Predicting the spatiotemporal characteristics of atmospheric rivers: A novel data-driven approach

Abstract

Atmospheric Rivers (ARs) are narrow bands of high-water vapor content in the low troposphere of mid-latitude regions through which most of the poleward moisture is being transported. ARs have been represented statistically as the regions of intense vertically integrated horizontal water vapor transport (IVT) in the atmosphere. These ARs have been found positively correlated with extreme precipitation and flood events at some coastal mid-latitude regions and thus have been linked to several socioeconomic implications. The robust and accurate forecasts of AR availability at a significant lead time can be a useful tool for managing AR-associated floods and water resources. To enhance the knowledge of data-driven methods for modelling nonlinear atmospheric dynamics associated with ARs, we have explored some popular deep-learning architectures for predicting AR availability. AR availability maps derived from the statistical characterization of IVT using ERA5 reanalyses data of ECMWF from the testing dataset are taken as ground truth for the prediction. The predictions of the models have been analyzed based on popularly adopted performance evaluation metrics structural similarity index measure (SSIM), mean square error (MSE), root mean square error (RMSE), and peak signal-to-noise ratio (PSNR). Our proposed autoencoder model outperforms the conventional convolutional neural network (CNN) and Conv-LSTM model. We have got comparatively higher scores (average) of SSIM (0.739) and PSNR (64.424) as well as lower scores (average) of RMSE (0.155) and MSE (0.025) for the predictions which signify the ability of our model to learn spatiotemporal features linked with AR-dynamics.