Predicting precipitation using dynamic distributed lag models in arid and sub-humid regions of South Africa

Abstract

Ocean characteristics have contributed to a series of unusual rainfall patterns and floods, leading to severe land degradation, loss of life and infrastructure in various regions. Modelling and prediction of precipitation using in-situ data and oceanographic variables is possible. There are limited studies to substantiate this approach in less-developed countries. This study aims to model and predict precipitation in the arid, semi-arid and sub-humid regions of South Africa using dynamic linear regression (DLR) models, with sea surface temperature (SST) anomalies, evaporation-precipitation differences, longwave radiation (lwRad), net surface heat flux and relative humidity as input variables. The prediction accuracy of the autoregressive integrated moving average model with extra data (ARIMAX) and dynamic distributed lag (DDL) models was compared on the mean monthly rainfall data for the period 2008 to 2022. The results highlighted that the DDL models predict better than the other ARIMAX models, with SST anomalies and lwRad having a significant contribution (p-values < 0.05). These models had the smallest root mean squared error (RMSE) values for the arid (8.27 mm), semi-arid (19.15 mm) and the sub-humid (26.77 mm) regions, indicating that DDL models are suitable tools for the prediction of precipitation in these regions. However, additional oceanographic predictors such as sea surface salinity, ocean heat content, and upper-ocean current patterns may further enhance precipitation prediction accuracy, particularly in regions with strong ocean-atmosphere coupling, such as coastal or monsoon-influenced areas.

HIGHLIGHTS • Rainfall distribution shows clear gradients across the arid, semi-arid and sub-humid zones. • Semi-arid and sub-humid regions share seasonal patterns with occasional heavy rainfall. • Oceanographic predictors, especially sea surface temperature (SST) anomalies and longwave radiation, drive rainfall variability. • Dynamic regression models improved rainfall prediction in multiple climatic regions. • Rainfall seasonal forecast will complement the decision support system for climatological, agricultural and hydrological management across the Southern African region.