A Domain-Invariant and Edge-Efficient Framework for Leaf Disease Classification

Abstract

Reliable leaf disease diagnosis is essential for improving agricultural productivity. However, deploying such systems in real-world environments remains challenging due to domain variations across different crops and imaging conditions. Models trained on controlled datasets often fail to generalize effectively when crop characteristics, environmental conditions, and image acquisition settings vary. Moreover, many traditional domain adaptation methods incur high computational costs, making them unsuitable for deployment on Tiny Machine Learning (TinyML) platforms with limited computational resources. To address these challenges, a domain-adaptive framework is proposed that integrates a custom Squeeze-and-Excitation Residual Network (SE-ResNet) encoder, the Simple Framework for Contrastive Learning of Visual Representations (SimCLR), and a lightweight Multi-Adversarial Domain Adaptation (MADA)-Lite mechanism. The SimCLR component enhances feature representation by learning instance-level similarities between augmented views of the same image, thereby improving feature separability and robustness without using disease class labels during representation learning. In addition, the redesigned MADA-Lite module employs a feature-confidence-guided adversarial alignment strategy to reduce cross-domain discrepancies while preserving computational efficiency. In the experimental setup, the PlantVillage dataset served as the source domain, while the RiceDiseases dataset served as the target domain. The proposed framework achieved a cross-domain accuracy of 69.98% (mean) while maintaining a compact model size of only 137.12 KB after 8-bit integer (INT8) quantization, demonstrating its suitability for deployment on ultra-low-power and resource-constrained edge devices. The proposed framework combines contrastive representation learning, lightweight adversarial domain alignment, and efficient model compression within a unified TinyML-oriented architecture. The effectiveness and robustness of the framework were further validated through extensive experimental evaluations, including baseline comparisons, multi-seed analysis, additional dataset evaluations, and comprehensive ablation studies. The complete source code, trained models, experimental configurations, and deployment-related files are publicly available in the GitHub repository.