Quality control of Alfalfa using FTIR Spectroscopy in combination Machine Learning: Optimisation of feature extraction algorithms and metaheuristic algorithm

This study aims to develop a robust, fast, and nondestructive quality control method for alfalfa. Its innovation lies in integrating FTIR with multi-strategy feature extraction and metaheuristic hyperparameter optimization into a unified workflow, thereby reducing spectral redundancy and eliminating instability and poor transferability caused by manual tuning. We collected 186 alfalfa samples (chopped material and pressed hay blocks) and applied SG smoothing, first derivative, MSC, and airPLS for preprocessing. On this basis, we enhanced conventional feature selection algorithms to form aSPA, eCARS, and OVMPCA to obtain information-dense feature subsets, while employing GA, CEO, and SFOA to optimize hyperparameters for PLSR and random forest (RFR). Using an 8:2 split and evaluation by R²/RMSE/RPD/SEP, the best workflow (SG first derivative + eCARS + GA-PLSR) achieved test-set R² values of 0.932 (crude protein), 0.984 (fat), 0.988 (starch), and 0.991 (neutral detergent fiber), with all RPD > 3.5 and all SEP < 0.5 %. This framework addresses key challenges such as the high dimensionality and collinearity of infrared spectra and spectral drift caused by sample and condition differences, significantly improving model stability and transferability, and it can be extended to quality control scenarios for other agricultural products.