DeepRivWidth: Deep learning based semantic segmentation approach for river identification and width measurement in SAR images of Coastal Karnataka

River width is an essential parameter for studying the river's hydrological process and has been widely used to estimate the river discharge. The existing approaches to measuring river width are based on remotely sensed imagery such as MODIS, Landsat to identify the river, and then estimate the river width. In this work, an alternate approach for river width estimation is proposed using the under-explored modality Synthetic Aperture Radar (SAR) images. SAR, unlike the traditional electro-optical sensors, can penetrate the clouds and can be used to collect the data in all weather conditions and even during the night. In this work, the river identification process is manifested as a binary semantic segmentation task in SAR images. For this, two state of the art deep learning algorithms (U-Net, DeepLabV3+) are utilized for river identification and subsequent width measurement. The proposed approach (DeepRivWidth) is used to estimate the width of the river of the Mangalore–Udupi region of Coastal Karnataka (India). These rivers originate or pass through Western Ghats (UNESCO world heritage site), and the proposed river width measurement approach could provide critical input for ecologists besides assisting efficient water management of the region. The estimated width is compared with the manually measured width, and significant improvement in the accuracy was obtained compared to existing river width measurement approaches. Besides, the performance evaluation of semantic segmentation approaches for river identification on a publicly available dataset provides valuable insights into segmenting rivers in SAR images.