Integrating EfficientNetV2 with guided filopic diffusion for enhanced rice leaf disease recognition.

Abstract

Rice production is integral to the agricultural sector of India; over 65% of the populations are dependent on rice as their major staple. The cultivation of rice sustains this important agricultural sector; yet, there are many challenges encountered by rice producers, one of which is several types of disease that negatively impact yield and quality. Due to the fact that rice leaf smut, brown spot and bacterial leaf blights are among the most important types of diseases that can significantly reduce the yield and quality of rice, it is important to be diligent when identifying these diseases using accurate and speedy methods on an annual basis for successful and sustainable production of rice crops. As technology advances there continue to be emerging technologies such as Deep Learning (DL) as applied in agriculture to identify diseases and therefore reshape the agricultural paradigm so as to address agricultural disease challenges more readily. This research proposes a previously undemonstrated approach for identifying Rice Leaf Disease using EfficientNetV2; a Diffusion Bounded Attention method for disease detection. The quality of the input imagery has been greatly increased using a Preceding Noise Reduction (PNR) using the Guided Filopic Diffusion (GFD) technique, retaining important characteristics of Rice Leaves (Leaf Texture) which are critical for disease classification within agricultural imaging. To evaluate the performance of our model we utilized the Dice Similarity Coefficient (DSC). This coefficient measures how much the predicted image areas representing disease overlap with the actual affected areas of the image. Therefore, DSC is a reliable way to evaluate model segmentation capability. The Rice Leaf Diseases Dataset we used to identify and classify Rice Leaf Diseases was very comprehensive. Our model achieved an accuracy rate of 98.92% and also attained the best recall, precision and F1 score.

Plant phenotyping relevance

イネ葉の病害症状を画像から検出・分類・セグメンテーションする手法が研究の中心であり、植物の病害状態を直接推定している。

Code and data availability

The paper's sole data asset is the public Kaggle Rice Leaf Diseases Dataset used for all experiments; no author code or models are deposited.