Avocado production is mostly confined to tropical and subtropical regions, leading to lengthy distribution channels that, coupled with their unpredictable post-harvest behavior, render avocados susceptible to significant loss and waste. To enhance the monitoring of 'Hass' avocado ripening, a data-driven tool was developed using a deep learning approach. This study involved monitoring 478 avocados stored in three distinct storage environments, using a 5-stage Ripening Index to classify each fruit's ripening phase based on their shared characteristics. These categories were paired with daily photographic records of the avocados, resulting in a database of labeled images. Two convolutional neural network models, AlexNet and ResNet-18, were trained using transfer learning techniques to identify distinct ripening indicators, enabling the prediction of ripening stages and shelf-life estimations for new unseen data. The approach achieved a final prediction accuracy of 88.8% for the ripening assessment, with 96.7% of predictions deviating by no more than half a stage from their actual classifications when considering the best side of the samples. The average shelf-life estimates based on the attributed classifications were within 0.92 days of the actual shelf-life, whereas the predictions made by the models had an average deviation of 0.96 days from the actual shelf-life.
Read full abstract