Prediction of soil available nitrogen using machine learning and digital mapping techniques in Northeast India

Nitrogen (N) is an essential nutrient for crop growth, influencing both agricultural productivity and ecosystem processes. Conventional fertilizer recommendations often rely solely on soil N availability at sampling points, lacking information on its spatial variability. High-resolution spatial data are therefore crucial for sustaining sericulture production. This study represents the first effort to develop a spatial prediction map of available N in the Lakhimpur district, using 160 soil samples along with local environmental covariates. Five machine learning algorithms—Random Forest (RF), Cubist, XGBoost, Support Vector Machine (SVM), and K-Nearest Neighbors (KNN)—were employed to predict available N in the plough layer (0–15 cm) at a 30 m resolution. Measured N values ranged from 102 to 580 kg ha⁻¹. Under random cross-validation, RF consistently outperformed other models (R²c = 0.96, RMSEc = 40.69 kg ha⁻¹ in calibration and R²v = 0.35, RMSEv = 70.44 kg ha⁻¹ in validation). In contrast, spatial cross-validation resulted in markedly lower accuracy across all models, with R² values varied between 0.02 and 0.27. Predicted N values varied between 172 and 438 kg ha⁻¹, while uncertainty, expressed as standard deviation, ranged from 101 to 111 kg ha⁻¹. The Dissimilarity Index (DI) and Area of Applicability (AOA) maps were used to delineate areas where model predictions are most reliable. Overall, the study demonstrates that machine learning combined with digital mapping provides a practical framework for predicting soil available N. The resulting high-resolution maps offer valuable insights for site-specific nutrient management and support precision agriculture planning in sericulture-based systems.