Abstract
In seasonal weather forecasting, the exact location within a study area determines the relationship between local predicted variables and global predictors. Both dynamic models and machine-learning approaches that define models for single grid points, can determine these relationships with high spatial granularity, but at high computational cost. To avoid the latter, clustering of predicted variables is often used in machine-learning approaches, which however sacrifices geographical resolution. In this paper, we present a machine-learning approach that is a hybrid between grid-point and cluster-based approaches (finres_S2S). This approach preserves geographical resolution, but at low computational cost, and is tested for monthly two-meter temperature and precipitation in Tanzania, and for a lead time of up to 6 months. The finres_S2S approach has a number of advantages over both the cluster and point approaches, including that of perform well compared to a commonly used forecasting dataset from a dynamic model. We find that the dominant predictors for the application area are associated with the El Nino Southern Oscillation, the Madden-Julian-Oscillation and the Indian Ocean Dipole.
© 2024 American Meteorological Society. This is an Author Accepted Manuscript distributed under the terms of the default AMS reuse license. For information regarding reuse and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).
