Non-destructive Classification of Harvested Terung Asam (Solanum lasiocarpum Dunal.) Utilising Thermal Imaging and Machine Learning Across Different Storage Days

Abstract

Thermal imaging has non-invasive qualities which have attracted the interest of researchers as it could improve agricultural methods, especially in the field of fruit classification based on their storage duration. This study explores the potential of thermal imaging and machine learning for the non-destructive classification of Terung Asam (Solanum lasiocarpum Dunal.) fruit at different storage durations. Nine thermal image parameters were analysed to monitor changes in fruit characteristics over day 0, day 7, day 14, day 21 and day 28. Key thermal image parameters, including major axis length (MajorAL), maximum intensity (MaxInt) and mean value within the region of interest (MeanROI), exhibited significant variations throughout the storage period, reflecting changes in fruit morphology and surface temperature associated with ripening and moisture loss. Correlation analysis revealed strong correlation between these parameters. The strongest correlation was found between MajorAL and MinorAL (r = 0.967) and between MaxInt and MajorAL (r = 0.962). Five machine learning models, i.e. Fine Decision Tree, Medium Decision Tree, RUSBoost Tree, Boosted Tree and Fine k-Nearest Neighbour (kNN), were evaluated for classification performance. Fine and Medium Decision tree achieved the highest classification accuracy at 86.7%, effectively distinguishing Terung Asam fruits based on storage duration. This study underlines thermal imaging as a reliable, non-invasive tool for post-harvest classification of Terung Asam fruit, improving storage monitoring and reducing waste. Future research should focus on deep learning integration to enhance classification performance over extended storage periods.