Creating Synthetic Weather Radar Images Using Convolutional Neural NetworksChristopher 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 Fieldsat 13 km Resolution
IR
Lightning
NumericalModel
CNN (Convolutional Neural Network)
Convolution RectifiedLinear Unit Pooling Dropout
Loss: MSEBackpropagation
Convolution RectifiedLinear Unit Pooling Dropout
Loss: MSEBackpropagation
Convolution RectifiedLinear Unit Dropout
Loss: MSEBackpropagation
Convolution RectifiedLinear Unit Pooling Dropout
Loss: MSEBackpropagation
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, 17Z0
0.2
0.4
0.6
0.8
1
Selected Overpasses from Summer 2015
Prob
abili
ty o
f Det
ectio
n
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
Feat
ure
Inte
nsity
32 filters learned in first convolutional layer
* Expa
nded
Cap
abili
ties Global Proof of Concept Forecasting Prototype
GALWEM Numerical Model
Portable Doppler Radar
Advanced Satellites
New
Dat
a So
urce
s
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.