INVESTIGATION OF SHORELINE EXTRACTION USING MACHINE LEARNING WITH X-BAND SAR IMAGES

Abstract

 High-frequency and wide-area coastal monitoring is expected to benefit from the use of Synthetic Aperture Radar (SAR), yet efficient methods for extracting coastlines from high-resolution SAR images have not been established. In this study, we developed a shoreline extraction method using DeepLabv3+, a convolutional neural network, applied to 1-meter resolution X-band SAR images. To explore applicability on coasts with structures such as offshore breakwaters, we created multiple models by varying the combinations of coastal images used for training. For coasts without structures, using SAR images of the specific coast for training enabled coastline extraction with an average error of 3 to 4 meters. Even for coasts with structures, incorporating SAR images of coasts without structures in the training reduced the average error in coastline extraction with a similar accuracy. These results demonstrate the potential of using high-resolution X-band SAR images for coastal observations.