Enhancing paddy leaf disease diagnosis -a hybrid CNN model using simulated thermal imaging

Abstract

• The study introduced the use of thermal rendering to convert rice leaf images into thermal representations, enabling the early detection of diseases based on temperature variations before visible symptoms appeared. • A comprehensive evaluation of 18 CNN models through transfer learning was conducted, followed by statistical analysis using Duncan's multiple range test (DMRT) to select the best-performing model, with Darknet53 emerging as the most effective. • The hybridized Darknet53 model achieved exceptional performance metrics with an accuracy of 99.43 %, sensitivity of 99.43 %, specificity of 99.81 %, precision of 99.43 %, and F1 score of 0.99, offering a reliable solution for plant disease detection. • The model's high efficiency and improved performance make it suitable for real-time agricultural applications, particularly for small-scale farmers, by offering a fast and accurate method for diagnosing rice leaf diseases and supporting sustainable farming practices. Rice, as a staple crop globally, requires proactive and accurate disease detection to ensure sustainable production. This study introduces a novel hybrid Deep Learning approach integrating thermal imaging and model hybridization for early and precise detection of rice leaf diseases. A dataset of 5,932 self-generated rice leaf images was augmented with simulated thermal images to capture subtle temperature variations indicative of early stress responses prior to visible symptoms. This novel use of thermal imaging enhances disease diagnosis efficiency and practicality. Eighteen Convolutional Neural Network (CNN) models were evaluated using transfer learning, with statistical analysis via Duncan's multiple range test (DMRT) identifying Darknet53 as the best-performing model, achieving an accuracy of 95.79 %, sensitivity of 95.79 %, specificity of 95.93 %, and an F1 score of 0.96. To further improve performance, Darknet53 was hybridized by replacing its dense layer with a Support Vector Machine (SVM), resulting in significant enhancements. The hybrid model achieved 99.43 % accuracy, 99.43 % sensitivity, 99.81 % specificity, and an F1 score of 0.99. These results highlight the model's potential for real-time deployment in agricultural applications, providing an efficient and reliable solution for small-scale farmers. This research underscores the value of integrating thermal imaging with Deep Learning for advancing crop disease management and offers a framework for addressing other crop pathologies.

Topics & Concepts