| description abstract | Visual detection and classification of water and waterbodies provide important information needed for managing water resources systems and infrastructure, such as developing flood early warning systems and drought management. But water itself is a challenging object for visual analysis because it is shapeless, colorless, and transparent. Therefore, detecting, tracking, and localizing water in different visual environments are difficult tasks. Computer vision (CV) techniques provide powerful tools for image processing and high-level scene analysis. Despite the complexities associated with water in visual scenes, there are still some physical differences, such as color, turbidity, and turbulence, affected by surrounding settings, which can potentially support CV modeling to cope with the visual processing challenges of water. The goal of this study is to introduce a new image data set, ATLANTIS Texture (ATeX), which represents various water textures of different waterbodies, and evaluate the performance of deep learning (DL) models for classification purposes on ATeX. Experimental results show that among DL models, EffNet-B7, EffNet-B0, GoogLeNet, and ShuffleNet V2×1.0 provide the highest precision, recall, and F1 score. However, by considering the training time, total number of parameters, and total memory occupied by these models, ShuffleNet V2×1.0 is presented as the most efficient DL network for water classification. Finally, results from this study suggest that ATeX provides a new benchmark to investigate existing challenges in the field of image analysis, in particular for water, which can help both water resources engineers and the computer vision community. | |
