Remote Sensing-based Identification Of Wheat Powdery Mildew At Leaf,Canopy And Reginonal Scales

Powdery mildew is an important disease that affects the healthy growth and safe production of wheat.In order to prevent and control the damage caused by diseases in a targeted manner,there is an urgent need for rapid,accurate,and quantitative analysis and assessment of the severity of diseases.Development of remote sensing provides a guarantee for obtaining multi-scale,multi-temporal,and multi-spectral remote sensing data of leaves,canopies,and regions,making disease monitoring more targeted.This thesis integrates imaging hyperspectral,non-imaging hyperspectral,satellite multispectral,elevation,and ground survey points and other data,completes identification of wheat powdery mildew on leaf,canopy,and regional scales.The main contents of this thesis are as follows:(1)Identification of powdery mildew in wheat based on hyperspectral data of leaves and canopies.Firstly,three dimensional-reduction algorithms including principal component analysis(PCA),random forest(RF)and successive projections algorithm(SPA)were used to select the most sensitive bands to PM;then,three diagnosis models were constructed by support vector machine(SVM),RF and probability neural network(PNN);finally,compared the total accuracy of models.The results show that the SVM model construed by the PCA dimensionality reduction has the best result,and the classification accuracy reaches 93.33%by a cross-validation method.It is an obviously improvement derived from original hyperspectral images.In order to verify the applicability and migration of the model,similar technical processes,same algorithms and models were adopted for wheat powdery mildew at canopy scale.The results showed that the classification effect of the RF algorithm combined with the SVM model was slightly better than that of other models,and the cross-validation result was 86.00%.This study can provide a method reference for non-contact quantitative estimation of powdery mildew and other crop diseases.(2)Improved method for preprocessing based on Landsat-8 OLI.Landsat-8 OLI(Operational Land Imager)has the characteristics of multiple bands,wide coverage,free distribution,etc.,and has been widely used in crop disease monitoring research.Among them,atmospheric correction and terrain correction are important steps of image preprocessing,and also a prerequisite for realizing quantitative remote sensing monitoring.a.According to the latest calibration parameters,apply the improved dark pixel method(COST)and improved COST atmospheric correction algorithm under the R language development environment to Landsat-8 OLI.The correction effect was compared and verified by the reflectance obtained by atmospheric correction(FLAASH)in ENVI,the results were evaluated by visual contrast,normalized vegetation index(NDVI),constant surface reflectance in different seasons,spectral curves of vegetation and water,and three texture features.The results show that the improved COST algorithm has brighter and clearer vision,the value distribution of NDVI is more concentrated,and the correlation coefficient(R)is the highest on the invariant surface.In addition,the texture complexity is more complicated,and the groove is deeper,which is slightly better than the other two algorithms.b.Combined with DEM data,a comparative analysis of two kinds of topographic correction is completed on the R platform.The effects are evaluated in three aspects: visual comparison,normalized evaluation of shade and sunny slope directions,and quantitative analysis of linear fitting.The results show that the Minnaert and Minaret-scs corrected images visually reduce the amplitude of terrain undulations,the linear fitting separation coefficient is small,and the correction effect is better.The atmospheric correction and terrain correction methods obtained in this study can be corrected with fewer parameters,without platform limitation,and high portability.(3)Comparison of regional scale remote sensing analysis methods for wheat powdery mildew based on Landsat-8 OLI and GF-1.Based on the completion of the screening of leaf and canopy-scale disease-sensitive features,model construction and comparison,and Landsat-8 OLI pretreatment,further regional-scale remote sensing analysis of wheat powdery mildew was carried out.After extracting the wheat planting area in the study area,the RF algorithm performs weight analysis on the primary selection factors of Landsat-8 OLI and the primary selection factors of GF-1.According to the ranking of weights,the first k features were selected as the input of SVM,deep forest and gradient boosting decision tree models.The results showed that DF and GBDT performed well in monitoring wheat powdery mildew at regional scale.For Landsat-8 OLI data,when the eight characteristics of soil regulated vegetation index,blue band reflectance,green band reflectance,short-wave infrared water stress index,structural-enhanced pigment index,red band reflectance,red edge and near-infrared area,and NDVI were used as input,the accuracy of the two models reaches the highest overall,90.1% and 91.3%.For GF-1 data,when the blue-band reflectance and red edge conbine near-infrared area are two features as input,the highest accuracy of the two models are respectively 88.2%,and 88.6%.The results showed that the application of DF and gradient lift decision tree for regional monitoring of wheat powdery mildew was better,which provided a monitoring model for large-scale remote sensing monitoring of crop diseases.