Advancing soil property prediction: integration of Vis-NIR sensors' data and machine learning models for rapid and accurate analysis

In Madhya Pradesh, agricultural soils face constraints, such as nutrient depletion, low fertility in Alfisol, drought and drainage issues in Vertisol, salinity, waterlogging and erosion-driven degradation. Sustainable soil management requires timely soil testing and spatial variability assessment, yet conventional methods are slow, costly and poorly accessible. This study examines visible–near infrared spectroscopy (Vis–NIR) with machine learning to estimate soil pH, electrical conductivity (EC) and soil organic carbon (SOC) from 2216 georeferenced surface samples. Laboratory results showed pH 4.55–8.39, EC 0.03–0.97 dS m⁻¹, and SOC 1.05–11.25 g kg⁻¹, with duplicates confirming high precision. Algorithms tested included partial least squares regression (PLSR), SVMR, ANN, RF, CatBoost, ELM, XGBoost and LightGBM. PLSR delivered the most accurate predictions for pH and SOC, while EC showed moderate predictability. Future research should combine mid-infrared spectroscopy, soil covariates, calibration and global spectral libraries to improve model transferability and support scalable soil health monitoring.