Evaluating Deep Learning-Driven Feature Selection Technique and Machine Learning Models for Non-destructive Oil Palm Ripeness Assessment Using Low-Cost Electrical Impedance Spectroscopy

The oil extraction rate of oil palm fruits decreases by 0.13% for every 1% of unripe fruit, emphasizing the need to harvest only ripe fresh fruit bunches (FFB). Spectroscopy has emerged as a reliable method for assessing FFB ripeness, but effective variable selection is essential due to high spectral dimensionality. This study presents a novel, non-destructive, and cost-effective method for determining oil palm fruit ripeness using electrical impedance spectroscopy. Impedance properties of 832 oil palm fruitlets at various maturity levels (unripe, ripe, and overripe) were measured within the 10–100 kHz frequency range and at 120 K resistance using a low-cost impedance analyzer. The spectra were analyzed with multivariate machine learning models, including Naïve Bayes, random forest (RF), and support vector machine, using variable selection methods of selectivity ratio, variable importance in projection, principal component analysis, and convolutional neural network (CNN). The CNN-RF approach yielded the best performance, with precision, sensitivity, specificity, and accuracy averaging 92.15%, 93.07%, 96.18%, and 92.77%, respectively. Thus, the electrical impedance spectroscopy coupled with CNN-RF multivariate calibration method offers a practical, low-cost, and reliable solution for the determination of FFB ripeness, achieving over 92% accuracy across all evaluation metrics.