Remote sensing monitoring of winter wheat powdery mildew based on wavelet analysis and support vector machine

Abstract: Wheat powdery mildew is one of the main serious diseases for winter wheat. A fast and accurate monitoring of the disease at a regional scale plays a vital role in reducing yield loss. Remote sensing data has great advantages over traditional data in disease monitoring, including simpler operation, more real-time and higher resolution. In this study, Chinese HJ-1A/1B data with high revisit frequency and 30 m spatial resolution was used to inverse Land Surface Temperature (LST), extract four-band reflectance data, and build seven vegetation indices. These indices should be filtrated to improve accuracy of the model due to redundancy of them. Then, we implemented screening features with the combination of Relief and K-mean algorithm. Relief algorithm which can provide the basis for feature evaluation, so features were ranked in descending order judged by feature weights in preparation for the next process. Clustering accuracy obtained by K-mean algorithm. According to the weight of the feature, the features clustered in turn to perform K-mean analysis. Then the cluster with the highest precision was picked out, and we finally got the normalized difference vegetation index (NDVI), Simple vegetation index (SR) and surface temperature (LST) as the feature set. Wavelet feature can decompose the data in multi-scale and multi-direction, which can highlight the sensitive factor of vegetation index to a certain extent. Forty wavelet functions were constructed from five scales and eight directions, and made them convolve with features. Because there were too many wavelet features after convolved, the independent T-test samples were used to obtain the most sensitive wavelet feature of disease and the corresponding wavelet kernel function. After this process, three features corresponding to vegetation indices were available. These three wavelet features were used as input variables of the model. Support vector machine is a kind of machine learning method based on statistical learning theory. Its core idea is to minimize the structural risk by mapping the input linear indivisible data to the high dimensional space, which makes the difference between different samples. The class interval is the largest while the intra-class interval is the smallest, then the hyper plane is constructed to classify data. The monitoring model of wheat powdery mildew in Jinzhou City of Hebei Province was established by using support vector machine (SVM) with three groups of features. The first group used twelve vegetation indices as the input variables of the model, which served as a control group. The second one used three features after feature selection and the third used three features of the wavelet transform. Then the monitoring precision of the three models was compared and analyzed. The experimental results showed that the overall accuracy and the kappa coefficient of the third model (called GaborSVM) were 86.7% and 0.583, respectively, performing better over the first model (60%, 0.286) and the second model (80%, 0.444). These results also showed that the combined method of wavelet analysis with SVM (GaborSVM) can be applied to large area disease monitoring based on satellite remote sensing image, and has important application value in improving the accuracy of disease monitoring.