FR1.L09.2 - ONBOARD RADAR PROCESSING CONCEPTS FOR THE DESDYNI MISSION
1. Onboard Radar Processing Concepts for the DESDynI Mission Yunling Lou, Steve Chien, Duane Clark, Joshua Doubleday, Ron Muellerschoen, and Charles Wang Jet Propulsion Laboratory California Institute of Technology Pasadena, California IGARSS 2010 26-30 July, 2010 Honolulu, Hawaii
7. Example Mission Scenario with Autonomous Sciencecraft Experiment Image taken by Spacecraft Event Detection No event Detected: Delete Image Event Detected Onboard Science Analysis Track a wide range of science events – floods, volcanoes, cryosphere, clouds,… Key Insight: No need to replicate ground science analysis – just detect activity ASE uses state of the art Machine Learning to detect events in the presence of noise
8.
9.
10. Onboard Product Examples: Forest Fire Extent July 2010 LHH , LHV , LVV Tracking forest fire extent with UAVSAR L-band polarimetric data products Fire scars
13. July 2010 Medium Fuel Load Forests Distribution of 10-hr Branch Fuel Load Onboard Product Examples: UAVSAR Fuel Products
14. July 2010 High Fuel Load Forests Distribution of 100-hr Branch Fuel Load Onboard Product Examples: UAVSAR Fuel Products
15. Onboard Product Examples: Earthquake Damage Assessment July 2010 By comparing two post-earthquake amplitude images, we are able to identify old features (green) that have been removed (perhaps damaged buildings) and new features (red) that have been built (perhaps tent cities) over a two-week period. Amplitude Change Detection with UAVSAR’s 16-day Repeat Pass Data over Port Au Prince Airport, Haiti Jan 27, 2010 Feb 13, 2010
16. Onboard Product Examples: Glacier Melting July 2010 L-band polarmetric image of the Kangerlugssuaq ice fjord in Eastern Greenland. The grounding line of the glacier is easily identifiable in the image LHH , LHV , LVV Grounding line
17. Input Data: Training, Cross Validation & Evaluation SVM Output Visual Imagery Low statistical correlation with input data (noisy), high visual feature correlation April 22, 2006 Courtesy Google Earth Labels for training and cross validation statistics. National Land Cover Data 2001, condensed to 4 coarse classes. Polarimetric data, incidence angle, etc: training and final classification input. Support Vector Machine Onboard Product Examples: Vegetation Classification with Support Vector Machine water dense veg. light veg. bare/ urban
18.
Hinweis der Redaktion
1-hour fuel: fallen needle, leaf needle, and small twigs – source of surface fire
10-hour fuel load: small branches and woody stems. Due to their resistance to drying and greater heat capacity, 10-hour fuels often do not combust in low-intensity surface fires. When moisture is low, however, 10-hour fuels can carry hot fires and help ignite larger (100- and 1000-hour) fuels. Ten-hour fuels are readily consumed when fuel moistures are low.
Larger downed woody debris is common 100-hour forest fuels. These fuels take longer to dry, deterring their consumption under most conditions. Likewise, 100-hour fuels are slow to gain moisture, so they can combust after prolonged drought, even with recent precipitation. When 100-hour fuels ignite they can burn for hours, in mixtures of flaming and smoldering combustion. Decay of 100-hour fuels can alter their response and makes them combust more readily than intact fuels.
NLCD: water(black), dense veg(dark gray), low veg(light gray), bare/urban(white) Training Samples: ~6000 randomly chosen samples (pixels) from many images in the area. Selected image is just one example image (500x500 pixels). SVM: multi-class strategy: one-vs-one; gaussian radial-basis kernel Inputs: hh, hv, vv (all in db scale, topography 'removed'), incidence angle, local incidence, 7x7 pixel average, 7x7 gaussian weighted average, 7x7 variance, hhvv phase, sin(phase), sin(local inc), sin(flat inc), rvi: 8hv / (hh+vv+2hv)