Identification And Differentiation Of Wheat Diseases And Insects With Multi-source And Multi-scale Remote Sensing Data

The identification and differentiation of different kinds of plant diseases and insects is an inevitable issue when monitoring plant diseases and insects with remote sensing technology. In this dissertation, the author, aim to this issue, three kinds of typical pests of winter wheat in north China-powdery mildew (PM), yellow rust (YR) and aphid (AH), were chosen as an example. Based on the observation data, such as the imaging/non-imaging spectrum, multi-source satellite images et al., the spectral response characteristics and differences of different diseases and insects at different levels were studied, then the discriminant models were constructed. These will be the basis for the diseases and insects monitoring and discriminating by remote sensing technology. The major contents and results in this dissertation are as follows:(1) A wheat diseases and insects discrimination method was proposed based on the hyperspectral imaging technology at leaf level, which including the segmenting of leaves and their background, leaf diseases/insects spot segmentation, and the discrimination of different pests combined with spectrum features and image geometry and texture features. In this method, the optimized spectral bands were selected in the first place. On the basis of spectral analysis, the 570 nm,680 nm and 750 nm bands were selected, and then an improved index-RTVI was proposed which could match these bands and was improved from the classic vegetation index TVI from the viewpoint of spectral geometry. The RTVI index was used as the characteristic to identify the pest spots from the normal area. In the aspect of plant diseases and insects discrimination, on the one hand, a spectrum ratio fingerprint characteristic was constructed based on the relative changes of spectrum; On the other hand, several imaging based geometry and texture characteristics were selected targetedly. Based on these, the author proposed a plant diseases and insects distinguish method, which could composite spectrum and image characteristics, and the accuracy based on the independent validation samples could reach to 90%. At the same time, to make the whole process automate, the method was integrated development to a software - PestDiscriminator V1.0, and which could takes the users more convenience.(2) At leaf level, the author proposed a method of selecting distinguishing features and constructing discriminant model for wheat diseases and insects based on the leaf spectrum. By comparing the original leaf spectrum and the characteristics between powdery mildew/stripe rust/aphids and their normal references, according to the results of independent t-test analysis, the sensitive wavelengths and spectral characteristics that has the ability to distinguish the three kinds of plant diseases and insects were obtained which including three regions of wavelengths at 666-683 nm,752-758 nm, and 1893-1905 nm, and four vegetation index-Dy, GI, NDVI, and PRI. In addition, the general process of the continuous wavelet feature extraction analysis was modified, which could carry out the function transformation from the parameter inversion to the classification, and obtained five wavelet features using in wheat diseases and insects discrimination at leaf level. On this basis, the selected original bands, vegetation indices and wavelet features were used to construct the plant diseases and insects discriminant models respectively based on both SVM and FLDA methods. The results showed that the vegetation index features based discriminant model was the best model and the discriminant accuracy could reach to above 80%.(3) At the canopy level, based on the feature selection results of the leaf level, some supplementary characteristics that more adapted to the vegetation information extraction of the canopy level were added. Then base on the canopy level experimental data, the sensitivity and distinguish ability of the above characteristics were tested. The results showed that four vegetation indices - Dy, WID550-750, PRI, and NPCI, and four wavelet features could pass the test. For the distinguishing models, the FLDA discriminant model that based on the vegetation indices showed the best effect with a discriminant accuracy of 75%, which is better than other types of features and algorithms. The results above showed that the sensitivity of the spectral characteristics and the distinction accuracies of different algorithms exist a big difference between the leaf and canopy levels. On the basis of the analysis above, in the form of canopy spectral data simulation, the responses and distinguish abilities to plant diseases and insects of the commonly used satellite remote sensing sensors and bands were evaluated. The results showed that the responses and distinguish abilities between different satellite sensors and channels performed consistently. And more bands showed the potential in diseases and insects monitoring and differentiation.(4) At the regional level, the wheat fields that infected by powdery mildew and aphid in Zhou Jiazhuang, Jinzhou city, Hebei province were chosen as an example. Based on the observation and survey data, the method that extracting crop growth and habitat characteristics from multi-source remote sensing data was studied, and as well as the plant diseases and insects monitoring and distinguishing method at regional level.The main processes include crop growth and habitat feature selection and extraction, identification abnormal growth area at regional level, and plant diseases and insects discriminant model construction. In terms of feature selection, combined with the results of canopy data analysis and collaborative satellite images and corresponding field survey data analysis, the characteristics that used to identify the abnormal growth area of winter wheat were selected, which including Yellow_WV2, NIR-2_ WV2, GNDVI_WV2, and LST_TM8. Based on the identified abnormal growth area, the characteristics used to distinguish different types of wheat diseases and insects were further selected including four crop growth characteristics-Green_WV2, Red_WV2, NIR-1_WV2, and NDVI_WV2, and five crop habitat characteristics-SIWSI_TM8, DSWI_TM8, Wetness_TM8, Greenness_TM8, and LST_TM8. Based on the above features and the relatively ratio value with normal reference, the discriminant model for powdery mildew and aphid was constructed, and the overall accuracy of the validation data could reach to 81%. The results showed that it is a feasible way that monitoring and distinguish plant diseases and insects by extracting crop growth and habitat information from multi-source remote sensing data.