Year 2000 GeoCover-Ortho Production Flow

Earth Satellite Corporation (EarthSat), under contract with NASA (Reference Number CARBON-0000-0184 – “Global Carbon Sink Ecosystem Spatial Change Image Orthorectification and Registration”), is producing an orthorectified Landsat-7 ETM+ image coverage of the earth’s entire landmass.

The products, hereafter referred to as the Year 2000 GeoCover-Ortho (GCO) products, consist of a suite of four product types:

  1. Individual Landsat-7 orthorectified scenes,
  2. Pansharpened color composites created from #1,
  3. Pansharpened color mosaics created from #2,
  4. Digitally compressed pansharpened color mosaics created from #3.

A description of each product type is given in Appendix A. From the list above, it can be seen that all three of the pansharpened products (numbers 2-4, above) are generated from the individual Landsat-7 orthorectified scenes (product #1). Therefore the geodetic accuracy of all four products is determined by the geodetic accuracy of the individual scenes.

Geodetic Accuracy

The geodetic accuracy (precise location of a pixel on the earth’s surface) is a function of three principal factors: 1) accuracy of the control points, 2) accuracy of the scene-to-scene tie points, and 3) accuracy of the digital elevation model. The accuracy of the control and tie points determine the accuracy of the photogrammetric bundle adjustment, and thus the accuracy of the ground-to-image transformation. During the orthorectification process, the accuracy of the digital elevation model affects the final geodetic accuracy of the orthorectified image.

GeoCover-Ortho Production

The GeoCover-Ortho production process for the Landsat TM products depends upon the ability to photogrammetrically triangulate large blocks of Landsat TM scenes in a single simultaneous block adjustment. The Photogrammetric blocks used to create the GeoCover-Ortho products generally consist of between 300 and 400 Landsat scenes. Figure #1 presents the general layout of the twenty seven GeoCover-Ortho photogrammetric blocks as they span the world’s land mass.

A high-level diagram of the orthorectification and mosaicing workflow followed within the GeoCover-Ortho production environment is presented in Figure #2. The following discussion will touch on the highlights of each of the processing steps.

Landsat TM Orthorectification Process

Orthorectification of the Landsat TM data is represented along the top portion of GeoCover-Ortho Work Flow diagram and consists of the following steps: Data Ingest, Tie Point Collection, Triangulation, and Orthorectification..

Data Ingest

Data ingest component consists of three primary data types; satellite image data, ground control points, and digital terrain models. The Landsat data is obtained on CD-ROM from EDC and various foreign ground receiving stations. The image data are ingested and stored on DLT tapes within an hierarchical storage management (HSM) system. The ground control points and digital terrain models, supplied by NASA solely for GeoCover-Ortho use, represent the some of the best available control for Landsat scale photogrammetry.

TM Tie Point Collection

In order to support photogrammetric bundling on a large number of individual Landsat scenes it is necessary to collect common (conjugate) points within the areas of overlap between juxtaposed scenes. Using these tie points it is possible to create a relative orientation of the complete block of Landsat scenes. By way of example, the Middle East photogrammetric block is displayed in figure #3. This block consists of a total of 351 Landsat scenes, only 34 of which contain ground control points (identified by the black triangles). There are over 600 overlap areas within the block. In order to facilitate the collection of conjugate points within the overlap areas, an automated image-to-image cross-correlation technique was developed which is specifically designed to provide a well distributed array of tie points across any given area of overlap.

