The exceptional geometry of Landsat 8 OLI/TIRS data provides an opportunity to improve the reference database used to precisely and accurately geolocate all Landsat 1-8 Level-1 data products. Landsat Science products inherit the geometry of Landsat Level-1 data products. Accurate geometry ensures that data pixels are aligned, and that the data can be used easily in time series analysis.
- Geometric Accuracy
- Levels of Processing
- Ground Control Points (GCPs)
- Spatial Performance of Landsat 8 Instruments
To assure Landsat Level-1 data are suitable for time-series analysis, products need to be co-registered. The root-mean-square error (RMSE) reported in the metadata (MTL.txt) file can be used to filter the precision and terrain corrected Level-1 data products to meet application specific requirements. Landsat processing levels and the related accuracies of each are described below.
Landsat scenes are processed to a Level-1 precision and terrain corrected product (L1TP), if possible. In the case of insufficient reference data, a systematic and terrain corrected L1GT or a systematic L1GS product will be created instead. L1GT products are created when the systematic product has consistent and sufficient locational accuracy to permit the application of a terrain model. L1GS products are created when the locational accuracy is not sufficient to apply terrain correction. Three primary reasons L1GS scenes are created include 1) insufficient number of ground control points, such as small islands or Antarctica, 2) opaque clouds that obscure the ground, or 3) locational errors greater than the search distance for ground control.
Levels of Processing
Precision and Terrain Correction (Level-1TP, L1TP)
Precision and Terrain Correction provides radiometric and geodetic accuracy by incorporating ground control points while employing a Digital Elevation Model (DEM) for topographic displacement. Geodetic accuracy of the product depends on the image quality and the accuracy, number, and distribution of the ground control points (GCP):
- Ground control points used for L1TP correction are currently based on the Global Land Survey (GLS) reference database. The reference database is being revised, using Landsat 8 Operational Land Imager (OLI) data within the GCP improvement plan.
- The elevation data used for relief displacement of the L1TP data include Shuttle Radar Topography Mission (SRTM), National Elevation Dataset (NED), Canadian Digital Elevation Data (CDED), Digital Terrain Elevation Data (DTED), Global 30 Arc-Second Elevation (GTOPO30), and Greenland Ice Mapping Project (GIMP) source DEMs.
- Precision fit and verification RMSE estimates are only available for L1TP products. The precision fit estimate (RMSE_Model) quantifies how well the control points used in the precision registration match the reference GCP database. The verification estimate (RMSE_Verify) of MSS and TM data quantifies how well the image matches an independent set of GCPs in the reference GCP database.
- The specification for L1TP product acceptance varies by sensor. The specification is rigid for Landsat 8, Landsat 7, and Landsat 4-5 Thematic Mapper (TM) data, many of which have excellent internal geometry. Given the poor internal geometry of the Multispectral Sensor (MSS) aboard Landsat 1-5, the use of ground control even for data with large RMSE was considered preferable, to creating a large proportion of data as L1GS with the internal geometry uncorrected.
The information provided in the metadata file can be used to evaluate the geodetic accuracy of the L1TP data product.
Systematic Terrain Correction (Level-1GT, L1GT)
Systematic Terrain Correction provides systematic, radiometric, and geometric accuracy, while employing a Digital Elevation Model (DEM) to correct for relief displacement:
- Landsat 7 scenes without sufficient control to produce L1TP images are processed to an L1GT.
- Landsat 8 scenes without sufficient ground point control to produce L1TP products are processed as L1GT. The accuracy of the L1GT systematic product approaches that of an L1TP product. Registration to the shared Ground Control Point reference data set improves the co-registration to the other Landsat sensors. For scenes where the reference database error exceeds 30 meters, the L1GT images will have better absolute accuracy than Landsat 8 L1TP data, but may not be co-registered to within 30 meters.
Landsat 7 and Landsat 8 data over Antarctica are processed to an L1GT, since it has not been possible to generate ground control in Antarctica suitable for the generation of an L1TP product. The Radarsat Antarctic Mapping Project Digital Elevation Model Version 2 (RAMP V2 DEM) is the terrain correction source for Antarctica.
