Research On Sea Background Suppression For Remote Sensing Images Based On Statistical Learning

Ocean remote sensing is one of important content of global environmental and security monitoring. How to extract sea target candidates from plenty of remote sensing images accurately has become a key research direction, among which sea background suppression is one of key technologies. It not only can improve signal-noise-ratio and contrast of images to determinate the positions of targets and extract their features exactly, but also can reduce the complexity of images processing and provide important information for sequence applications, for instance, targets classification and distinguishing. However, due to the great influence of many factors, such as weather condition, illumination, marine environment, shooting angle and so on, effective suppression for sea background with varied morphologies, fluctuation patterns, distribution characteristics, et al, has become a research problem for ocean remote sensing technology. In order to extract targets from remote sensing images exactly, further research and development for new effective sea background suppression method is the urgent need. The paper deeply analyzes the statistic distribution characteristics of sea background in large number of optical remote sensing images and a set of sea background suppression methods with remarkable results are proposed.Specifically, major contributions of this dissertation are summarized as follows:1. The extended dark prior channel method(EDCP) and the patch-based dark channel prior method(PB-DCP) are proposed for remote sensing multi-spectral images clound removal. How good the dark channel prior on remote sensing multi-spectral images is verified and we extend the dark channel prior from three-channel color images to multi-channel remote sensing images. Then the E-DCP cloud removal method is proposed. We also analyze the spatial variability of atmospheric light with optical radiation characteristics and PB-DCP cloud removal method is proposed.2. The statistical characteristics of sea background in satellite high-resolution remote sensing images are derived. Several classical statistical distribution models, including Gaussian, Rayleigh, Weibull, Gamma and Alpha stable, are adopted to fit the spatial intensity distribution of a single sea background image, the intensity distribution in every pixel of a series of background images, collected from the same sea area, with similar fluctuation pattern, and their magnitude distribution in every frequency point. Their K-S hypothesis testing results are also given. It presents that the spatial intensity distribution and the intensity distribution in every pixel are uncertainty, but the amplitude distribution in every frequency point is accord with some statistical properties. These analysis results provide a forceful evidence for the sea background suppression filters and methods designing.3. A new Gaussian statistical learning sea background suppression method, GBSF, is proposed for high-resolution remote sensing images in frequency domian. The time shift invariance and consistency of magnitude spectrum of sea background samples for satellite high-resolution remote sensing images are derived. Gaussian statistical distribution is adopted to model the magnitude in every frequency point of sea background samples and an ideal sea background suppression filter(BSF) is generated. The ringing effect and ghost phenomena caused by the sharp cut-off characteristic of ideal sea background suppression filter are solved by using of Gaussian kernel function smoothing, and then an effective Gaussian kernel function smoothing sea background suppression filter(GBSF) is achieved. Finally, the two sea background suppression methods based on magnitude Gaussian statistical learning are proposed to suppress sea background and extract target candidates. Experimental results show that, compared with the existing methods, the spectral Gaussian statistical learning method can effectively suppress sea background in images with different spatial resolution, fluctuation patterns and targets.4. A new multi-model statistical optimization sea background suppression method, GMBSF, is proposed for high-resolution remote sensing images in frequency domian. It analyzes the fitting goodness of the statistical distribution models for the magnitude spectrum in every frequency point and the multi-model optimized statistical distribution for the magnitude spectrum of sea background samples is given. Based on optimized statistical distribution model, a Gaussian kernel function smoothing multi-model statistical optimization sea background suppression filter(GMBSF) is designed and the sea background suppression method based on magnitude multi-model statistical optimization is proposed. The experimental results shows, compared with Gaussian statistical learning sea background suppression method, multi-model statistical optimization sea background suppression method can suppress sea background and extract target candidates from remote sensing images more effectively and accurately.5. A new multi-Gaussian statistical learning sea background suppression method, MGBSF, is proposed for aerial ocean remote sensing images. It analyzes spatial distribution characteristic of sea background in aerial ocean remote sensing images. A multi-Gaussian model characteristic cover learning method is proposed to obtain the number of Gaussian components adaptively and model sea background precisely. Then a sea background suppression method with high performance is presented based on multi-Gaussian model for aerial ocean images.In the end of this dissertation, we summarize our work and discusse the problems need to be researched further.