Climate Data Downscaling Through Single Image Super-Resolution

The impact of climate change is comprehensive,and its negative impact is often much more concerned.Climate change,global temperature rising,has a great impact on our living life,such as economic system and infrastructure.Our vital natural resources,including crops,freshwater and coastal systems,are vulnerable to rising temperatures and other worse climates.What's more,people living in poor areas are more likely to be influenced because of the intensity,duration and frequency of extreme weather.Scientists and the most common people have the right to reliable meteorological data for risk assessment.Earth System Models(ESM)are models run on lots of supercomputers which is based on physics and respond to the changes in greenhouse gas in atmosphere.ESM outputs the principal data such as temperature,wind,precipitation,humidity,pressure etc.Computationally demanding of the ESM limits the spatial resolution and the spatial resolutions are too coarse which result in the inaccuracy of the climate projections.The resolutions are course for some physical process like convection,which always generates heavy rainfall.Local scale statistical downscaling has been shown bad accuracy or reliability and difficult to migrate from a region to others.The spatial downscaling learns a mapping from low resolution to high resolution.But the traditional spatial downscaling methods needs high resolution data which means the regions with little observation data at high resolutions,always the poorest regions which are most impact by the climate change,will be unable to get the downscaled climate data.With the development of convolutional neural network and deep learning,researches on image super-resolution have progressed so much.We propose that using deep learning methods to downscale the climate data with image super-resolution.Referring to the spatio-temporal nature,for example,the precipitation is an image consisting of precipitation with longitude and latitude.So we can treat the climate data as images and using the image super-resolution methods to downscale the climate data.And experimentally,we train a super-resolution model and get satisfying result on test set,in other words,we train the model on regions full of high resolution observation data and test on regions which have little high resolution observation data.This solves the problem that poor regions often have little high resolution observation data.Also,the factors which influence the climate change are often in different resolutions.So using the image super resolution methods to make the factors in unified resolutions are meaningful and will help the projections be more accurate and detailed.