Metaheuristic Optimization for Deep Learning in Plant Disease Detection: A Hybrid Approach

Abstract

This study investigates metaheuristic hyperparameter optimization for deep learning–based plant disease detection across two datasets: Dataset A (1,530 images; three classes: Healthy, Powdery, Rust) and a large multi-crop corpus evaluated in a binary Healthy/Diseased setting with an 80/20 training–validation split. A hybrid optimizer is proposed that interleaves Dragonfly Algorithm (DA) for population-wide exploration with Firefly Algorithm (FA) for elite intensification (DA–FLA), and is applied to five pretrained CNN backbones (DenseNet, VGG19, InceptionV3, MobileNet, Xception). All models are trained under an identical 50-epoch protocol. On Dataset A, DenseNet provides the strongest baseline (accuracy/macro-F1 = 0.9733/0.9735), which rises to 0.9800/0.9800 with DA–FLA tuning. On the large-scale binary corpus, Xception and DenseNet perform competitively (≈0.9846 macro-F1 and 0.9838 macro-F1, respectively), while the optimized Xception attains 0.9924 accuracy and 0.9913 macro-F1. A one-way ANOVA with Tukey HSD confirms significant performance differences (p < 0.001), with optimized Xception outperforming all comparators. The hybrid search introduces modest training overhead but leaves inference cost essentially unchanged. Results demonstrate that balancing global exploration with local exploitation yields reproducible, statistically supported gains, advancing accurate and efficient plant disease diagnostics suitable for mobile/edge deployment and supporting early intervention and sustainable farming practices.