Dynamic Monitoring Of Major Wheat Diseases Based On Multi-source And Multi-temporal Remote Sensing Analysis

Strengthening the investigation of major wheat diseases is of great significance in guiding the scientific prevention and control of diseases and the sustainable agricultural development.Compared with the traditional crop survey methods,remote sensing technology provides an important supplementary method.The key to achieve accurate,real-time,rapid,and large-scale quantitative discrimination and monitoring of crop diseases is to efficiently use abundant remote sensing data and method basis.This study focuses on the three main wheat diseases of stripe rust,powdery mildew,and Fusarium head blight.The cycle mechanisms and the difference of the environmental conditions of the different crop diseases were both considered.The advantages of multi-source and multi-temporal remote sensing data were integrated,including unmanned aerial vehicle(UAV)hyperspectral imagery and multi-source satellite remote sensing imagery such as Sentinel-2,Landsat-8,and GF-1 in different disease periods.The discrimination models of the different crop disease and pest stresses were firstly constructed.On the premise of accurately discriminating different diseases and pests,the crop disease monitoring models with high precision and high specificity were then established.Finally,from a practical perspective,the crop disease severity estimation models were established.These models provide the technical support for the scientific control and ecological management of the crop diseases.In this study,the main research contents and results are as follows:(1)For the remote sensing discrimination of crop diseases and pests,considering the simultaneous of different disease and pest stresses in the field,two different situations such as the simultaneous occurrence of wheat powdery mildew and aphid and the simultaneous occurrence of wheat stripe rust and Fusarium head blight were as the case study.The discrimination methods of different crop diseases and pests were studied based on the remote sensing features integrating crop growth condition and environmental information extracted from the multi-source and multi-temporal satellite remote sensing imagery.Firstly,based on the bi-temporal remotely sensed features characterizing growth conditions and environmental factors calculated by using two Landsat-8 scenes,an approach which combined with the synthetic minority oversampling technique(SMOTE)and back propagation neural network(BPNN)was developed to discriminate wheat powdery mildew and aphid.The results illustrated that the overall accuracy of the SMOTE-BPNN model based on the remotely sensed features integrated crop growth information and environmental factors was 10.9%higher than that of the model based on the traditional single-date crop growth indices.By coupling the multi-temporal Sentinel-2 ? Landsat-8 and GF-1 imagery,an approach by combing with principal component analysis and k-nearest neighbors(k NN)was then developed to discriminate wheat stripe rust and Fusarium head blight.The results illustrated that the overall accuracy of the PCs-based k NN model was 11.0% higher than that of the SFs-based k NN model.(2)For the remote sensing monitoring of crop diseases,combining with the occurrence characteristics of different crop diseases,wheat stripe rust and Fusarium head blight were as the case study.The monitoring methods of the specific crop diseases at the field scales were studied by using the disease sensitive typical vegetation indices and wavelet features obtained from multi-temporal UVA hyperspectral imagery.Firstly,based on BPNN method and the disease sensitive vegetation indices extracted using multi-temporal UAV hyperspectral images,a monitoring model was developed to monitor wheat stripe rust at the field scales.The results illustrated that the overall accuracy of the BPNN monitoring model based on the vegetation indices reached 95.7%,which increased by 11.6% than that of the k NN monitoring model.Meanwhile,the monitoring accuracies of the BPNN method for the stripe rust in the three disease occurrence periods were all higher than that of the k NN method.Then,based on the six optimal wavelet features extracted by using continuous wavelet analysis(CWA)from ASD hyperspectral data,an identification approach for Fusarium head blight in wheat ears was developed by using Fisher linear discriminant analysis(FLDA).The results illustrated that the model provided an overall accuracy of 88.7%.On this basis,based on the five optimal wavelet features obtained from the bi-temporal UAV imagery by using CWA,a monitoring model was developed to monitor wheat Fusarium head blight at the field scales using k NN method.The results illustrated that the model provided an overall accuracy of 86.7%.(3)For the remote sensing evaluation of the disease severity,considering the limitations of single-date and single-source satellite remote sensing imagery in characterizing the occurrence of crop diseases,wheat powdery mildew and stripe rust were as the case study,by integrating the crop disease cycle mechanisms and the complementary advantages of the multi-source and multi-temporal satellite remote sensing imagery,the response relationship between the remote sensing features of the disease occurrence critical periods and disease severities was established,and the crop disease severity evaluation methods were studied by using the data mining algorithms.Firstly,based on the typical vegetation indices of the optimal key periods obtained by combining the multi-temporal Landsat-8 imagery and a backward stepwise elimination method,an approach using k NN was developed to evaluate wheat powdery mildew severity at the regional scales.The results illustrated that the overall accuracy of the multi-temporal images based k NN approach was 7.7% higher than that of the traditional disease severity evaluation method based on the single-date image.By combining the normalized two-stage remotely sensed features which obtained from Sentinel-2 and Landsat-8 imagery and SMOTE-BPNN method,an approach was then developed to evaluate wheat stripe rust severity at the regional scales.The results illustrated that the overall accuracy and G-means of the evaluation approach of wheat stripe rust severity based on the normalized two-stage remotely sensed features were 20.0% and 26.8% higher than that of the traditional disease severity evaluation method based on a single-date image,respectively.The results of this study demonstrate that the multi-source and multi-temporal remote sensing imagery integrating crop growth and environmental parameters can be effectively applied to the discrimination of different crop diseases and pests,improving the discrimination accuracies significantly;the multi-temporal UAV hyperspectral imagery combing with a proper spectral analysis algorithm and classification technology can potentially realize the specific monitoring of crop disease at the field scales;the coupled multi-source and multi-temporal remote sensing imagery during the disease occurrence key periods can characterize the spatial distribution of the of crop disease occurrence severity more accurately.