Study On The Curvelet Based Information Extraction And Classification Method Of Marginal Ice Zone In SAR Imagery

Sea ice covers about 7% of the earth's surface.The marginal ice zone(MIZ)is the transition region from consolidated ice to open water,where is typically defined as the region between 80% and 15% of sea ice concentration(SIC).It is a very dynamic region,with intense biological activity,intense ocean-ice-atmosphere interaction.The MIZ can accelerate the melting of sea ice and promote the growth of new sea ice,meanwhile,MIZ has a big impact on the save of human activities.Therefore,the accurate information extraction and estimation of the MIZ covering has become increasingly important in recent years toward understanding changes environment and human activities.SAR(Synthetic Aperture Radar)can provide high resolution and multi-polarization imagery,which is able to capture the details of sea ice characteristics,especially where the ice concentration is low.Therefore,it becomes an important way to learn and study on MIZ.However,the difficulty of SAR image processing and the complex situation of MIZ surface,the methods to accurate extract the information of MIZ is restricted.Therefore,the paper based on the multi-scale and multi-orientation of curve-like features of MIZ in RADARSAT-2 dual HH and HV SAR imagery,the curvelet transform is chosen because it can decompose the curve information into multi scale and multi orientation and give an optimal expression of curve singularity.The paper applied the method of curvelet transform and combing with other image processing methods to the information extraction and detection of MIZ in SAR imagery.It is achieved two accurate ice edge detection methods especially at where sea ice concentration is low,a dynamic feature extraction method of sea ice and MIZ estimation method,and a pixel-level MIZ automatic detection method.The main research contents and results are as follows:(1)The paper provided a novel algorithm of ice edge detection in MIZ,at where the sea ice concentration is low in SAR imagery by combing the curvelet transform and active contour.The method according to the coefficients and energy distribution features in middle scales of curvelet domain,to determine regions in the image likely to contain the ice edge and enhance the contrast of two classes and the homogeneity inside ice.These regions are then joined using the active contour method to obtain the final ice edge.Through comparison of the ice edge with that from image analysis chart,it is demonstrated that the proposed method can detect the ice edge effectively in SAR images.Compare with the ice edge from passive microwave,the accurate has a big improvement.(2)A novel ice edge detection method was proposed,which is based on the edge features in MIZ.The method,firstly,using Canny edge detection method on the curvelet middle-scale filtered SAR images,the curve-like features in various of contrast and backscatter of MIZ can be unique to edge information.Then the ice edge is the boundary of the rich edge information and no edge region.Since the edge information is discrete,in order to remain the shape and position of ice edge,a multi scale region based active contour in curvelet domain method is designed to keep the regional consistency while refining the position gradually.The results show that the method can detect the ice edge position effectively.The mean distance to the ice edge ground truth is much closer than the ice edge from passive microwave sea ice concentration.Compared by using sliding window method,this can guarantee the high accurate results while greatly reducing the computation time.(3)Through analyzing the relationship between dynamic feature of sea ice and curve-like features in SAR imagery,the paper provided a dynamic feature extraction method in SAR imagery,which also can be used to estimate the MIZ covering.By using curvelet,the curve singularity in spatial domain of MIZ is transformed to point singularity.The curve distribution can be transformed into texture features in curvelet domain.Based on the statistic feature and curvelet co-occurrence matrix energy of curvelet coefficients neighborhood,a dynamic feature is extracted by calculating the relationship of that two features.Since the MIZ is a dynamic region,the extracted dynamic feature can be used to estimate the MIZ by given an appropriate threshold.Compare the MIZ from image analysis chart,it can be seen that the dynamic feature can estimate the MIZ effectivelly,the highest accuracy of the result is achieved to 96%.The accuracy improved about 30%~40% compared with the MIZ from passive microwave sea ice concentration.The method can be used as the first step of MIZ detection,and enlarge the application of SAR in sea ice analysis.(4)A pixel level supervised classification method into MIZ and non-MIZ in SAR imagery is proposed.The result can identify the MIZ with a high precision.Firstly,the features of image patch are the statistical features of curvelet subbands and curvelet co-occurrence matrix features with different steps in multi scales,and a SVM(Support Vector Machine)classification method is chosen to train the model and classify the image into MIZ/non-MIZ.Secondly,utilizing the difference among the same feature in curvelet domain from different size of patch with the same center pixel,a decision tree is chosen to determine the MIZ pixels from SVM result into MIZ or no-MIZ further.The results show that the features in curvelet domain are useful in identifying the MIZ and the results can classify the image into MIZ/non-MIZ effectively.The the SVM result can identify most of the MIZ and the decision tree method can correct the result at the boundary of two class.The accuracy of result has a large improvement which can achieve to 97%.The paper analyzed the curve-like features of MIZ.Since the curvelet transform has advantages of multi-scale,multi-orientation,position parameter and has the most optimal expression of curve,it is chosen to extract information and classification of MIZ in SAR imagery.The paper provided two ice edge detection methods,a dynamic feature extraction method and a pixel level MIZ detection method.The results give stronger and more reliable parameter in the study of MIZ than using passive microwave sea ice concentration,while it enlarges the application of sea ice interpretation of SAR.