Downscaling general circulation models for spatial data using interpolation and machine learning model

Abstract

Climate change is now a global problem which can impact on a local scale. It doesn’t only impact on pattern of climate in long term, but also has chained effect on human life and nature in many sectors, which could be endangers. To reduce the severity of the impact from the change, future climate prediction from General Circulation Models (GCMs) is essential. However, the prediction is in the coarse spatial scale and should not be used directly for local scale projection. So, downscaling, a process to increase spatial resolution, is required. Nevertheless, some area, may not be able to use the traditionally method according to insufficient of observed data, also Thailand. Hence, this study aims, to propose a high accuracy GCMs downscaling process under the limitation in quantity of the observed data of Thailand. To explore the proper one, six combination methods of three interpolation techniques- Inverse Distance Weighing (IDW), Triangular Interpolation Network (TIN) and Kriging and two machine learnings – Artificial Neural Networks (ANN) and Gradient Boosting Regression Tree (GBRT) are performed on downscaling output into 1 km spatial resolution. The result from each combination is evaluated for both accuracy and validity. From the evaluation, IDW-ANN shows the best performance with scores of all metrices below 0.1 for downscaling under the limitation of data quantity.