Artificial intelligence for sustainable farming with dual branch convolutional graph attention networks in rice leaf disease detection.

Abstract

Rice is susceptible to various diseases, including brown spot, hispa, leaf smut, bacterial leaf blight, and leaf blast, all of which can negatively impact crop yields. Current disease detection methods encounter several challenges, such as reliance on a single dataset that diminishes accuracy, the use of complex models, and the limitations posed by small datasets that hinder performance. To overcome these challenges, this paper presents a novel hybrid deep learning (DL) approach for classifying rice leaf diseases. The proposed model leverages two distinct datasets: the Rice Leaf Diseases Dataset and the Rice Disease Images Dataset. It enhances image quality through two advanced techniques: Upgraded Weighted Median Filtering (Up-WMF) to minimize noise and Aligned Gamma-based Contrast Limited Adaptive Histogram Equalization (AG-CLAHE) to improve image contrast. Features from these images are extracted using methods such as Discrete Wavelet Transform (DWT), Gray Level Run Length Matrix (GLRLM), and deep learning-based VGG19 features. To optimize model performance, the most significant features are selected using the Bio-Inspired Artificial Hummingbird (BI-AHB) method, which streamlines complexity. Classification of rice diseases is conducted using a new model known as the Dual Branch Convolutional Graph Attention Neural Network (DB-CGANNet). This model demonstrates remarkable performance, achieving 98.9% accuracy on rice leaf disease dataset and 99.08% on Rice diseases image, surpassing existing techniques. The proposed methodology enhances disease detection accuracy, facilitating improved management of rice crops and contributing to increased agricultural productivity.

Plant phenotyping relevance

イネ葉の病徴を画像から分類する深層学習手法の開発が中心であり、植物の病害状態を直接推定する画像ベース表現型解析に該当する。

Code and data availability

The paper uses two public rice leaf disease image datasets (cited as refs 31 and 32) as inputs, but the supplied blocks contain no availability statements, repository names, or URLs for the datasets or any authors' code. Analysis was done in Python with no code deposit or checkpoint release mentioned. The datasets are,

No evidence-backed public reproduction asset is currently recorded.