Soil salinization is a major global problem, particularly in arid and semi-arid regions, threatening agriculture and land sustainability. High-resolution spatial prediction is essential for effective land management, yet strong spatial variability and diverse remote sensing (RS) data complicate regional-scale modeling. In particular, the influence of input variable combinations on model performance remains insufficiently explored, hindering the development of accurate and efficient prediction frameworks. In this study, 6,045 soil salinity samples from three soil depths were collected across 15 major oasis irrigation districts in southern Xinjiang, covering a total area of 2.1 x 10(7) hectares. Based on these data, we systematically evaluated the performance of five soil salinity mapping models, e.g., multiple linear regression, random forest, Cubist, quantile random forest, and quantile neural network, across five nested predictor sets. These predictor sets integrated diverse datasets encompassing soil properties, climate, topography, RS indices, soil degradation, and hydrogeological conditions. Results revealed that soil salinity exhibited notably strong spatial variability, with coefficients of variation ranging from 161 % to 293 %. Among all models and input scenarios, quantile random forest consistently achieved the best performance, with root mean square error of 2.64 similar to 4.29 g/kg and coefficient of determination values of 0.64 similar to 0.79. Notably, all models performed the worst under the input scenario excluding wind-related variables, underscoring the crucial role of wind in salt transport and redistribution in southern Xinjiang. In contrast, adding more variables beyond a key subset yielded a negligible impact on model accuracy. Dominant driving factors of soil salinization varied by region and scale, indicating that no single predictor set is universally optimal. Therefore, a nested variable modeling approach that balances prediction accuracy and computational efficiency is recommended. This study provides a practical and scalable framework for high-resolution soil property prediction at regional scales.