Deep learning for automated river-level monitoring through river-camera images: an approach based on water segmentation and transfer learning

Abstract. River-level estimation is a critical task required for the understanding of flood events and often complicated by scarcity available data. Recent studies have proposed to take advantage large networks river-camera images estimate river levels but, currently, utility this approach remains limited as it requires amount manual intervention (ground topographic surveys water image annotation). We developed an using automated semantic segmentation method ease process river-level from images. Our based on application transfer learning methodology deep neural designed segmentation. Using datasets series extracted four cameras manually annotated observation event rivers Severn Avon, UK (21 November–5 December 2012), we show that algorithm able automate annotation with accuracy greater than 91 %. Then, apply our year-long same observing Avon (from 1 June 2019 31 May 2020) compare results nearby river-gauge measurements. Given high correlation (Pearson's coefficient >0.94) between these measurements, clear automation could allow straightforward, inexpensive events, especially at ungauged locations.