Deep Learning for CONUS-Wide Prediction of Thunderstorm Wind in Real-time

In just the last three years, the field of deep learning has exploded. Deep learning is a subset of machine learning in which models are trained on high-dimensional native data, rather than scalar features computed from the native data. For example, deep learning can be trained with complete images, rather than spatial statistics extracted from images.

The most documented success of deep learning has been AlphaGo (Silver et al. 2016), which beat the European champion of Go in October 2015, then the world champion in May 2017. Experts predicted that computers would not beat the best humans in Go for another decade. Deep learning has also been used to generate realistic but novel human faces (Wang et al. 2017), perform as well as dermatologists in diagnosing skin cancer (Esteva et al. 2017), and outperform cardiologists in diagnosing heart arrhythmias (Rajpurkar et al. 2017).

Deep learning has also found a few applications in meteorology. Convolutional neural networks (CNNs) have been used by Elhoseiny et al. (2015) to classify photographs as cloudy or sunny; by Wang et al. (2016) to predict spatial maps of sea-ice concentration from satellite images; and by Liu et al. (2016), Mahesh et al. (2018), and Kunkel et al. (2018) to detect features such as synoptic fronts, atmospheric rivers, and tropical cyclones.

We expand the domain of deep learning to severe wind generated by thunderstorms. Specifically, we use CNNs to predict the probability of damaging straight-line wind (> 50 kt or 25.7 m s-1) on a 1-km CONUS-wide, at various lead times up to 90 minutes.

Our predictor fields are radar images from the Multi-year Reanalysis of Remotely Sensed Storms (MYRORSS) and forecast soundings from the Rapid Refresh (RAP) model. Our verification data (“ground truth”) consist of local storm reports and surface wind observations from the Meteorological Assimilation Data Ingest System (MADIS), Oklahoma Mesonet, and 1-minute METARs.

Before training the CNN, we process the input data in four ways. First, we identify and track storm cells, using the tracking algorithm of Homeyer et al. (2017). Second, we link observations to nearby storm cells, which allows each storm cell to be labeled as “severe” (produces one or more wind gusts > 50 kt) or “non-severe”. Third, for each storm object (one storm cell at one time step), we extract a storm-centered radar images. This includes multiple variables (composite reflectivity, vertically integrated liquid, maximum estimated hail size, etc.), each of which becomes a “channel” in the image. Fourth, we interpolate the RAP sounding to the time and location of each storm object.

Following standard practice, the top layers in the CNN (those nearest to the input radar images) alternate between convolutional and pooling layers. Each convolutional layer L runs a moving window (“filter”) over its input image, and the subsequent pooling layer spatially aggregates information from L. The input to the top convolutional layer is an actual radar image, and the input to each subsequent convolutional layer is a “feature map,” created by running one or more convolutional-pooling-layer pair over the radar image. With each successive convolution and pooling, the feature map becomes smaller (fewer pixels) and represents larger-scale features from the original radar image. After all the convolutional and pooling layers, there is a traditional (fully connected) neuron layer. (See Figure in Wang et al. [2016] for an illustration of this architecture.) The input to this layer consists of the convolved radar image (now flattened to a 1-D vector, which represents features discovered in the radar image at various scales) and statistics (CAPE, CIN, storm-relative helicity, etc.) computed from the interpolated RAP sounding.

Finally, forecast probabilities from the CNN are projected to a 1-km CONUS-wide grid. Specifically, for each storm cell and time horizon (0-15, 15-30, 30-45, 45-60, and 60-90 minutes), a swath is created by extrapolating the storm location along its current motion vector (with some uncertainty). Then the CNN probability is attached to all grid points inside the swath.

We will compare the performance of this model to our non-deep-learning approach, which was operationalized in the 2017 Hazardous Weather Testbed, and discuss operationalization plans for the current approach.

Elhoseiny, Mohamed, Sheng Huang, and Ahmed Elgammal. "Weather classification with deep convolutional neural networks." Image Processing (ICIP), 2015 IEEE International Conference on. IEEE, 2015.

Esteva, Andre, et al. "Dermatologist-level classification of skin cancer with deep neural networks." Nature 542.7639 (2017): 115.

Homeyer, Cameron R., Joel D. McAuliffe, and Kristopher M. Bedka. "On the development of above-anvil cirrus plumes in extratropical convection." Journal of the Atmospheric Sciences 74.5 (2017): 1617-1633.

Kunkel, Kenneth, J.C. Biard, and E. Racah. “Automated detection of fronts using a deep learning algorithm.” Conference on Artificial and Computational Intelligence and its Applications to the Environmental Sciences, American Meteorlogical Society, 2018.

Liu, Yunjie, et al. "Application of deep convolutional neural networks for detecting extreme weather in climate datasets." arXiv preprint arXiv:1605.01156 (2016).

Mahesh, Ankur, M. Prabhat, and W. Collins. “Assessing uncertainty of deep learning techniques that identify atmospheric rivers in climate simulations. Conference on Artificial and Computational Intelligence and its Applications to the Environmental Sciences, American Meteorlogical Society, 2018.

Rajpurkar, Pranav, et al. "Cardiologist-level arrhythmia detection with convolutional neural networks." arXiv preprint arXiv:1707.01836 (2017).

Silver, David, et al. "Mastering the game of Go with deep neural networks and tree search." Nature 529.7587 (2016): 484-489.

Wang, Lei, et al. "Sea ice concentration estimation during melt from dual-pol SAR scenes using deep convolutional neural networks: A case study." IEEE Transactions on Geoscience and Remote Sensing 54.8 (2016): 4524-4533.

Wang, Ting-Chun, et al. "High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs." arXiv preprint arXiv:1711.11585 (2017).