Increasing amounts of data, together with more computing power and better machine learning algorithms to analyse the data, are causing changes in almost every aspect of our lives. This trend is expected to continue as more data keep becoming available, computing power keeps improving and machine learning algorithms keep improving as well. Flood risk and impact assessments are also being influenced by this trend, particularly in areas such as the development of mitigation measures, emergency response preparation and flood recovery planning. Machine learning methods have the potential to improve accuracy as well as reduce calculating time and model development cost. It is expected that in the future more applications will become feasible and many process models and traditional observation methods will be replaced by machine learning. Examples of this include the use of machine learning on remote sensing data to estimate exposure and on social media data to improve flood response. Some improvements may require new data collection efforts, such as for the modelling of flood damages or defence failures. In other components, machine learning may not always be suitable or should be applied complementary to process models, for example in hydrodynamic applications. Overall, machine learning is likely to drastically improve future flood risk and impact assessments, but issues such as applicability, bias and ethics must be considered carefully to avoid misuse. This paper presents some of the current developments on the application of machine learning in this field and highlights some key needs and challenges.

Classical approaches to flood hazard are obtained by the concatenation of a recurrence model for the events (i.e., an extreme river discharge) and an inundation model that propagates the discharge into a flood extent. The classical approach, however, uses “best‐fit” models that do not include uncertainty from incomplete knowledge or limited data availability. The inclusion of these, so called epistemic uncertainties, can significantly impact flood hazard estimates and the corresponding decision‐making process. We propose a simulation approach to robustly account for uncertainty in model's parameters, while developing a useful probabilistic output of flood hazard for further risk assessments via the Bayesian predictive posterior distribution of water depths. A Peaks‐Over‐Threshold Bayesian analysis is performed for future events simulation, and a pseudo‐likelihood probabilistic approach for the calibration of the inundation model is used to compute uncertain water depths. The annual probability averaged over all possible models’ parameters is used to develop hazard maps that account for epistemic uncertainties. Results are compared to traditional hazard maps, showing that not including epistemic uncertainties can underestimate the hazard and lead to non‐conservative designs, and that this trend increases with return period. Results also show that the influence of the uncertainty in the future occurrence of discharge events is predominant over the inundation simulator uncertainties for the case study.

This paper presents the advantages and opportunities for rapid preliminary intervention screening to enhance inclusion of green infrastructures in regional scale stormwater management. Stormwater flooding is widely recognised as a significant and worsening natural hazard across the globe; however, current management approaches aimed at the site scale do not adequately explore opportunities for integrated management at the regional scale at which decisions are made. This research addresses this gap through supporting the development of stormwater management strategies, including green infrastructure, at a regional scale. This is achieved through upscaling a modelling approach using a spatially explicit inundation model (CADDIES) coupled with an economic model of inundation loss (OpenProFIA) to support widescale evaluation of green infrastructure during the informative early-stage development of stormwater management strategies. This novel regional scale approach is demonstrated across a case study of the San Francisco Bay Area, spanning 8300 sq km. The main opportunity from this regional approach is to identify spatial and temporal trends which are used to inform regional planning and direct future detailed modelling efforts. The study highlights several limitations of the new method, suggesting it should be applied as part of a suite of landscape management approaches; however, highlights that it has the potential to complement existing stormwater management toolkits.

Computational simulators of complex physical processes, such as inundations, require a robust characterization of the uncertainties involved to be useful for flood hazard and risk analysis. While flood extent data, as obtained from synthetic aperture radar (SAR) imagery, have become widely available, no methodologies have been implemented that can consistently assimilate this information source into fully probabilistic estimations of the model parameters, model structural deficiencies, and model predictions. This paper proposes a fully Bayesian framework to calibrate a 2D physics-based inundation model using a single observation of flood extent, explicitly including uncertainty in the floodplain and channel roughness parameters, simulator structural deficiencies, and observation errors. The proposed approach is compared to the current state-of-practice generalized likelihood uncertainty estimation (GLUE) framework for calibration and with a simpler Bayesian model. We found that discrepancies between the computational simulator output and the flood extent observation are spatially correlated, and calibration models that do not account for this, such as GLUE, may consistently mispredict flooding over large regions. The added structural deficiency term succeeds in capturing and correcting for this spatial behavior, improving the rate of correctly predicted pixels. We also found that binary data do not have information on the magnitude of the observed process (e.g., flood depths), raising issues in the identifiability of the roughness parameters, and the additive terms of structural deficiency and observation errors. The proposed methodology, while computationally challenging, is proven to perform better than existing techniques. It also has the potential to consistently combine observed flood extent data with other data such as sensor information and crowdsourced data, something which is not currently possible using GLUE calibration framework.

Computational simulators of complex physical processes, such as inundations, require a robust characterization of the uncertainties involved to be useful for flood hazard and risk analysis. While flood extent data, as obtained from synthetic aperture radar (SAR) imagery, has become widely available, no methodologies have been implemented that can robustly assimilate this information source into fully probabilistic estimations of the model parameters, model structural deficiencies, and model predictions. This paper proposes a fully Bayesian framework to calibrate a 2D physics-based inundation model using a single observation of flood extent, explicitly including uncertainty in the floodplain and channel roughness parameters, simulator structural deficiencies, and observation errors. The proposed approach is compared to the current state-of-practice Generalized Likelihood Uncertainty Estimation (GLUE) framework for calibration and with a simpler Bayesian model. We found that discrepancies between the computational simulator output and the flood extent observation are spatially correlated, and calibration models that do not account for this, such as GLUE, might consistently mispredict flooding over large regions. The added structural deficiency term succeeds in capturing and correcting for this spatial behavior, improving the rate of correctly predicted pixels. We also found that binary data does not have information relative to the magnitude of the observed process (e.g. flood depths), raising issues in the identifiability of the roughness parameters, and the additive terms of structural deficiency and observation errors. The proposed methodology, while computationally challenging, is proven to perform better than existing techniques. It also has the potential to consistently combine observed flood extent data with other data such as sensor information and crowd-sourced data, something which is not currently possible using GLUE calibration framework.