MSF-Transformer-based spatial modeling of soil total nitrogen in Hengduan Mountains using multi-source remote sensing covariates: a case of Kangding City, Sichuan Province, China

Background Accurate mapping of soil total nitrogen (STN) spatial distribution is critical for effective nitrogen management and precision agriculture, particularly in complex high-altitude terrains where extreme topographic variations result in highly heterogeneous nutrient distribution. Although soil organic carbon (SOC) and soil organic matter (SOM) are often strongly correlated with STN, modeling STN directly is justified because it integrates both organic and inorganic nitrogen pools, thereby more directly representing plant-available nitrogen. Furthermore, in high-altitude agro-ecosystems with strong vertical climatic gradients, STN exhibits unique spatial dynamics that warrant direct and independent modeling. However, capturing the nonlinear spatial dependencies of STN using traditional methods remains challenging. To address this, we focused on Kangding City in the Hengduan Mountains, a region where topographic complexity makes conventional STN mapping particularly difficult. We integrated 51 multi-source auxiliary variables—including spectral, topographic, meteorological, and anthropogenic features—and proposed a novel deep learning framework termed MSF-Transformer to map the STN distribution in this region for the year 2022. Results We systematically evaluated classical models (random forest, extreme gradient boosting, and multi-layer perceptron). While these models demonstrated moderate performance, with the best-performing RF model achieving an R 2 of 0.764, RMSE of 0.529, and MAE of 0.408, the MSF-Transformer demonstrated superior performance on independent test data. The MSF-Transformer combines two key components: a multi-scale feature (MSF) fusion module that adaptively integrates heterogeneous inputs, and a Transformer encoder that models sample-wise spatial relationships. With this design, the MSF-Transformer achieved the highest accuracy (R 2 = 0.780). The predicted STN for 2022 ranged from 1.88 to 5.51 g/kg (mean: 4.37 g/kg), showing higher values in northern and southern areas. Feature importance analysis revealed that STN distribution is primarily driven by elevation (ELEV), potential evapotranspiration (PET), soil types (ST), and distance from center (DISTCC), exhibiting nonlinear responses. Conclusions The results demonstrate that the MSF-Transformer model not only improves prediction accuracy but also offers enhanced interpretability and robust generalization for spatial soil property mapping, providing a powerful tool for precision agricultural management.