MobileUNet Based Flood Region Segmentation in UAV Images for Developing Early Flood Warning and Disaster Management System

To support people with information on natural disasters, numerous early warning systems have been developed. Most of these systems employ a suitable computer-algorithm to process the data received from various sensors and its outcome help to provide necessary alert when there is severe abnormality. This study aims to develop a deep-learning (DL) based automatic flood region segmentation system from the digital images collected using the unmanned-Areal-Vehicle (UAV). This process help to detect the severity of the flood level in a chosen location. To reduce the implementation complexity and to enhance the detection accuracy, this work considered the MobileUNet (MU) scheme. The different phases of this tool includes; image and mask collection and resizing, training the MU with image and mask, segmentation of the flood region, and executing the pixel wise comparison between segmented binary section and the mask to computer the necessary metrics. The outcome achieved with MU is compared with chosen UNet models and the overall result of this study confirms that the proposed MU helps to achieve a better Jaccard (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$94.41\pm 0.13$</tex>), Dice (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$98.11\pm 0.06$</tex>), and accuracy (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$98.49\pm 0.02$</tex>) compared to other chosen DL-methods. This confirms that the implemented technique provides better result on the chosen database.