Exploring Multi-Features in UAV Based Optical and Thermal Infrared Images to Estimate Disease Severity of Wheat Powdery Mildew

Remote sensing based on unmanned aerial vehicle (UAV) is a non-destructive way for wheat powdery mildew (WPM) detection in the field management and crop protection. However, WPM causes complex symptoms and impacts on wheat plants, such as reducing pigment, losing biomass and hindering normal growth. There are great challenges on UAV-based estimation on the disease index (DI) of WPM stress with such kinds of different symptoms during the infected status. Thus, this study aimed to explore the potential multi-features in the UAVbased optical and thermal infrared information to indicate the WPM impact and estimate DI. UAV multispectral and thermal imagery data were acquired continuously in the field during the early, middle, and late infection stages after artificially inoculated with fungal pathogens at the Institute of Plant Protection, Chinese Academy of Agricultural Sciences, Xinxiang, China in 2022. The multi-features between healthy and infected plots were analyzed, mainly including (i) spectral reflectance calculated by mean (labeled as CS) and percentile methods (labeled as LS) from all pixels; (ii) vegetation indices (VIs) constructed from CS and LS; (iii) three-band texture combination indices (TTCI); and (iv) canopy temperature (CT). The optimal variables for DI estimation were determined by the Pearson correlation analysis and a recursive feature elimination algorithm. Multiple linear regression was used to construct DI estimation models based on single and fused types features. Results showed spectral features calculated by LS were more suitable for detecting WPM impacts than CS because more marked differences were observed. Only normalized difference red edge index, plant senescence reflectance index, and Meris terrestrial chlorophyll index could consistently capture changes induced by the disease in the middle to late stages. The proposed TTCI were able to distinguish infection changes compared to spectral and CT features as early as 9-30 days. The combination of spectral features calculated by LS, TTCI2, and CT provided the highest estimation accuracy of DI, with 0.90 of coefficient of determination (R2) and 8.28 % of root mean square error (RMSE). Compared with CS, LS, TTCI2, LS + TTCI2, the R2 was increased by 36 %, 27 %, 7 %, and 3 %, and the RMSE was decrease by 44 %, 39 %, 18 %, and 10 %, respectively. This study analyzed the ability of optical and thermal infrared information to detect the WPM impact and compared the potential of features in estimating DI, which provided a UAV-based remote sensing method for disease prevention and control in the field.