Deep Learning Methods for Flood Mapping: A Review of Existing Applications and Future Research Directions

Abstract

Abstract. Deep Learning techniques have been increasingly used in flood risk management to overcome the limitations of accurate, yet slow, numerical models, and to improve the results of traditional methods for flood mapping. In this paper, we review 45 recent publications to outline the state-of-the-art of the field, identify knowledge gaps, and propose future research directions. The review focuses on the type of deep learning models used for various flood mapping applications, the flood types considered, the spatial scale of the studied events, and the data used for model development. The results show that models based on convolutional layers are usually more accurate as they leverage inductive biases to better process the spatial characteristics of the flooding events. Traditional models based on fully-connected layers, instead, provide accurate results when coupled with other statistical models. Deep learning models showed increased accuracy when compared to traditional approaches and increased speed when compared to numerical methods. While there exist several applications in flood susceptibility, inundation, and hazard mapping, more work is needed to understand how deep learning can assist real-time flood warning during an emergency, and how it can be employed to estimate flood risk. A major challenge lies in developing deep learning models that can generalize to unseen case studies and sites. Furthermore, all reviewed models and their outputs, are deterministic, with limited considerations for uncertainties in outcomes and probabilistic predictions. The authors argue that these identified gaps can be addressed by exploiting recent fundamental advancements in deep learning or by taking inspiration from developments in other applied areas. Models based on graph neural networks and neural operators can work with arbitrarily structured data and thus should be capable of generalizing across different case studies and could account for complex interactions with the natural and built environment. Neural operators can also speed up numerical models while preserving the underlying physical equations and could thus be used for reliable real-time warning. Similarly, probabilistic models can be built by resorting to Deep Gaussian Processes.

Transferability of data-driven models to predict urban pluvial flood water depth in Berlin, Germany

“…Both U-Net and RF models were skillful in predicting water depth within the training domain (minimum NSE=0.6). Contrary to the hypothesis that deep learning algorithms were superior to shallow machine learning algorithms (Bentivoglio et al, 2022), the results suggested that the RF models outperformed the U-Net models for predictions within the training domain. However, we found that the high performance of RF models was largely owed to overfitting: outside of the training domains, RF models exhibited a substantial performance loss for all considered metrics (NSE, RMSE, and CSI).…”

“…In contrast to U-Net models, TWI was not among the most important predictive features for the RF models. The estimated best predictive features from the U-Net and RF models were not the same but the results agree with the findings in the literature that TWI (Jalayer et al, 2014;Seleem et al, 2021;Bentivoglio et al, 2022),SDepth (Zhang and Pan, 2014;Seleem et al, 2021) and altitude (Zhang and Pan, 2014;Seleem et al, 2021Seleem et al, , 2022 are indicators for urban flood-prone areas.…”

“…However, shallow machine learning algorithms have not been systematically challenged in terms of transferability for urban flood modelling. A data-driven model that generalizes outside the training domain is still a major challenge in literature (Bentivoglio et al, 2022). While previous studies tried to examine the transferability of CNN in space to predict flood water depth under certain limitations (Löwe et al, 2021;Guo et al, 2022) and use transfer learning techniques to improve the CNN performance outside the training domain to map flood susceptibility (Zhao et al, 2021), such efforts have been examined neither for RF models nor for surrogates of physical numerical 2D hydrodynamic models.…”

The future of coastal monitoring through satellite remote sensing

“…Image segmentation, whereby each pixel is classified into predetermined classes (e.g., 'sand,' 'water,' 'vegetation'; see Figure 3), facilitates spatially explicit mapping of coastal environments. Image segmentation methods have been useful for estimating landcover types (Buscombe and Ritchie, 2018), backshore morphology (Mao et al, 2022), roughness for flood models (Sherwood et al, 2021), mapping flood extents (Erdem et al, 2021, Bentivoglio et al, 2021, flood damage (Adriano et al, 2021), and economic impacts (Donaldson and Storeygard, 2016). Segmentation and classification methods can also be slightly modified into so-called "image regression" methods, which provide continuous estimates of quantities of interest (e.g., wave height - Buscombe et al, 2020) or water depth (Collins et al, 2020).…”

Flood Modeling and Prediction Using Earth Observation Data

“…This success can be largely attributed to the generalizing power of CNNs toward spatiotemporal patterns in the data, specifically including contextual signatures, closely mimics human interpretation given enough training data. CNNs have already been successfully applied to various problems in the earth observation domain, including water and flood mapping using either optical or radar data (e.g., Bentivoglio et al 2021;Wieland and Martinis 2019;Li et al 2019a, b;Bonafilia et al 2020;Nemni et al 2020;Katiyar et al 2021;Bai et al 2021;Helleis et al 2022).…”