Spatiotemporal Gaussian Process Regression Advances Early Detection of Precipitation Emergence

Detecting long-term precipitation trends is difficult because internal variability obscures externally forced signals. Conventional methods, such as multimodel means (MMM), often suppress regional signals and inflate uncertainty. We develop Gaussian Process Regression (GPR) models that combine observations with Coupled Model Intercomparison Project Phase 6 (CMIP6)-based priors to predict long-term precipitation anomalies. We test three variants: a univariate GPR (UGPR) using only temporal covariances, a circulation-aware GPR (CGPR) incorporating circulation variability, and a spatio-temporal GPR (SGPR) that incorporates spatial and temporal covariances. Using leave-one-out cross-validation across 33 CMIP6 models, we compare the predictive skill, uncertainty reduction, and emergence detection of these methods. All three GPR frameworks substantially outperform the CMIP6 MMM, reducing global predictive errors by approximately 60% and narrowing uncertainty by about 12%, with broadly comparable global-mean performance among the GPR methods. Differences between GPR variants emerge primarily at regional scales. In particular, SGPR advances the detection of forced precipitation changes by up to 15 years relative to UGPR in regions including the Middle East, Northwestern Africa, Eastern Europe, Japan, and parts of South America, while CGPR yields more limited and regionally heterogeneous improvements. These results indicate that while GPR provides robust improvements, explicitly accounting for spatiotemporal covariances yields additional regional benefits for early detection of precipitation change. Our findings highlight the importance of regional covariance structure for early detection and underscore the value of flexible Bayesian frameworks for dynamically integrating observations into climate change detection and prediction.