Systematic Correction (Level-1GS, L1GS)
Systematic Correction provides systematic radiometric and geometric corrections, which are derived from data collected by the sensor and spacecraft.
- Landsat scenes processed as L1GS do not have sufficient geodetic accuracy to include in image-to-image analysis without further image-specific evaluation and registration.
- Landsat ETM+ geometric accuracy of the systematically corrected product should be within 125 meters 90 percent of the time for low-relief areas at sea level based on pre-fit estimates. Error increases as distance and elevation increase from low relief areas.
- Landsat TM geometric accuracy for L1GS products should be within 700 meters, 90 percent of the time for low-relief areas at sea level based on pre-fit estimates.
- Landsat MSS geometric accuracy for L1GS products is substantially worse than later sensors. Both the internal geometry and locational accuracy will require manual registration of the images
Landsat TM and MSS images may be offset from its correct spatial location by thousands of meters, preventing the use of terrain correction for systematic products.
The success rate for creating L1TP products varies by sensor, but also by cloud cover. The table below displays each cloud-cover class and lists the proportion of images that process to an L1TP, those that fallback to an L1GT or L1GS after a failed attempt to produce an L1TP (clouds or poor ephemeris data), and the proportion that are planned to produce an L1GT (night, Antarctica or insufficient land features). These values were generated from the most recent Landsat Product Generation System (LPGS) software version.
Figure 1. Level-1 Products Registration Success by Cloud Cover, based on LPGS release version 12.8.0, August 2016
*Landsat 1-7: All path/row combinations are separated internally into two groups; path/rows that will undergo precision and terrain correction, and path/rows that will not be precision and terrain corrected. If precision and terrain correction is attempted and successful, the scene becomes an L1TP. If precision correction is unsuccessful, the scene becomes an L1GT FB (fallback). If precision and terrain correction is not applied, the scene will become an L1GS (or L1GT for Landsat 7).
*Landsat 8: All path/rows are separated into two groups; path/rows that have produced at least one L1TP scene, and path/rows that have never produced an L1TP scene. Precision and terrain correction is attempted on all Landsat 8 scenes. If precision correction fails from the path/row group that has produced at least one L1TP, the scene becomes an L1GT FB (fallback). If precision correction fails from a path/row that has never produced an L1TP (which is likely), then the scene will become an L1GT (no fallback).
RMSE Distribution Plots
The distribution of the RMSE for each sensor shows significant improvement as sensor and spacecraft technologies evolve.
Figure 2. Landsat 8 OLI Collection 1 L1TP RMSE: Less than 12 meters in 92.4 percent of the data
Figure 3. Landsat 7 ETM+ Collection 1 L1TP RMSE: Less than 12 meters in 96.5 percent of the data
Figure 4. Landsat 4-5 TM Collection 1 L1TP RMSE: Less than 12 meters in 96.2 percent of the data
Figure 5. Landsat 1-5 MSS Pre-Collection L1T RMSE: Less than 12 meters in 0.2 percent of the data
Ground Control Points (GCPs)
Ground Control Points (GCPs) are defined as points on the surface of the earth of known location used to geo-reference Landsat Level-1 imagery. The Landsat Ground Control Point Search allows you to extract ground control point binary files over your area of interest.
Updates to GCPs
In 2014, improvements to GCP files began – these include removing outdated files, and creating new, time-specific and seasonal GCP’s. The GCP Improvement Plan provides an overview of the goals and plans for the effort. Because the updated GCPs will be more accurate, users may consider re-ordering previously downloaded images, as new GCPs may allow scenes that once could only be processed to systematic L1GT products to be processed as a precision and terrain corrected L1TP.
February 2014: GCPs were created for the following WRS-2 paths/rows: 138/41, 151/30, and 172/43.
