SAR Image Change Detection Algorithm Research

As a result that synthetic aperture radar is not affected by atmospheric conditions and cloud cover and other related conditions, therefore, multi-temporal SAR images technology plays a more and more important role in production. This paper mainly discuss and study in depth on the SAR image change detection technology, two phase SAR image difference generation, the automatic division on the difference image and SAR image change detection based GPU parallel aspects. This paper proposes some effective algorithm to solve the SAR image change detection some practical problems.First, this paper presents a neighborhood-based ratio operator to generate difference image for change detection in SAR images. The proposed neighborhood-based ratio operator produces difference image by combing gray level information and spatial information of neighbor pixels. It is inspired by the ratio operator and is possible to reduce the negative influence of multiplicative noise of SAR images. The performance comparisons of the proposed operator with traditional ratio operator and log-ratio operator indicate that the neighborhood-based ratio operator is superior to these traditional methods, and produces better detection results.Then, an unsupervised threshold method based on a novel one-side fitting strategy is proposed for change detection in SAR images. The threshold strategy of one-side fitting is proposed by analyzing the inherent law and characteristic of SAR image change detection task. Two one-side histogram probability statistics models are designed for the one-side fitting strategy. Then, the combination of related model estimation strategy with expectation-maximization parameter estimation technique is adopted to obtain the probabilistic statistics models of changed and unchanged areas, respectively. Finally, the threshold is determined automatically by Bayesian decision rule. Experimental results on real SAR image datasets show that the proposed methods outperform other three state-of-the-art threshold methods in finding the optimal threshold. It is also shown that the proposed methods are robust to the problem of overlap between changed area and unchanged area.After the proposed unsupervised threshold algorithm, for further research this paper propose the hidden Markov random field (HMRF) in SAR image change detection. But most of methods about HMRF adopt crisp models and neglect the fuzzy aspect, so it is difficult to find a satisfactory result in complex situations. In this paper we propose a novel HMRF model base on fuzzy approach for change detection of SAR Images. The main contribution of this work is to offers a fuzzy energy function of the HMRF model, combining the advantages of HMRF model, which has the ability of constructing the spatial coherency model, with the enhanced flexibility obtained by the fuzzy approach. In this way, the fuzzy energy function can effectively less the accumulation error caused by hard partition on the iterative process. Our experiment results on widely-used related SAR image datasets with reference image compared to the obtained with a classical hard HMRF model and popular threshold algorithm, and the proposed algorithm produces better detection results.Because of the computation challenge from a large number of remote sensing image, and rapid processing and decisions is required by many applications of multi-temporal remote sensing. So we need to develop a parallel algorithm which fully integrates the advantages of software and hardware to satisfy the demand. In this paper, the algorithm PLog-FLCM, which is based on multi-temporal remote sensing change detection analysis on GPU, is proposed. The parallel characteristic and implementation details of the proposed PLog-FLICM algorithm are introduced, which is implemented on AMD Accelerated Parallel Processing (APP) SDK v2based on OpenCL. Experiments on standard data sets demonstrate that the proposed PLog-FLICM algorithm is superior to other algorithm, and has great accelerating capacity and meanwhile performs equivalently to benchmark algorithms.