Postprocessing Subseasonal Ensemble Rainfall Forecasts in the Pacific Islands Using Process-Informed and Statistical Techniques

Extreme weather events can have devastating impacts on communities. Subseasonal rainfall forecasts may alleviate these impacts by providing alerts multiple weeks ahead to inform planning for potential hazards like droughts and floods. In the Pacific region, subseasonal forecasts have the potential to perform well due to proximity to the El Niño Southern Oscillation (ENSO), but they also face special prediction challenges in part due to the small spatial scale of many islands. Several postprocessing techniques are available to bias correct and downscale raw 100km scale predictions from numerical weather ensembles, including both statistical and process-informed techniques. Process-informed subsampling, for instance, has been proposed as one potential method to enhance subseasonal rainfall forecast skill by selecting individual members from a large ensemble of dynamic forecasting model outputs based on their ability to represent potential predictors of rainfall (Kowal et al. 2024). This method has shown promise in sea surface temperature-driven regions like Central America, but it remains an open question how this approach compares with statistical options to postprocess large ensemble forecasts. Here we evaluate how three distinct postprocessing techniques compare within the Pacific region, using Fiji, Solomon Islands, and Kiribati as test cases. We test two statistical techniques implemented in XCast: a linear method that uses Canonical Correlation Analysis, and a nonlinear method that is an extended version of an extreme learning machine algorithm applied to probabilistic forecasts (EPOELM). We then compare these with process-informed subsampling to bias correct subseasonal rainfall forecasts one to four weeks ahead. Results showcase spatiotemporal variability in skill across the postprocessing techniques and highlight the role the seasonal evolution of ENSO plays in the relative performance of these techniques.