| Popis: |
Landslide due to heavy rains and earthquakes is major hazards to human life and property. Applications for rapid detection and mapping of the damage situation and extent using earth observation satellite imageries are expected. Especially, Synthetic Aperture Radar (SAR) imagery is effective due to the capabilities of cloud penetration and is independent of solar illumination. It was, however, difficult to extract landslide areas in SAR images accurately using the traditional methods. Therefore, we tried to extract landslide areas using Convolutional Neural Networks (CNNs), which are being used for computer vision. We adopted U-Net, one of the CNNs, for Landslide extraction. The U-Net enables accurate segmentation from a small amount of training data. We verified the landslide extraction with U-Net, using the collapsed areas caused by the 2018 Hokkaido Eastern Iburi Earthquake that occurred on September 6, 2018. Landslide extraction was performed using pre- and post-event X-band COSMO-SkyMed imageries. For pre-processing, we performed multi-looking, radiometric calibration, and ortho-rectification using 10 m DEM data. The U-Net was trained for 100 epochs with a mini- batch size of 24, 32, and 40. Two types of dataset were prepared for the model input, that is, (1) pre- and post-event COSMO-SkyMed amplitude and the ratio of pre- and post-event COSMO-SkyMed amplitude, (2) pre- and post-event COSMO-SkyMed amplitude and slope. As a result, the optimal value of the F-measure (70.9%) was obtained with the dataset (1) using 128 × 128 strides and batch size of 32. Topographic factor (slope) did not improve landslide extraction in this study. |
