Convolutional Neural Networks for Disease Detection in Cocoa Pod: A Roboflow Approach (#2058)

Authors

López-Arévalo, Oscar Alessandro

Reyes-Duke, Alicia María

Abstract

This paper addresses the development of a convolu- tional neural network (CNN) model capable of detecting diseases in cocoa fruits, specifically black pod and moniliasis, using images collected in the field and processed with the Roboflow platform. These diseases represent a significant challenge for farmers due to the economic losses they generate, highlighting the importance of early and accurate detection. 2,000 representative images were collected, adjusted with saturation variations (+25% and -25%) and capture distances (10 cm and 30 cm), which were used to train specialized neural networks. The developed models achieved outstanding metrics, exceeding 95% accuracy and recall in the detection of both diseases in the mixed network. Among the designed networks, the network focused on black cob showed a performance higher than 96%, while the network for moniliasis obtained slightly lower, but satisfactory results of 91%, highlighting the relevance of representative data and iterative training to optimize the model performance. The conclusions highlighted the effectiveness of the model developed, the relevance of the quality and diversity of the data collected, and the positive impact that this technology can have on agricultural management.