Metaheuristic-Optimized Self-Supervised Hybrid CNN–Transformer Framework for Plant Leaf Disease Classification
Automated plant leaf disease classification still faces three practical challenges: dependence on labeled data, weak transfer from controlled images to field conditions, and difficult selection of architecture and training hyperparameters. This study proposes a metaheuristic-optimized self-supervised hybrid CNN–Transformer framework that integrates local feature extraction, global attention, self-supervised pretraining, and cooperative ant colony optimization–particle swarm optimization (ACO–PSO). A lightweight CNN encoder extracts lesion, texture, and edge features, while a Vision Transformer models long-range spatial relations. Contrastive learning and masked prediction are used during self-supervised pretraining. ACO searches discrete architectural variables, whereas PSO tunes continuous training variables. The fitness function balances validation macro-F1, model size, and inference latency. In the based computational evaluation, the proposed cooperative ACO–PSO model achieved 99.72 ± 0.26% internal accuracy, 99.40 ± 0.31% internal macro-F1, and 91.06 ± 0.69% external macro-F1. The optimized model used 4.98 million parameters and required 6.45 ms per image. Ablation analysis increased external macro-F1 from 87.63 ± 0.36% for the supervised hybrid baseline to 91.11 ± 0.23% for the full model. The selected configuration used a patch size of 16, two Transformer blocks, four attention heads, and 64 CNN output channels. These results show that combining dual self-supervision with cooperative mixed-variable optimization can improve label-efficient plant disease classification while controlling model complexity.
Authors
- Rajaa Daami Resen
Institutions
- University of Information Technology and Communications (IQ)
Publication Details
- Journal
- Journal of Al-Farabi For Engineering Sciences
- Published
- 2026-11-14
- DOI
- https://doi.org/10.59746/bqqbph54
- Primary Topic
- Smart Agriculture and AI
- Type
- article
- Field-Weighted Citation Impact
- 0.00