Comparação de arquiteturas de redes neurais convolucionais para a detecção de doenças foliares do tomateiro

Abstract

The automated recognition of leaf diseases is one of the main challenges of Agriculture 4.0, requiring methodologies that integrate agronomic knowledge, image collection for monitoring, and advanced machine learning techniques. This study aims to perform a comparative analysis of different convolutional neural network (CNN) architectures applied to the detection of leaf diseases in tomato plants, including target spot. Three widely recognized architectures were used: ResNet-50, Inceptionv3, and VGG-16, exploring combinations of hyperparameters such as learning rate, optimizers, and the use of weight decay. Additionally, activation maps were employed to identify relevant visual patterns that influence the models’ decisions. The results show that ResNet-50 achieved the best overall performance and stability, followed by Inception-v3, while VGG-16 exhibited greater sensitivity to training configurations. Through this analysis, we aim to understand the impact of these variations on network performance, providing valuable insights for improving models applied to crop protection management and precision agriculture.