Using metaheuristic algorithms to improve the estimation of acidity in Fuji apples using NIR spectroscopy

This study focuses on the spectrochemical estimation of pH and titratable acidity (TA) of apples of Fuji variety at different stages of ripening. A novel approach is proposed for near-infrared (NIR) spectral analysis using hybrid machine learning methods that combine artificial neural networks (ANN) and metaheuristic algorithms. One hundred twenty samples were collected at three ripening stages and spectral data within two bands of NIR were extracted from each sample to predict the acidity properties. Alternatively, the 4 most effective wavelengths were also selected using a hybrid of ANN and the cultural algorithm. The experimental results prove that the models using spectral bands and the 4 most effective wavelengths are comparable, with a correlation coefficient, R, of 0.926 for the prediction of pH and 0.925 for TA using spectral bands, while for the second approach the R obtained were 0.924 and 0.920 for pH and TA, respectively. The models could not accurately predict extremely high or low pH and TA values, due to the clusters that formed after regression. However, for a classification problem in low/high acidity, both approaches were able to achieve a high accuracy of 100% for pH and 99.2% for TA.