Studies On Methods Without Despeckling For Change Detection In Bi-temporal SAR Imagery

Change detection is the process of identifying differences in the state of an objector phenomenon observed at different times, and optical images have gotten widelystudied and applied. However, the optical sensor suffers from the Sunshine andAtmosphere conditions, which limits the applications of change detection. On thecontrary, Synthetic Aperture Radar (SAR) has the capability of working under theall-time weather conditions, and its acquired images have been applied into many fieldsof national security and economy. But the speckling phenomenon is always companyingwith SAR images, due to the drawbacks of imaging mechanism. The existence ofspeckle leads to the degradation of objects’ properties, even sometimes, it submerges thetrue objects and then makes a badly undesired effect on the following SAR imagesinterpretation.Usually, an unsupervised SAR image change detection is composed of three steps:preprocessing, generation of the difference image and threshold selection. For twoco-registered and corrected SAR images, the main aim of preprocessing is to reduce theeffect of the speckle noise to the following acts. In general, preprocessing is alwaysdone by a filter. However, this would lead to the loss of detail information, and then thequantitative and visual performances would be affected. With the aim of keeping theperformance to be comparable and meanwhile avoiding the use of the filter, the thesisworks on the study of change detection methods for SAR imagery. The main works areas follows:(1) To avoid reducing the speckle noise, and meanwhile overcome the limitation ofselecting the distribution model, first, characteristics of the difference image (DI)are integrated with an interactive segmentation method not referring to anydistribution assumption, to generate change detection maps corresponding todifferent “seeds”, then a voting competition strategy is used to fuse those results togive the final change detection map. During segmenting, the feature of each pixel isset as a vector consisted of the corresponding intensities in the DI and each scalerepresentation of the DI given by stationary wavelet transform (SWT). This kind offeatures and the decision level fusion make our proposed method robust to thespeckle noise. Results on real SAR datasets obtained under the situation that there isno despeckling preprocessing to SAR images confirm the effectiveness of ourmethod.(2) Threshold-based methods have been applied to the field of change detection due totheir simply implementations. But most need despecking, or the performance is a bitsatisfactory. To improve the performance of a threshold method, a novel thresholdway is used to replace the traditional threshold way, and a fusion strategy based onthe markov random field (MRF) model is emplyed to fuse different results. Due to the change of thresholding way and the smoothness of the MRF model, theproposed method can have the performance comparable without despeckling. Resultson real SAR datasets obtained under the situation that there is no despeckling preprocessingto SAR images confirm the effectiveness of our method.(3) As we all know, the use of the local information is helpful for suppressing the effectof the speckle, and the Markov random field (MRF) model is the more widely usedone. But this kind of approaches still generated the final change detection map at thepixel level, which means that there are noisy points, holes in connected componentsand jagged boundaries in the change detection map. To handle the defect, the MRFmodel is replaced with region of interest to introduce the local information. Thisstrategy can also generate the result at the region level, by combining with the act ofsearching all the connected regions and viewing each connected region as aprocessing unit. Results on real SAR datasets obtained under the situation that there is nodespeckling preprocessing to SAR images confirm the effectiveness of our method.(4) To avoid the limitation of a threshold-based method suffering from the goodness-fitbetween the selected and the true distribution models, a threshold method based onextracting transition region is tried. This method does not refer to distributionassumption, but is still affected from the speckle, like other threshold-based methods.Thus, an improved method of extracting transition region is proposed and a noveldifference image is constructed. Due to the new difference image and the strategy ofextracting transition region, the proposed method can have a comparableperformance under the situation that no despeckling is done. Results on real SARdatasets obtained under the situation that there is no despeckling preprocessing to SARimages confirm the effectiveness of our method.(5) The threshold-based methods and the MRF-based ones are two popular kinds ofmethods for SAR imagery change detection. The threshold-based method has thefollowing disadvantages, such as distribution assumption, ignorance of the localrelationship, etc. The MRF-based method always has the boundaries of change areasoversmoothed thanks to the nature characteristics of the MRF model. Aiming attheir owned defects and advantages, a hybrid method based on fusion is proposed.Due to the use of a fustion strategy inspirited by region growing method, theproposed method does not need the despeckling preprocessing, and meanwhile hasthe quanlitative performance improved. Results on real SAR datasets obtained under thesituation that there is no despeckling preprocessing to SAR images confirm theeffectiveness of our method.