Research On The Methods Of Wheat Fusarium Head Blight And Powdery Mildew Monitoring Using Remote Sensing Technology At Different Scales

Remote sensing technology provides an important alternative to traditional method in monitoring crop disease spatially.However,there are many challenges and limitations in existing methods.An important point is how to find spectral response characteristics of crop diseases at different scales,and how to combine these spectral features with multi-source data to develop a robust and reliable monitoring model.The present study takes fusarium head blight and powdery mildew of winter wheat for example.We tried to understand the spectral response characteristics of these diseases at different scales,we also tried to combine the spectral response characteristics and image texture characteristics to monitor the severity of these disease.Besides,we tried to pick the appropriate meterological feaures and growth condition features of these diseases and extend all these features to regional scale to develop a crop disease monitoring model.The major contents and results in this dissertation are as follows:1.At wheat ear scale,the hyperspectral data of ears were analyzed and the bands that were sensitive to fusarium head blight severity were selected.Derivative spectral features,continuous removal transformed spectral features,hyperspectral vegetation indices and wavelet features were calculated based on these sensitive bands.Our results showed that wavelet features had higher correlation coefficient with disease severity than other features,and the disease severity estimating model with wavelet features performed better than the model without wavelet features.2.At the canopy scale,the spectral response characteristics of fusarium head blight and powdery mildew were analyzed,and the disease severity monitoring models were developed.For fusarium head blight,the spectral reflectance of the wheat canopy in the range of 350–1000 nm were analyzed and we found that the sensitive bands at Zadoks stage 65(anthesis stage)and Zadoks stage 75(milk stage)were different.Hyperspectral features with these sensitive wavebands were examined and compared for the detection of fusarium head blight using partial least square regression.The results showed that,the optimal hyperspectral feature for fusarium head blight detection was Structure insensitive pigment index(SIPI)in Zadoks stage 65,with R2 as 0.53,and Triangular vegetation index(TVI)in Zadoks stage 75 with R2 as 0.64.For powdery mildew,the bands that were sensitive to powdery mildew severity were selected,and a Powdery Mildew Damage Index(PMDI)was developed.Our results showed that PMDI had high correlation coefficient with disease severity.3.At the field scale,the hyperspectral bands,vegetation indices and texture features were extracted using UAV hyperspectral image.NRI?MCARI?MSR?GI?TVI?TCARI and Band50(650nm)were selected using backward elimination feature selection,and these features were used to develop a disease monitoring model by applying an optimized BP-neural network.Besides,the infection maps of field at anthesis stage and milk stage were produced using optimized BP-neural network.The result indicated that the development of wheat fusarium head blight in milk stage was rapid when the control measures had not been taken.4.At the regional scale,firstly,a powdery mildew monitoring methodology based on probabilistic model was established to monitor powdery mildew occurrence of wheat in Guanzhong Plain,Shaanxi Province.The monitored map of powdery mildew was compared with results of monitoring model developed using classification and regression tree(CART)and random forests(RFs).The results showed that the overall accuracy of the probabilistic model was 81.25%.Then,hyperspectral data at the canopy scale was integrated to simulate Sentinel-2 multispectral reflectance using the relative spectral response(RSR)function of the sensor.Then,many differential and ratio combinations of Sentinel-2 bands that were sensitive to wheat fusarium head blight severity were selected,and REHBI was established based on these basic vegetation indexes.Finally,REHBI was applied to monitoring wheat fusarium head blight in the wheat planting areas of Changfeng and Dingyuan counties from Sentinel-2 imagery.Generally,REHBI performed better in disease monitoring than other indices,and the overall accuracy was up to 78.6%,and the kappa coefficient was 0.51.In addition,to mitigate small training samples problem in crop disease monitoring,an instance-based transfer learning method,i.e.,Tr Ada Boost,was applied to improve the monitoring accuracy with limited field samples by using auxiliary samples from another region.By taking into account the representativeness of contributions of auxiliary samples to adjust the weight placed on auxiliary samples,an optimized Tr Ada Boost algorithm,named Op Tr Ada Boost,was generated to map regional wheat powdery mildew.Op Tr Ada Boost was tested using a dataset with 39 study area samples and 106 auxiliary samples.The overall monitoring accuracy was 82%,and the kappa coefficient was 0.72.The result indicated that Op Tr Ada Boost performed better than other algorithms that are commonly used to monitor wheat powdery mildew at the regional level.