Deep learning for climate studies

Abstract

Numerical Weather Prediction is the basis for weather forecasting which require interpretation by experts to generate weather forecasts for a local area. NWPs predictions have very low spatial resolution like 30-50 kms or more. For disaster management, agriculture, etc we often need weather prediction for local regions, in some cases for a particular location. Deep learning models like CNN can be used to downscale the global NWPs predictions to produce local forecasts. In this project, I have used different convolutional neural network models to interpret numerical weather prediction model data. Also, different architectures are compared against each other to find which gives the best performance. Here the measure of the performance is mean absolute error (MAE). We show that CNNs can learn certain configurations of the atmospheric pressure system and connect them with wind speed and visibility. There is a possibility CNN-based models can be used to automatically generate derived products, in addition to numerical weather model interpretation.