Unlocking the Potential for Landsat 7 Data
By Dr. Philip Cheng, Dr. Thierry Toutin, and Dr. Victor Tom

Introduction
After a series of long delays, the Landsat 7 satellite was successfully lifted into orbit on April 15, 1999. Landsat 7 data consists of six visible infrared (IR) bands (30-m), two thermal infrared bands (60-m at low and high gains), and one panchromatic band (15-m). This data is now available for online viewing and purchase via the Internet from both the U. S. Geological Survey (USGS) web site, and from the National Aeronautics and Space Administration (NASA) web site. Large portions of the Earth's landmass are being imaged every sixteen days. The satellite has the capability to collect as many as 450 scenes daily. These images are used to monitor nature's mysteries-from the movement of ice streams in Antarctica to the temperature of active volcanoes in Hawaii, and even the effects of urban growth on metropolitan areas around the world.
      The release of Landsat 7 data has had a significant impact on the geomatics industry in two principal areas. First, Landsat 7 data is sold for only the cost of reproduction and distribution-US$475 for each raw data scene, or US$600 per scene for system corrected data (L1R - radiometrically corrected, L1G - both geometrically and radiometrically corrected). This price is extremely low in comparison to other commercial forms of satellite data, thus making Landsat 7 data one of the most practical sources of quality satellite data. This distribution approach virtually ensures that satellite imagery will soon be used by more people than ever before, and for a wider range of applications. Secondly, the addition of a new panchromatic band sensor, which covers an area of approximately 180km x 180km, should broaden the role of Landsat 7 data to include new areas of operation such as monitoring urban growth, updating maps, integrating cartographic features with GIS data, and so on. In fact, Landsat 7 offers nine times the coverage of a SPOT 10-m panchromatic scene.
      Many questions arise regarding the use of the new Landsat 7 data. For example, how will Landsat 7 data be geometrically corrected so that it can be used as an image map? Which level of data should be used for the correction? How can panchromatic data be fused with the infrared data, yet preserve the spectral characteristics of each band? How does the resolution of Landsat 7 data compare with other satellite data? The answers to these questions, and more, will help end-users decide which satellite data to purchase for their applications.

Geometric Correction of Landsat 7 Data
The process of geometric correction is also known as ortho-rectification. During this process, data is corrected according to the user's ground reference system. Because the resulting output is planimetrically correct, the resulting ortho-images can then be used as maps. The two most common geometric correction methods are based on polynomial models, or parametric models.
      The polynomial method is a very simple, but outdated, method for correcting images. This method does not correct distortions that are introduced during image acquisition, nor does it take into account terrain relief distortions. This deficiency limits the polynomial method's use to small- and flat terrain areas. The polynomial method also requires numerous ground control points (GCPs), and only corrects locally around these GCPs. Since this method usually produces low-fitting accuracy (many pixels), it is not recommended for most applications, including GIS.
      A parametric model is a mathematical representation of the physical law of the transformation between the image and ground spaces. It corrects the entire image globally and also takes into consideration the distortions due to terrain. Since most parametric models have the advantage of high modeling accuracy (sub-pixel), it is the recommended method for achieving the best results.

Which Level of Correction Data?
Landsat 7 is distributed in HDF, GEOTiff and Fast formats with several levels of corrections. Level 0R is only corrected for scan-line direction and band alignment. Level 1R is level 0R with radiometric correction applied. No geometric correction has been applied to either level; however, the data consists of broken 16-line strips due to the fore- and reverse direction scanner, and it is very difficult to correct for the discontinuity without special software. Level 1G is radiometrically and geometrically corrected and does not have this discontinuity problem. Level 1G is, therefore, the one that is most recommended. Sample images of different levels can be seen on the Landsat 7 web page at http://landsat7.usgs.gov.

Landsat 7 Geometric Correction Example
To test the Landsat 7 data, a level 1G HDF format data set of the state of California was used. It took only a few days to acquire the data once it was ordered, downloaded through ftp at a cost of US$600. The data covers an area from longitude 117W to 118W, and from latitude 32.5N to 34.25N. This area has an elevation range of from zero to 3300 meters.
      PCI OrthoEngine Satellite Edition software was used for testing. This software supports reading of the data, GCP collection, geometric modeling, ortho-rectification, and either manual or automatic mosaicking.
      The parametric model used inside the OrthoEngine software was developed by the author at the Canada Centre for Remote Sensing (CCRS), Natural Resources Canada. This model is based on principles related to orbitography, photogrammetry, geodesy and cartography. It reflects the physical reality of the complete viewing geometry and corrects distortions that occur due to the platform, sensor, Earth, and cartographic projection. It has been successfully applied with few GCPs (3-6) to VIR data (Landsat, SPOT, IRS, MOS), and also to SAR data (ERS, JERS, SIR-C and RADARSAT). Based on good quality GCPs, the accuracy of this model was proven to be within one-third of a pixel for VIR images, and one resolution cell for SAR images.
      Fifteen GCPs and 15 independent checkpoints (ICPs) were collected from 1:24,000-scale USGS scale maps from both panchromatic and infrared data. It should be noted that the size of the infrared data is half the size of the panchromatic data. Hence, it is possible to collect GCPs from the panchromatic data and use them for the infrared data by multiplying the pixel and line positions by two. For the panchromatic data, the root mean square (RMS) residuals of the GCPs from the CCRS parametric model are 7.6m in X direction and 5.3m in Y direction. RMS errors of the ICPS are 6.7m and 5.2m in X and Y directions respectively.
      To ortho-rectify the data using the rigorous model and DEM, 30m resolution DEM data was obtained from the USGS ftp site. In addition, 1:100,000-scale DLG vectors were obtained from the same ftp site. These vectors overlay perfectly on the resultant ortho-rectified image. Figures 1A and 1B show a sample of ortho-rectified panchromatic and infrared images, respectively overlaid with their vectors in the area of Laguna Beach, Calif.

