A Light-weight CNN with integration of Explainable AI to detect Guava Leaf Disease

Abstract

This thesis is focused on the implementation and evaluation of a deep learning system that was specifically trained to be used for detection of guava leaf diseases. The designed system is based on a lightweight CNN model, i.e., M-Net; it combines XAI techniques to enhance interpretability. The goal was to create a model that guarantees reliable and efficient performance and easy interpretability for easy identification of diseases and timely action by farmers. A dataset comprising the images of guava leaves was used for the research that comprised two locations located in Ashulia and Savar, Dhaka, each image being identified on the basis of one of the seven precise disease classes. To begin with various pretrained Convolutional Neural Network (CNN) models, such as VGG19, DenseNet201, InceptionV3, ResNet152V2, and MobileNetV2 were tested for the task of guava leaf diseases detection. While MobileNetV2 had showed remarkable performance, M-Net performed on high drag, with 99.04% accuracy – a value that makes it more appropriate for rapid disease identification on mobiles in resource-limited settings. To enhance the transparency of the system, SHAP and Grad-CAM techniques were implemented and gave farmers a clear idea of how the model decides on recommendations leading to trust. For the system, an optimized web application for mobile devices was created allowing farmers to upload leaf images within a short time and receive the results of disease classification instantly. This study shows that using deep learning with XAI has the great promise to improve decision-making in agriculture particularly where expert help is lacking. The project indicates a positive innovation on the use of AI technology in managing crops diseases in the field.