Temporal context of lightweight network model for detecting boats approaching the tsunami early warning system

Abstract

The tsunami early warning system (TEWS) is a device that detects potential tsunamis. However, a boat that approaches TEWS is a source of communication disturbance. A convolutional neural network (CNN), as part of intelligent computer vision, is one solution for detecting boats and providing a warning to move away from the TEWS area. Water segmentation and refinement-temporal (WaSR-T), as the current advanced CNN network, exhibits impressive performance in detecting object obstacles in the marine domain, although it requires a powerful computational device. In the paper, we propose a modification of WaSR-T, replacing the most computationally intensive stages with a lightweight version called lightweight WaSR-T. On the proposed lightweight WaSR-T, the previous encoder of WaSR-T was replaced with MobileNetV3, and some feature layer maps were reduced as input to the decoder. For training and validating the lightweight WaSR-T, the image dataset representing the open sea and our extended dataset from Indonesia's ocean region were used. Based on the quantitative results and evaluation of the computational load, the sensitivity to detect a boat for WaSR-T and lightweight WaSR-T is 95.71% and 90.00%, respectively. The lightweight WaSR-T required less memory at 32.57%, resulting in a 0.0761% reduction in total processing time compared to the original WaSR-T. Therefore, our proposed lightweight WaSR-T is promising for use as the central part of an intelligent maritime computer vision system in TEWS.