Leveraging advanced machine learning with diffuse reflectance spectroscopy for modelling and forecasting soil organic carbon content in Biswanath district, Assam

Abstract

Diffuse reflectance spectroscopy (DRS) holds promise for predicting soil organic carbon (SOC) levels, but its adoption in Indian soil laboratories remains limited. This study aimed to assess whether spectral techniques could effectively replace traditional chemical methods for measuring SOC without compromising quality in tropical soils of Upper Brahmaputra Valley Zone of Assam. Additionally, it sought to identify the most effective prediction strategies. The study utilized a dataset of 300 samples collected from Biswanath district in Assam to develop spectroscopic models for predicting SOC content. This involved careful consideration of sample stratification and preprocessing techniques. SOC levels were determined using wet combustion followed by titration as the analytical-chemical method. Both reflectance (Ref) and Kubelka-Munk spectra were processed using three preprocessing techniques: Savitzky-Golay, Standard Normal Variate, and Multiplicative Scatter Correction. Among the various preprocessing techniques and methods, the most accurate estimation of SOC content was achieved by applying the Standard Normal Variate (SNV) preprocessing technique to reflectance spectra and using partial least square regression (PLSR). The validation parameters for this best-performing prediction model were as follows: R2 at 94.1, RPD at 3.45, RPIQ at 5.34, and RMSE at 0.08. In conclusion, the study demonstrated that SOC predictions through spectral techniques, specifically SNV preprocessing of reflectance data via PLSR, could effectively replace SOC measurements obtained through routine chemical methods for soils in Biswanath district, Assam.