Improving Integrated Vapour Transport Forecasts over the Himalayas Using a Convolutional Neural Network for Better Atmospheric River Prediction

Sheikh Imran Fayaz, Munir Ahmad Nayak, Adnan Kaisar Khan

Extreme precipitation over the Himalayas is often linked to Atmospheric Rivers (ARs) interacting with its unique and complex topography. The topographic complexity and sparse observational data pose a challenge for numerical weather prediction models. We find that the widely used Global Forecast System (GFS) exhibits systematic errors in high-magnitude Integrated Vapor Transport (IVT), and its predicted AR structure and direction show significant mismatches with observations. Our work proposes a modified convolutional neural network model, based on a previously developed ARcnn, for IVT over South Asia, including the Himalayas. ARcnn significantly improves the error metrics, root mean square error (RMSE), bias in high-IVT, and directional mean angular error (MAE), for both 24-hour lead and 7-day lead forecasts over the Himalayan region. The results in this study indicate improved IVT forecasts across the entire study area and more specifically over the Himalayas. Further, the ARcnn model was able to correct IVT that made it possible to detect the ARs missed in GFS forecasts for both lead times. Thus, these results provide compelling evidence of ARcnn’s powerful postprocessing capability and its potential for use in early prediction tools for IVT and AR.