Floods are natural disasters that can lead to loss of life or the destruction of infrastructure and the natural environment. A number of recent studies have reported an increase in floods globally, with some regions being affected more than others. Increases in floods have given impetus to the research of flood mapping and monitoring, which has resulted in the improvement and development of new technologies and techniques for modelling these events.
The most popular approaches to generating flood hazard maps rely on modelling fluvial hydraulics. Such models can simulate stream current according to one direction (1D) or two directions (2D) of motion. Some methods implement the full three-dimensional (3D) solutions. However, given their complexity, 3D hydraulic modelling is rarely performed for operational flood hazard mapping. Although hydraulic modelling is the best and most reliable approach for deriving detailed flood line maps, its use is limited due to the extensive input data required. This includes: (1) detailed topographic data (of river channels and catchments) to construct the model geometry; (2) bulk flow data to provide time-varying boundary conditions; (3) an estimate of the roughness of the river channel and surrounding areas (e.g. detailed land cover maps); and (4) a source of validation data (e.g. historical water level and gauge station records). The main obstacle in conducting hydraulic simulations is the collection of these input datasets, especially for large areas or in rural areas where streamflow records are rarely kept. Consequently, most areas exposed to the hydraulic hazard have never been modelled.
The lack of information on hydraulic simulations has stimulated the development of simplified methods that rely on topographic variables and basin morphologic features to provide an indication of flood hazard (Samela et al. 2015). These methods, often referred to as geomorphometry, have become more attractive and adept at estimating flood hazard, especially given recent increases in the availability of digital elevation models (DEMs). Recent examples of geomorphometry-based flood hazard assessments include Degiorgis et al. (2013); Samela et al. (2015); and Wang et al. (2015). Many of these studies have shown that DEM-based flood hazard modelling can provide results that are reasonably accurate (even comparable to hydraulic simulations), although some limitations still persist.
Machine learning (also known as artificial intelligence) algorithms use samples of known identity to classify instances of unknown identity and have the ability to incorporate many types of predictors (e.g. spectral, terrain, and even nominal data such as soil types) in the model building process. Although much work has been done on the use of machine learning and remotely sensed imagery for mapping land cover, relatively little research on the use of geomorphometry and machine learning for modelling flood risk has been published. Some notable exceptions include Wang et al. (2015) who used random forests for generating flood susceptibly maps in Malaysia; and Tehrany et al. (2014) who employed support vector machines for mapping flood-prone areas in Ethiopia. These ground-breaking studies have opened up a new, reliable and scientifically defendable approach to flood hazard index development.
Geosmart's flood hazard model
Geosmart, in collaboration with the Centre for Geographical Analysis (CGA) at Stellenbosch University, developed a flood hazard model by combining DEM derivatives and machine learning. Thousands of locations representing high and low flood risk (sourced from 2D hydrological model outputs) were collected throughout South Africa and used as input to random forest and decision tree algorithms. A large set of 24 derivatives were generated from the SUDEM5 and used as predictor variables. The derivatives included height above nearest drainage, distance to nearest drainage, slope gradient, terrain wetness index and local upslope area (flow accumulation). The result is a model with 84% accuracy (i.e. the model correctly predicted a specific location’s flood hazard 84 out of 100 times). This compares favourably with Tehrany et al. (2015) who achieved overall accuracies of between 63% and 84% when using support vector machines (SVM) for generating flood susceptibility maps in Malaysia.
The flood hazard model was implemented for the Western Cape Province of South Africa to produce a flood hazard map showing areas where flooding will likely occur during extreme rainfall events. It is currently being employed by the Western Cape Disaster Management Centre for flood mitigation planning and disaster management. Below are some examples of flood hazard maps in the Cape Town area.
The model has also been implemented elsewhere in South Africa and can be generated anywhere in the world. The accuracy of the model is dependent on the quality of the DEM used as input. Ideally, a high-resolution DEM, such as the SUDEM5 or DSMSA2, should be employed.