Comparative Evaluation of Conventional Classification Techniques and Deep Learning Models in the Detection of Fungal Phytopathogens in Agricultural Crops (Avaliação Comparativa de Técnicas Convencionais de Classificação e Modelos de Deep Learning na Detecção de Fitopatógenos Fúngicos em Cultivos Agrícolas)

Resumo

https://doi.org/10.5281/zenodo.20183961

This project performed a comparative analysis between classical Machine Learning models and a Convolutional Neural Network (CNN) for the binary classification of rust in plant leaves, evaluating performance in both controlled and unseen (PlantDoc) data scenarios. The study went beyond traditional accuracy analysis by introducing a decision framework based on Predictive Efficiency (IEP) and Real-Time Viability (IVTR) indices. Results demonstrated that CNN superiority is not absolute without calibration: only after applying Fine-Tuning techniques with weighted loss did the model achieve maximum diagnostic reliability (F1-Score of 0.68 for the target class). Simultaneously, a hybrid approach, utilizing the CNN for feature extraction and Logistic Regression for classification, emerged as the most balanced solution, presenting the study's highest IVTR. Conversely, aggressive dimensionality reduction pre-processing, such as binarization, severely degraded predictive performance. The central conclusion is that optimal model selection depends on deployment constraints: the fine-tuned CNN is mandatory for clinical precision, while hybrid architectures offer the best trade-off for edge devices.