Deep Learning Approach for Cotton Plant Disease Identification and Severity Evaluation

Abstract

Cotton productivity plays a crucial role in the global agricultural economy; however, various leaf diseases significantly threaten crop yield and fiber quality. Early and accurate disease detection is essential for effective crop management, yet traditional inspection methods are time-consuming, labor-intensive, and dependent on expert knowledge, often leading to inconsistent results. Conventional machine learning approaches also face limitations in real-world agricultural environments due to variations in lighting conditions, complex backgrounds, and similarities between disease symptoms. To address these challenges, this research proposes an intelligent framework called Cotton Plant Disease Identification Using ResMobNet with Attention-Guided Localization and Severity Analysis (CPDI-RMN). The proposed system integrates advanced image preprocessing, hybrid feature extraction, deep learning classification, and attention-based localization to create a comprehensive disease detection framework. Initially, cotton leaf images are collected from a comprehensive dataset and preprocessed through image resizing, noise removal, contrast enhancement, and Min–Max normalization to improve visual quality and ensure stable model training. Data augmentation techniques such as rotation, flipping, zooming, and brightness adjustment are applied to enhance dataset diversity and improve model robustness against overfitting. For feature enhancement, contour visualization and geometric feature representation are combined with texture analysis using the Gray-Level Co-occurrence Matrix (GLCM) and Laplacian filtering. The core of the framework is the ResMobNet hybrid architecture, which integrates ResNet-50, EfficientNet-B3, and MobileNet-V2 to capture multi-scale spatial and texture features while maintaining computational efficiency. Gradient-Weighted Class Activation Mapping (Grad-CAM) is employed to generate attention maps for disease localization, followed by segmentation to isolate infected regions. Disease severity is then quantified by calculating the percentage of infected leaf area and classifying it into mild, moderate, and severe categories. Experimental results using five-fold cross-validation demonstrate that the CPDI-RMN model achieves 98.85% classification accuracy, outperforming CNN, ANN, ResNet, and MobileNetV2 models. Additionally, the attention-based localization achieves 96.8% Intersection over Union and 98.0% Dice Score, indicating highly accurate disease region detection. Overall, the proposed framework provides a reliable and scalable solution for intelligent cotton disease monitoring and supports precision agriculture through data-driven crop management.

Plant phenotyping relevance

ワタ葉画像から病変領域を抽出し、感染面積率に基づいて病害重症度を定量化する画像解析手法が研究の中心であるため、植物フェノタイピング手法として含める。

Code and data availability

The paper describes a cotton disease detection framework using an external cotton leaf disease dataset, but provides no public URL, repository, or availability statement for the dataset, code, or trained models. The dataset is only cited via reference [5] (a prior publication), and no authors' public asset links appear

No evidence-backed public reproduction asset is currently recorded.