Application of machine learning in operational flood forecasting and mapping

Abstract

Considering the computational effort and expertise required to simulate 2D hydrodynamic models, it is widely understood that it is practically impossible to run these types of models during a real-time flood event. To allow for real-time flood forecasting and mapping, an automated, computationally efficient and robust data driven modelling engine - as an alternative to the traditional 2D hydraulic models - has been proposed. The concept of computationally efficient model relies heavily on replacing time consuming 2D hydrodynamic software packages with a simplified model structure that is fast, reliable and can robustly retains sufficient accuracy for applications in real-time flood forecasting, mapping and sequential updating. This thesis presents a novel data-driven modelling framework that uses rainfall data from meteorological stations to forecast flood inundation maps. The proposed framework takes advantage of the highly efficient machine learning (ML) algorithms and also utilities the state-of-the-art hydraulic models as a system component. The aim of this research has been to develop an integrated system, where a data-driven rainfall-streamflow forecasting model sets up the upstream boundary conditions for the machine learning based classifiers, which then maps out multi-step ahead flood extents during an extreme flood event. To achieve the aim and objectives of this research, firstly, a comprehensive investigation was undertaken to search for a robust ML-based multi-step ahead rainfall-streamflow forecasting model. Three potential models were tested (Support Vector Regression (SVR), Deep Belief Network (DBN) and Wavelet decomposed Artificial Neural Network (WANN)). The analysis revealed that SVR-based models perform most efficiently in forecasting streamflow for shorter lead time. This study also tested the portability of model parameters and performance deterioration rates. Secondly, multiple ML-based models (SVR, Random Forest (RF) and Multi-layer Perceptron (MLP)) were deployed to simulate flood inundation extents. These models were trained and tested for two geomorphologically distinct case study areas. In the first case of study, of the models trained using the outputs from LISFLOOD-FP hydraulic model and upstream flow data for a large rural catchment (Niger Inland Delta, Mali). For the second case of study similar approach was adopted, though 2D Flood Modeller software package was used to generate target data for the machine learning algorithms and to model inundation extent for a semi-urban floodplain (Upton-Upon-Severn, UK). In both cases, machine learning algorithms performed comparatively in simulating seasonal and event based fluvial flooding. Finally, a framework was developed to generate flood extent maps from rainfall data using the knowledge learned from the case studies. The research activity focused on the town of Upton-Upon-Severn and the analysis time frame covers the flooding event of October-November 2000. RF-based models were trained to forecast the upstream boundary conditions, which were systematically fed into MLP-based classifiers. The classifiers detected states (wet/dry) of the randomly selected locations within a floodplain at every time step (e.g. one hour in this study). The forecasted states of the sampled locations were then spatially interpolated using regression kriging method to produce high resolution probabilistic inundation (9m) maps. Results show that the proposed data centric modelling engine can efficiently emulate the outcomes of the hydraulic model with considerably high accuracy, measured in terms of flood arrival time error, and classification accuracy during flood growing, peak, and receding periods. The key feature of the proposed modelling framework is that, it can substantially reduce computational time, i.e. ~14 seconds for generating flood maps for a flood plain of ~4 km2 at 9m spatial resolution (which is significantly low compared to a fully 2D hydrodynamic model run time).