Transcript Document

TEAM
Dr. Chuck O’Hara, PI
Dr. Roger King, PI
Eric Kolstad, Geospatial Analyst
Sunil Reddy Repaka, Civil Engineer
Initial Assessment of QuickBird Image Tiles for
I-10 Project Corridor Mapping and Evaluation
Overview
Initial observations
The proposed highway corridor paths for CSX railroad
relocation have potential impact on large portions of
southern Mississippi. To gauge factors in construction
planning such as environmental impact, high resolution
spatial data is requisite for detailed study of route
alternatives. Through a cooperative agreement with
NASA, multi-spectral and panchromatic QuickBird
imagery at 0.61m and 2.44m resolution respectively is
being collected for this region. As this beta image set
arrives, our initial focus will be on several tasks:
1) Arranging an overall map of image “footprints,”
showing large-scale placement of specific tile sets and
respective overlap between sets; 2) Manually outlining
individual tiles’ boundary edges to assess artifacts or
gaps that may be introduced by processing steps such
as clipping and cell division; and 3) Assigning attribute
metadata (filename, image type, acquisition date, etc.)
incorporating a rough estimate of image quality (e.g.
sharp/clear, cloudy, haze) to individual tile panels.
The first data set (6 ~4.7Gb DVDs of single- and multiband data) includes an overlapping panchromatic
image set, detailed in the large composite map at left.
Complete images acquired at fixed intervals are visible
along the QuickBird trajectory; each image has roughly
3.9km horizontal and 3.7km vertical overlap.
Individual images are received UTM/WGS84 projected
and split into a grid of 9-20 image tiles (~32Mb (multispectral) or 128Mb (panchromatic) each). These tiles
may not be completely matched; depending on
interpolation algorithm used for zoomed-in detail
viewing (e.g. via ArcView or ERDAS) tile edges can
display “black” borders, where occasional pixels
translate to null data. Further discussion and
interpretation is needed. Prior to receipt, all images &
tile sets were clipped to the larger study area defined
for satellite tasking. No-data regions are mapped
beyond that border; some tiles are thus blank, partially
or completely void of image data. The GeoTIFF
images supplied (16-bit format) required specific
extensions to view without conversion. Retention of
pre-computed pyramid layers for this format also
presented technical difficulties with COTS software.
However, with few exceptions all appropriate scene
images could be reviewed. At present basic quality
information is attributed for multispectral image tiles
e.g. "clear", "haze", "clouds", "blank" (containing no
valid data), and panchromatic image quality is marked
similarly, e.g. “clouds”, “NC” (no clouds) or “blank”.
Cloud layers are apparent within several tiles in the
northeast section of the map composite.
MAY23 _16 491 3
FEB1 2_1 639 46
FEB1 7_1 645 00
MAY23 _16 491 4
FEB1 2_1 639 48
FEB1 7_1 645 03
FEB1 2_1 639 51
FEB1 7_1 645 06
(1,1)
(1,2)
MAY23 _16 492 1
FEB1 2_1 639 54
(2,1)
(2,2)
(3,1)
(3,2)
Table 1: Example attribute table for highlighted
image tile set (valid image area blocks shown).
FEB1 7_1 645 00
Figure 1: Panchromatic coverage area showing onlap and side lap regions. Blue, green and red denote
successive acquisition paths. E-W overlap is not present between February and May images.
QuickBird Imagery for Dead Tiger Creek, Mississippi, 7.5' Quad
DigitalGlobe QuickBird
Panchromatic Imagery
DigitalGlobe QuickBird
Multispectral Imagery
USGS DOQQ’s
Team
Location Map
Image Tiles Used
Dr. Roger L. King, PI
Dr. Charles G. O’Hara, PI
Sunil Reddy Repaka, Civil Engineer
Eric Kolstad, Geospatial Analyst
Procedure
Pan
Multi
02feb17164500-p2as_r2c2-000000041925_01_p005.tif
02feb17164500-p2as_r2c3-000000041925_01_p005.tif
02feb17164500-p2as_r2c4-000000041925_01_p005.tif
02feb17164500-p2as_r3c2-000000041925_01_p005.tif
02feb17164500-p2as_r3c3-000000041925_01_p005.tif
02feb17164500-p2as_r3c4-000000041925_01_p005.tif
02feb17164503-p2as_r1c2-000000041925_01_p006.tif
02feb17164503-p2as_r1c3-000000041925_01_p006.tif
02feb17164503-p2as_r1c4-000000041925_01_p006.tif
02feb17164503-p2as_r2c2-000000041925_01_p006.tif
02feb17164503-p2as_r2c3-000000041925_01_p006.tif
02feb17164503-p2as_r2c4-000000041925_01_p006.tif
02feb17164503-p2as_r3c2-000000041925_01_p006.tif
02feb17164503-p2as_r3c3-000000041925_01_p006.tif
02feb17164503-p2as_r3c4-000000041925_01_p006.tif
02may23164914-p2as_r1c1-000000041926_01_p002.tif
02may23164914-p2as_r2c1-000000041926_01_p002.tif
02may23164914-p2as_r3c1-000000041926_01_p002.tif
02may23164914-p2as_r4c1-000000041926_01_p002.tif
02feb17164503-m2as_r1c2-000000041925_01_p006.tif
02feb17164503-m2as_r1c3-000000041925_01_p006.tif
02feb17164503-m2as_r1c4-000000041925_01_p006.tif
02feb17164503-m2as_r2c2-000000041925_01_p006.tif
02feb17164503-m2as_r2c3-000000041925_01_p006.tif
02feb17164503-m2as_r2c4-000000041925_01_p006.tif
02feb17164503-m2as_r3c2-000000041925_01_p006.tif
02feb17164503-m2as_r3c3-000000041925_01_p006.tif
02feb17164503-m2as_r3c4-000000041925_01_p006.tif
02may23164917-m2as_r1c1-000000041926_01_p003.tif
02may23164914-m2as_r1c1-000000041926_01_p002.tif
02may23164914-m2as_r2c1-000000041926_01_p002.tif
02may23164914-m2as_r3c1-000000041926_01_p002.tif
02feb17164500-m2as_r2c2-000000029365_01_p001.tif
02feb17164500-m2as_r2c1-000000029365_01_p001.tif
Multispectral and panchromatic QuickBird images acquired through the NASA
scientific data purchase have been used to create mosaiced image products. Individual
QuickBird scenes were delivered as tiled image sets due to memory and file size limits.
