A Machine Learning Based Multi-model ensemble Approach to reconstruct the historical monthly precipitation over Oklahoma using NOAA's SPEAR dataset

Abstract

General Circulation Models (GCMs) are important tools in simulating and projecting future precipitation at the decadal scale. However, it is inevitable that simulation and projection errors and uncertainty exist in GCMs, hindering their applications for regional water resources planning. Different post-processing tools are available to address the uncertainty issues associated with GCMs and to utilize these tools better for regional water resources planning. For example, a multi-model ensemble (MME) could reduce uncertainties from different GCMs and help reduce the model biases from a single model. In this study, we employed multiple Machine Learning algorithms (MLs) to combine ensemble members from NOAA’s Seamless System for Prediction and EArth System Research (SPEAR) to reconstruct historical monthly precipitation over Oklahoma during a study period (1981-2014). The employed MLs include Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Support Vector Machines (SVM), and Classification And Regression Trees (CART). The performances of the employed MLs are benchmarked with Simple Model Averaging (SMA), Bayesian Model Averaging (BMA), and Reliability Ensemble Averaging (REA). Our result echoes previous studies where the raw precipitation simulation from SPEAR presents significant simulation bias and marginal simulation skills. Different spatial and seasonal patterns of the simulation bias and skill are also observed over our study region. All the employed multi-model averaging techniques have delivered better performances than any single ensemble member from SPEAR. The employed MLs have outperformed SMA, BMA, and REA, which is evident from the reduction of bias and skill improvement. In general, this study highlights future applications of other data-driven techniques in post-processing the multi-model simulation from GCMs.