A Lightweight EfficientNet-B0 Framework for Real-Time, Mobile-Deployable Cotton Leaf Disease Classification

Abstract

Cotton is a principal cash crop and a cornerstone of the textile-driven economies of several South Asian countries, yet its productivity is persistently undermined by foliar diseases that are difficult to diagnose accurately and in a timely manner using manual field inspection. Although deep learning has substantially advanced automated plant disease recognition, most reported systems are validated on a narrow set of disease categories, trained and tested under controlled imaging conditions, and evaluated without regard to the computational constraints of on-device, real-time use by smallholder farmers. This study presents a lightweight, end-to-end deep learning framework for cotton leaf disease classification built around a fine-tuned EfficientNet-B0 backbone and validated on the SAR-CLD-2024 dataset, comprising 9,137 field-acquired images spanning seven classes: bacterial blight, curl virus, herbicide damage, leaf hopper damage, reddening, variegation, and healthy leaves. The framework combines geometric, photometric and viewpoint-based data augmentation with transfer learning and systematic hyper parameter tuning to improve robustness to lighting, orientation and background variations common in unconstrained field photography. In addition to classification accuracy, the trained model was converted and quantized for on-device inference and integrated into an Android mobile application that supports both live camera capture and gallery-based inference for offline, real-time disease diagnosis. Experimental findings show that the proposed framework competes with or outperforms recently reported cotton-disease classification studies while remaining substantially lighter than comparable architectures, enabling practical deployment. This is a preliminary methodological and empirical report; larger-scale field validation and formal ablation analysis are identified for immediate future work.