Decadal Climate Prediction (DCP) Models Skill Over Southeast Asia

Abstract

Decadal Climate Predictions (DCPs) have emerged as a crucial component of the latest Coupled Model Intercomparison Project (CMIP6), offering valuable insights into climate variability over 10-year periods. These predictions are particularly significant for understanding Pacific Ocean variability and its influence on Southeast Asia climate patterns. This study evaluates the predictive capabilities of six General Circulation Model (GCM) DCPs through an Ensemble model approach, focusing on surface temperature and precipitation predictions across Southeast Asia. The evaluation utilizes the dcpp-A hindcast product and employs multiple assessment metrics, including Anomaly Correlation Coefficient (ACC), Root Mean Square Error (RMSE), Normalized Standard Deviation (NSD), and Taylor Diagram Skill Score (TDSS) to conclude the model skill objectively. The analysis encompasses 51 hindcast datasets from 1960 to 2010, with ERA5 reanalysis data serving as the reference benchmark. The evaluation was conducted for single-year evaluation, Lead Year 2 (LY2) to Lead Year 9 (LY9). Multi-year evaluation named FHDecad (average of LY2-LY5), SHDecad (average of LY6-LY9), and Decad (average of LY2-LY9). Surface temperature predictions show robust temporal consistency with strong ACC values and minimal RMSE across lead years, primarily driven by anthropogenic factors. While spatial distribution is generally uniform across Southeast Asia, the northern mainland subregion (MLSEA) exhibits lower performance with reduced ACC values and elevated RMSE. The maritime continent subregion (MCSEA) presents a complex picture - stronger correlation coefficients but high RMSE and underestimated variability. These precipitation prediction challenges likely stem from varying model capabilities in capturing regional climate dynamics and internal variability patterns, which significantly influence rainfall compared to temperature. Precipitation predictions vary considerably across temporal and spatial domains, with peak performance during Lead Year 2 (particularly in winter months) and correlation skill in LY8-LY9. The decadal segmentation approach improves hindcast skill, with FHDecad consistently outperforming SHDecad according to TDSS metrics.