Aplicación Web Basada en Aprendizaje Profundo para la Clasificación Multiclase de Enfermedades Foliares en Cultivos Agrícolas

Foliar diseases are one of the main factors that limit agricultural productivity by reducing crop quality and yield. Traditional identification methods, based on visual inspection, are costly, time-consuming, and prone to human error, which highlights the need for automated solutions. This study aim...

Descripción completa

Detalles Bibliográficos
Autor: Torres Talaverano, Luz Elena
Tipo de recurso: artículo
Estado:Versión publicada
Fecha de publicación:2025
País:Perú
Institución:Universidad Nacional Mayor de San Marcos
Repositorio:Revistas - Universidad Nacional Mayor de San Marcos
Idioma:español
OAI Identifier:oai:revistasinvestigacion.unmsm.edu.pe:article/32350
Acceso en línea:https://revistasinvestigacion.unmsm.edu.pe/index.php/rpcsis/article/view/32350
Access Level:acceso abierto
Palabra clave:Deep learning
Convolutional neural networks
Transfer learning
Xception
Web application
Aprendizaje profundo
Redes neuronales convolucionales
Aprendizaje por transferencia
Aplicación web
Descripción
Sumario:Foliar diseases are one of the main factors that limit agricultural productivity by reducing crop quality and yield. Traditional identification methods, based on visual inspection, are costly, time-consuming, and prone to human error, which highlights the need for automated solutions. This study aims to develop a web application based on deep learning for the multiclass classification of foliar diseases in apple, maize, pepper, potato, and tomato crops. The approach employed transfer learning with fine-tuning, using pretrained convolutional neural network architectures and training models on a dataset of 35,725 images distributed across 23 classes. Among the evaluated architectures, Xception achieved the best performance, with an accuracy of 98.1% and a macro F1-score of 97.6%. This model was integrated into a web application structured in a layered architecture based on the client-server model, where tests with real images confirmed its practical validity and demonstrated stable inference times below 0.1 seconds per image. The results show that the proposed system constitutes a decision-support tool for crop management with potential application in real agricultural scenarios.