Detection study of early-stage classification of rice diseases using a hyperspectral multi-feature fusion model (PMA-VRNet) driven by the vegetation index RBDI

Abstract

Rice Blast (RB) is a highly destructive fungal disease that causes millions of tons of rice yield loss worldwide every year. Therefore, using precise indicators to quantify diseases and achieve early detection is crucial for establishing a preventive plant protection system. However, there is currently a technological gap in quantitative research on early grading and detection of RB. This study employs Unmanned Aerial Vehicle (UAV) hyperspectral remote sensing technology to acquire hyperspectral images of rice canopies, and then preprocesses them to extract spectral, textural, and structural features. To address the limitation of traditional vegetation indices (VI) in reflecting the spatial structural characteristics of vegetation, the rice blast indices RBDI1 and RBDI2 are constructed based on vegetation cover (FVC), effectively quantifying rice disease indices (DI) and improving the accuracy and spatial resolution of disease identification. Furthermore, the Parallel Multi-Head Attention VGG-ResNet (PMA-VRNet) model was proposed, which integrates multiple feature data such as RBDI, Texture Features (TF), and Canopy Coverage (CC) through a parallel Multi-Head Attention mechanism. This model deeply integrates the high semantic feature extraction capabilities of VGG with the deep residual learning advantages of ResNet, achieving high-precision grading detection of rice diseases in the early stage. The study has shown that the PMA-VRNet model demonstrates excellent feature learning capabilities in data fusion, particularly on the sensitive wavelength dataset selected by the BS-CARS algorithm, achieving a detection accuracy of OA = 93.5 % and Kappa = 91.86 %. Compared with the comparative model, OA improved by 1.17 %-3.00 % and Kappa improved by 0.62 %-3.76 %. Additionally, SMOTE data augmentation is better than the original data, and the performance of combined feature modeling is better than single-feature modeling. Among them, the model combining VI_BS-CARS, TF, and CC achieves the highest accuracy (OA = 94.5 %, Kappa = 93.12 %). This study provides an efficient and feasible technical solution for accurately monitoring of RB using UAV hyperspectral images.

Plant phenotyping relevance

UAVハイパースペクトル画像からイネの病害指数・病害重症度を定量化し、特徴抽出とモデル性能を検証する方法開発・技術評価が研究の中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.