Improving Atmospheric River Forecast Over Himalayas using Convolutional Neural Network

Long, narrow zones of the Integrated Vapor Transport (IVT) in the lower troposphere are known as Atmospheric Rivers (ARs). ARs are major causes of heavy rain, and they are often associated with serious cases of flooding. For instance, the 2014 Kashmir flood and the 2013 Uttarakhand flood are linked to Himalayan ARs. Therefore, ARs are important in causing extreme weather and risk of floods in the Himalayan region. Thus, skillful prediction of ARs can be helpful in better severe weather risk management. The most widely accepted metric for identifying ARs is IVT as it integrates moisture content and its transport. Although the Global Forecast System (GFS) forecasts IVT globally, it is shown to suffer from systematic error over the West Coast of USA, especially for high magnitude IVT, and also fails in the accurate spatial organization of AR events. Recently, Chapman et al. (2019) proposed a Convolutional Neural Network (CNN) to the enhance the skill of GFS IVT forecasts in mid-latitude areas on the West Coast. However, the model lacks correction of IVT direction, which is critical in defining the precipitation produced from an AR upon impacting a mountain barrier. In addition, there is no machine learning model that is specifically designed for the Himalayan region. This work modifies the Chapman CNN architecture, in the South Asian region, incorporating the Himalayan region for correcting both the magnitude and direction of GFS IVT. In our work we take Modern Era Retrospective Analysis for Research and Applications Version 2 (MERRA-2) as a proxy to ground truth. The model significantly reduced various metrics, such as the Root Mean Square Error (RMSE) and Mean Angular Error (MAE), in comparison to GFS IVT, in the Himalayan region and in the entire study domain. When the model was tested for AR events, its performance significantly improved the AR forecast. These advances show that the model offers a powerful deep learning framework for AR prediction as compared to the raw GFS baseline.