Comparative Analysis On Typhoon Center Location Methods Based On Multi-source Satellite Data

The typhoon is extremely destructive and poses a serious threat to the safety of human life.Accurately predicting the typhoon path and monitoring the location of the typhoon is conducive to effective early warning and prevention of disastrous weather,and has important instruction and reference significance for climate and weather research.This paper combines satellite cloud image typhoon cloud system boundary detection with satellite scatterometer wind field positioning,and comparatively studies the positioning effects of two wind field positioning methods at different stages of typhoon development.The monitoring and research work of the typhoon center in this paper is as follows:First,the typhoon cloud system is identified based on the Faster Region Convolutional Neural Networks(Faster RCNN)algorithm,and the location of the typhoon boundary is monitored.According to the China Meteorological Administration(CMA)best path data set,the positive sample typhoon cloud in the satellite cloud map is marked,and the training data set and the test data set are produced.This paper compares the positioning errors of the two size frames of 20 pixel by 20 pixel and 100 pixel by 100 pixel.Based on the 20 pixel by 20 pixel size frame,the average deviation of longitude is 0.227 °,and the average deviation of latitude is 0.235 °;based on the 100 pixel by 100 pixel size frame The recognition accuracy is higher,the average deviation of longitude is 0.152 °,and the average deviation of latitude is 0.148 °.Then select the Visual Geometry Group Network-16(VGG16)model for pre-training.After Region Proposal Network(RPN)and Fast Region Convolutional Neural Networks(Fast RCNN)process classification and regression,the classification probability of the foreground(typhoon cloud)and background(non-typhoon cloud)and the regression position of the typhoon cloud are output.And calculate the accuracy rate and recall rate to evaluate the accuracy of identification and detection.The detection rate of typhoon identification for different structures is high or low.The accuracy and recall rate of all typhoons are 94.3% and 88.4%,among which there are eye typhoons.The algorithm is easier to learn features and recognition,and the accuracy rate and recall rate are 100%.However,the rotation characteristics of the non-eye typhoon are not obvious,and its accuracy and recall rate are 91.6% and 82.6%.Secondly,the space-time weighted interpolation method is adopted to construct the HY-2A/HSCAT-A and Cross-Calibrated Multi-Platform Ocean Winds(CCMP)mixed wind field,and the authenticity test of the mixed wind field is made by comparing the National Data Buoy Center(NDBC)buoy wind data.The average deviation of the overall wind speed in the mixed wind field is 0.90m/s,the root mean square error of the wind speed is large,the root mean square error of the wind speed is 2.74m/s;the average deviation of the overall wind direction is 2.63 °,and the root mean square error of the wind direction is 22.83 °.Then,based on the dual-data fusion wind field and the HY-2B/HSCAT-B single observation wind field,the vorticity field and the divergence field are constructed respectively,and then the composite field is established.Based on the Faster RCNN algorithm to identify the location of the detected typhoon,the wind field positioning methods,including high wind speed threshold and wind direction positioning method and high wind speed threshold and compound field positioning method,are used to conduct typhoon formation,maturity and recession positioning research and result evaluation.The results of the study found that: comparing the positioning results of the two methods in the single-star observation wind field and the dual-data fusion wind field,it is found that the threshold of high wind speed in the single observation wind field and the positioning error of the composite field positioning method are smaller,the average deviation of the total distance is 13.9km,the root mean square error of the total distance is 22.8km,and the high wind speed in the double data fusion wind field.Compared with similar composite field positioning methods,the longitudinal positioning accuracy is improved by 63%,and the latitude positioning accuracy is similar.The positioning effect of the threshold and the wind direction positioning method is better,the average deviation of the total distance is 15.7km,and the root-mean-square error of the total distance is 26.9km.