Development of a new method and an AI algorithm to predict the postharvest ripening time of avocados based on hyperspectral images

The accurate assessment of avocado ripening is essential for optimizing post-harvest logistics and reducing waste. This study explored the use of hyperspectral imaging to estimate the ripening time of Hass avocados, with an emphasis on transitioning to multispectral imaging for real-time and commercial applications. A total of 24 datasets were generated using different preprocessing methods and sampling techniques. The results showed that normalization at 811 nm effectively reduced illumination effects, and that predefining the sampling regions enhanced classification accuracy. To optimize the spectral efficiency, two band selection algorithms were developed. The local forward selection algorithm, which dynamically filters bands, achieved 92.3 % balanced accuracy using nine spectral bands, whereas the micro lambda-by-lambda algorithm, which selects the most informative pair of spectral bands, achieved 89.9 % accuracy, thus confirming the value of the development of low-cost multispectral sensors. Classification performance was evaluated on a range of validation and labeling methods. Most misclassifications lay within a one-day deviation from the actual ripening date. These findings can thus provide the groundwork for automated product sorting in packing houses that can improve supply chain efficiency and minimize post-harvest losses, and overall point to the feasibility of real-time multispectral sensing for agricultural applications . • Reduce the number of spectral bands for predicting avocado ripeness time. • Propose novel band selection methodology for high-resolution hyperspectral data. • Compare spectral preprocessing methods with avocado hyperspectral data. • Design low-cost multispectral sensors for real-time ripeness prediction. • Enable transition from lab-based imaging to packing-house implementation.