Integração de horticultura de precisão e inteligência artificial na classificação pós-colheita de morangos

Abstract

Strawberry fruit quality prediction and grading currently rely primarily on manual and destructive methods, compromising efficiency and standardization in the post-harvest process. The use of digital images and machine learning algorithms emerges as a promising alternative for automating these steps accurately and non-invasively. This study aimed to evaluate machine learning-based models to develop non-destructive methods for measuring strawberry fruit weight and commercial grading. To this end, fruits of eight strawberry genotypes were produced in a greenhouse, harvested, and analyzed between September and October 2023 at UFLA, Lavras, Minas Gerais. Each fruit was photographed in a standardized studio with a mobile device, and the images were processed for segmentation, background removal, and extraction of morphological attributes (e.g., area, perimeter, shape factor). In parallel, the fruits were evaluated in the laboratory for weight, diameter, and quality parameters at scale. Eight supervised learning algorithms—SVM, RF, DT, KNN, MLP, XGB, LGBM, and MLR—were applied to predict strawberry mass and classify it. Analyses were performed in Python on Google Colab. The models were trained on 70% of the data and tested on the remaining 30% using k-fold cross-validation. For mass prediction, the models were evaluated using R², RMSE, MAE, and d, and for classification, using the metrics accuracy, F1-score, recall, and precision. For fruit mass prediction, in the testing phase, the SVM model demonstrated the best results for both datasets. KNN performed well on the manual dataset, and XGB performed similarly to SVM for the image dataset. The LGBM model achieved the best performance in training but showed a decline in testing, suggesting overfitting. For both datasets, the MLP model performed less well. The use of models for fruit classification achieved accuracy ranging from 98% to 96% for manual data, with emphasis on the SVM, XGB, LGBM, and DT models. For image data, all models achieved accuracy above 0.91, except for MLP, which achieved 0.83. Although manual data provided greater stability and accuracy for the models tested, the variables extracted from the images demonstrated potential for predicting and classifying strawberry fruits. This reinforces the use of standardized images as a non-destructive, efficient, and scalable tool for estimating fruit mass and classifying them, paving the way for practical applications in automated monitoring, with gains in agility and cost reduction, contributing to advances in precision agriculture and supporting the automation of mass prediction for strawberry crops.