Enhancing soil nutrient prediction accuracy using a robust UAV-based modeling framework in dryland environments

Accurate mapping of soil properties is vital for optimizing fertilizer use and crop production, especially in arid regions with saline soils. Conventional ground-based soil evaluation methods often overlook soil nutrient variability. Remote sensing (RS) via Unmanned Aerial Vehicles (UAVs) offers a high-resolution alternative. To address this gap, this study proposes three frameworks: (i) the Site-Specific Model (SSM) for local prediction of soil properties, (ii) the Land Use Model (LUM) for evaluating arable and fallow land, and (iii) the Integrated Site Model (ISM), which combines both approaches. To evaluate these frameworks, high-resolution UAV spectral indices were analyzed alongside 92 soil samples from four sites in southern Morocco. Variable selection was performed using Recursive Feature Elimination (RFE), and outlier detection was performed using the Isolation Forest algorithm. The measured nutrients included Soil Organic Matter (SOM), Total Nitrogen (Total N), available Phosphorus (P₂O₅), Potassium (K₂O), Nitrate (NO₃), and Ammonium (NH₄). Each framework was assessed using four machine learning algorithms: Random Forest (RF), Cubist, K-Nearest Neighbors (KNN), and Partial Least Squares (PLS). Model performance was evaluated with 5-fold and Leave-One-Group-Out (LOGO) cross-validation, using R², Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Monte Carlo uncertainty analysis. The results indicate that: (i) RF achieved R² values between 0.79 and 0.94 with low errors for all soil nutrients across all frameworks; (ii) ISM explained more variance (R² = 0.89 to 0.93) but exhibited higher errors, demonstrating that high R² alone does not ensure precision; and (iii) LUM identified land use as a key determinant, with greater accuracy for arable land, suggesting that management practices are significant predictors. Overall, the RF-SSM combination emerged as the most reliable method for predicting soil nutrient levels, underscoring the importance of selecting a framework tailored to specific sites and land management practices to support sustainable agriculture.