Data Set Name: National Geophysical Data Center TerrainBase Global DTM Version 1.0
The TerrainBase global digital terrain model contains a complete matrix of land elevation and ocean depth values for the entire world gridded at 5-minute intervals. NGDC/WDC-A developed the model using the best public domain data available at the time of publication.
This version of TerrainBase, called "beta test release 1.0," is the first of an ongoing sequence of global terrain models to be produced and disseminated by NGDC/WDC-A. Subsequent releases will contain higher quality data and will supersede all previous versions. The frequency of updates and level of quality enhancements will be dependent upon the contribution of new source data.
The global model was developed by integrating all of the data sets described in the previous chapter into a single model. Model development was a three-step process that entailed regridding each individual regional model to 5-minutes, then patching each regridded model together into a single global model, and, finally, smoothing a few significant discontinuities at the boundaries of adjoining models. Each of these steps is discussed in detail in the following sections. The GRASS (version 4.1) geographic information system was used for all data processing.
A term to describe the type of height values represented in the model (such as mean, mode, point, etc.) cannot be assigned to the TerrainBase model since it is comprised of a variety of source data. The type of terrain height represented by these source data sets varies from one model to the next. For example, the FNOC model has modal heights, while the USA model uses point heights, and the Europe model has mean heights. Consequently, it is not possible to assign a single term to represent all of the height values given in the TerrainBase model.
To satisfy interested users, however, a term that can be used which is common to all of the source data is best estimate of height. That is, the data represent the best estimate of height for each 5-minute cell. Although such a term has minimal scientific value, it does emphasize the fact that the data only represent reasonable estimates of the height of the terrain.
The first step in model development entailed regridding each regional model into 5-minute grids. Twenty-six regional and worldwide terrain models were used to create the TerrainBase model. Each of these input models is described individually in detail in the previous chapter. A 5-minute grid spacing was chosen for this model since this is the highest global resolution that can be supported by the individual data sets.
Two types of processes were used to regrid the source files. For source models that had grid spacing smaller than 5-minutes, a matrix filter was used to average adjoining cells into 5-minute cells. Averaging entailed taking the non-weighted mean of all input cells that fell within the boundaries of the 5-minute output cell. For some models, this process entailed point sampling the data to a smaller grid cell spacing that is an even divisor of 5-arc-minutes and then averaging these cells to create 5-minute cells. For source models that had a grid spacing larger than 5-minutes, a linear interpolation process was used to regrid the data to 5-minutes.
Averaging with a matrix filter was also used to convert 5-minute models from a grid intersection registration to a center of cell registration. Users should note that models that have been re-registered in this manner are somewhat smoothed in comparison to the original model. These include only the Australia, and Northwest Territories models.
Each of the individual input models is listed in Figure 5.1 of the TerrainBase User Manual along with the type of processing that was used in regridding to 5 minutes.
After each regional model was regridded to 5 minutes, it was integrated with the other DTMs into a single file of continuous global land and ocean coverage. The models were combined using a technique in GRASS called "patching" which permits the integration of gridded files with arbitrary spatial coverage. This technique generates an output model by assembling multiple input models in a prioritized order. The highest quality input models were assigned the highest priority, and the poorest quality models were assigned the lowest priority. Assembly takes place by using the highest priority input model and filling in areas of no data coverage with data from the model that is second in order. Areas of no data coverage are denoted by cells containing a numeric zero value. Next the third model is applied to fill in any areas not covered by the first and second model. This process continues until all models have been applied in their specified order. The advantage of this process is that it enables the highest quality input models to cover as much area of the output model as possible and using the lowest quality models only to fill in gaps where better quality data is not available. A example of this scheme is shown in the TerrainBase User Manual.
The order in which the input models were integrated is provided in the list below. Coverage of input files as they exist in the final global model are shown in Figures 5.2 through 5.9 of the user manual. U.S. coastal bathymetric data was not used in the global model due to substantial offsets between it and the ETOPO5 bathymetry data.
