Study On The Methods Of Intelligent Extraction Of Seismic Damage From Remote Sensing Images

Earthquake is one of the worst natural disasters for human beings. As earthquakeprediction is still a worldwide scientific problem, the accurate and rapid damageassessment is crucial for emergency rescue after a destructive earthquake. Because ofthe excellent characteristics such as fast acquisition, multi-sensor, all-day observation,large area coverage, being informative and independence on ground natural conditions,remote sensing could access to the earthquake disaster information from the groundafter the devastating earthquake, which has been playing an important role inearthquake relief.The main methods of seismic damage information extraction from thepost-earthquake imagery are artificial visual interpretation and automaticinterpretation with computers. Because of the drawbacks such as heavy workload,long processing cycle, being affected by people’s experience, artificial visualinterpretation has reduced the advantages of the processing with remote sensingtechnology. As the great improvement of the image processing technology, automaticinterpretation has also the advantages of the short processing cycle and highproduction efficiency, it has not been applied widely recently because of its lowprecision results and high requirements for the remote sensing imagery.In this thesis, a scheme is studied to improve the algorithm of intelligent, rapidand automatic seismic damage information extraction from optical imagery forearthquake losses assessment. The main content includes three aspects: thefundamental theories and the development of automatic methods of seismic damageinformation interpretation from remote sensing imagery, the algorithm improvementof change detection for damage extraction by multi-temporal imageries on the pixellevel, and the scheme improvement based on the OBIA (object based image analysis)for seismic damage extraction.First, a review of the development and current status of automatic methods withcomputer for the seismic damage extraction has been introduced. The advantagesand disadvantages of the methods based on pixel and OBIA classification have also been summarized. The possible ways to improve the algorithms of above two methodswere also analyzed.Second, the improvement of the algorithm for the seismic damage extractionfrom remote sensing imagery on pixel-level has been investigated. Based on digitalchange detection from multi-temporal high resolution RS images, a regional searchingoptimization (RSO) method is proposed to improve correlation coefficient changedetection. The method has also been applied to the experiment of extraction ofbuilding damage in the partial area of Dujiangyan City, Sichuan, caused by2008Ms8.0Wenchuan, Sichuan, China earthquake, from the pre-and post-earthquakeoptical remote sensing imageries. Distinguished from the conventional method, thedamage level of individual building was identified as one unit based on its outlinevector. The Kappa index, indicating the classification accuracy, showed high precisionand better robustness.Third, an improved method based on object-oriented damage extractionalgorithms from remote sensing image was developed. An improved OBIA wasproposed whose aim is to determine the optimal segmentation parameters by Kappaindex statistics, after analyzing the shortage of present OBIA methods such as ENVIFX, Definiens eCognition. The improved OBIA algorithm mainly depends on theimage segmentation and classification by the method of MeanShift and SupportVector Machine (SVM) respectively, and then the classification accuracy Kappa indexwould be calculated automatically. The best Kappa value would correspond to theoptimal segmentation parameters and then get the best classification resultrepresenting the result of the optimization process. This effectiveness compared withtraditional OBIA was demonstrated in Dujiangyan city and Yushu County withpost-earthquake VHR (very high resolution) optical imagery.Three advantages have been shown in the experiment:(1) the optimalsegmentation parameters and the best classification result could be calculatedautomatically and intelligently;(2) high efficiency, message passing interface(MPI)was applied in this improved algorithms;(3) better portability and open source, thismethod could been operated on cross-platform such as Linux and Windows system.Finally, conclusions and prospects of the two improved algorithms were summarized. Although great progress has been achieved, there are some problems andshortcomings:(1) the improved RSO method on pixel-level has a higher accuracy andbetter robustness, but a low versatility and efficiency because of the process ofexcessive human interaction.(2) The improved method based on OBIA would dependon samples for classification which need artificial interaction and thehigh-performance equipment. Overall, the results of this research work in this paperachieved the desired goals, and expected to further study and apply in practice.