Inversion model for severity of powdery mildew in wheat leaves based on factor analysis-BP neural network

Abstract: The detection of crop health under disease stress is an important study in precision agriculture. In order to understand characteristics and disease severity of wheat leaves under stress of powdery mildew, we conducted artificial inoculation experiment of wheat powdery mildew to test winter wheat leaf's spectral reflectance under powdery mildew with different severity degree in different growth phases using hyper-spectrometer. We chose 2 susceptible cultivars to powdery mildew, i.e. Yanzhan 4110 and Yumai 34, and 2 medium resistant cultivars to powdery mildew, i.e. Aikang 58 and Zhengmai 366. Using artificially inoculation method, we measured the spectra of wheat leaves of different varieties at different levels of incidence and growth stages, and investigated the disease severity of each leaf. We analyzed the relationship between the conventional spectral characteristic parameters, the ratio index, the normalized index and the disease severity of powdery mildew, simulated disease severity of wheat leaf powdery mildew using the factor analysis-back propagation neural network (FA-BPNN) method, and evaluated the its fitting accuracy and applicability. The results showed that with the aggravation of disease severity of wheat powdery mildew, spectral reflectance increased in visible bands of 350-760 nm, while spectral reflectance obviously decreased in near infrared bands of 760-1 050 nm. Conventional spectral parameters, PSRI (plant senescence reflectance index), MCARI(modified chlorophyll absorption in reflectance index), SIPI(structure insensitive pigment index) and RGRcn(red green ratio chlorophyll content), had a better fitting effect of disease severity of wheat blades than others, whose coefficients of determination (R2cal) and root mean square error (RMSEcal) in calibration set were 0.776, 0.769, 0.757 and 0.712, respectively, and 8.68, 8.82, 9.05 and 9.16, respectively, and the fitting equations of RGRcn and SIPI had higher RMSEcal than PSRI and MCARI. Statistical analysis showed that in validation set, prediction model of RGRcn was the best with the RMSEval of 7.67, followed by PSRI with the RMSEval of 11.64. Combining fitting and testing performances, RGRcn and PSRI were good retrieval models of wheat leaf powdery among conventional spectral parameters. The best two-band vegetation index that was correlated with wheat powdery mildew between 400 and 1000 nm wavelength was located in band combination of 605-630 and 520-550 nm, 645-690 and 710-1000 nm for the ratio index, and in band combination of 650-685 and 710-1000 nm for the normalization index, and the coefficients of determination (R2cal) ranged between 0.70 and 0.80. These band combinations had lower RMSEcal, which was lower than 10.0. ratio index(RI) (670, 855) and normalized difference vegetation index (NDVI) (680, 880) were the best two-band vegetation indices, the R2cal were 0.764 and 0.765, the RMSEcal were 8.91 and 8.89, and the RMSEval were 7.62 and 7.21, respectively. So these band combinations of ratio indices and normalized indices were better than PSRI and RGRcn as a whole. According to the correlation analysis, we obtained the sensitive bands, 400-415, 450-500 and 590-695 nm. We further refined the sensitive bands using factor analysis and obtained the new sensitive bands, which were 415, 485-495 and 620-640 nm. Therefor, factor analysis could be used as a new type of band extraction method. The critical factors of sensitive bands, the accumulated contribution rate of which was more than 99% at each period of filling stage, were extracted using factor analysis and as the input of BPNN. The number of critical factors was the number of nodes in the hidden layer. Disease severity of wheat leaves at different periods was the output of BPNN. The results showed that the BPNN simulation could greatly improve the estimation accuracy of disease severity of wheat leaf powdery mildew, with the R2val higher than 0.80 at each growth period, and especially the R2val being up to 0.922 at middle filling stage. The R2val of the whole filling stage had been greatly improved, and the RMSEval and relative error in validation set (REval) had been reduced, which were 0.872, 7.84 and 7.56%, respectively. Therefore, compared to the above 2 methods, the FA-BPNN method can greatly improve the inversion precision of wheat leaf powdery mildew and has the applicability to the whole filling stage. It is of great significance to precise prevention and control of disease.