Downscaling for Climate Data in Indonesia Using Image-to-Image Translation Approach

An accurate climate prediction is crucial for policymakers to anticipate the impact of climate change on various sectors. The Global Climate Model (GCM) is a primary tool to predict global climate condition in the future. Unfortunately, GCM can only define climate condition over a wide area. To get data in a more specific region, we commonly use a downscaling method to transform the coarse GCM data to a finer resolution. Conventionally, the Regional Climate Model (RCM) is used for this purpose, but it is less accurate for a complex topography domain and inefficient computationally. In recent years, many studies have shown that deep learning has become a powerful solution to solve image-to-image translation problems. Meanwhile, the climate data used for downscaling can be represented in a spatial two-dimensional form, similar to 2- $\mathbf{D}$ images. This similarity enabled us to use a deep learning technique for downscaling of climate data. In our work, we attempted to use a deep architecture intended for image-to-image translation, $\mathbf{Pix}2\mathbf{Pix}$ for this purpose. We implemented downscaling of two variables of regional climate data in Indonesia, i.e. surface temperature and precipitation, using topography and five climate data variables produced by a GCM as the input. The five input variables were specific humidity, surface air pressure, air temperature, eastward $(u)$ wind, and northward (v) wind. From our experiments using 100 training data and 100 testing data, we obtain NRMSE and mean SSIM of 0.0038 and 0.75 for surface temperature and 1.20 and 0.29 for precipitation.