Probabilistic Detection of Atmospheric Rivers Across Climate Datasets and Resolutions with Neural Networks

Atmospheric rivers (ARs) are long, narrow filaments of moisture that can alleviate drought or cause intense precipitation. Because of ARs' significance to human systems, we propose using neural networks to detect ARs in different resolutions, datasets, and input fields. With more than fourteen threshold-based detection algorithms, the Atmospheric River Tracking Method Intercomparison (ARTMIP) project has developed a catalogue of AR labels in the MERRA2 reanalysis dataset, and we use these catalogues as a training dataset for the neural networks. We demonstrate the importance of two methods of training the networks. First, perceptual loss functions avoid grid-cell level comparisons between AR predictions and labels, so they allow the network to detect ARs across resolutions. Second, style transfer is a method traditionally used to convert a photorealistic image into the style of a painting, and we adapt it to allow neural networks to identify ARs in climate models, even though they are only trained on reanalysis data. Using C20C climate projections, we compare the climate change signal identified by neural networks to that identified by ARTMIP algorithms. Using a perturbation experiment, we validate that the neural network has learned the deeper spatial structure of ARs and is not simply replicating a threshold-based algorithm. Finally, we examine the neural networks' output on a range of input fields: Integrated Vapor Transport, Integrated Water Vapor, and the infrared window of satellite data, and we demonstrate that the neural networks' probabilistic detections match the uncertainty across the ARTMIP algorithms.