Implementing a space-time rainfall model for the Sydney region
Rainfall is highly variable in space and time and is the driving process to the water dynamics of a catchment. Small and urban catchments alike demonstrate quick response times to rainfall input, hence it is important to accurately model rainfall at a high resolution in space and time. Whilst there are numerous models capable of this task at a single point, very few models are capable of the equivalent in both space and time.
This paper overviews current space-time rainfall models capable for use as a continuous input to catchment risk studies and investigates the Spatial Neyman-Scott Rectangular Pulse (SNSRP) model in detail. The SNSRP is a spatial extension of the Neyman-Scott Rectangular Pulse model at a single point. The SNSRP is therefore parsimonious and has the advantage of being calibrated solely to rain gauges. However, the idealised structure of the model, having rain-cells represented as simple cylindrical volumes of rain and without advection, imposes numerous limitations.
The SNSRP model has six parameters: storm arrival, cell arrival, cell radius, cell lifetime and two cell intensity parameters. The parameters are calibrated using least-squares fits to statistics that are based on hourly and daily rain-gauge data: mean, standard deviation, skewness, auto-correlation and cross-correlation. The model proceeds by simulating cells representing a cylindrical volume of rain that are clustered in time over the specified region.
The SNSRP model has had limited applications in recent literature, mostly in European catchments for simulating rainfall at a set of points. Consequently, this paper considers an Australian case study and implements a spatial regression of the scale parameter in order to simulate rainfall-fields over the entire region of interest. The interests of the paper are twofold: (i) to assess the SNSRP under an Australian climate and (ii) to consider the suitability of the simulated spatial images.
Firstly, the SNSRP is calibrated to a large network of daily and pluviograph gauges over metropolitan Sydney. The model is rigorously assessed against a range of spatial and temporal statistics from previous cast-studies, including a comparison of the SNSRP temporal statistics with the same statistics for the point Neyman-Scott model. Additionally, the calibration procedure and parameter correlation is discussed.
Secondly, the simulated images are contrasted with observed radar images and spatial statistics for the same Sydney region. Particular attention is given to the smoothness of the simulated spatial images, the spatial correlation and lack of modelled cell advection. Conclusions regarding the limitations of simulated images are however restricted to visual and statistical properties only. Further research is required to determine the impact of discrepancies between simulated and observed images on a flood or risk study.