An innovative approach to predict atmospheric rivers: Exploring convolutional autoencoder

Abstract

Atmospheric rivers (ARs) are filamentary regions of high moisture content in mid-latitude regions through which most of the poleward moisture is being transported. These ARs carry a huge amount of water in the form of vapor and thus landfalling of these ARs may bring either a beneficial supply of water or may create hazardous flood situations and thus cause damage to life and property. These regions have been statistically characterized as intense integrated water vapor transport (IVT) regions in the troposphere based on various thresholds of magnitude, direction, and geometry. To enhance the knowledge of data-driven methods for modelling nonlinear atmospheric dynamics associated with ARs, a first ever study with data-driven methodology incorporating a Deep Learning architecture, Autoencoder has been proposed. While training the proposed model, the Adam optimizer was used to reduce the mean squared error loss and was optimized using the Rectified Linear Unit (ReLU) and Sigmoid activation functions. The prediction results of the availability of ARs at next frames by the Autoencoder were assessed by popularly used performance evaluation metrics structural similarity index metrics (SSMI), mean squared error (MSE), root mean squared error (RMSE), and peak signal to noise ratio (PSNR). We have got comparatively higher scores (average) of SSIM (0.739) and PSNR (64.422) and lower scores (average) of RMSE (0.155) and MSE (0.0247) for AR prediction from our model which signifies the accuracy of the proposed Autoencoder in capturing AR dynamics. The findings of the study could be useful in giving important insights to incorporate Deep Learning models for forecasting ARs at significant lead time and consequently reducing the risk and increasing the resilience of AR flood prone regions.