Wind Field Retrieval For Typhoon Based On Partial Differential Equation And Machine Learning

China is one of the countries, which are worst affected by typhoon, in the world. When typhoon comes, it causes huge economic losses in the southeast coastal area of our country and people's lives are threatened. It relates to the vital interests of the people and our country that how to deal with typhoon and how to minimize economic losses when it lands. So accurately forecasting becomes an important measure of disaster preventing. The information of wind field reflects the changes of typhoon, so choosing kinds of intelligent algorithms reasonably to invert typhoon wind field can improve the accuracy of typhoon forecast. This paper researches the method of typhoon wind field retrieval based on the techniques of partial differential equation, machine learning and so on. The main work of this paper includes four aspects as follows:(1) Typhoon eye extraction by partial differential equations. The first step of wind field retrieval is to extract useful data from static satellite images, but the gray level information of typhoon cloud image and typhoon intensity is linked closely, especially the powerful changes of gray level at the eye-wall embody the typhoon intensity. Therefore, to extract gray level information accurate or not is related to the processing precision of next steps directly. The contrast of original typhoon cloud image is poorer due to various reasons and it is not good for typhoon eye extraction. Therefore, we design a new non-linear gray transformation function based on arc-tangent curve, and construct quality evaluation function of typhoon cloud image by using image information entropy and standard deviation, then we use the differential evolution algorithm to achieve the parameter of non-linear gray transformation algorithm. The function designed has only one parameter, comparing with other traditional gray transformation functions which have many parameters, this function's computation has been reduced. The experimental results show that the method designed can enhance the typhoon cloud image's global contrast effectively, and is good for segmentation of typhoon eye in the next steps. According to the typhoon cloud image which has been pre-enhanced, this paper compares with many kinds of common segmentation algorithm, and choose the partial differential equation segmentation based on geodesic active contour model to segment the typhoon eye, and has got a good segmentation effect.(2) Research of eyed typhoon wind field retrieval method based on support vector machine. This part mainly includes two aspects:1)Establishing mathematical model between typhoon size and the radius of maximum wind from static infrared satellite images:First we use the non-linear gray transformation method proposed in part (1) to enhance the infrared typhoon image for subsequent typhoon eye segmentation. Then we use partial differential equation to extract typhoon eye, and calculate the size of the eye. And then we get the radius of maximum wind according to the distance from the coldest cloud top to the warmest point in the eye-wall. Finally we establish the model between the size of typhoon eye and the radius of maximum wind by using support vector machine (SVM) to reflect the characteristics of eyed typhoon wind field 2) Establishing the two-dimension surface wind field of eyed typhoon:First we use SVM to establish the model between the eye-wall gray information and maximum wind speed, and then we use this model with linear interpolation to estimate the speed of typhoon wind field, so we get the two-dimension wind field of eyed typhoon.(3) No eye typhoon wind field retrieval:We use SVM to establish the multiple relationship model between the critical wind radius R34, R50 and maximum wind speed, latitude, life history of typhoon. When calculating the two-dimensional wind field of non-eyed typhoon, we first construct the eye-wall manually according to the size of eyed typhoon images from the same typhoon, and then it does the same next steps like the eyed typhoon images. The proposed algorithm has the characteristics of achieving easily, high precision and fast running speed, especially in establishing models by SVM, comparing with other common fitting methods, SVM is not only running fast, but also getting lower error.(4) Typhoon motion vector calculation based on local gray coding and wavelet multi-resolution analysis. A complete wind field not only includes establishing wind field feature model and calculating wind speed, but also includes calculating the motion vector of typhoon. Traditional gray level matching method is very accurate in computing the motion vector, but the only drawback is its low running rate. Because of this drawback, this paper realizes image matching by combining wavelet multi-resolution analysis with a local gray level coding method, and has not only got a good effect of computing vector, but also has improved the algorithm's running rate.