TJ4.5 Spatial Structure Evaluation of Unsupervised Deep Learning for Atmospheric Data

Tuesday, 9 January 2018: 11:45 AM
Room 7 (ACC) (Austin, Texas)
David John Gagne II, NCAR, Boulder, CO; and S. E. Haupt and D. W. Nychka

Handout (15.4 MB)

The weather is influenced by atmospheric processes operating on a wide range of spatial scales. These multiscale processes create spatial structures within gridded atmospheric data that may be linked with high-impact weather events. If statistical and machine learning models can learn how to represent these spatial structures, they may be able to extract useful predictive information from them. The field of deep learning has developed some novel approaches for creating abstract, multiscale representations of spatial and temporal data for improving the accuracy of prediction tasks. Many of these approaches require very large amounts of labeled data, but some unsupervised deep learning methods can learn spatial representations without large labeled datasets. Generative adversarial networks (GANs), an unsupervised deep learning method that optimizes a pair of neural networks using an adversarial training approach, have been able to produce realistic synthetic images. However, they have primarily been evaluated either subjectively or on their based on their accuracy in semi-supervised learning problems. GANs also have many possible parameter settings and network structures, but the choices for these parameters have not been rigorously justified. In this study, we statistically evaluate how well a large number of GAN configurations represent a variety of spatial covariance structures in both spatial random fields with prescribed spatial covariances and numerical simulations of thunderstorms and their surrounding environment. With the spatial random fields, we found that GANs can represent exponential covariances with different length scales and can represent domains with spatially-varying length scales. With the storm dataset, we have found that GANs can produce synthetic examples that replicate the correlations among different variables, such as the radar reflectivity and the surface temperature and wind fields. The choice of parameter settings and network structure does have a large impact on the quality of these representations.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner