84th AMS Annual Meeting

Thursday, 15 January 2004: 11:45 AM
A non-linear fuzzy set technique for combining precipitation forecasts
Room 605/606
Brian P. Mackey, Florida State University, Tallahassee, FL
Poster PDF (188.5 kB)
This study, based on fuzzy set theory, describes a non-linear classifier system that optimally weights member model precipitation forecasts to produce a more skillful end product. The focus is on the one to five day prediction of extreme precipitation events across the globe in real time. In addition, a regional case study of the Limpopo River basin floods of 2000 is also carried out with attention given to both meteorological and hydrological predictions.

As an extension of the multiple linear regression superensemble technique, this fuzzy set method, in essence, acts as a smooth switching regression model. Within the technique, precipitation regimes are classified as either "low" or "high" for each of the five member NWP models at each of the 46 464 global grid points from 55 degrees S to 55 degrees N. By utilizing a training period, a regression model is then formulated for each regime. Then, for each time step within the training, a set of model weights are determined by calculating how effectively the different member model forecasts satisfy a given regime. These weighting coefficients are then applied to an independent forecast data subset during the forecast phase.

The benchmark observed analysis used in training and verifying the global precipitation forecasts consists of both TRMM (Tropical Rainfall Measuring Mission) and SSM/I (Special Sensor Microwave Imager) data. The multi-model five-day global forecasts of precipitation from five different cooperating weather services are initialized at 1200 UTC each day. Those centers and models involved in this collaboration are as follows: (1) Bureau of Meteorology Research Center (BMRC) Atmospheric Model (BAM), Australia; (2) Japan Meteorological Agency (JMA) Global Spectral Model (GSM); (3) National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS); (4) Naval Research Laboratory (NRL) Navy Operational Global Atmospheric Prediction System (NOGAPS); and (5) Recherche en Prévision Numérique (RPN) Global Environmental Multiscale (GEM) Model, Canada.

Testing this methodology during the month of February 2003 for the global domain described above demonstrates the superiority of the fuzzy set method over the five member models, thereby making it a viable choice for combination forecasting. Equitable threat scores reveal that the greatest improvement is for the higher thresholds (25, 35, 50, and 75 mm). The bias scores are close to unity except for the lower thresholds where overestimation occurs (2 and 5 mm).

Furthermore, a regional-scale case study is performed over southeastern Africa within the Limpopo River basin. Here, in February 2000, Tropical Cyclone Leon-Eline devastated the region with very severe inland flooding. Rainfall accumulations of over 500 mm in four days led to a significant rise in the Limpopo River, especially downstream over Mozambique. This flood catastrophe -- the worst in 50 years -- severely damaged Mozambique's infrastructure and displaced some 2 000 000 people.

In order to investigate the predictability of streamflow response to this extreme precipitation event, the rainfall forecasts described above are used as inputs to a spatially-distributed, physically-based hydrology model. Together with topographic, soil, and vegetation data as well as atmospheric forcing, the terrestrial hydrology model is able to simulate the effects of spatial heterogeneities via the use of physically significant model equations and parameters. This model consists of three inter-related modules: a land surface model (LSM) where the water and energy budgets are computed; a one-dimensional surface flow routing model (SRFM); and a two-dimensional lateral subsurface flow routing model (LSFRM).

Preliminary results show that the hydrology simulation captured well the effect of the heavy rain over the basin. At the outlet, the peak streamflow discharge is very close to what was actually observed.

Supplementary URL: