Research On Post-earthquake Building Collapse Information Extraction Based On Pléiades Satellite Images

Earthquake,especially medium-strong earthquake,often causes a devastating impact on human life and Production activities.Rapidly obtaining information of building damage after a destructive earthquake can contribute to earthquake disaster evaluation and emergency rescue work Satellite remote sensing technology can provide strong support for earthquake emergency response due to its advantages of large coverage range,fast speed acquisition and less restriction by ground conditions.In order to quickly and accurately obtain the building earthquake damage information after earthquake,in this paper we used the Pléiades satellite image data after the Ludian earthquake in Yunnan to extract the building collapse information in two ways,which are image element-based method and object-oriented classification method.The main contents are as follows:(1)Collecting and analyzing and multi-source data related to the Ludian earthquake.Primary data were acquired with high-resolution panchromatic imagery(0.5m)and multispectral imagery(2m)provided by the Pléiades satellite on 6 August2014.The ancillary data include;1)Google Earth image with a spatial resolution of0.125 m on 14 March 2013 and 20 August 2014,respectively;2)ALOS-12.5m DEM;3)Basic geographic information data such as ground control points and surface water system.By analysing the Pléiades satellite image data,we known the data has advantages of high spatial resolution,temporal resolution and wide coverage,the data requirements for post-earthquake building damage information extraction can be met.(2)Extracting post-earthquake building damage information.In the extraction based on pixel-based classification,after selecting samples,the five methods of Maximum Likelihood method,Neural Network method,Support Vector Machine,CART Decision Tree method and Random Forest method were applied to extract and classify the post-earthquake information from remote sensing images in the study area.In the building information extraction based on object-oriented classification,samples loading and multi-scale segmentation were firstly carried out,and the post-earthquake information extraction and classification were carried out by three methods,namely the Support Vector Machine,the CART Decision Tree methods and the Random Forest method.The eight results were eventually derived based on the eight methods.(3)Post-classification processing and accuracy assessment.The accuracy of the classification was improved by speckle processing,and a thematic map of post-earthquake feature information was produced;the high-resolution Google Earth image data was used as a reference data source for accuracy assessment by the confusion matrix.The assessment results show that: Among the results based on pixel-based classification,the overall accuracy for all categories of features obtained by the five methods of Maximum Likelihood,Neural Network,Support Vector Machine,CART Decision Tree and Random Forest is 76.97%,77.05%,78.23%,81.11% and 87.04% respectively.The average mapping accuracies of fully damaged houses,partially damaged houses and non-damaged houses obtained by the five methods are 61.49%,55.25%,58.45%,66.37% and 78.82%,respectively,and the average user accuracies are 65.86%,58.56%,60.20%,67.60% and 79.09%,respectively;Among the classification results under the object-oriented approach,The overall accuracy for the all categories of features obtained by the three methods of Support Vector Machine,CART Decision Tree and Random Forest is 83.92%,88.70%and 92.81% respectively,and the average mapping accuracy for fully damaged houses,partially damaged houses and non-damaged houses obtained by the three methods is75.28%,83.33% and 88.13% respectively,and the average user accuracies is 77.88%,80.66% and 86.56%.The study results show that:(a)Pléiades satellite remote sensing data has practical significance for post-earthquake disaster information extraction due to its high spatial resolution,high temporal resolution,wide a coverage and accessible;(b)The accuracy of object-oriented classification is higher than that based on pixel-based classification;(c)Multi-scale segmentation and sample selection have a great influence on the accuracy of the classification results;(d)Among all the experimental solutions,the best results can be obtained by using the Random Forest method of the object-oriented classification to extract building damage information.The conclusions and research ideas in this paper can provide some technical and methodological reference for the extraction and assessment of post-earthquake building damage information.