Work

NIRS and Machine Learning Integration for Dynamic Monitoring of Nutritional Properties in Citronella Residues during Solid-State Fermentation

Year: 2026
Type: article
Abstract: Solid-state fermentation (SSF) can improve the nutritional value of agro-industrial residues by enriching microbial protein and modifying fiber structure. However, routine monitoring of these dynamic changes by conventional chemical methods is time-consuming. This study evaluated the feasibility of combining near-infrared spectroscopy (NIRS) with chemometrics and machine learning for rapid estimation of gross energy (GE), crude protein (CP), and crude fat (ether extract [EE]) in citronella residues during SSF. Citronella residue was fermented for 28 days under four white-rot fungal treatments and an uninoculated control (n = 60). NIR spectra (1000–2500 nm) were preprocessed using multiplicative scatter correction (MSC), and informative wavelengths were selected using the successive projections algorithm (SPA). Predictive models were developed using partial least squares regression (PLSR), support vector regression (SVR), and light gradient boosting machine (LightGBM), with a 3:1 split for calibration and prediction samples (45/15 samples). CP, the principal indicator of protein enrichment during SSF, can be effectively monitored using NIRS, with PLSR yielding the highest predictive accuracy among all models (R² = 0.72, RMSE = 0.29%, and MAE = 0.23%). The ability to monitor CP through NIR enabled the rapid evaluation of nutritional upgrading. GE showed moderate predictive performance, with the best results from LightGBM (R² = 0.55, RMSE = 192.34 J/g, and MAE = 146.02 J/g). The poor predictive capability of EE (best R² = 0.15) reflects the variability and low concentrations of EE, the weak and non-unique spectral signatures of EE, and strong spectral absorption in the sample matrix. Therefore, these data support the use of NIRS to rapidly evaluate CP yields and changes in the dynamic range of CP during SSF, but indicate that EE will continue to have limited value as a NIRS-based predictor in this highly complex fermented matrix.
Language: English
Field: Chemistry
Open Access status: gold
Estimated APC paid: $300
OpenAlex ID: W7170164779
FWCI:
Cites: 43
Cited by:
Related to: 10