Using UAV hyperspectral imagery and deep learning for Object-Based quantitative inversion of Zanthoxylum rust disease index

Abstract

Zanthoxylum rust (ZR) poses a significant threat to Zanthoxylum bungeanum Maxim.(ZBM) production, impacting both the yield and quality. The lack of current research on ZR using unmanned aerial vehicle (UAV) remote sensing poses a challenge to achieving precise management of individual ZBM plant. This study acquired six UAV hyperspectral images to create a ZR inversion dataset . This dataset, to our knowledge, is the first dataset for remote sensing deep learning (DL) of ZR using UAV. To facilitate automated extraction of individual ZBM plant and the quantitative inversion of ZR disease index (DI), we introduced the object-based quantitative inversion framework (OQIF). OQIF achieved high accuracy in recognizing ZBM (average precision at an intersection over union threshold of 0.5 was 90.0 %). Remarkably, OQIF demonstrates outstanding quantitative inversion results for ZR DI (R2 = 0.90, RMSE = 3.97, n = 8166). For DI < 10, the RMSE was 2.48, showcasing early detection capability. Our research has significant implications for ZBM cultivation and precision management, pioneering object-based quantitative inversion for tree diseases and yield estimation, with potential for early ZR detection.

Keywords

• UAV
• Hyperspectral
• Deep Learning
• Quantitative Inversion
• Zanthoxylum Rust
• Disease index