Avaliação comparativa de arquiteturas de redes neurais convolucionais na detecção e classificação de distúrbios em folhas de macieira

Abstract

This work proposes a Computer Vision system based on Deep Learning for the automatic classification of apple leaf disorders, aiming to diagnose phytopathogens and nutritional deficiencies. From a public dataset, 32,384 patches were generated and categorized into five classes (Scab, Glomerella, Potassium and Magnesium deficiencies, and Herbicide damage). The data were balanced via random undersampling and efficiently processed in an ETL pipeline using TFRecords. The experimental protocol evaluated five Convolutional Neural Networks (VGG16, Res-Net50, DenseNet121, MobileNetV3Large, and EfficientNetB0) trained with a two-stage Transfer Learning approach. Validation employed a blind test partition (10% Hold-out) alongside stratified cross-validation (5-Fold Cross-Validation). The MobileNetV3Large model achieved the best performance, recording an overall accuracy of 93.36% and a macro F1-Score of 91.87%, followed by EfficientNetB0 (92.96% accuracy). The greatest inferential difficulties occurred in distinguishing between Magnesium and Potassium deficiencies due to colorimetric similarities, while the accuracy for fungal diseases exceeded 95%. It is concluded that integrating optimized models with rigorous statistical validation establishes an automated and reliable tool for precision agriculture monitoring.