Ordering of Source Data used to Compile the TerrainBase Global Model
1. Italy | 10. America | |
2. Austria | 11. Northwest Territories, Canada | |
3. Netherlands | 12. Andes Mountains and Peru-Chile Trench | |
4. Madagascar | 13. Brazil Cerrados | |
5. Haiti | 14. Australia | |
6. U.S.A. | 15. Japan | |
7. Greenland | 16. ETOPO5 bathymetry | |
8. Africa | 17. FNOC topography | |
9. Europe |
The next step in model development was testing for significant errors and artifacts. A number of data spikes were identified in this process and methods were developed for correcting them. Several dozen spikes were corrected using an automated method that replaced each erroneous cell with the mean of the neighboring eight cells. Several other spikes were corrected by modifying obvious errors such as replacing missing negative signs or replacing missing zeros. Contact NGDC for a listing of corrected cells.
One artifact of producing a global model using the method described above is vertical misalignment between the borders of regional models. Vertical offsets which break the continuity of the terrain surface represented by the model have been detected along many of the boundaries where regional models adjoin. Fortunately, nearly all of the offsets are small and are well within the vertical tolerance of the 5-minute global model. The only region where the offsets are significant are in South America along the boundaries of the Brazil Cerrados DTM and in Antarctica at the southern edge of the ETOPO5 bathymetry model. Offsets along these boundaries range from tens to several hundreds of meters.
In South America, the main cause of the offsets is the lower data accuracy in the areas where the input models adjoin. Offsets are most prominent in the Amazon Basin, Paraguay, and Southern Brazil, where the quality of the source models appears relatively poor. In these areas, the density of source data used to generate the regional models are low, and consequently the quality of the models in these areas is poor. The low density of source data manifests itself as areas of low grid detail in the FNOC model and as broad undulating spurious features in the Brazil model. Where the two models adjoin, the data show practically no agreement between each model. Since there is a lack of good comparative data for the area, it is not yet possible to determine which model (FNOC, Andes, or Brazil) is more accurate along the boundaries.
A few different methods were tested to reduce the discontinuities between data sets in South America. The only method that worked satisfactorily was to smooth the data along the boundaries using a matrix filter. A 5x5 filter was applied to an area that extended 5 cell widths from each boundary. The boundaries where the filter was applied are shown in Figure 5.10. The filter has the effect of replacing each cell with the non-weighted mean of itself and the surrounding 24 neighboring cells to create a highly smoothed output grid. Figure 5.11 shows the before and after appearance of a portion of the smoothed boundary which illustrates the abrupt offset being converted to a smoothly tapered offset.
Such a modification is merely cosmetic, since filtering does not improve the reliability of the data. Boundary offsets are undesirable from a scientific standpoint whether they are smoothed or not. The only desirable option for correcting these offsets is to replace the data along the boundaries with more accurate values. Unfortunately, such a correction is not possible at this time due to the lack of reliable data for the area. Hopefully, better quality data can be obtained to alleviate this problem in future releases of TerrainBase.
Discontinuities near Antarctica were not smoothed due to the nature of the offsets in this area. Below 78 S, where the ETOPO5 bathymetry model terminates, there are no other bathymetric data available. Ocean cells south of 78 S were filled with numeric zeros which denote null values. This produces a broad, flat, zero-elevation terrace extending from 78 S to the Antarctic coast. At 78 S, a sharp drop-off occurs where the shelf meets the edge of the bathymetry model. Smoothing this boundary would have little purpose since it would entail modifying the bathymetry data, in part, to conform with the null data. In addition, smoothing would also have minimal impact on the cosmetic appearance of this discontinuity. This problem only affects the southernmost coastal areas of the Ross and Weddell Seas.
