Tushar Kumar, Seema Pooniyan, Kamal Kishore Yadav, Gajanand Jaat and Pawan Kumar Verma
Abstract
Soil physical properties - texture, aggregate stability, moisture content, and bulk density - determine water storage, nutrient cycling, and crop rooting behaviour, yet conventional measurement of these properties at survey scale is slow and reagent-intensive. Diffuse reflectance spectroscopy (DRS), operating across the visible-near infrared (Vis-NIR: 350-2500 nm) and mid-infrared (MIR: 4000-400 cm-¹) regions, offers a practical alternative: rapid, non-destructive analysis capable of characterizing hundreds of samples per day. In practice, however, prediction accuracy depends less on the sensor alone and more on two analytical decisions made before any model is built - how raw spectra are pre-processed, and which calibration model is applied. This review examines both decisions in detail, drawing on published calibrations for clay, sand, silt, soil moisture, bulk density, and aggregate stability. Pre-processing methods- Savitzky-Golay smoothing,standard normal variate correction, and first and second derivative transformations - consistently improve prediction accuracy, but their effects are property- and soil-specific. No universally optimal sequence has been established. Among modelling approaches, partial least squares regression (PLSR) remains the standard for physical property prediction, while support ector regression (SVR) and ensemble methods offer measurable gains in heterogeneous, non-linear settings. The evidence points clearly toward iterative, property-specific optimization rather than fixed analytical recipes.