Generating High-Resolution Climate Change Projections Using Super-Resolution Convolutional LSTM Neural Networks

Abstract

Generating projections of climate change through extreme indices such as precipitation and temperature is crucial to evaluate their potential impacts on critical infrastructures, human health, and natural systems. However, current Earth System Models (ESMs) run at spatial resolutions of hundreds of kilometers which is too coarse to analyze localized impacts. To tackle this issue, statistical downscaling is a widely employed technique that uses historical climate observations to learn a coarse-resolution to fine-resolution mapping. Traditional statistical methods are inefficient in downscaling precipitation data and vary significantly in terms of accuracy and reliability since local climate variables such as precipitation are dependent on non-linear and complex spatio-temporal processes. To capture both spatial and temporal variabilities, we develop a Super-Resolution based Convolutional Long Short Term Memory Neural Network and test the robustness and predictability of this model on monthly precipitation data in China. We integrate original climate data from an ESM and perform downscaling on precipitation at $(1.25^{\circ}\times 0.9^{\circ})$ to $(0.25^{\circ}\times 0.25^{\circ})$. Experimental data indicates that our Convolutional LSTM model performs the best compared to existing methods in terms of mean squared error, relative bias, and correlation coefficient.