Figure #2 GeoCover-Ortho Work Flow

Triangulation

Triangulating such large blocks of Landsat scenes presents two major hurdles: 1) The cost of collecting accurate control on a global scale mandates that the adjustment process be able to be performed with a minimum of control, and 2) The triangulation technique needs to be stable when performing adjustments across the curved surface of the Earth. In order to meet these requirements, EarthSat has developed a new Landsat block adjustment approach called “MosPoly”. The patent pending MosPoly technique offers the unique capability to simultaneously adjust large blocks of Landsat data (single blocks comprising between 300 and 400 scenes) with less than 15% of the scenes containing ground control points. Furthermore, a non-linear approach to the bundling process, allows the adjustments to be made across the Earth’s spherical surface. Figure #3 displays the control scene density and distribution of a typical GeoCover-Ortho block. The MosPoly user interface allows the analyst to interactively change the weightings of the individual tie and control points within the block. A typical block will consist of well over 30,000 individual tie points and 250 individual control points. By iteratively tuning the weights of these individual points it is typically possible to achieve sub-pixel (RMS) residuals.

Orthorectification (Ortho TM)

The result of the triangulation process is, for each individual Landsat scene, a mathematical mapping between the Earth’s geoid and row/column space of the raster Landsat image. The GeoCover-Ortho image is then produced by simply stepping through the ground space at the final orthoimage projection (UTM / WGS84) and, with the elevation being provided by the digital elevation model, then projecting the X,Y,Z location of the given output image pixel to a fractional pixel location (row/column) within the input Landsat TM image. A pixel interpolation technique, either nearest neighbor or cubic convolution, is then be used to calculate the intensity value to assign to given output image pixel within the final orthorectified image. In the case of the individual scene GeoCover-Ortho products, a nearest neighbor interpolation is use so as to preserve as much of the original spectral information as possible. When creating the orthorectified Landsat TM data for the GeoCover-Ortho mosaic products, a cubic convolution interpolation is used in order to best present the data’s spatial information content.

Figure #3 Middle East Block

1) Positional Accuracy Specification:

The objective here is to establish the contractual specification for the circa 2000 ortho products, both with respect to the ground (absolute) and between the 1990 and 2000 orthos (relative). During the October meeting of the Landsat Global Data Working Group (LGDWG), it was agreed that given a choice between optimizing the absolute or relative accuracy of the 2000 product, the science community would choose to maximize the relative accuracy (coregistration) between the two image data sets (see section 2.c of the minutes of the October meeting, attached). This being the case, the contractual accuracy specification would logically be specified as the relative coregistration accuracy between the 1990 and 2000 orthos. EarthSat anticipates that the vast majority of the orthos will have relative accuracies of better than one multispectral pixel (i.e. better than 28.5 meters RMSEr). However, due to changes in surface features over the 10 year time period, and to the fact that both data sets have been required to be nearest neighbor resampled, EarthSat believes that the specification should be set at 1.4 multispectral pixels (i.e. 40.0 meters RMSEr).

If the relative accuracy is specified as 40.0 meters RMSEr, and the absolute accuracy of the circa 1990 orthos meets (and largely beats) the 50 meter RMSEr specification, the absolute accuracy of the circa 2000 orthos would be the power sum [RMSEr(diff) = sqrt ( RMSEr(1) ^ 2 + RMSEr(2) ^2)] of the two independent specifications. Therefore, the expected absolute accuracy (image to ground) of the circa 2000 orthos would be 64.0 meters RMSEr [sqrt ( 50 ^2 + 40 ^2)].

The circa 1990 coverage consists of approximately 7,600 path/row areas. However, the 2000 coverage is currently set for 8,500 path/row areas. Thus there are approximately 900 circa 2000 scenes (largely over island areas and Antarctica) for which circa 1990 ortho do not exist. Clearly, when there is no preexisting 1990 orthos available, other source of control must be used. Where NIMA is able and willing to provide control, the LS-7 panchromatic band will be sent to St. Louis for control purposes. In other locations only the LS-7 ephemeris will be used. In such cases, if the path/row is a single isolated scene, then at least two acquisitions of the same path/row will be required to establish a verifiable ephemeris-based position.

Where NIMA control has been passed directly to the LS-7 panchromatic data, the absolute positional accuracy is expected to be equal to, or better than, 25.0 meters RMSEr. In the case of Landsat-7 ephemeris, the absolute accuracy would be equal to, or better than, 75.0 meters RMSEr.

