Deep Learning Applied to Climate Data Downscaling
Climate change and variability have tremendous impact on natural and human systems, and climate forecasts have the potential to influence policies and planning in a many fields such as agriculture, ecological preservation and resource management. Climate modeling is a complex function of natural and human factors, and phenomena of interest to climate science researchers are typically the products of extremely complex relationships between many variables. Global climate models (GCMS) are used to capture these complex physical processes and components of the Earth's climate system. These GCMS are computationally expensive, limiting their use to predict large scale weather systems at a coarse resolution (e.g., 100km x 100km grids). This coarse level of modeling is insufficient to make accurate regional weather predictions because global models do not account for the effect of geographical variations. Higher resolution regional estimates are obtained from the coarser GCMS through a process known as downscaling. This project investigates the use of state of the art machine learning algorithms to identify and model the nonlinear relationships between global models and regional estimates of weather parameters. Specifically, we explore the use of deep learning architectures to discern patterns in atmospheric changes given large sets of observed data. Some architectures we look at in particular are the Deep Belief Network (DBN) and the Convolutional Neural Network (CNN).