Machine Learning and NIR Spectroscopy Quantify Soil Carbon and Nitrogen.

Traditional methods for analyzing soil composition can be slow and costly. This research explores Near-Infrared (NIR) spectroscopy as a faster, cheaper, and non-destructive alternative for quantifying carbon (C) and nitrogen (N) content in specific soil types, namely Oxisols and Inceptisols. The study leverages a portable MyNIR device to collect spectral data, which is then processed using machine learning. The methodology involved a thorough evaluation of different data preprocessing techniques, with the Savitzky-Golay filter and a robust outlier removal method proving most effective. Various validation strategies were also compared to ensure model reliability. The core of the predictive system uses stacking ensemble learning, combining Partial Least Squares (PLS), Support Vector Regression (SVR), and Ridge regression as base models, with a linear regression meta-model. The resulting models achieved strong performance metrics (R2, RMSE, MAE, RPD > 2.0) and low overfitting, validating the potential for rapid and accurate C and N quantification.