LightET-FusionNet as a Lightweight Deep Ensemble Model for Scalable Maize Disease Classification

Accurate detection and classification of maize leaf diseases is essential for maintaining crop health and improving yield outcomes. Traditional convolutional neural networks struggle to identify lesions of varying sizes and shapes, often missing small disease spots or misclassifying large spread lesions. This project proposes an advanced hybrid approach combining Adaptive Multi-Scale Dynamic Receptive Field Network (AM-DRFNet) and Dual Dynamic Attention Mechanism (DDAM) to enhance disease classification performance. AM-DRFNet employs multi-dilated convolutions with a spatial gating network to provide pixel-wise adaptive receptive fields, effectively capturing both small and large lesion patterns. DDAM further refines feature representation through sequential channel and spatial attention modules, focusing on the most discriminative features and highlighting lesion regions. The system is implemented as a web-based application using HTML, CSS, JavaScript, and Python Flask, offering modules for user registration, login, disease classification, and history tracking. The dataset employed consists of maize leaf images collected from multiple disease categories, including blight, common rust, gray leaf spot, and healthy leaves. The models are trained and evaluated using performance metrics such as accuracy, precision, recall, and F1-score, demonstrating superior performance compared to baseline CNN models. This research contributes a scalable and interpretable framework for multi-class leaf disease classification, addressing challenges of variable lesion sizes and improving feature localization for enhanced model accuracy.

Keywords: Maize leaf disease, AM-DRFNet, dual attention, pixel-wise receptive field, multi-scale lesions, feature refinement, channel attention, spatial attention, leaf classification, deep learning.