Bio-Inspired Optimization of Convolutional Neural Networks for Enhanced Maize Disease Detection in Precision Agriculture

Resumen

Objective: This study develops an automated, image-based system for early detection of four key maize diseases: Puccinia sorghi, Cochliobolus carbonum, Bipolaris maydis, and Exserohilum turcicum using Convolutional Neural Network (CNN) optimized through bio-inspired algorithms.

Design/methodology/approach: A dataset of 17,280 high-resolution images across six disease stages was preprocessed and used to train CNNs. Two metaheuristic algorithms, Spider Monkey Optimization (SMO) and Squirrel Search Algorithm (SSA), were applied to optimize weights and hyperparameters. An 80/20 training-validation split was used, and performance was assessed with standard classification metrics.

Results: The SMO-optimized CNN outperformed SSA, achieving 95.14% accuracy versus 89.74%. SMO also yielded better precision, recall, and F1-scores, showing strong performance even in distinguishing visually similar symptoms.

Study Limitations/Implications: SMO’s computational demands may limit its usability in low-resource settings. Some classification confusion persisted, highlighting the need for improved feature extraction and broader datasets for increased generalization.

Findings/conclusions: CNNs optimized with SMO provide a robust tool for maize disease diagnosis, reducing analysis time and enabling more precise crop management. Future work will explore hybrid optimization methods to enhance scalability and real-time application in precision agriculture.