Detection of Diseases in Maize Plants by Analyzing Foliar Images Using Machine Learning Techniques

Abstract

The objective of this research was to develop a system capable of detecting diseases in maize leaves through the analysis of foliar images, using machine learning techniques to support early diagnosis in the agricultural sector. A custom dataset of field-captured images was built, and a preprocessing workflow was applied that included data augmentation, segmentation, feature extraction, and normalization. Subsequently, three models: EfficientNetB0, DenseNet121, and EKNN were trained and evaluated to determine the architecture with the best classification performance. The results showed that model performance varied according to how each approach processed visual features, with the EKNN model achieving the highest overall accuracy of 94.33%, outperforming EfficientNetB0 (89.70%) and DenseNet121 (88.04%). The CNN-based architectures achieved adequate classification in diseases with well-defined patterns but presented limitations when dealing with visually similar lesions. In contrast, the EKNN model, which relies on segmentation and enhanced feature extraction, achieved the best overall performance, demonstrating the importance of preprocessing in diagnostic accuracy. Finally, the selected model was integrated into a functional web application, validating its practical utility as a tool for the early detection of diseases in maize leaves. This research demonstrates that machine learning can effectively assist farmers and agricultural technicians in the efficient identification of plant diseases, contributing to improved productivity and better decision-making in the field.