The Hybrid Deep Learning Approach for Cotton Leaf Disease Detection and Management Using Fine-Tuned VGG16 and Inception v3 Models

Cotton Leaf Disease Detection

Abstract

Agricultural losses cause significant barriers to the practice of sustainable farming, and the only way to prevent them is through earlier interventions. This paper introduces a novel technique of Hybrid Deep Learning Method (HDLM) to augment the classification accuracy of the diseases by using a voting system at the output layers on the tuned VGG16, Inception v3, ResNet-50, and MobileNet-v2 models. The model was trained and tested on a structured dataset of 3,000 images, which upon manual curation, were categorized into Aphids, Armyworm, Bacterial Blight, Powdery Mildew, Target Spot, and Healthy Leaves. Out of the 3,000, 2,400 were labeled for the model to learn on and 600 were held for validation. In order to improve the model's external predictive ability, strategies for data augmentation were used. The proposed method of HDLM achieved an accuracy of 98.56% which is a significant improvement over the benchmark models, namely AlexNet, DenseNet-121, ResNet-50, LeNet-5, and a 7-layer CNN, which report an accuracy between 90 and 95 percent. The model also features a disease management recommendation system which provides users with relevant strategies proven to aid farmers and agricultural constituents. The research demonstrates the value addition offered by ensemble deep learning to the field of plant disease detection. In the future, the focus is to broaden the model’s adaptability and effectiveness across different agricultural settings by expanding the dataset and utilizing more sophisticated augmentation methods.