A Multi-scale/step Object Oriented Image Analysis Scheme for Landslide Recognition

• Main Authors: Jyun-PingJhan, 詹鈞評
• Other Authors: Jiann-Yeou Rau
• Format: Others
• Language: en_US
• Published: 2012
• Online Access: http://ndltd.ncl.edu.tw/handle/96585630982045147960

碩士 === 國立成功大學 === 測量及空間資訊學系碩博士班 === 100 === Taiwan was attacked by Typhoon Morakot during 5-10 Aug. 2009, which brings a huge amount of cumulative rainfall and caused lots of landslides in the mountainous area, such as in the Xiao Lin village, Lu San spring region, and Teng Zhi area. In this study for the purpose of performing multi-scale/step object-based landslide detection, we utilize images acquired from different platforms through Formosat-2 satellite, an airplane, and an unmanned-aerial vehicle (UAV) that equipped with DSLR digital camera as well as digital elevation model (DEM) and digital surface model (DSM) derived from airborne LiDAR system. In traditional image analysis methods, pixel-based classification generally causes pepper and salt noise that will decrease the overall accuracy. Hence, in this study based on object-based image analysis (OBIA) we proposed a multi-scale/step landslide detection scheme. For landslide classification, considering image may contain vegetation, buildings, roads, cloud, shadow, and landslides, six feature indices are derived from spectral, topographical, and object shape information, such as vegetation index, brightness, slope, object height model (OHM), density, and contrast. In the multiresolution segmentation, image is segmented through a root-scale that is determined by the image spatial resolution and the thresholds for feature indices are determined manually by selecting some training samples. In the first, the preliminary landslide results are classified as landslide seeds and then we adopt region growing on seed objects with 1.5 times of the root-scale to expand the landslide regions. In the final, a down scale segmentation is applied on the expanded objects with a 0.75 times of the root-scale to remove some small patches errors. However, considering that landslides may be covered by shadow, we also develop a shadow landslide detection method based on the same multi-scale/step classification idea to reduce the omission errors. After evaluation accuracy assessment through three stages of landslide detection, we conclude that the proposed scheme can optimize the results and get a maximum value of overall accuracy. Meanwhile, after shadow landslide detection the omission error is reduced. It is a major contribution of this study, particular for images acquired during the winter season or has low sun elevation angles.