Methods For Information Extraction Of Wheat Disease Based On Multi-source Remote Sensing Data

With the emerging of various types of remote sensing data, the monitoring and forecasting of crop disease through remote sensing is becoming a very important and effective means of obtaining disease information in field condition, and will be an alternative to mannully inspection. However, to achieve this goal, a most important question is how to pick and apply appropriate methods, to exert the benefit from those data at different scales. The present study takes yellow rust and powdery mildew of winter wheat for example. We tried to understand the underline mechanism of spectra, pick the appropriate spectral features and extend those rules and features to regional scale for disease monitoring and forecasting. We applied spectroscopic and imaging techniques in disease detecting at leaf and canopy scale. On the other hand, airborne hyperspectral data and onbroad multispectral images were used for monitoring and forecasting of wheat diseases at regional scale. The major contents and results in this dissertation are as follows:(1) At leaf scale, we analyzed the spectral response of powdery mildew of winter wheat and its underline mechanism with the aid of leaf spectroscopy data, leaf spectral imaging data as well as biochemical and biophysical measurements. Based on such understanding, we examined the performance of several purposely selected spectral derivative features, continuous removal features and 20 commonly used vegetation indices for disease detection. Apart from those conventional spectral features, the continuous wavelet analysis (CWA) was also employed and compared for its capability in disease information extraction. The disease severity was linked with spectral features through both continuous and discrete manner by partial least square regression (PLSR) and fisher linear discrimination analysis (FLDA), respectively. Our results showed that according to cross-validation, the overall accuracy and coefficient of determination R2 are 91% and 0.86, respectively, and the normalized root mean square error (NRMSE) is below 0.20.(2)At canopy scale, on the one hand, the characteristics of spectral response for powdery mildew was studied. On the other hand, we also conducted a study on how to differentiate the disease from those common nutrient stresses. An important observation in this study is the broad-band vegetation indices, such as SR and NLI, exert great potential in detecting powdery mildew, which thus facilitat the implementions at regional scale. For the study of differentiating the disease and nutrient stresses, a series of normalization of hyperspectral reflectance for the diseased spectra and the nutrient stressed spectra were undertaken prior to the comparison. The normalization processes were implemented to minimize the effects of differences in illuminating conditions, measuring dates, cultivars and soil backgrounds on the target spectra. Taking advantage of an independent t-test, the responses of a total of 38 commonly used spectral features for the yellow rust disease and different forms of nutrient stresses were examined at 5 major growth stages. Four vegetation indices, including PRI, PhRI, NPCI and ARI were identified as those persistently having a response to yellow rust at 4 out of 5 growth stages. However, further independent t-test analysis showed that all the four vegetation indices also responded to several nutrient stresses, except for the Physiological Reflectance Index (PhRI), which was only sensitive to the yellow rust disease at all growth stages.(3) At regional scale, We attempted to extend the spectral features from leaf and canopy scale to regional scale based on the multi-temporal HuanJing satellite images. A method which combines the PLSR and the mixture tuned matched filtering (MTMF) was proposed in this study. A field investigation was conducted at Tongzhou and Shunyi district in the suburban area of Beijing. The ground survey data and the corresponsing multi-temporal images facilitate the extraction of disease information from both spectral and temporal dimensions. Four algorithms, including spectral information divergence (SID), spectral angle mapper (SAM), PLSR and mixture tuned matched filtering (MTMF), are examined for their capability in disease monitoring. The performance and traits of those algorithms were evaluated and compared. On this basis, we proposed a noval method which is a combination of MTMF and PLSR. Higher accuracies were achieved as the overall accuracy (OAA), average accuracy (OA) and kappa coefficient are 0.78,0.71 and 0.59, separately. As for the spatial distribution pattern in field for powdery mildew, theχ2 test based quadrat analysis and the landscape analyzing software FRAGSTATS were adopted. The results show that the powdery mildew tends to have an assembly pattern in the entire region, whereas presents a relatively sparse distribution pattern locally. This distribution pattern can be a basis for the management and prevention of powdery mildew.(4)For diseae forecasting at regional scale, we proposed a method that relies on the optical channels from HJ-CCD and thermal channel from HJ-IRS. The logistic regression model was used for the prediction of disease occurrence probability. Three parts of information are used as inputs, including the spectral features that represent the growth status of crop, the soil content and the land surface temperature (LST) from HJ-IRS that represent the environmental characteristics. With the aid of both datasets for calibration and validation, the probability of the disease occurrence was in agreement with the predicted results given by the model in general. The overall accuracy for ground survey points and the segmented occurrence map were 72.22% and 71%, respectively. In addition, we also discussed the impact of different cut values to the accuracy from the balance of accuracy and cost.(5) We presented an approach of constructing a spectral knowledge base (SKB) of diseased winter wheat plants, as a solution when the ground survey data is lacking. This method takes the airborne images as a medium and links the disease severity with band reflectance from environment and disaster reduction small satellite images (HJ-CCD) accordingly. An empirical retrieving model was adopted for relating the PHI spectra with the disease severity, whereas the spectra of PHI pixel was simulated to a HJ-CCD pixel through an relative spectral response (RSR) function. Through a matching process with a SKB by Mahalanobis distance or spectral angle (SA), we estimated the disease severity with a disease index (DI) and degrees of disease severity. The proposed approach was validated against both simulated data and field surveyed data. The SKB performed poor when quantifying disease severity by DI, with NRMSE of only 0.46 and 0.55 for both matching methods. On the other hand, the SKB performed much better when quantifying disease severity by discrete degrees, with overall accuracy of 0.77 and kappa coefficient of 0.58.