Landsat 7 Data Fusion for IR Enhancement
The availability of a 15m panchromatic band, in conjunction with 30m IR bands, affords the opportunity to fuse panchromatic and IR data to create an effective 15m IR image. The concept of fusion for multiband images is not new. Ever since the first remote sensing satellite, Landsat MSS data (bands 4, 6 and 7) has been spatially enhanced (from 240m to 80m resolution) by using weighted high frequency information from band 5 at 80m resolution. Band 5 was chosen as the reference due to its high scene contrast and edge content. Other techniques used different weighting coefficients for the panchromatic band and IR mulitbands. It produced an overall SPOT image that resembled a 10m color infrared photograph. Another common approach is the RGB-IHS transformation, where the high-resolution panchromatic band replaces the intensity channel derived from the lower resolution multispectral channels. Although all of these techniques yield enhanced imagery that appears to be sharper, these techniques destroy the spectral characteristics of the data.
      Atlantic Aerospace has developed a new technique founded on local modeling principles. The rationale for using a local modeling approach on satellite imagery is based on several factors. In general, edges are manifestations of object boundaries, and occur wherever there is a change in material type, lighting or topography. Since most objects exhibit broad spectral reflective characteristics, edges are evident in the panchromatic band and in most IR bands, stronger in some and weaker in others. Edge locations are the constant, and variations occur both in contrast strength and in polarity.
      Atlantic's enhancement technique uses local correlation of edges to enhance edges in the lower resolution IR data, wherever there is a corresponding edge in the panchromatic band. This enhancement is done one IR band at a time, using the same panchromatic band as a high-resolution reference. Low spatial frequency information is automatically preserved, since the technique is only adding high spatial frequency information to the edges. The enhanced data is entirely consistent with the input data. It is possible to restore the original data by blurring the enhanced data down to 30m resolution. Edges are enhanced wherever there are correlated edges in the panchromatic band. IR edges that have no corresponding edge in the panchromatic band are left untouched. The enhanced result can be used not only for visualization, but also for detailed analysis such as classification. Furthermore, classifiers that have been trained on 30m multispectral data can be used directly on 15m enhanced data as well. This technique was integrated into PCI software. Figures 2A, 2B and 2C respectively show the original IR data, the enhanced data using RGB-HIS transformation, and the enhanced data using Atlantic's enhancement technique. All data was resampled at 15m resolution using cubic convolution. It can be seen from the figure how the RGB-IHS method destroyed the spectral characteristics of the data, while Atlantic's method enhanced the resolution of the IR data and preserved the spectral characteristics of the data.

Sample Satellite Data
To illustrate the difference in resolution among satellite sensors, different satellite data of the same area are shown in this article. The satellite data includes Landsat 7 (15m panchromatic band and 30m infrared bands), SPOT (10m panchromatic and 20m multiband), IRS-1C (5.8m panchromatic), RADARSAT (12.5m pixel spacing), and ERS (12.5m pixel spacing). All data were orthorectified at 10m resolution using the same parametric model inside PCI's OrthoEngine software. Additionally, a 1:60,000-scale aerial photo scanned at 300 dpi was also used.
      Although the Landsat 7 panchromatic band is not as sharp as either the SPOT panchromatic or the IRS-1C panchromatic, the wide coverage of the data and its extremely low price make this data a very attractive option for a number of general-use or low-scale mapping applications.

Summary
Landsat 7 data provides a new opportunity for users to take advantage of satellite imagery. The data is economical, and the area it covers is extensive. To correct the imagery as an image map, a rigorously accurate model is required. A special data fusion technique, one that preserves spectral characteristics, can also be used to enhance infrared data.

About the Author:
Dr. Philip Cheng is a senior software engineer at PCI Enterprises, Richmond Hill, Ont., Canada. His e-mail address is [email protected]. Dr. Thierry Toutin is a senior research scientist at the Canada Centre for Remote Sensing, Natural Resources Canada. His e-mail address is [email protected]. Dr. Victor Tom is a research scientist at Atlantic AeroSpace, Boston, Mass. His e-mail address is t[email protected].

Back