[U-04] Geospatial Applications for Societal Benefits

[U04-P06] Mapping Disaster Areas Using Satellite Images, Web Processing Service and Machine Learning Algorithms

*Joel Bandibas¹、Shinji Takarada¹ (1.Geological Survey of Japan, National Institute of Advanced Industrial Science and Technology)

キーワード：Mapping Disaster Areas, Convolutional Neural Network, Satelliate Images, Web Processing Service

Satellite images are very useful for efficiently mapping large areas affected by natural disasters. Conventional methods for identifying damaged areas using satellite images involved the determination of spectral and spatial patterns of damaged land covers using parametric statistics. Satellite images covering disaster areas often have high spectral frequency where pixel values belonging to a spectral class are not normally distributed. This makes the mapping of disaster areas using parametric statistics less accurate. This paper presents supervised and unsupervised methods to identify areas affected by natural disasters using Artificial Neural Network (ANN) and Convolutional Neural Network (CNN), respectively. The first method involves land cover change detection using satellite images taken before and after the occurrence of disasters. This method involves three steps. First, the satellite image taken before the occurrence of the event is segmented into several spectral classes using a simple moving average algorithm; second, a 3-layered ANN is trained using the values of neighboring pixels for each spectral class within a defined pixel window, and third, the final change detection process is implemented by feeding the trained ANN with pixel values from the same location of the two satellite images taken before and after the event, generating two ANN outputs. The difference between the two outputs above a set threshold signifies that the location is damaged. The second method uses CNN, a type of deep learning-based approach for the identification of damaged areas. CNN uses filters and learns what characteristics in the filters are most important in representing damaged land covers. Sample patches of satellite images covering damaged areas are used to train CNN. Trained CNN is then used to map areas affected by natural disasters.
WebGIS has recently been the preferred platform for rendering and processing geospatial data. It is linked to Web Map Service (WMS) and Web Processing Service (WPS) for rendering and processing geospatial data online, respectively. In this study, WebGIS is developed, and WPS for the two mapping methods is formulated using PyWPS, C++, and GDALDataset C++ API. The developed system has been successfully tested to map areas damaged by landslides in the Southern Philippines.