Abstract:
|
Climate model integration combines information across multi-model ensembles into a single climate system prediction. However, existing integration approaches based on model averaging can dilute fine-scale spatial information and incur bias from rescaling low-resolution climate models. Instead, we propose a statistical approach, called NN-GPR, using Gaussian process regression (GPR) with a deep neural network based kernel. NN-GPR requires no assumptions about the relationships between models, no interpolation to a common grid, no stationarity assumptions, and automatically downscales as part of its prediction algorithm. We show that NN-GPR produces highly skillful surface temperature and precipitation forecasts by preserving geospatial signals at multiple scales and capturing inter-annual variability. NN-GPR tends to have lower test set mean squared errors (MSE), lower continuous ranked probability scores (CRPS), and higher structural similarity index measures (SSIM) than global models, model averages, regression predictions, and regional climate models. Furthermore, our predictions show high accuracy and uncertainty quantification skill in regions with high variability.
|