CLASSIFICAÇÃO DE DANOS FOLIARES USANDO DEEP LEARNING

Abstract

Soybeans, an essential source of protein for human and animal nutrition, are especially vulnerable to attacks by defoliating insects. Manual collection and identification of insect forms present in crops is a time-consuming and error-prone practice. Considering the positive impact of monitoring techniques in cultivated fields, the use of mobile devices along with computer vision tools has been gaining prominence in precision agriculture.

Unlike other studies that focus solely on the development of models for predictive detection or diseases, this work proposes an approach based on the indirect classification of clues through leaf damage, in addition to integrating the model into a mobile application as a proof of concept, demonstrating its practical application.

We compared three neural network architectures, ResNet50, VGG16, and InceptionV3, to classify two soybean phrases (green beetle and soybean caterpillar). The three architectures performed very well in classification with accuracy above 90%, but VGG16 showed the best performance, with 95.75%. The results demonstrate the ability of deep learning models to recognize insects based on the damage they cause to pages and highlight the potential for integration between deep learning models and mobile applications, thus providing a viable alternative for monitoring guidelines in soybean crops.