Machine learning framework for investigating biases in a regional climate model simulated surface soil moisture

Development of climate change mitigation and adaptation strategies, particularly for engineering systems, require information on the future evolution of various climatic variables/loads, including those related to soil moisture, which are mostly obtained from transient climate change simulations performed with high-resolution climate models. Developing credible high-resolution models, particularly in the identification of biases and their causalities for subsequent improvements, is computationally expensive. This study presents an efficient machine learning framework to assist in the causal analysis of biases in climate model simulated fields at significantly reduced computational costs. The framework consists of a two-step approach. The first step involves the development of a random forest (RF) powered emulator, trained on observed data of the climate variable of interest and related predictors. The second step involves, emulations of the climate variable of interest with the RF model, developed in step one, by replacing the observed predictors with those from the climate model. The assumption is that comparing these emulations with that of a reference emulation driven by all observed predictors can shed light on the contribution of respective predictor biases to the biases in the variable of interest. The proposed framework is applied to understand biases in the regional climate model GEM (Global Environmental Multiscale) simulated surface soil moisture (SSM), for the April-September period, over a domain covering part of north-east Canada. Two approaches to build random forest (RF) models are examined in step 1: a domain-based single model and a collection of grid cell-based models. In the absence of gridded observed predictors and target variable (i.e., SSM), those derived from a reanalysis product, ERA5, are used to train and validate the RF models. Given the superior performance of the grid-cell based models, compared to the domain-based one, the reference SSM required in step 2 is generated from these. Comparison of emulations with ERA5 predictors replaced with those from GEM with the reference emulation helps to quantify the contribution of predictor biases to SSM biases in GEM; Biases in water availability, relative humidity and 2-m temperature are mostly responsible for the biases in GEM simulated SSM, but vary in space, with bias contributions being important over regions where the respective predictor was identified to influence SSM based on a predictor importance analysis. The study thus demonstrates the ability of the proposed framework in narrowing down the sources of biases in climate model simulated field – with slightly reduced skill for heavily perturbed GEM simulation as the predictors in this case come from outside of the predictor space used to train the RF model, which can inform targeted climate model developments, thereby reducing computational costs

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.