Study On The SAR Image Classification Methods Of The Yellow River Ice

The rivers located in the area north of 35°N latitude in China and in the Qinghai-Tibet Plateau,are easy to form ice in winter.During the period of the breakup of the ice in these rivers in spring,ice flood disasters often occur,especially in the Inner Mongolia section of the Yellow River.In order to reduce or even avoid the impact of ice flood disasters,it is necessary to obtain timely and accurate ice information.The emergence of Polarimetric Synthetic Aperture Radar technology greatly enhances the ability of imaging radar to obtain target information.This provides prerequisite and technical support for the realization of larger-scale,higher-efficient,high-resolution,all-day,all-weather and higher-precision river ice monitoring.Based on the full polarimetric SAR data of the study area provided by spaceborne satellite RADARSAT-2,the corresponding classification methods were studied for river ice classification,including supervised classification based on prior knowledge and unsupervised classification in this thesis.At the same time,the thickness inversion method was studied based on the classification results and the measured data.The main research and work can be divided into the following aspects:1.Aiming at the problem that the traditional minimum distance supervised classification method does not consider the relationships between various categories,resulting in poor classification results,an improved mindistance supervised classification method based on Gaussian mean was proposed.In order to enlarge the differences in the centers of various categories and weakening the adverse effects of categories with large range on categories with small range,normalized Gaussian weighting coefficients were introduced to process each kind of centers for improving classification accuracy.And the classification accuracy of the proposed method are evaluated by means of the river ice optical images in the target area and confusion matrix.In addition,the effectiveness of the supervised classification method proposed in this thesis was demonstrated by comparing with the classification results of several classic supervised classification methods.2.In view of that the initial cluster center is randomly generated in the traditional unsupervised classification methods based on clustering methods,which leads to the problem of low classification accuracy,a decision tree secondary classification method based on FCM clustering algorithm was studied.Firstly,the initial classification result was obtained based on FCM clustering algorithm.Then,the back scattering coefficient,polarization scattering entropy and depolarization power were used as decision basis for secondary classification of river ice.By combining the ground truth information and the confusion matrix,the final classification result was verified and analyzed,which shows the practicability of the unsupervised classification method proposed in this thesis.3.The field measurement experiments in the Yellow River were carried out,the relevant information of different types of river ice in the study area was obtained,and the types of the river ice were divided into Thermal ice(Ti),Frazil ice(Fi)and Consolidated ice(Ci),according to their evidently differences in physical characteristics and distribution position.And the relevant data were recorded separately.The correlations between polarization parameters and river ice thicknesses were discussed by combining the measured data.On this basis,a multi-parameters thickness inversion method was proposed.Two polarization parameters were utilized for inversion of river ice thickness simultaneously.Accuracy evaluation of the thickness inversion was presented,the accuracy and efficiency of this method for the inversion of river ice thickness in the Yellow River were verified.