Cotton Leaf Disease Detection and Classification Using Lightweight CNN Architecture

Abstract

Cotton is one of Bangladesh’s most valuable agricultural plant. Plant diseases are primarily caused by pest insects and pathogens, and if not handled promptly, they reduce productivity on a big scale. The main goal of this research is to classify 3 diseased class and one healthy class of cotton leaves using a deep learning-based lightweight CNN architecture. Despite having fewer parameters, the suggested model beats the pre-trained transfer learning models VGG16 and VGG19 in terms of accuracy. These model’s classification accuracy was only about 96.80% & 95.92% respectively, whereas the proposed model’s average classification accuracy is 99.42%. Additionally, the proposed model’s precision, recall, F1-score is about 99% and Area Under Curve (AUC) score is 99.89%, which is a sign of the model’s excellent performance.Parameters & size on the disk of the proposed model is much lesser than the compared transfer learning models. Using the Gradient Weighted Class Activation Mapping (Grad-CAM) approach, class activation maps were made to illustrate the disease identified by the proposed model, and a heatmap was created to denote the region responsible for categorization.