Exploring the potential of dimensionality-reduction techniques for impact-based flood mapping

Flood nowcasting at detailed, fine spatiotemporal scales is crucial for the deployment of reliable warning systems, especially in built-up environments where the majority of socio-economic activity is concentrated. These environments are also characterized by significant complexities that require sufficient detail, up to street level. The derivation of flood maps for early warning systems can be organized via three main pillars: a) nowcasting of rainfall at high spatiotemporal resolution, typically obtained from weather radars; b) deployment of physics-based mechanistic simulators, typically based on 2D Shallow Water Equations; c) utilization of High-Performance Computing (HPC) facilities to handle the associated significant computational effort and make practically feasible the computational process. However, even with such infrastructure, there are still limitations mainly arising from: a) model errors, either related with the epistemic or the deep uncertainty of real-world randomness; b) the required simulation time which can still be prohibitive for the development of operational nowcasting tools, especially for large case study areas. The first limitation is addressed through impact-based approaches, in which uncertainties are compensated through the translation of the natural variables derived by the model (i.e. water depths and flow velocities) into classified hazard zones. With respect to the second limitation, surrogate modelling, and particularly the relevant Machine Learning (ML) techniques, promises a potential remedy to the high computational burden, since it enables the development of fast emulators based on the results derived by the mechanistic (accurate, yet slow) simulators. However, the high spatiotemporal variability of flood-related variables, as exhibited in detailed scales increases significantly the dimensionality of the problem, hampering the application of such techniques in real-world operational conditions. To address this, herein we explore the use of dimensionality-reduction techniques such as, Single Value Decomposition (SVD) and Principal Component Analysis (PCA), which are widely employed, for similar purposes, in the domain of data science. The feasibility of such methods is investigated via impact-based flood maps derived by a detailed mechanistic simulator in real-world conditions.