The IRI probabilistic seasonal climate forecast is based on the U.S. National Oceanographic and Atmospheric Administration (NOAA)’s North American Multi-Model Ensemble Project (NMME). It includes the output from the ensemble seasonal prediction systems of NOAA’s National Centers for Environmental Prediction, Environment and Climate Change Canada, NOAA/Geophysical Fluid Dynamics Laboratory, NASA, NCAR and COLA/University of Miami. Ensemble mean seasonal forecast anomaly maps for each NMME model can be found on this NOAA CPC web page.
The output from each NMME model is re-calibrated prior to multi-model ensembling to form reliable probability forecasts. The ensemble mean precipitation (or temperature) from each individual NMME model, on an interpolated 1-degree latitude-longitude grid, is used in conjunction with extended logistic regression (ELR; Wilks 2009, Meteorol. Appl. DOI: 10.1002/met.134) to produce probability forecasts from each individual model; these probabilities are then averaged together with equal weight to create a multi-model forecast probability (Vigaud et al. 2017, DOI: 10.1175/MWR-D-17-0092.1). In the case of precipitation, the final probability maps are smoothed spatially with a 9×9 point Gaussian smoother.
The ELR is trained using hindcast ensemble-means and observed precipitation (or temperature) data over the 1982-2010 period. The observed precipitation data is from the CPC-CMAP-URD dataset, while the GHCN-CAMS dataset is used for observed temperatures; both datasets are given on the same 1-degree latitude-longitude grid. Observed tercile-category occurrences are used to train the ELR. All the training and forecast data can be found in the IRI Data Library: http://iridl.ldeo.columbia.edu/SOURCES/.Models/.NMME/
The probability of exceedances (or non-exceedances) for the flexible forecast format are computed analytically as a function of the threshold and the ELR coefficients at each gridpoint. This is done separately for each model, with the values then averaged across models with equal weight.
Verification are available in terms of cross-validated RPSS and reliability diagrams, constructed from NMME hindcasts. Verification of the real-time forecasts will also be provided through the IRI’s standard Real Time Seasonal Climate Forecast Verifications pages using a variety probabilistic skill metrics. Please note that the skill scores are based on forecasts made with IRI’s two-tier system at 2.5-degree spatial resolution prior to April 2017, and with the NMME-based system at 1-degree resolution from April 2017 onward.
About the dry mask: To address the discontinuity in observed precipitation between 0 and non-zero seasonal cumulated amounts, and the fact that GCMs tend to have very few non-rainy days compared to observation due to their tendency to simulate too many rainfall events with intensities overall too low (Goddard et al. 2001), ELR forecast are only produced when at least 10% of the training sample are non-zero.
Recommended reference for citation:
Kirtman, B., D. Min, J. Infanti, J. Kinter, D. Paolino, Q. Zhang, H. van den Dool, S. Saha, M. Mendez, E. Becker, P. Peng, P. Tripp, J. Huang, D. DeWitt, M. Tippett, A. Barnston, S. Li, A. Rosati, S. Schubert, M. Rienecker, M. Suarez, Z. Li, J. Marshak, Y. Lim, J. Tribbia, K. Pegion, W. Merryfield, B. Denis, and E. Wood, 2014: The North American Multimodel Ensemble: Phase-1 Seasonal-to-Interannual Prediction; Phase-2 toward Developing Intraseasonal Prediction. Bull. Amer. Meteor. Soc., 95, 585–601, doi: 10.1175/BAMS-D-12-00050.1.
The specific NMME models used are as follows. Please see http://iridl.ldeo.columbia.edu/SOURCES/.Models/.NMME/ for details.