Multivariate sensitivity analysis for a large-scale climate impact and adaptation model

Oluwole Oyebamiji, Christopher Nemeth, Paula A. Harrison, R. W. Dunford, George CojocaruJournal of the Royal Statistical Society Series C (Applied Statistics) 2023 · DOI

TL;DR

This paper introduces a novel, efficient Bayesian approach for global sensitivity analysis of complex, large-scale models with multiple outputs. By employing a multivariate sparse Gaussian process emulator and compactly supported correlation functions, the method significantly reduces computational burden, enabling accurate analysis of high-dimensional climate impact models.

What Problem Does This Paper Address?

Climate impact and adaptation models, like the IMPRESSIONS Integrated Assessment Platform (IAP2), are crucial for informing policy decisions. However, these models are often computationally intensive and involve a large number of input variables and outputs. Understanding how variations in input parameters influence model outputs (sensitivity analysis) is vital for uncertainty quantification and developing robust adaptation strategies. Traditional sensitivity analysis methods struggle with the high dimensionality and computational cost associated with such models, especially when dealing with multiple, correlated outputs. This paper aims to bridge this gap by providing an efficient and accurate method for global sensitivity analysis of these complex, multivariate systems.

Methods

The core of the proposed methodology lies in the Multivariate Sparse Gaussian Process (MSGP) emulator. This approach replaces the computationally expensive original model with a surrogate model. Key innovations include:

  • Multivariate Gaussian Process: This allows for modeling multiple correlated outputs simultaneously, unlike traditional methods that analyze each output independently.
  • Compactly Supported Correlation Functions: These functions are crucial for creating sparse covariance matrices. This sparsity drastically reduces computational requirements for matrix operations (inversion, determinant), making the analysis feasible for large datasets.
  • Adaptive Metropolis-within-Gibbs (AMWG) Algorithm: This MCMC technique is used for efficient posterior sampling of model parameters, speeding up convergence.
  • Parallel Implementation: The computational load is distributed across multiple processing nodes to further accelerate the analysis.
  • Vector Projection for Generalised Indices: The paper extends existing Sobol sensitivity indices to the multivariate case using vector projection, which accounts for correlations between outputs.

Datasets

The primary dataset used for evaluation is from the IMPRESSIONS Integrated Assessment Platform version 2 (IAP2). This platform models the impacts of climate and socio-economic change across various landscape sectors in Europe. The study uses a subset of 24 input parameters and 18 selected output variables from the IAP2 model. The data was aggregated to EU-wide and regional levels for analysis.

The paper also uses two test cases:

  1. Sobol g-function: A standard, non-monotonic, non-additive function used for testing global sensitivity analysis methods.
  2. Arctangent temporal function: A multivariate test case with 100 outputs and two independent input variables.

Key Insights and Ideas

  • Sparsity is Key for Scalability: The use of compactly supported correlation functions to induce sparsity in covariance matrices is a game-changer for handling large, high-dimensional problems. It allows for efficient computation without sacrificing accuracy.
  • Multivariate Analysis Matters: Analyzing multiple outputs simultaneously, rather than independently, can reveal more nuanced relationships and interactions between inputs and outputs, leading to a more comprehensive understanding.
  • Computational Efficiency is Achievable: The proposed MSGP approach, combined with parallelization and efficient sampling, demonstrates that rigorous sensitivity analysis of complex, computationally demanding models is indeed feasible.
  • Identifying Key Drivers: The analysis successfully identifies critical input variables that significantly influence various climate impact indicators, providing valuable insights for policy and adaptation planning.

Key Results

  • The MSGP emulator demonstrated good performance, with a high proportion of variance explained (P ranging from 0.34 to 0.96) and low root mean squared errors (ρ ranging from 0.21 to 0.83) across various sparsity levels.
  • The method significantly reduced computational time compared to direct model evaluation, enabling a more thorough sensitivity analysis.
  • For the IAP2 model, key drivers identified for various outputs include changes in agricultural yields, population, GDP, and food imports. Variables like water savings due to behavioral change and diffuse pollution reduction had negligible effects, suggesting they could be fixed to simplify models.
  • The analysis revealed that interactions between input factors play a significant role in determining model outputs, particularly for indicators like biodiversity vulnerability.

Limitations and Future Work

  • Spatial Aggregation: The study used spatial aggregation to reduce data dimensionality. The authors note that the effect of this aggregation on MSGP and sensitivity analysis outcomes is not fully clear.
  • Spatio-temporal Analysis: Performing a fully spatio-temporal sensitivity analysis for the IAP2 model was deemed practically infeasible.
  • Future Work: The authors suggest exploring the proposed method for spatiotemporal data without aggregation. They also propose investigating the implications of spatial aggregation further and demonstrating the method for spatiotemporal sensitivity analysis.

Summary updated

Abstract

Abstract We apply a new efficient methodology for Bayesian global sensitivity analysis for large-scale multivariate data. A multivariate Gaussian process is used as a surrogate model to replace the expensive computer model. To improve the computational efficiency and performance of the model, compactly supported correlation functions are used. The goal is to generate sparse matrices, which give crucial advantages when dealing with large data sets. The method was applied to multivariate data from the IMPRESSIONS Integrated Assessment Platform version 2. Our empirical results on Integrated Assessment Platform version 2 data show that the proposed methods are efficient and accurate for global sensitivity analysis of complex models.
BibTeX
@article{oyebamiji2023multivariate,
  author = {Oluwole Oyebamiji and Christopher Nemeth and Paula A. Harrison and R. W. Dunford and George Cojocaru},
  title = {{Multivariate sensitivity analysis for a large-scale climate impact and adaptation model}},
  journal = {Journal of the Royal Statistical Society Series C (Applied Statistics)},
  year = {2023},
  doi = {10.1093/jrsssc/qlad032},
}
Figure