The individual sets were merged in ENVI to recreate the larger scenes. Erdas Imagine
was used to composite and clip these to USGS 1:24,000 Topographic Quadrangle Map
boundaries.
Panchromatic Imagery:
Image resampling was performed using Nearest Neighbor to retain band combination
intensities. Inter-scene color balancing and histogram matching was configured with
default manual options. Feathering was used to smooth transitions between image
scenes. The output area was specified by an AOI outlining the Dead Tiger Creek Quad
region.
Multispectral Imagery:
A similar procedure was utilized to create the multispectral image mosaics for the Dead
Tiger Creek area. Color balancing and histogram matching proved more difficult for
multiple bands, but acceptable results were achieved with the parabolic (automatic)
surface method.
PROJECT
TEAM
CSX Corridor Relocation
Dr. Chuck O’Hara, PI
Dr. Roger King, PI
I-10 Corridor: QuickBird Image Accuracy Assessment
Eric Kolstad, Geospatial Analyst
Sunil Reddy Repaka, Civil Engineer
Wade Givens, Research Associate
Fixed (known) reference points
Introduction
When evaluating different imagery sources for potential to attribute features (such as
roads/highways, land cover type etc.) it is important to assure spatial accuracy in classification.
To assess image alignment and feature locations in this context, image-identifiable pixel locations
must be compared to their real-world counterparts via ground truth techniques. When
georeferencing images, precision image-identifiable control positions are of key importance.
Potential bias due to errors in the measurement process must be kept to a minimum.
Error sources include instrument accuracy, differences due to sensor position and/or
orientation, and user error. These and other inherent errors can be estimated and either eliminated
or substantially reduced. What remains thereafter should indicate the discrepancies in horizontal
X and Y location of image pixels vs. observed real-world positions due to (potentially systematic)
satellite position variance and/or common translational offset. Such discrepancies can thus be
described and potentially corrected for.
To achieve high accuracy control, we plan to utilize NGS monument sites within the High Accuracy Reference Network
(e.g. yellow triangles in Fig. 1). Only those points with A- or B- level accuracy will be considered. “A HARN is a
statewide or regional upgrade in accuracy of NAD 83 coordinates using Global Positioning System (GPS) observations. …
Horizontal A-order stations have a relative accuracy of 5mm +/- 1:10,000,000 relative to other A-order stations.
Horizontal B-order stations have a relative accuracy of 8mm +/- 1:1,000,000 relative to other A-order and B-order
stations.” [http://www.ngs.noaa.gov/faq.shtml] HARNs consist of regularly spaced (20-100km), easily accessible control
points with clear horizons for satellite reception. [http://www.ngs.noaa.gov/PC_PROD/WorkShops/PPT/SPCS/sld070.htm]
A differential GPS system incorporating a base station equipped with a radio or cellular phone link can provide RealTime Kinematic (RTK) surveying with accuracy of +/-10cm. This base station will be set up on the A- and B-order HARN
monuments with precisely known locations, establishing correctional factors applied to unknown point measurements
within a 5-mile radius. Field notes include marker ID, antenna height above ground, location description, # satellites
visible, date/time and weather conditions. Antenna height is a key factor in proper horizontal control; spheroid datum
calculations benefit from if not rely upon this information for centimeter-level accurate positioning. Roving receiver
measurements should be recorded similarly, with attention to operating requirements such as minimum acquisition time to
achieve desired centimeter-level accuracy.
Control point selection
Prior to fieldwork, a carefully chosen network of sample points was created. Sets of points are clustered near NGS
monuments within a radius defined by the limits of the GPS receiver and base station. (approx. a 5-mi. radius at present).
