Multimodal sensing approach using remote and onsite data for estimating the pre-harvest ripening stage of Hass Avocado with machine learning algorithms

Abstract

In recent years, avocado has gained significant global importance due to its nutritional benefits, and rising consumer demand, becoming a staple in health-conscious diets. This growing interest has also raised concerns about environmental sustainability, and as a result, efforts are being made to promote more sustainable farming practices while meeting the rising demand. In this study, we present a tool designed to enhance the efficiency of the pre-harvest process and improve avocado quality. We propose a multimodal sensing scheme integrating three different data sources: a portable multispectral system for in situ measurements, satellite imagery and, onsite environmental sensors to estimate the fruit ripening stage. This combined remote and onsite yielded a high correlation with dry matter content, considered here as the reference indicator of avocado ripening, across three consecutive harvest seasons. The performance of various machine learning techniques was evaluated using different combinations of these datasets. Notably, the artificial neural network (ANN) model achieved the highest accuracy (0.74) and recall (0.96) for predicting the overripe avocado class. Therefore, ANN model was extended to regression models, where all of them have demonstrated high predictive accuracy, with R2 coefficient ranges from 0.81 to 0.91. The online data achieved the highest coefficient (0.91), providing a slightly better performance compared to the offline model. Nonetheless, predictions based solely on multispectral data remain valuable, particularly when online data are unavailable.