Spring Boot vs Quarkus the ultimate battle - DevoxxUK
IGARSS-SAR-Pritt.pptx
1. Automated Registration of
Synthetic Aperture Radar
Imagery to LIDAR
IGARSS 2011, Vancouver, Canada
July 24-29, 2011
Mark Pritt, PhD Kevin LaTourette
Lockheed Martin Lockheed Martin
Gaithersburg, Maryland Goodyear, Arizona
mark.pritt@lmco.com kevin.j.latourette@lmco.com
2. Problem: SAR Image Registration
Registration of SAR and optical imagery is difficult.
Features appear different.
Different viewpoints and illumination conditions cause difficulties:
SAR layover does not match optical foreshortening.
Shadows do not match.
Conventional techniques rely on linear features.
But these features can be rare and noisy in SAR imagery.
MSI
image
SAR
image
2
3. Solution
Our solution is image registration to a high-resolution
digital elevation model (DEM):
A DEM post spacing of 1 or 2 meters yields good results.
It also works with coarser post spacing.
Works with terrain data derived from many sources:
LIDAR: BuckEye, ALIRT, Commercial
Stereo Photogrammetry: Socet Set® DSM
SAR: Stereo and Interferometry
USGS DEMs
3
4. Methods
Create a predicted image from the DEM, illumination
conditions and sensor model estimate.
Register the predicted and the actual images.
Refine the sensor model.
Predicted SAR Image SAR Image
4
5. Methods (cont)
The same approach works for SAR and optical sensors.
Projection into the imaging plane is similar.
Layover in SAR images is similar to occlusion in optical images.
Radar shadow is similar to optical shadow.
SAR Optical
Sensor Sensor
Layover Scene Shadow Occlusion Scene Shadow
5
6. Methods (cont)
To register SAR and optical images, use the DEM as
the “bridge”.
Generate a predicted “DEM” image for each SAR and optical image.
Register the predicted images to the actual images.
This neatly bypasses the problem of direct SAR-optical registration.
SAR Image DEM MSI Image
6
7. Example 1: SAR-LIDAR Registration
COSMO-SkyMed SAR
Image of Mosul, Iraq BuckEye LIDAR DEM
Post Spacing: 1 meter
Area: 100 km2 Absolute Accuracy: 1.5 m (CE90)
21,000 x 20,000 pixels
7
9. Results (cont)
Predicted SAR Image from DEM and
Estimated SAR Camera Model
Flicker with previous slide
9
10. Results (cont)
Normalized Cross-Correlation Image
Between Predicted and Actual Images
Flicker with previous slide
10
11. Results: Zoom
Note the
SAR layover
COSMO-SkyMed SAR Image and shadow
11
12. Zoom (cont)
Note the
SAR layover
Predicted SAR Image from DEM and shadow
Flicker with previous slide
12
13. Zoom (cont)
Cross Correlation
Flicker with previous slide
13
14. Registration Accuracy
NCC Registration Tie Points
Best shift:
Δx = 16.76m
Δy = 4.27m
After least-squares fit to shift-only registration function
with RANSAC outlier removal, 4572 tie points remained.
14
15. Registration Accuracy (cont)
Error Propagation
Statistic x y
Mean Residual 0 pixels 0 pixels
Sigma Residual 0.948 pixels 0.981 pixels
RMSE 1.364 pixels
Circular Error
Propagated to DEM 1.48 m (CE90)
This includes
Circular Error the geospatial
Propagated to Ground 2.1 m (CE90) errors in the
DEM and the
registration.
CE90 = circular error 90%
15
22. Example 2: SAR-MSI-LIDAR Fusion
Waterton, COSMO-
Colorado SkyMed
SAR
Ikonos
MSI
BuckEye
LIDAR DEM
BuckEye Lidar: March 2003 (4.1 x 5.2 km, 0.75-m post spacing)
Ikonos: July 9, 2001 (1-m GSD). COSMO SkyMed SAR: Oct 31, 2008 (0.5-m GSD)
22
27. Conclusion
We have introduced a new method for registering
SAR images with other sensor data:
LIDAR, Digital Elevation Models, Optical Images, MSI
It works by image registration to a high-resolution
DEM.
It does this by generating a predicted image from the DEM and
sensor model estimate.
It then registers the predicted and actual images and refines the
sensor model estimate.
Accuracy: 1-2 m CE90
Our approach also extends to the case where no DEM
is available:
DEM can be generated from stereo EO or interferometric SAR.
27
28. Conclusion (cont.)
For an extension to Video Geo-registration:
Pritt, M & LaTourette, K., Stabilization and Georegistration of Aerial Video
Over Mountain Terrain by Means of LIDAR.
FR1.T08.4
28
Hinweis der Redaktion
A noted hard problem in a flat area, it is extremely difficult if not impossible over mountainous or urban terrainMany conventional techniques attempt to perform a 2D-2D registration between SAR/EO images, and while this may be sufficient in flat/planar scenes, the techniques will fail in mountainous or urban terrain. Since each image is a 2D representation of a 3D scene, the perspective distortions induced by the terrain must be accounted for.
Solves problems ranging from Cross sensor registration, Radar/Optical/Infrared, including over rugged and urban terrain, and true orthorectification of SAR images.
Required input: (1) SAR Image to be registered, (2) Estimate of the image collection geometry and Image Formation Processing and (3) A high resolution DEM.From the SAR image metadata, we can create a mapping from the 3D world space of the DEM into the 2D pixel space of the SAR image. To simulate Radar backscatter, we use a weighted average of Lambertian and Specular shading, accounting for Layover and Shadow.Use the Sensormodel to render the SAR-shaded DEM, and register the simulated and actual images together. Since we have accounted for perspective distortions, layover, etc., there is no need for a scale/rotation invariant registration such as SIFT/SURF…instead a Normalized Cross-Correlation based method is ideally suited.The resulting registration function is then used to properly update the Sensor-camera model.
0.75 cm COSMO-SkyMed Spotlight-mode SAR image, processed in RMA-INCA (Range Migration Algorithm – Imaging Near Closest Approach)
After fitting the tie points to a registration function, we then perform an error propagation analysis by applying the registration function to our tie points, computing the residuals and various other statistics.
We first note that the average residuals in x- and y- are both zero, indicating that no bias is present, similarly the fact that our standard deviations in x- and y- are so close indicates that our shift only approach was appropriate. Had an affine or higher order function been necessary, we would expect to see skewed or larger values.The geographic location of each pixel in the image is computed to within 2.1 meters with 90% confidence.