These new GCPs allowed Landsat 8 scenes that once processed only to systematic Level-1 (L1GT) product to process a precision and terrain corrected Level-1T (L1TP). Additionally, GCPs that were contained within large bodies of water were removed from use during Level-1 processing. Although the presence of these points typically has little effect on the quality of the geometric registration of the L1TP data, the removal of these GCPs indicated that generating L1TP products was possible.
August 2014: Phase 1
Phase 1 updated the accuracy within a number of GCPs in high priority areas, known to have large offsets in the original points. The new GCPs allowed scenes that once created only a systematic Level-1GT product to process a precision Level-1TP. 171 paths/rows covering the following areas were updated:
- Island and coastal locations where meager coverage existed (lack of reference data in the original triangulation, few GCPs available, etc.)
Desert regions where shifting dunes caused existing GCPs to become obsolete and other anomalous locations in which an apparent bias remained.
Phase 1 updates were implemented in the following Landsat Product Generation System (LPGS) releases (noted in the MTL file):
- Landsat 8: PROCESSING_SOFTWARE_VERSION = “LPGS_2.4.0”
Landsat 1-7: PROCESSING_SOFTWARE_VERSION = "LPGS_12.5.0"
The GCP Improvement Plan provides the triangulation results for the paths/rows improved in Phase 1.
April 2015: Phase 1-Software Fix
In April 2015, a software issue was fixed. The GCP image chips used to perform the precision correction for L1TP products are in a UTM projection. When the set of GCPs to use is identified for a specific scene, some of the chips might be in different UTM zones – so some of the image chips are reprojected to get all the chips into the correct UTM zone for the scene being processed.
It was discovered that the chip reprojection software had a rounding bug that added a bias to the image chip location. On average, the bias would be roughly half of a 30m pixel. The impact was considered minor enough that no data has been reprocessed to correct this. The impact should always be less than half a pixel and usually much, much less than that (depending on how many chips with the location bias end up being used in the precision solution).
Landsat 8: PROCESSING_SOFTWARE_VERSION = “LPGS_2.5.0” for this correction.
January 2016: Phase 2
Phase 2 includes 1,151 WRS-2 paths/rows, covering island areas and inland regions where the estimated absolute accuracy of the original Global Land Survey 2000 (GLS2000)-based ground control varied between 50-75 meters. This update affected:
- Middle to low latitude regions
Specific updates to GCPs used over Australia in order for that region to be more consistent with the Australian Geographic Reference Image (AGRI).
The metadata file (MTL.txt) delivered with the reprocessed Level-1 scenes include a new field, to indicate newly processed data:
- The GROUND_CONTROL_POINTS_VERSION parameter will display version number “3”: [GROUND_CONTROL_POINTS_VERSION=3]
The FILE_DATE will be updated with a date after January 11, 2016
May 2016: Phase 3
Phase 3 updated 918 path/rows, covering high latitude arctic regions where the existing GLS2000-based ground control points were found to contain errors of 50 meters or more. In addition to correcting the existing GCPs and adding new GCPs where the existing coverage is sparse or absent, this update implements improvements to the digital elevation model (DEM) used for Greenland, and for the islands of Svalbard and Jan Mayen Land. This DEM update uses more accurate data recently created by the Greenland Ice Mapping Project (GIMP) and the Norwegian Polar Institute (NPI).*
*The new Greenland DEM data are described in this publication:
Howat, I., A. Negrete, and B. Smith. 2015. MEaSURES Greenland Ice Mapping Project (GIMP) Digital Elevation Model, Version 1. The 90-meter resolution data set was used. Boulder, Colorado USA. NASA National Snow and Ice Data Center Distributed Active Archive Center. doi:http://dx.doi.org/10.5067/NV34YUIXLP9W. Accessed 29 January 2016.
*The new NPI DEMs for Svalbard and Jan Mayen Land are described in the following publications:
Norwegian Polar Institute. (2014). Terrengmodell Svalbard (S0 Terrengmodell). The 50-meter resolution data set was used. Tromsø, Norway: Norwegian Polar Institute. https://data.npolar.no/dataset/dce53a47-c726-4845-85c3-a65b46fe2fea Accessed 10 February 2016.
