Water Body Detection in High-Resolution SAR Images With Cascaded Fully-Convolutional Network and Variable Focal Loss

Deep-Learning-Based Multispectral Satellite Image Segmentation for Water Body Detection

“…It also facilitates timely flood protection planning and water quality control for public safety and health [2]. To achieve an insightful analysis of water systems in cities, automated accurate water body detection is the first and fundamental stage to provide pixel-level identification of water regions [3], [4].…”

“…Since its launch in 2015, the Sentinel-2 satellite has provided publicly available multi-spectral imagery that has been widely employed in land-cover applications [5], [6], [7]. It offers one of the most suitable data sources for timely urban hydrological monitoring and analysis due to its neardaily update frequency compared to higher-resolution remote sensing data such as Very High Spatial Resolution (VHR) [8] and Synthetic Aperture Radar (SAR) [4]. Thus, in this paper, we investigate the use of 10 meter resolution multi-spectral data from Sentinel-2 due to its potential for urban hydrological applications that require frequently updated data in their analysis process.…”

“…Deep convolutional neural network (DCNN) models have become popular for water body detection in recent years [14], [15], [4], [3], [16]. DCNN-based semantic segmentation networks employed for remote-sensed water detection in urban hydrological applications include fully convolutional networks (FCNs) [17], upsampling pyramid networks (UPNs) [4] and DenseNet [18].…”

“…Deep convolutional neural network (DCNN) models have become popular for water body detection in recent years [14], [15], [4], [3], [16]. DCNN-based semantic segmentation networks employed for remote-sensed water detection in urban hydrological applications include fully convolutional networks (FCNs) [17], upsampling pyramid networks (UPNs) [4] and DenseNet [18]. The advantage of these models is that they are able to extract more distinctive feature representations compared to traditional water index features, thus enabling improved water body detection.…”

“…Multi-spectral imagery should support further improved water body segmentation compared to using only RGB channels due to the additional information contained in the extra bands that cover a wider part of the electromagnetic spectrum, while the resulting higher-dimensional data can be reduced through appropriate methods [19]. However, in recent work [4], [20] the use of multiple bands did not prove to be very effective. In the Kaggle Satellite Imagery Feature Detection challenge [20], methods using all of the available 20 channels (1 panchromatic channel with pixel resolution of 0.31m, RGB channels (0.31m), 8 multi-spectral bands (1.24m) and 8 short-wave infrared bands (7.5m)) achieved only insignificant improvements compared to those employing RGB bands only.…”