Comparison of precipitation datasets for representing flood characteristics in Malaysia

Reliable assessment of flood risk is very important for mitigating the disastrous impacts of floods. Since extreme precipitation is the most common cause of floods, accurate spatio-temporal precipitation data is crucial for flood risk assessment. Limited availability of gauge observations makes flood risk assessment challenging in Southeast Asian countries like Malaysia. In such cases, various gridded precipitation datasets developed using data sources such as satellite, reanalysis and gauge observations are of vital importance, however, the differences in the data sources and methods used to derive these datasets lead to significant uncertainty regarding the choice of dataset for a specific purpose. For flood risk quantification over a region where rainfall and streamflow data exhibit significant spatial dependence, it is important to ensure that the use of the chosen dataset results in an adequate representation of flood characteristics observed in the region. This is an important consideration in the development of flood catastrophe models widely used to quantify flood risk in terms of monetary losses in the insurance and reinsurance industry. 

At Impact Forecasting, Aon’s catastrophe model development team, the key undertaking in this study is to identify a suitable gridded daily precipitation dataset for modelling flood risk in the Southeast Asian region using Malaysia as a case study. Comparisons are made among six datasets (namely IMERG, CHIPRS, ERA5, ERA5-Land, CHELSA and APHRODITE) regarding their representation of the characteristics of historical flood events in Malaysia. While pluvial flood events are directly determined by the precipitation datasets, streamflow data is needed to represent fluvial flood events. However, observed streamflow data is available only at a few locations in Malaysia. In such situations, rainfall-runoff models can be forced with precipitation data to generate simulations of streamflow. For this purpose, we use the Impact Forecasting rainfall–runoff (IFRR) model, a spatially distributed (gridded) adaptation of the HBV model to generate daily streamflow simulations at 10kmx10km grids in Malaysia.  

We first compare the general characteristics of the precipitation datasets such as the total accumulated rainfall, number of wet days, length of wet spells and spatial correlation. The accuracy of the daily streamflow simulations at locations where observed streamflow data is available is evaluated using the Kling-Gupta Efficiency (KGE). Next, we apply an innovative clustering-based method to extract pluvial and fluvial flood events from the precipitation and simulated streamflow data respectively, and then merge dependent events. The resulting sets of flood events derived from each dataset are compared in terms of characteristics such as frequency, severity, duration, and spatial extent. The ability of the datasets to represent some of the severe flood events are evaluated using available data.