NIR and Machine Learning-Based Rapid Monitoring of pH and Moisture in Citronella Residue Fermentation

Abstract

The moisture content and pH level are the primary parameters influencing the efficacy of the solid-state fermentation (SSF) process of feed ingredients derived from agricultural residues. Due to the potential for contamination and process failure, the traditional methods for detecting changes in moisture content and pH level during the SSF process are unfeasible. Consequently, there is an urgent necessity to develop alternative techniques that yield highly accurate results without being time-consuming or labor-intensive. One of the most promising sensing techniques for in-line applications is near-infrared (NIR) spectroscopy. This study employed both classical and advanced machine learning (ML) models based on NIR spectra to develop a predictive model for moisture content and pH level in the thermophilic SSF process of citronella residues (CR) feed for ruminant livestock using different white-rot fungi. Principal component analysis (PCA) was utilized on the NIR spectra to extract relevant features for input into the ML models. Among the models evaluated, support vector regression (SVR) demonstrated the highest predictive accuracy (R2p of 1.00 for both moisture content and pH level), outperforming light gradient-boosting machine (LightGBM) and random forest (RF). Although SVR achieved the highest predictive accuracy, LightGBM offers practical advantages, including faster training, lower computational demand, and better scalability for large datasets. With competitive predictive performance (R2p of 0.95 for moisture and 0.87 for pH), LightGBM provides a strong alternative for applications requiring real-time or resource-efficient deployment. In conclusion, integrating NIR spectroscopy with ML offers a promising pathway for intelligent and real-time monitoring in large-scale SSF applications, contributing to sustainable valorization of agricultural residues into high-quality ruminant feed.