QuickBird image pixels are roughly 0.6-0.7m (panchromatic) or 2.4-2.88m (multispectral). Points chosen should be
identifiable at a sub-pixel level; for accuracy assessment, our focus is primarily on panchromatic imagery.
Ideal control points are:
- within surveyable distance of a known reference point
- within accurate measurement limits of equipment incl. base station
- permanent features (e.g. roads rather than cropland)
- clear and distinct spots, such as a near-perpendicular road edge/driveway or road/railway intersection
- features of reasonable scale vs. pixel size (e.g. small roads or driveways instead of footpaths)
- unobscured by nearby objects, overhanging trees et al.
- in a fairly open area (best: 15-degree azimuth in several directions, to facilitate GPS satellite locks and minimize
multipath interference)
Some control points will not be viable in the field for varied reasons. In practice, a larger number of points (min. 10-15
per image tile) should be chosen to ensure representative sample size. Individual NGS markers encompass some set of
valid points within a radial distance. In some instances two NGS markers’ viable coverage areas will overlap. Collecting
readings for common points with a base station at different NGS monuments is recommended, as this may reveal sources of
error due to procedures involved in differential GPS measurement (rather than the imagery source).
Figure 1: Several HARN geodetic markers w/ nearby ground-control points.
After selection, control points’ coordinates (in geographic or UTM projection) and IDs can be exported for rough
navigation via handheld GPS. It has been found advantageous to have a laptop with multispectral and/or panchromatic
image sets along with target GCS monument and control point locations at hand for field reference. Initial GPS results
(esp. when observed inside a vehicle) were not always sufficient to clearly indicate correct ground location for given points.
PROJECT
CSX Corridor Relocation
Dr. Chuck O’Hara, PI
Dr. Roger King, PI
TEAM
I-10 Corridor: QuickBird Image Accuracy Assessment
Eric Kolstad, Geospatial Analyst
Sunil Reddy Repaka, Civil Engineer
Wade Givens, RSTC Extension
CE90 determination
ð
4106
4113
4107
ð
ð
ð
4103
ð
4097
The CE90 standard is commonly used to assess horizontal accuracy in map and
image products. Established by the National Map Accuracy Standard (NMAS) in
1947, the CE90 measure states that for a set of well-defined points, 90% will fall
within a specified radial distance. Greenwalt and Schultz (1968) state that if X
and Y errors are part of a bivariate normal distribution and the X and Y errors are
equal, independent, and zero-mean, then:
4101
ð
ð
4102
CE90 = 2.1460 * RMSEx
RMSEx is the root mean square summary of X error for a given point:
4096
ð
RMSEx = sqrt(∑ (Ximage - Xcontrol) 2 / n)
[RMSEy is found similarly]
ð
4094
Figure 2: GPS coords. (black) and distance to
image-based control pts (red vectors, 10x actual)
When X and Y errors are not be equal, however, a linear combination of RMSEx
and RMSEy is said to be more accurate (Greenwalt and Schultz, 1968):
RMSEc = 0.522* RMSEmin + 0.4778* RMSEmax
where RMSEmin is the minimum value of RMSEx and RMSEy, and RMSEmax is
the higher value. This linear combination was shown to be a good estimate of the
circular error equivalent value for RMSEmin to RMSEmax ratios between 0.6 and
1.0. It was also found that a reasonable approximation for RMSEc is given by
RMSEc ≈ 0.5 * (RMSEmin + RMSEmax)
Figure 4: Radial plot for pictured points with
standard and bias-adjusted CE90 radii. Inner
circles show potential CE90 for same-day data.
Where RMSEmin / RMSEmax lies between 0.6 and 1.0, the above equation has
been used by the NSSDA in the U.S. spatial data standard (Federal Geographic
Data Committee, 1998). In terms of RMSEc, the estimate for CE90 looks the
same as in the first equation (ref: NASA SSC Geometric Accuracy Assessment of
DigitalGlobe QuickBird Imagery) .
RMSEc
RMSE min
When RMSEmin / RMSEmax is between 0.6 - 0.2,
RMSE max RMSEmax
RMSEc must be adjusted using an interpolated value
1.0000
1.0000
from statistical data relating RMSEmin / RMSEmaxto
0.8165
0.9063
0.6547
0.8197
RMSEc / RMSEmax, according to the table at right
0.5000
0.7323
(Greenwalt and Schultz, 1968).
0.3333
0.2294
0.1005
0.0
Figure 3: GPS readings (yellow circles) + offset to imagebased coordinates (green vectors). Solid blue shows to-scale
displacements for main point set in test group.
0.6327
0.5727
0.5274
0.5151
The limited test dataset has thus far approached
DigitalGlobe’s estimated positional accuracy of ~23m.
(with a bias-adjusted CE90 calculated to be 25.25m).
While the majority of data points were acquired in a similar timeframe, outlying
data (e.g. two W/SW- rather than NE-oriented offsets) seems correlated with
satellite position on different dates. An extensive multi-county dataset will be
acquired for detailed analysis of error sources & potential for bias correction.
Figure 5: Absolute errors for sample points in this
regional area.