An End‐To‐End Flood Stage Prediction System Using Deep Neural Networks

Abstract

Floods are on the rise globally with the frequent record-breaking events occurring during the past few years in the US alone. These extreme events pose a considerable threat to human life and result in destructive damage to property, communities, and the built environment (e.g., Phillips et al., 2018). The south and the southeast US have experienced frequent storms with annually, on average, more than 85 named and unnamed thunderstorms (NWS, 2020). These events happened in quick succession (∼2 weeks apart) and produced catastrophic flooding in wide geographic areas (∼1,000 km swath) and within short timespans (less than a 48-hr period; Donratanapat et al., 2020). Successive flood events can even lead to higher costs in terms of repairing and rebuilding destroyed buildings and critical infrastructures (CIs) due to a lack of early warning systems (e.g., Donratanapat et al., 2020;Field et al., 2012;Hinkel et al., 2014). This necessitates the importance of detecting flood magnitudes ahead of the event to protect communities and CIs. The flood stage is the height of the water surface in a stream gaging station, not the height throughout the stream. A vast amount of research has been conducted to develop different tools and test their reliability in predicting near real-time flood stage estimation (Krzysztofowicz et al., 1994).

A Comprehensive Review of Methods for Hydrological Forecasting Based on Deep Learning

“…In general, the model built using a one-dimensional CNN does not fully exploit the advantages of the CNN's convolution. At this time, the integration of CNN and LSTM [43] techniques can bring out the advantages of each and improve the accuracy of the whole end-to-end model, which is valuable for engineering applications. Unlike CNNs, there is a notable abundance of research in the literature focusing on RNN usage in hydrological prediction.…”

“…Nonetheless, alternative architectures, such as hybrid CNN-LSTM algorithms, have demonstrated potential in elevating the precision of flood prediction outcomes [42]. Windheuser et al [43] proposed a fully automated end-to-end image detection system using the fusion of multiple deep neural networks, including the CNN and LSTM models, to predict flood levels from two USGS gauging stations, Columbus River and Sweetwater River, Georgia, USA. Their experimental results demonstrated that the proposed model predicted NSEs of historical water gauge height data of 85% for 6 h, 96% for 12 h, 96% for 24 h, and 95% for 48 h at Columbus station, and that the short-term NSEs were greater than 83% at Sweetwater Creek Station.…”

Bed Topography Inference from Velocity Field Using Deep Learning

“…Since our input/output is of the same size, we reworked the model by changing the convolution function parameters and using the model as a regression problem. The U-net has been used for segmentation [44][45][46], classification [47][48][49], and the regression problem [50][51][52].…”

Assessment of a Machine Learning Algorithm Using Web Images for Flood Detection and Water Level Estimates

“…For example, ref. [8] proposed the use of a fully automated end-to-end image detection system to predict flood stage data using deep neural networks across two US Geological Survey (USGS) gauging stations. The authors made use of a U-Net convolutional neural network (CNN) on top of a segmentation model for noise and feature reduction to detect the water levels.…”