U-Net Based Water Region Segmentation for LAPAN-A2 MSI

Abstract

In this work, we analyzed the performance of a deep learning-based segmentation method in extracting water regions from multispectral imageries (MSI) taken by LAPANA2 microsatellite. The interested water regions include open seas and the river as well as their branches. The capability of detecting and segmenting the water component on LAPAN-A2 MSI is important as the satellite was dedicated to support maritime surveillance missions on Indonesian waters. Therefore, this capability will help future water object detection to encapsulate its region of interest, i.e., water. The segmentation has been performed by employing a state-of-the-art deep learning-based method, U-Net, using 696 training images. This method is considered due to its capability to provide promising accuracy without requiring an extremely extensive amount of training dataset. Based on our experiment, the trained U-Net has shown a satisfying result with an accuracy of 89.13% as measured using Intersection over Union (IoU) metric.