Machine learning-based estimation and spatial mapping of soil phosphorus and potassium using sentinel-2 and...

This study evaluates the effectiveness of integrating remote sensing data with field observations and machine learning techniques for estimating and mapping available phosphorus (Pav) and exchangeable potassium (Kex) in the Gonbad Kavous region. A total of 394 soil samples collected from the surface layer (0-15 cm depth) were analyzed for Pav and Kex concentrations. Sentinel-2 satellite imagery, together with environmental covariates, was used as predictor variables to model the spatial distribution of soil nutrients. Four machine learning algorithms Random Forest (RF), Support Vector Machine (SVM), Boosted Regression Trees (BRT), and Generalized Linear Model (GLM) were trained and validated for digital soil nutrient mapping. Among these models, RF achieved the highest predictive performance for both Kex (R² = 0.79, RMSE = 13.39) and Pav (R² = 0.83, RMSE = 2.60), outperforming the other approaches. The SVM model also demonstrated satisfactory performance in capturing spatial variability, while GLM showed comparatively lower accuracy. The results confirm the strong potential of combining Sentinel-2-derived spectral information with machine learning algorithms for high-resolution digital soil mapping. The generated spatial distribution maps of Pav and Kex provide valuable insights for soil fertility assessment and can support precision agriculture, nutrient management planning, and sustainable land management practices in semi-arid agricultural regions. Importantly, this study highlights the robustness of ensemble-based learning methods, particularly Random Forest, for predicting soil nutrient variability using multi-source geospatial data.