Spatiotemporal Density Correction of Multivariate Global Climate Model Projections using Deep Learning (2411.18799v2)

Abstract: Global Climate Models (GCMs) are numerical models that simulate complex physical processes within the Earth's climate system and are essential for understanding and predicting climate change. However, GCMs suffer from systemic biases due to simplifications made to the underlying physical processes. GCM output therefore needs to be bias corrected before it can be used for future climate projections. Most common bias correction methods, however, cannot preserve spatial, temporal, or inter-variable dependencies. We propose a new semi-parametric conditional density estimation (SPCDE) for density correction of the joint distribution of daily precipitation and maximum temperature data obtained from gridded GCM spatial fields. The Vecchia approximation is employed to preserve dependencies in the observed field during the density correction process, which is carried out using semi-parametric quantile regression. The ability to calibrate joint distributions of GCM projections has potential advantages not only in estimating extremes, but also in better estimating compound hazards, like heat waves and drought, under potential climate change. Illustration on historical data from 1951-2014 over two 5x5 spatial grids in the US indicate that SPCDE can preserve key marginal and joint distribution properties of precipitation and maximum temperature, and predictions obtained using SPCDE are better calibrated compared to predictions using asynchronous quantile mapping and canonical correlation analysis, two commonly used bias correction approaches.

• The paper introduces a deep learning method using conditional density estimation and Vecchia approximation to correct biases in multivariate global climate model data.
• The method successfully aligns calibrated climate data with observations, significantly improving the capture of extreme values and variable dependencies compared to traditional techniques.
• The approach efficiently preserves crucial spatio-temporal and cross-variable dependencies, demonstrating superior numerical performance for various statistical properties important for downstream applications.

Conditional Density Estimation with Neural Networks for Bias Correction of Multivariate Climate Model Data

The paper discusses a novel approach to the bias correction of multivariate climate data obtained from Global Climate Models (GCMs). GCMs are instrumental for simulating and predicting climate dynamics; however, they often suffer from biases due to simplifications in modeling complex physical processes. Traditional bias correction methods frequently fall short in preserving vital dependencies across temporal, spatial, and multiple climate variables. This paper introduces a methodology utilizing conditional density estimation to enhance bias correction specifically for daily precipitation and maximum temperature datasets.

A distinctive feature of this method is the integration of the Vecchia approximation and semi-parametric quantile regression (SPQR). The Vecchia approximation allows the model to retain spatial, temporal, and inter-variable dependencies by expressing multivariate responses as products of univariate conditional distributions. SPQR leverages neural networks to approximate these conditional densities efficiently.

Methodology

The researchers apply their SPCDE method to historical climate data from two distinct regions in the United States over a substantial period (1951-2014). For practical implementation, they calibrate bias using data from 1951-2000 and validate the results using data from 2001-2014. This meticulous approach ensures that the model captures prevailing trends and dependencies between variables like maximum temperature (TMAX) and precipitation (PRCP).

Key Findings

• Marginal and Joint Distributions: The method successfully aligns the calibrated climate model data with observed data, significantly improving over asynchronous quantile mapping and canonical correlation analysis (CCA), particularly in capturing extreme values and dependencies.
• Spatio-Temporal Dependence Preservation: The approach efficiently retains spatio-temporal and cross-variable dependencies, critical for applications in hydrology and other sectors reliant on comprehensive climate data analyses.
• Numerical Performance: The method exhibits superior performance in terms of reduced bias and uncertainty in estimating various statistical properties such as the proportion of zero precipitation days, 0.95 quantiles, and auto and cross-correlations of TMAX and PRCP.

Implications and Future Directions

The implications of this paper are profound. By incorporating neural networks in bias correction, the method addresses limitations associated with traditional techniques while offering scalability. The Vecchia approximation ensures computational feasibility, an asset when dealing with significant climate datasets across vast geographical regions.

A speculative trajectory for future research involves enhancing this approach by integrating annual models that can handle higher data volumes, thereby improving model accuracy. Additionally, directly modeling the joint distribution of variables without decomposition could lead to further improvements in bias correction fidelity.

In conclusion, this research paves the way for more reliable climate projections by deftly combining modern statistical methodologies with principles of geostatistics and machine learning, setting a foundation for advancements in climate model bias correction strategies.