Research On Object - Oriented High - Resolution Remote Sensing Image Building Damage Information Extraction

With the improvement of image spatial resolution and the development ofremote sensing information extraction technology, remote sensing technology isbecoming an effective means to quickly get earthquake information, earthquakeemergency and rapid damage assessment. High-resolution remote sensing images canobtain abundant surface details. Traditional pixel-based classification methods can notfully exploit images spectral, geometry, texture and context information. In addition,it has some other limitations such as low classification accuracy, slow speed and soon. Therefore pixel-based classification methods can not meet the needs of quickearthquake damage information extraction. The object-oriented image analysismethods provide a new way of thinking for extraction of damage information usinghigh spatial resolution remote sensing images. In this paper, the object-oriented imageanalysis methods are used to extract the building damage information fromhigh-resolution remote sensing images. Research work and achievements are asfollows:1) The advantages and disadvantages of domestic and foreign major existingimage segmentation algorithms are first analyzed. Then, the description ofobject-oriented multi-resolution segmentation’s algorithms and processes is detailed.At the same time, the selected basis and principles of segmentation parameters suchas band weights, color factor, shape factor, compactness, smoothness andsegmentation scale are in-depth analyzed.2) A method to select the optimal segmentation scale according to the qualityfunction of segmentation is proposed based on the principle of “image objecthomogeneity of maximum and heterogeneity of minimum”. Experiment proves thatthe quality function of segmentation is a good indicator of heterogeneity andhomogeneity and an efficient method for optimal segmentation scale’s choice.3) Based on the acquisition of image objects knowledge, objects’ featureselection policies and other related issues in the rule base of knowledge andmembership functions are analyzed in this paper. Then the mathematical model and classification principles of nearest neighbor classification and fuzzy classification arestudied. A complete fuzzy classification system including building fuzzy sets withfuzzy input variables, fuzzy sets combined by fuzzy logic and anti-fuzzyclassification is built.4) Typical construction zone in Jiegu Town, Yushu post-earthquake Quickbirdhigh spatial resolution image is selected as experimental data for damage informationextraction. According to the characteristics of different objects, it applied the optimalsegmentation scale to image segmentation. So a multi-resolution hierarchical networkstructure has been built. Then the knowledge base of fuzzy rules is bulit based onimage objects’ features, which included spectral, geometry, texture and spatialcharacteristics of topological relations. Following, the different levels of buildingdamage are successfully extracted using object-oriented approach. Finally, theaccuracy classification stability and optimal classification results evaluation isexecuted from the perspective of classification membership of fuzzy concept.Experimental results show that the object-oriented method for building damageinformation extraction based on the segmentation at optimal scales can meet theneeds of fine extraction.