The following table summarizes there accuracy specifications.

EarthSat LS-7 Year 2000 Global Mosaic –

Absolute and Relative Accuracy Specifications of the circa 2000 ortho products:

Ground Control Used for Correction
NIMA control / Ephemeris
control / 1990 Geocover Control (image matching)
1990 Geocover
Available / Absolute accuracy
(RMSEr, meters) / N/A / N/U / 64.0
Relative accuracy
(RMSEr, meters) / N/A / N/U / 40.0
1990 Geocover
Not Available / Absolute accuracy
(RMSEr, meters) / 25.0 / 75.0 / N/A
Relative accuracy
(RMSEr, meters) / N/A / N/A / N/A

N/U – “Not Used”; N/A – “Not Available”

What if? – NIMA providing all new control:

NASA has requested that EarthSat also provide absolute and relative accuracy specifications for the case where NIMA control, rather than the circa 1990 orthos, is used to control the 2000 products. It is important to note that NIMA has not signed up to providing such control, and that it would be a very big job – greater than the total effort expended by NIMA on the circa 1990 program.

With NIMA passing their control to the LS-7 panchromatic data, EarthSat believes that the absolute positional accuracy of the 2000 ortho product would be 25.0 meters RMSEr. Given the 50.0 meters RMSEr specification of the circa 1990 product, this would predict a 55.9 meter RMSEr relative accuracy between the circa 1990 and 2000 ortho products.

Even though this would provide a more accurate absolute positioning, it would reduce the image-to-image registration (which is critical for change analysis) by almost 40% [ (55.9 – 40) / 40 ].

Figure #1 – GeoCover-Ortho Production Flow
Appendix A - Landsat-7 ETM+ Year 2000 GeoCover-Ortho Deliverable Products:

Table A-1: Individual Landsat-7 Scene (9-bands) deliverable specification:

ETM + Scene Processing / Orthorectified scenes; the coverage of the output scene is defined by that of the input Landsat-7 scene.
Number of spectral bands / All nine LandSat-7 ETM+ bands (1 panchromatic, 6 multispectral, and 2 thermal).

Resampling interpolation method

/ Nearest neighbor (no interpolation).
Pixel size / 14.25, 28.5, and 57.0 meters for the panchromatic, multispectral, and thermal, respectively.
Image format / GeoTIFF; nine individual GeoTIFF files – one for each band (1 panchromatic, 6 multispectral, and 2 thermal).

Metatdata format

/

ECS Intermediate Standard; FGDC compliant

Projection / UTM – zone determined by scene center.

Datum/Spheroid

/ WGS 84

Horizontal control

/ Geocover orthorectified 1990 TM scenes where available (reference contract NAS 13-98046).
Where Geocover orthorectified TM scenes are not available, horizontal control shall be established using
US Government provided control. Definitive ephemeris will be used to supplement the control
Vertical control / Digital Elevation Models (DEM), where available, used by EarthSat for the production of Geocover orthorectified 1990 TM scenes (reference contract NAS 13-98046). This DEM was produced by EarthSat using government provided control and GTOPO30.
Where above DEM is not available, EarthSat shall use best available selected from USGS 1-arcsecond DEM, DTED Level-1, and GTOPO30.
NOTE: CLIN 0006 covers all DEM used in production of these products.
Positional accuracy / (1) Scenes for which 1990 Geocover data is available
  • Absolute accuracy: +/- 64 m net horizontal RMSE
  • Accuracy relative to 1990 Geocover scene: +/- 40 m net horizontal RMSE
(2) Scenes for which 1990 Geocover data is not available
  • Absolute accuracy: +/- 25 m net horizontal RMSE (for scenes for which NIMA -provided L-7 panchromatic band control is available)
  • Absolute accuracy: +/- 75 m net horizontal RMSE (for scenes for which NIMA -provided L-7 panchromatic band control is not available).

