Multimodel classification and regression technique for the statistical downscaling of temperature

Human activity has increased the amount of carbon dioxide and other greenhouse gases emitted into the atmosphere, causing climate change. As a result, rising temperatures have wide-ranging consequences on water management. This study proposes downscaling daily temperature based on a modified Classification and Regression Technique with an ensemble machine learning (EML) approach at the Woodstock station in the Upper Thames River basin. The GCM Canadian Earth System Model (CanESM5) from Coupled Model Intercomparison Project-6 is used. The CanESM5 model simulated variables are used as predictors and observed baseline daily temperature as predictands. The Regression-based single machine learning (Support Vector, Tree-based and Gaussian Process Regression) and EML based statistical downscaling are applied and compared. The variable temperature states are determined using Gaussian Mixture Model clustering, and the Light Gradient Boosting Model (LightGBM) is used to classify future temperature states. Results showed that applying the EML boosted the performance by 2–25% compared to single models. The temperature states for the two projected climate scenarios (SSP126 and SSP585) were simulated by selected best-performing single and EML model combinations for the near (2026–2050) and far future (2076–2100). The findings demonstrate that the future projected temperatures may rise 1–3 °C for both scenarios and are less volatile than the observed baseline temperature. Overall, the study indicates that the ensemble approach-based downscaling combining several single models have considerably improved the performance and was more reliable.