Mid infrared spectroscopy and machine learning for predicting soil properties in Inceptisols of the Indo Gangetic plain

Mid-infrared (MIR) spectroscopy offers the potential for rapid soil assessment, particularly when predictions are adequately evaluated against laboratory measurements set. In this study, we compared partial least squares regression (PLSR), random forest (RF), and support vector machine (SVM) regression to predict ten physical and chemical properties of Inceptisols from Haryana, Uttar Pradesh, and Bihar in the Indo-Gangetic plains of India. A total of 450 surface soil samples were collected from arable land at 0–15 cm depth, and the property-specific datasets used for modeling include 434–447 observations after removing outliers and samples affected by measurement or sampling errors. Laboratory MIR spectra were related to particle size distribution, gravimetric water retention at field capacity (FC) and permanent wilting point (PWP), soil organic carbon (SOC), pH, electrical conductivity (EC), and available P and K. Approximately 70% of the observations were used for calibration and 30% for held-out validation. Overall, PLSR performed best for most properties. In validation, PLSR yielded R² = 0.84 and RMSE = 2.89% for clay, 0.75 and 3.28% for silt, 0.71 and 4.14% for sand, 0.74 and 1.52% for PWP, and 0.70 and 2.37% for FC. PLSR also predicted SOC with R² = 0.72 and RMSE = 0.11%, and pH with R² = 0.73 and RMSE = 0.46 pH units. SVM achieved slightly higher R² values for EC and available K, although predictions of EC and extractable nutrients remained weak. These findings support using MIR spectroscopy as a screening tool for selected soil properties within the calibration domain; however, the validation analysis does not demonstrate transferability to new regions, instruments, or field measurement conditions.