Remote Sensing Detection Of Wheat Stripe Rust Using Reflectance Spectrum And SIF

Wheat stripe rust has the characteristics of strong epidemic and great harm,which seriously affected the yield and quality of wheat Compared with traditional field survey sampling,remote sensing technology has the obvious advantages of fast,macro,and non-destructive,which makes it possible to monitor wheat stripe rust in a large area.After wheat is infected by stripe rust,the photosynthetic capacity and pigment content of crops will change.Solar-induced chlorophyll fluorescence(SIF)is more sensitive to changes in crop photosynthetic physiology,while reflectance spectra are more affected by crop biochemical parameters.Based on this,in order to improve the detection accuracy of wheat stripe rust this study combines reflectance spectrum with canopy SIF data to detect wheat stripe rust.The main research contents of this article are as follows:(1)This paper estimated the SIF data of wheat canopy through two algorithms of radiance and reflectance,Then based on the correlation analysis method,the indexes were selected which to have a significant correlation with the disease index(DI)of wheat stripe rust as the characteristic variable for model construction.The results show that 11 differential spectral indexes and 6 chlorophyll fluorescence indexes that were significantly related to the wheat stripe rust DI and were selected as the characteristic variables of the model.Among the six reflectance chlorophyll fluorescence indexes the correlation between the chlorophyll fluorescence intensity estimated by ?440/?690 and the DI of wheat stripe rust was the highest,and the correlation coefficient reached 0.7189.Among the 11 first-order differential spectral indexes the correlation between SDr/SDb and the DI of wheat stripe rust was the highest,and the correlation coefficient reached-0.8064.(2)This paper improves the traditional discrete binary particle swarm optimization(DBPSO)algorithm from two aspects:inertia weight and particle update method.The modified discrete binary particle swarm optimization(MDBPSO)was used to select the characteristic variables from total spectral reflectance.The MDBOSO algorithm is compared with the correlation coefficient(CC)analysis method and DBPSO algorithm when extract the feature variables.The results show that the number of feature variables screened by CC analysis,DBPSO algorithm,and MDBPSO algorithm are 50,320,and 56,respectively.The number of iterations before and after improvement has been reduced from 395 to 156.The OFV value decreased from 0.145 to 0.127.In addition,compared with the CC method,the MDBPSO algorithm has less redundancy in the feature variables.(3)In this paper,the first-order differential spectral and first-order differential spectral synergize with canopy SIF data were used as independent variables to construct a wheat stripe rust disease severity estimation model,and a comparative analysis of different data sources and different modeling methods through cross-validation methods.The research results show that no matter what modeling algorithm is used,the accuracy of the estimated model is improved to some extent after adding SIF data to the independent variables.The average RMSE between the predicted DI value of the random forest(RF)model and the measured DI value increased by 9%.The average RMSE between the predicted DI value of the back propagation(BP)neural network model and the measured DI value increased by 10%,.The average RMSE between the predicted DI value of the partial least squares(PLS)model and the measured DI value increased by 20%.Therefore,adding SIF data to the reflectance data can improve the detection accuracy of wheat stripe rust.(4)In terms of model algorithms,whether the differential spectral index is used as an independent variable or the differential spectral index synergize with canopy SIF data as independent variable,the RF model exhibits a better estimation effect than the other two models.The results show that when using differential spectral index as independent variables,The RMSE between the predicted DI value of the RF model and the measured DI was reduced by 21%and 35%compared to BP and PLS,respectively.When using differential spectral index synergize with canopy SIF data as independent variables,The RMSE between the predicted DI value of the RF model and the measured DI was reduced by 21%and 26%compared to BP and PLS,respectively.Therefore,regardless of whether the canopy SIF data is added to the model characteristic parameters,the prediction accuracy of the RF is superior to the BP neural network and the PLS.When using full-band data and SIF data to monitor wheat stripe rust,regardless of the feature selection algorithm,the accuracy of the model constructed by the RF algorithm is higher than that of the BP neural network algorithm.The RMSE of the validation set(RMSEv)is reduced by an average of 22%compared with the BP neural network.In terms of feature selection algorithms,The RMSEv between the predicted DI value of the MDBPSO-BP model and the measured DI was increased by 21%and 11%compared to CC-BP and DBPSO-BP.The RMSEv between the predicted DI value of the MDBPSO-RF model and the measured DI was increased by 18%and 11%compared to CC-RF and DBPSO-RF.In summary,when using full-band data and canopy SIF data to monitor wheat stripe rust,MDBPSO-RF is a suitable algorithm for establishing the estimation model.