Table A-2: Individual Landsat-7 (pan-sharpened) deliverable specification:

ETM + Scene Processing / Pan-sharpened scenes; the coverage of the output scene is defined by the orthorectified Landsat-7 scene.
Number of spectral bands / Three; ETM+ 7,4,2 as R,G,B.
Image format / Single 24-bit GeoTIFF file

Metadata format

/

ECS Intermediate Standard; FGDC compliant

Resampling interpolation method

/ Cubic Convolution

Pan-sharpening method

/ TBD (EarthSat to Propose, NASA to approve based on examination of a pre/post sample)
Pixel size / 14.25 meters (same as 9-band product)
Projection / UTM (same as 9-band product)
Datum/Spheroid / WGS 84 (same as 9-band product)
Positional accuracy / (1) Scenes for which 1990 Geocover data is available
  • Absolute accuracy: +/- 64 m net horizontal RMSE
  • Accuracy relative to 1990 Geocover scene: +/- 40 m net horizontal RMSE
(2) Scenes for which 1990 Geocover data is not available
  • Absolute accuracy: +/- 25 m net horizontal RMSE (for scenes for which NIMA -provided L-7 panchromatic band control is available)
  • Absolute accuracy: +/- 75 m net horizontal RMSE (for scenes for which NIMA -provided L-7 panchromatic band control is not available).

Table A-3: Landsat-7 mosaics (pan-sharpened) deliverable specification:

Individual mosaic coverage / 3 x 3 degrees with overlap on all sides
Number of spectral bands / Three; ETM+ 7,4,2 as R,G, B.
Image format / Single 24-bit GeoTIFF file

Metadata format

/ ECS intermediate level and FGDC compliant metadata. NASA will provide the initial ECS metadata template and EarthSat will respond with ECS metadata content for a sample dataset. The metadata content is subject to review and approval by NASA before production can take place.
Resampling interpolation method / Cubic Convolution

Pan-sharpening method

/ TBD (EarthSat to Propose, NASA to approve)
Pixel size / 14.25 meters (same as 9-band product)
Projection / UTM (same as 9-band product)
Datum/Spheroid /

WGS 84

Positional accuracy / (1) Scenes for which 1990 Geocover data is available
  • Absolute accuracy: +/- 64 m net horizontal RMSE
  • Accuracy relative to 1990 Geocover scene: +/- 40 m net horizontal RMSE
(2) Scenes for which 1990 Geocover data is not available
  • Absolute accuracy: +/- 25 m net horizontal RMSE (for scenes for which NIMA -provided L-7 panchromatic band control is available)
  • Absolute accuracy: +/- 75 m net horizontal RMSE (for scenes for which NIMA -provided L-7 panchromatic band control is not available).

Table A-4: Compressed Landsat-7 mosaics (pan-sharpened) deliverable specification:

Individual mosaic coverage

/ 5 x 6 degrees with overlap on all sides
Compression Technique / MrSID, Lizard Tech, Inc.
Compression ratio / 30:1
Image format / SID file

Metadata format

/

ECS intermediate level and FGDC compliant metadata. NASA will provide the initial ECS metadata template and EarthSat will respond with ECS metadata content for a sample dataset. The metadata content is subject to review and approval by NASA before production can take place.

Resampling interpolation method / Cubic Convolution
Pixel size / 14.25 m (same as 9-band product)
Projection / UTM (same as 9-band product)
Datum/Spheroid / WGS 84
Positional accuracy / (1) Scenes for which 1990 Geocover data is available
  • Absolute accuracy: +/- 64 m net horizontal RMSE
  • Accuracy relative to 1990 Geocover scene: +/- 40 m net horizontal RMSE
(2) Scenes for which 1990 Geocover data is not available
  • Absolute accuracy: +/- 25 m net horizontal RMSE (for scenes for which NIMA -provided L-7 panchromatic band control is available)
  • Absolute accuracy: +/- 75 m net horizontal RMSE (for scenes for which NIMA -provided L-7 panchromatic band control is not available).