Predicting Fertilizer Quantity Using Geo-Spatial and Physicochemical Properties of Soil

Abstract

Nepal's economy is highly dependent on agriculture, yet smallholder farmers benefit far less than they could because they rely on traditional practices or one-size-fits-all recommendations that ignore local soil and topographic variation. To address this, precision farming must be adopted nationwide, and the precise use of fertilizers for the right crop type is a critical part of this transition. This study presents a novel approach to recommend fertilizer quantities for the three major cereal crops, rice, maize, and wheat, including their OPV and hybrid subtypes, by modeling the problem as a multi-output regression task. The model uses the soil's physicochemical properties (pH, organic matter, total nitrogen, available phosphorus, available potassium, sand, silt, parent soil, clay, zinc, and boron) and geospatial properties (latitude, longitude, and elevation) as inputs. Around 386,000 soil data points from all provinces of Nepal were collected from the National Soil Science Research Center of Nepal Agricultural Research Council through a custom data extraction and transformation pipeline. A comparative analysis of Random Forest, XGBoost, and LightGBM was conducted against Linear and Ridge Regression baselines. Tree-based ensemble models consistently outperformed linear approaches, with LightGBM, XGBoost, and Random Forest achieving R² values of 0.976, 0.975, and 0.972, respectively, while both linear models only achieved 0.755. 10-fold cross-validation on a small subset suggests that performance across individual folds is consistent, rather than being overfit to the data. The proposed framework offers a scalable, data-driven path toward site-specific fertilizer recommendation across Nepal's diverse topography.