In this study, we construct a model to extract flooded areas from SAR (Synthetic Aperture Radar) images using semantic segmentation and clarify its accuracy. DeepLab v3+ is used as the semantic segmentation model, and images of the tsunami-flooded area of the Great East Japan Earthquake taken by Pi-SAR2 are used for training and validation. In addition, the accuracy of detecting flood areas is improved by combining SAR images and DEM (Digital Elevation Model) information for the input of semantic segmentation.

 To improve the classification of gravel using the classification model which was developed based on the convolution neural network model, this study focused on the amount of information in the training datasets. There are two pieces of training in this study. The first case of pieces of training was conducted with the datasets which included pictures that changed brightness added to normal datasets. In another case of training, training datasets were created with different resolutions of original pictures. The results showed that the addition of pictures that were decreased and increased brightness improved training and test accuracy. To increase of picture's resolution that was used in the process of making the training datasets improved training and test accuracy. From the results of the application for aerial photos, changing brightness did not improve the classification of gravel but increasing the picture's resolution improved the classification accuracy of gravel.