[Paper Review] Statistical Downscaling of Model Projections with Multivariate Basis Graphical Lasso

Ayesha Ekanayaka, Emily L. Kang+2 more|arXiv (Cornell University)|Jan 31, 2022

Climate variability and models4 citations

TL;DR

This paper proposes a novel statistical downscaling method using Multivariate Basis Graphical Lasso (BGL) to enhance spatial resolution of climate model projections, specifically downscaling CMIP6 sea surface temperature (SST) data from ~100 km to 1 km. The method improves predictive accuracy by 8% over state-of-the-art interpolation-based approaches and uniquely provides full probabilistic uncertainty estimates across space and time.

ABSTRACT

We describe an improved statistical downscaling method for Earth science applications using multivariate Basis Graphical Lasso (BGL). We demonstrate our method using a case study of sea surface temperature (SST) projections from CMIP6 Earth system models, which has direct applications for studies of multi-decadal projections of coral reef bleaching. We find that the BGL downscaling method is computationally tractable for large data sets, and that mean squared predictive error is roughly 8% lower than the current state-of-the-art interpolation-based statistical downscaling method. Finally, unlike most ofthe currently available methods, BGL downscaling produces uncertainty estimates. Our novel method can be applied to any model output variable for which corresponding higher-resolution observational data is available.

Motivation & Objective

• To address the resolution gap between coarse global climate model (GCM) outputs (~100 km) and the fine-scale spatial resolution needed for ecological impact studies.
• To develop a computationally efficient statistical downscaling method that captures complex spatio-temporal dependencies in climate data.
• To provide meaningful uncertainty estimates for downscaled projections, which are often missing in existing statistical downscaling techniques.
• To demonstrate the method’s effectiveness in downscaling sea surface temperature (SST) for coral reef bleaching risk assessment in the Great Barrier Reef region.
• To establish a framework applicable to any model output variable with corresponding high-resolution observational data.

Proposed method

• The method employs Multivariate Basis Graphical Lasso (BGL), a framework designed for high-dimensional, non-stationary multivariate spatial processes with complex inter-variable dependencies.
• BGL models the residual field between coarse GCM projections and high-resolution observations using a low-rank basis representation to reduce computational burden.
• It estimates a sparse precision matrix over spatial locations and variables, enabling efficient modeling of conditional independence structures and spatial dependence.
• The approach uses a penalized likelihood framework with a graphical lasso penalty to induce sparsity in the inverse covariance matrix, improving interpretability and stability.
• Uncertainty quantification is achieved through a full Bayesian hierarchical model, providing predictive distributions at each spatial location and time point.
• The method is validated using CMIP6 SST projections and high-resolution MUR SST observations, with comparisons to interpolation-based and laGP-based downscaling methods.

Experimental results

Research questions

• RQ1 Can a statistical downscaling method based on BGL outperform standard interpolation-based downscaling in terms of predictive accuracy for SST projections?
• RQ2 Does the BGL method effectively capture spatial and temporal dependencies in climate data, particularly in regions with complex coastlines?
• RQ3 Can the BGL framework provide reliable, spatially varying uncertainty estimates for downscaled climate variables, unlike most existing methods?
• RQ4 How does the BGL method compare to alternative probabilistic downscaling approaches, such as those based on local approximate Gaussian processes (laGP), in terms of mean squared error and structural similarity?
• RQ5 To what extent does the BGL method improve downscaling performance in ecologically sensitive regions like the Great Barrier Reef?

Key findings

• The BGL downscaling method reduces mean squared predictive error (MSE) by approximately 8% compared to the current state-of-the-art interpolation-based method.
• The BGL method achieves the lowest overall MSE across all seasons, with a 21.4% reduction in MSE compared to the GCM baseline and a 12.6% reduction compared to the standard interpolation method.
• The BGL method produces the highest structural similarity index (SSIM) values, with an overall average SSIM of 0.859, indicating superior preservation of spatial structure in downscaled SST maps.
• The method significantly reduces MSE along coastal regions, where interpolation-based methods often fail due to complex spatial gradients.
• Unlike standard downscaling, which assumes constant uncertainty, the BGL method provides time-varying, spatially heterogeneous uncertainty estimates, enabling more realistic probabilistic forecasts.
• The BGL method is computationally tractable for large-scale Earth system model data, making it suitable for global or regional climate applications.