Multi-Input Deep Learning Model with RGB and Hyperspectral Imaging for Banana Grading

Abstract

Grading is a vital process during the postharvest of horticultural products as it dramatically affects consumer preference and satisfaction when goods reach the market. Manual grading is time-consuming, uneconomical, and potentially destructive. A non-invasive automated system for export-quality banana tiers was developed, which utilized RGB, hyperspectral imaging, and deep learning techniques. A real dataset of pre-classified banana tiers based on quality and size (Class 1 for export quality bananas, Class 2 for the local market, and Class 3 for defective fruits) was utilized using international standards. The multi-input model achieved an excellent overall accuracy of 98.45% using only a minimal number of samples compared to other methods in the literature. The model was able to incorporate both external and internal properties of the fruit. The size of the banana was used as a feature for grade classification as well as other morphological features using RGB imaging, while reflectance values that offer valuable information and have shown a high correlation with the internal features of fruits were obtained through hyperspectral imaging. This study highlighted the combined strengths of RGB and hyperspectral imaging in grading bananas, and this can serve as a paradigm for grading other horticultural crops. The fast-processing time of the multi-input model developed can be advantageous when it comes to actual farm postharvest processes.