enhance soil texture prediction. They may provide useful reference information for regional soil surveys, soil type identification, and land suitability analysis.