Moisture content prediction in durian husk biomass via near infrared spectroscopy coupled with aquaphotomics and explainable machine learning

Abstract

Accurate determination of moisture content is essential for energy efficiency and biomass management for fuel materials such as durian husk. Traditional methods of determining biomass moisture content are time-consuming and require specialized expertise, posing challenges for continuous monitoring. To address this limitation, this study applies Near Infrared Spectroscopy (NIRS) combined with machine learning models to rapidly and accurately assess moisture content. Both linear Partial Least Squares Regression (PLSR) and non-linear approaches were used, including Support Vector Machines (SVM), Artificial Neural Networks (ANN), and Extreme Gradient Boosting (XGB). The application of preprocessing techniques, notably the Savitzky-Golay second derivative (SD) and Standard Normal Variate (SNV), significantly augmented the predictive performance, highlighting the importance of data preprocessing in spectral analysis. Synthetic spectral augmentation using Gaussian noise revealed that while SVM and ANN exhibited near-perfect performance, SVM demonstrated quantifiable reliability. This study also demonstrates SVM as the most sensitive and reliable method for detecting and quantifying moisture content in durian husk. This research contributes novel insights to biomass analysis, highlighting the benefits of integrating NIRS and feasibility of explainable machine learning techniques to identify water related spectral parameters to advance aquaphotomics, thereby advancing rapid and accurate biomass characterization. • NIRS coupled with SVM demonstrates potential for quick moisture content analysis in durian husk biomass. • Explainable machine learning can be used to identify and characterize the water spectral pattern. • Traditional random sampling outperforms KS and SPXY methods when calibration data is limited. • SHAP provides clear explanations for model differences, enhancing the selection and transparency of SVM, ANN, XGB models.