Study On Information Extraction Of Landslides Hazard Of Object-oriented Based On High-resolution Remotely Sensed Images

As we all know that the 8.0 magnitude earthquake happened in Wenchuan county of Sichuan Province had a great impact on most regions in Sichuan, but southern regions in Gansu province also have, experienced a heavy hazard that the earthquake induced a series of massive landslides, collapse, debris flow and other geological disasters. To complete the assessment of landslides quickly and efficiently and provide technique support and theoretical basis for rescue work of landslides, the information extraction of landslides is needed. Only when we know the distribution, characteristics, scale of landslides and the damaged degree on buildings, road, farmland, bridge and so on, rescue operation and prevention of landslides could be launched immediately. In this letter, a semi-automatic approach based on object-oriented for the information extraction of landslide using very high resolution satellite images is introduced.The object-oriented method used object level as analytical unit with several systemized spectral, textural measurements, topological relationship, the context semantic information and so on. Compared with pixel-based classification technique, the object-oriented method has great advantages such as lower salt-and-pepper noise and better classification accuracy.Quick Bird remotely sensed data with spatial resolution of lm in 2013 and another IKONOS satellite image with spatial resolution of lm in 2007 was selected to the extraction information of landslides hazard by object-based hierarchical classification combined with DEM data with resolution of 30m. The typical region called Wudu District in southern region of Gansu province was selected as the study region. Many landslides were distributed in this region.The main conclusions are as follows:Firstly, for the seek of the optimal parameter values, the image segmentation parameter quantitative experiments were performed. After several experiments, the optimal segmentation parameter values suitable for different terrain types were determined based on visual interpretation method. After that, spectral signature, geometric characteristics and texture feature were analyzed to establish an vector space model for the information extraction.Secondly, the object-oriented change detection method was also completed to obtain the change result using the information of residence in between 2007 and2013.The change information of residence also reflected indirectly the damaged residence due to landslides triggered by earthquake.Thirdly, the fuzzy classification was adopted to extract the hazard-affected information of landslides including the residence, road and vegetation. The obtained road and vegetation information under the threat of landslides hazard would provide technical support for the vulnerability assessment of landslides.Finally, support vector machine method was applied to extract the landslides.200 landslide samples and 300 samples of non-landslide were used for data training and to select the optimal attributive character combination from different kinds of characteristics such as spectral signature, texture properties and so on. Then the assessment of classification accuracy is discussed and the confusion matrix method was selected for the quantitative evaluation. The practical survey data proved that the total accuracy of the information extraction of landslide hazard was above 85%.In conclusion, the object-oriented method proved to be favorable for the information extraction of landslides and could satisfy the practical requirement for providing the basis for the rapid assessment of hazard vulnerability of landslide hazard.