Estimating soil organic carbon by integrating mid-infrared spectroscopy with explainable machine learning

Organic carbon is a crucial soil parameter that significantly influences soil fertility, structural stability, microbial activity, and overall agricultural productivity. Traditional methods for measuring organic carbon are often time-consuming, labor-intensive, and expensive. This study explores the use of spectroscopy integrated with machine learning as a non-destructive and efficient alternative. Specifically, the Partial Least Squares Regression (PLSR) and Random Forest (RF) models were employed to predict soil organic carbon (SOC) based on mid-infrared spectroscopy (MIR) data within the spectral range of 600–4000 cm−1. The RF model was optimized through hyperparameter tuning using the Grid-Search method, achieving high predictive accuracy with a coefficient of determination (R2P) value of 0.98 and a root mean square error (RMSEP) of 1.45%. Furthermore, explainable artificial intelligence (XAI) based Shapley Additive Explanation (SHAP) technique was utilized to interpret the model and evaluate the contribution of specific wavelengths to the predictions. The results underscore the effectiveness and increasing relevance of combining explainable machine learning with spectroscopy for reliable and practical applications in the agricultural sector.