Implementation of Deep Learning Approach with Attention Mechanism for Multiclassification for Plant Disease Detection

Abstract

Introduction: The agricultural industry is greatly affected by the essential job of disease identification from leaf images of plants. The ability to recognize these diseases through a simple interface or machine learning model can empower farmers with better preparation strategies.

Objectives: The goal of this study is to build a Convolutional Neural Network (CNN) model that can identify plant diseases more reliably.The model leverages the MobileNetV2 network and incorporates an attention mechanism to improve feature extraction and prediction accuracy.

Methods: The proposed CNN model was built on the MobileNetV2 architecture, with modifications to include an attention mechanism for better feature extraction. The study experimented with increasing the number of layers in the CNN and tested various activation functions, including hard-sigmoid, to determine their impact on the model's specificity, sensitivity, F-measure, recall, and Matthews correlation coefficient.

Results: The introduction of an attention mechanism and the increase in CNN layers significantly enhanced model performance. For tomato plants, specificity improved from 0.47 to 0.99, and accuracy increased from 71% to 99%. Apple plants saw accuracy improvements from 62% to 98%. Considering specificity, sensitivity, F-measure, recall, and Matthews correlation coefficient, among the activation functions that were investigated, hard-sigmoid fared the highest.

Conclusions: The study demonstrates the effectiveness of deep learning, particularly CNNs enhanced with attention mechanisms, in accurately identifying plant diseases from leaf images. The advancements in model architecture and activation function selection significantly improve prediction accuracy, offering a powerful tool for agricultural disease management.