Flood risk often needs to be estimated at locations where no gauge data is available, as most of the world’s catchments are ungauged. For example, in order to produce our Global Flood Event Set, we make use of around 150,000 catchments, of which only 18,000 are gauged. As a result, to predict river flow in an ungauged catchment, we need to be able to find a set of gauged catchments that are most similar to any chosen ungauged catchment of interest.
This process is known as regionalisation and is the key step that allows us to assess flood risk at any given point in the world. Different methods for regionalisation are available, ranging from simple empirical equations to more complex methods involving geostatistics or machine learning.
The choice of method depends partly on the hydrological model. When hydrological models have lots of parameters that need to be calibrated, it’s better to transfer the parameters as a set; some of the model parameters are highly correlated and a high value of one parameter can offset the low value of another parameter.
We use a method in which the streamflow of the ungauged catchment of interest is calculated based on parameters directly transferred from the most similar donor catchments.
JBA’s method
Our method for regionalisation involves the following steps, enabling us to effectively identify flood risk across ungauged areas:
Step 1: Extract catchment attributes
The first step is to delineate the catchment boundaries. It is an essential process that has been fine-tuned, first using a custom drainage algorithm applied on a high-speed computer, followed by a quality control step carried out by GIS specialists. The result is a set of many thousands of delineated catchments.
We then use global-scale datasets to extract attributes for each catchment: which climate zone it belongs to, what type of land cover and soil type is dominant, how much average rainfall each receives and each catchment’s average elevation.
Step 2: Set up IHACRES
The next step is to set up a rainfall-runoff model for each catchment where gauge data is available. We use the well-tested IHACRES model (Identification of unit Hydrographs and Component flows from Rainfall, Evaporation and Streamflow data).
To calibrate the IHACRES models, we need both the gauged stream flow data together with meteorological data describing precipitation and temperature. For this, we use a global gridded dataset which has daily time resolution.
An IHACRES model has 13 parameters (8 if no snowmelt is modelled) which need to be calibrated against observed data to transform the meteorological inputs into a river discharge series. An automatic optimisation is used, utilising a particle swarm approach, to estimate the best possible combination of parameter values by comparing the modelled to the observed discharge.
Our chosen measure of model fit is the Kling-Gupta Efficiency score (2) (KGE). A KGE score of zero indicates a model with a poor fit whereas a score of one would be a perfect fit. The average KGE score for our 18,000 IHACRES models is 0.6 which indicates a good overall fit to the observed data (see Figure 1 below).
Pictured Above: Performance of IHACRES models.
Step 3: Produce a Global Catchment Database
We have created the JBA Global Catchment Database (GCD) by combining catchment attributes and hydrological models. This database consists of attributes, calibrated IHACRES models and hydroclimatic data for more than 18,000 gauged catchments. We are continually extending this resource as new catchment data becomes available.
Step 4: Select similar donor catchments
We use a nearest neighbour clustering algorithm to select a set of the most similar donor catchments from our Global Catchment Database to the ungauged catchment of interest. This is done via a machine learning method designed to find similarities between objects in large data sets. Based on this comparison, a similarity score is calculated and used to select donor catchments.
Step 5: Produce estimates of streamflow
Finally, we run the donor catchments with the ungauged catchment of interest’s input data (precipitation and temperature) and use a weighted average of the output from these models to obtain the final estimate of streamflow at the ungauged catchment of interest.
Improving identification of flood risks
Our method for predicting streamflow in ungauged catchments is an essential tool for the Global Flood Event Sets, which allows us to achieve a high density of points and to better identify risk across the world.
We presented this method at the European Geoscience Union (EGU) 2018 General Assembly in Vienna and you can find more information here. If you want to learn more about our Global Catchment Database, Global Flood Event Sets or catastrophe models, get in touch today.
References
Jakeman, A.J., Littlewood, I.G., Whitehead, P.G., (1990). Computation of the instantaneous unit hydrograph and identifiable component flow with application to two small upland catchments.
Journal of Hydrology 117, 275–300.
Gupta, H. V., Kling, H., Yilmaz, K. K., & Martinez, G. F. (2009). Decomposition of the mean squared error and NSE performance criteria: Implications for improving hydrological modelling. Journal of Hydrology, 377(1), 80-91.
