A Lightweight Hybrid Deep Feature and Discriminant Learning Framework for Early Cotton Disease Diagnosis
DOI:
https://doi.org/10.62643/ijerst.2026.v22.n2.pp1-9Keywords:
Cotton Leaf Disease Detection, Convolutional Neural Network, ResNet-50, Deep Learning, Agricultural IoT.Abstract
Recent studies report that cotton leaf diseases lead to yield losses of up to 30–40% when not detected at early stages, while automated vision-based systems achieve diagnosis accuracy exceeding 90% compared to manual inspection. Despite these advancements, existing models exhibit limitations such as reduced class-wise recall and high computational complexity when handling subtle early-stage symptoms. The motivation of this work is to develop an accurate, reliable, and efficient system for early cotton disease detection to support timely intervention and improve agricultural productivity. Manual diagnosis depends heavily on expert knowledge and remains time-consuming, subjective, and inconsistent under varying environmental conditions, while early symptoms often remain visually indistinguishable, leading to delayed treatment. To overcome these challenges, a lightweight Convolutional Neural Network (CNN)-based ensemble framework is proposed for early-stage cotton disease classification. A curated cotton leaf dataset comprising multiple disease categories and healthy samples is pre-processed and divided into training and testing sets to ensure unbiased evaluation. Baseline comparisons are conducted using Bagging, Light Gradient Boosting Machine (LGBM), and a CNN–Random Forest (RF) hybrid model. The proposed framework utilizes ResNet-50 for deep feature extraction, capturing both low-level texture and high-level disease-specific patterns, followed by Linear Discriminant Analysis (LDA) for classification to enhance class separability. Experimental results demonstrate that the ResNet-50 with LDA model achieves superior accuracy, precision, recall, and improved class-wise discrimination, particularly for early-stage diseases such as Aphids, Army worm, Bacteria blight, Powdery mildew, and Target spot. The lightweight architecture also reduces computational cost and inference time, ensuring suitability for real-time deployment in resourceconstrained agricultural environments.
Downloads
Published
Issue
Section
License

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.