Norwegian Polar Institute. (2014). Terrengmodell Jan Mayen (J0 Terrengmodell). The 25-meter resolution data set was used. Tromsø, Norway: Norwegian Polar Institute. https://data.npolar.no/dataset/e2b2417e-9926-4519-b6a9-7eefb3bb1012 Accessed 10 February 2016.
Spatial Performance of Landsat 8 Instruments
During prelaunch testing, performance analysis was conducted on the spatial response of Landsat 8's Operational Land Imager (OLI) and Thermal Infrared Sensors (TIRS). This page briefly describes how this performance was measured. Relevant data were received from Ball Aerospace for the OLI sensor, and the TIRS sensor data were collected at NASA Goddard Space Flight Center (GSFC). This spreadsheet provides the measured point spread function values for each spectral band in each instrument: .xlsx ( 514 KB)
Subsequent on-orbit analyses have not detected significant changes to OLI or TIRS spatial performance since launch.
OLI Point Spread Functions
OLI spatial performance was measured prelaunch by scanning edge targets across the OLI detectors in the along- and across-track directions.
These scans were performed in discrete steps so that the target was stationary at each sub-pixel location; so the measurements did not contain detector integration effects.
Edge response measurements were made in all spectral bands and across the full OLI field of view. Figures 6 and 7 respectively show sample across-track and along-track edge response profiles.
Figure 6. Sample across-track edge response profiles for five consecutive detectors in the first OLI sensor chip assembly (SCA01 detectors 55 through 59) in the red band. The vertical axis is normalized response (dimensionless) and the horizontal axis is target position in units of microradians. Note the sequential detector arrangement in the across-track direction.
Figure 7. Sample along-track edge response profiles for five consecutive detectors in the fourth sensor chip assembly (SCA04 detectors 245 through 249) in the cirrus band. The vertical axis is normalized response (dimensionless) and the horizontal axis is target position in units of microradians. Note the aligned even detector then odd detector arrangement in this direction. OLI even and odd detectors are vertically offset/staggered, creating this effect.
The edge response measurements for each band were differentiated to construct band-average along-track and across-track line spread functions.
Detector integration effects, computed based upon the nominal detector integration times, were then analytically included in the along-track line spread functions to reflect on-orbit performance.
The OLI spatial response for each band is represented in terms of two separable along- and across-track line spread functions which can be combined to construct a two-dimensional point spread function. The line spread functions are sampled at 4 microradian intervals (2 microradians for the panchromatic band) providing approximately ten samples per pixel.
Figure 8 shows a plot of the along- and across-track line spread functions for OLI band 3 (green band).
Figure 8. Across-track and along-track line spread functions for OLI band 3 (green). Note the effect of detector integration in the along-track function. The OLI multi-spectral detectors integrate for approximately 85 percent of the sample time.
TIRS Point Spread Functions
TIRS spatial performance was measured prelaunch using images of a disk target approximately 16 pixels in diameter.
The TIRS focal plane consists of three sensor chips, each of which contains a two-dimensional detector array from which two science rows in each spectral band are sampled during normal imaging operations. The instrument also has a diagnostic mode in which all detector rows can be read out. It is this mode that was used for spatial characterization. Two-dimensional images of the disk target were acquired as the target was shifted in 0.2 pixel increments across the focal plane.
A sample disk image is shown in Figure 9. The resulting images of the disk at a range of sub-pixel offsets were then used to estimate the parameters of a two dimensional model of the TIRS spatial response for each spectral band.
Figure 9. Sample TIRS image of 16-pixel disk target.
A two dimensional formulation was adopted to allow for an optical transfer function that may not be separable or symmetric. The fitted models were then used to calculate the two dimensional point spread function for each band. The TIRS point spread functions were generated in two dimensions (i.e., the functions were not assumed to be separable) at 0.1 pixel sampling, corresponding to 10 meter ground sampling.
A plot of the point spread function for the 10.8 micron band (band 10) is shown in Figure 10.
Figure 10. Plot of the 2D point spread function for the TIRS 10.8 micron band (band 10).