Creating Synthetic Weather Radar Images Using Convolutional Neural Networks Christopher J. Mattioli, Mark S. Veillette, Haig Iskenderian, Eric P. Hassey, Patrick M. Lamey: Massachusetts Institute of Technology, Lincoln Laboratory, Lexington, MA Synthetic Weather Radar • The Synthetic Weather Radar (SWR) capability is a system that creates radar-like mosaics of precipitation intensity and echo tops beyond the range of current weather radar • At its core, this capability makes use of a Convolutional Neural Network (CNN) trained with truth sources over land • Its architecture is that of a directed acyclic graph (DAG) where multiple output layers contribute to the loss VIS 1 Channel at 1 km Resolution 4 Channels at 4 km Resolution 9 Densities at 2 km Resolution 10 Model Fields at 13 km Resolution IR Lightning Numerical Model CNN (Convolutional Neural Network) Convolution Rectified Linear Unit Pooling Dropout Loss: MSE Backpropagation Convolution Rectified Linear Unit Pooling Dropout Loss: MSE Backpropagation Convolution Rectified Linear Unit Dropout Loss: MSE Backpropagation Convolution Rectified Linear Unit Pooling Dropout Loss: MSE Backpropagation Concatenation Convolution Convolution • • • N { Objective Loss: MSE Backpropagation *MSE: Mean Squared Error Current extent of weather radar cannot represent hazardous weather offshore indicated by lightning flashes (white markers). SWR applied to the same case is able to fill in the precipitation outside the range of the radar. Data Sources for Synthetic Weather Radar Regions without weather radar have varying areas and times of data coverage. The challenge is to combine these data sources to create a radar-like depiction of weather. Fixed Weather Radar Surface Observations Low-Earth Orbiting Satellite Lightning Strike Locations Numerical Model Geostationary Satellite Global Expansion and Forecasting for DoD MIT Lincoln Laboratory routinely performs quantitative evaluations of the synthetic weather radar against land-based NEXRAD radar and space-based radar from NASA’s Global Precipitation Measurement (GPM) mission Core Observatory satellite. Quantitative Capability Evaluation SWR Probability of Detection vs. GPM NEXRAD Radar Synthetic Radar Comparison Comparison Probability of Detection (POD) = #Hits / (#Hits + #Misses) 05/21, 13Z 06/02, 19Z 06/03, 08Z 06/03, 18Z 06/06, 07Z 06/06, 17Z 07/13, 20Z 07/14, 06Z 07/20, 19Z 07/23, 18Z 07/25, 18Z 07/28, 17Z 08/09, 13Z 08/10, 12Z 08/18, 11Z 08/20, 11Z 08/20, 19Z 08/21, 08Z 08/25, 18Z 08/25, 19Z 08/26, 09Z 08/27, 18Z 08/28, 18Z 08/29, 17Z 08/30, 17Z 0 0.2 0.4 0.6 0.8 1 Selected Overpasses from Summer 2015 Probability of Detection GPM Radar Synthetic Radar GPM Limited Weather Radar over Caribbean The Caribbean region represents over a million square kilometers of offshore U.S. airspace and supports nearly 17% of all U.S. outbound passengers. Air traffic demand in the Caribbean is expected to grow rapidly by 5-6% per year over the next decade, and the lack of weather radar coverage in this airspace will continue to be a major hurdle for safe and effective air traffic control. ZHU ZIX ZMA ZNY Non-Radar ADS-B Challenges International Coordination Degraded NEXRAD Destruction of Puerto Rico’s NEXRAD Learned Features for Visible Satellite “Peakyness” Detector “Bottom Edge” Detector “Clear Sky” Detector “Cumulus” Detector High Low Feature Intensity 32 filters learned in first convolutional layer * Expanded Capabilities Global Proof of Concept Forecasting Prototype GALWEM Numerical Model Portable Doppler Radar Advanced Satellites New Data Sources DISTRIBUTION STATEMENT A. Approved for public release. Distribution is unlimited. This material is based upon work supported by the Federal Aviation Administration and Department of the Air Force under Air Force Contract No. FA8702-15-D-0001. Any opinions, findings, conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Federal Aviation Administration and Department of the Air Force. © 2019 Massachusetts Institute of Technology. Delivered to the U.S. Government with Unlimited Rights, as defined in DFARS Part 252.227-7013 or 7014 (Feb 2014). Notwithstanding any copyright notice, U.S. Government rights in this work are defined by DFARS 252.227-7013 or DFARS 252.227-7014 as detailed above. Use of this work other than as specifically authorized by the U.S. Government may violate any copyrights that exist in this work.