Research On Sea-land Segmentation Algorithms Of Clutter Scene Based Upon Fusion Of Multiple Measures

Marine surveillance radar operating in scanning mode monitors a complex area of nearly 100 kilometers in radial,with ocean,land and island reefs included.The returns of marine radar always consist of ground echo,sea clutter,echoes from targets,echoes from island and the ubiquitous noise of receiver.In the case of low grazing angle or high spatial resolution,the intensity of sea clutter is sometimes lower than noise in some ocean areas.Therefore,the purpose of sea-land segmentation of marine surveillance radar is to separate the illuminated backgroud into land area,sea clutter dominant area and noise dominant area according the different characters of the detection scene in radar echoes.The quality of land-sea segmentation directly affects the selection and performance of subsequent detection methods.The existing land-sea segmentation methods are mainly based on a single measure extracted from radar echoes,which leads the low segmentation quality.In this thesis,we will focus on the use of multiple measures extracted from radar echoes to reflect the differences between ocean,land and noise,and ealize the sea-land-noise three-dimensional segmentation of detection scenes through the fusion of multiple measures.The main contents of this thesis are as follows:Firstly,the image processing theory involved in this thesis is briefly introduced,reviews the application of machine learning in image processing,emphatically K-NN algorithm in supervised learning,and analyses its advantages and disadvantages.Then the mathematical morphology method in image processing is introduced,and the application of binary morphology and gray morphology in image processing is briefly described.The function and processing effect of each operator are demonstrated by examples.Secondly,an effective sea-land segmentation method is implemented in this thesis.The essential difference between sea clutter and ground clutter is analyzed based on the principle of radar and the measured data.It is shown that the original difference is Doppler bandwidth,other than Doppler shift and initial phase.Therefore,we propose an initial phase-Doppler shift invariant distance to characterize the difference between the echoes of each spatial resolution cell.Under this measure,the echo vectors of some land and sea areas are selected as training sets according to prior information,and K-NN classifier is used to segment the detection scenes.Finally,the initial classification results are further processed by mathematical morphology to obtain land-sea segmentation results.The proposed method can correctly divide the detection scene into ocean and land areas without setting any threshold.Finally,a method of sea-land-noise three-dimensional segmentation based on K-NN classifier is proposed in this thesis.According to the different characteristics of sea clutter,ground clutter and noise,the feature vectors are composed of five features: non-Gaussian measure,whiteness,Doppler bandwidth,intensity and noise ratio.The differences among the resolution cells are characterized by Markov distance between the feature vectors.Under this measure,K-NN classifier is used to segment the detection scenes.The final scene segmentation is obtained by morphological filtering.The noise area is successfully removed from the detection scene,and more refined scene segmentation is realized.The segmentation results of measured data verified the validity of proposed K-NN segmentation method.