*TERRAINBASE_help TerrainBase Global Terrain Model Summary Documentation (modified from TerrainBase CD-ROM) Description Data Set Name: National Geophysical Data Center TerrainBase Global DTM Version 1.0 File Name: TBASE.BIN Coverage Area Coverage: Worldwide Coordinate Coverage: 90 N to 90 S x 180 W to 180 E Topography/bathymetry/both: Both Grid Structure Cell Dimensions: Latitude: 05'00" Longitude: 05'00" (that is, 5-arc-minute latitude-long grid) Cell Registration: Center of cell Total Grid Rows: 2160 Total Grid Columns: 4320 Data Characteristics Elevation Units: Meters Elevation Type: varies by (perhaps within) data source Projection: Latitude/longitude (no projection) Vertical Accuracy: variable Horizontal Accuracy: variable Null Land Value Designator: N/A Null Ocean Value Designator: 0 File Structure Byte binary data (no headers, no carriage returns or line feeds) Byte ordered for IBM-compatible PCs Byte swapping needed for some computers, such as many UNIX workstations, such as Apple Macintoshes Origin of the file: 90 deg. North, 180 deg. West Data scan: eastward accross columns; downward by row. Source Data Developer: Lee W. Row, III and David Hastings National Geophysical Data Center and World Data Center-A for Solid Earth Geophysics Boulder, Colorado U.S.A. Model Development Overview The TerrainBase global digital terrain model contains a complete matrix of land elevation and ocean depth values for the entire world gridded at 5-minute intervals. NGDC/WDC-A developed the model using the best public domain data available at the time of publication. This version of TerrainBase, called "beta test release 1.0," is the first of an ongoing sequence of global terrain models to be produced and disseminated by NGDC/WDC-A. Subsequent releases will contain higher quality data and will supersede all previous versions. The frequency of updates and level of quality enhancements will be dependent upon the contribution of new source data. The global model was developed by integrating all of the data sets described in the previous chapter into a single model. Model development was a three-step process that entailed regridding each individual regional model to 5-minutes, then patching each regridded model together into a single global model, and, finally, smoothing a few significant discontinuities at the boundaries of adjoining models. Each of these steps is discussed in detail in the following sections. The GRASS (version 4.1) geographic information system was used for all data processing. A term to describe the type of height values represented in the model (such as mean, mode, point, etc.) cannot be assigned to the TerrainBase model since it is comprised of a variety of source data. The type of terrain height represented by these source data sets varies from one model to the next. For example, the FNOC model has modal heights, while the USA model uses point heights, and the Europe model has mean heights. Consequently, it is not possible to assign a single term to represent all of the height values given in the TerrainBase model. To satisfy interested users, however, a term that can be used which is common to all of the source data is best estimate of height. That is, the data represent the best estimate of height for each 5-minute cell. Although such a term has minimal scientific value, it does emphasize the fact that the data only represent reasonable estimates of the height of the terrain. Regional Model Regridding The first step in model development entailed regridding each regional model into 5-minute grids. Twenty-six regional and worldwide terrain models were used to create the TerrainBase model. Each of these input models is described individually in detail in the previous chapter. A 5-minute grid spacing was chosen for this model since this is the highest global resolution that can be supported by the individual data sets. Two types of processes were used to regrid the source files. For source models that had grid spacing smaller than 5-minutes, a matrix filter was used to average adjoining cells into 5-minute cells. Averaging entailed taking the non-weighted mean of all input cells that fell within the boundaries of the 5-minute output cell. For some models, this process entailed point sampling the data to a smaller grid cell spacing that is an even divisor of 5-arc-minutes and then averaging these cells to create 5-minute cells. For source models that had a grid spacing larger than 5-minutes, a linear interpolation process was used to regrid the data to 5-minutes. Averaging with a matrix filter was also used to convert 5-minute models from a grid intersection registration to a center of cell registration. Users should note that models that have been re-registered in this manner are somewhat smoothed in comparison to the original model. These include only the Australia, and Northwest Territories models. Each of the individual input models is listed in Figure 5.1 of the TerrainBase User Manual along with the type of processing that was used in regridding to 5 minutes. Data Integration After each regional model was regridded to 5 minutes, it was integrated with the other DTMs into a single file of continuous global land and ocean coverage. The models were combined using a technique in GRASS called "patching" which permits the integration of gridded files with arbitrary spatial coverage. This technique generates an output model by assembling multiple input models in a prioritized order. The highest quality input models were assigned the highest priority, and the poorest quality models were assigned the lowest priority. Assembly takes place by using the highest priority input model and filling in areas of no data coverage with data from the model that is second in order. Areas of no data coverage are denoted by cells containing a numeric zero value. Next the third model is applied to fill in any areas not covered by the first and second model. This process continues until all models have been applied in their specified order. The advantage of this process is that it enables the highest quality input models to cover as much area of the output model as possible and using the lowest quality models only to fill in gaps where better quality data is not available. A example of this scheme is shown in the TerrainBase User Manual. The order in which the input models were integrated is provided in the list below. Coverage of input files as they exist in the final global model are shown in Figures 5.2 through 5.9 of the user manual. U.S. coastal bathymetric data was not used in the global model due to substantial offsets between it and the ETOPO5 bathymetry data. Ordering of Source Data used to Compile the TerrainBase Global Model 1. Italy 10. America 2. Austria 11. Northwest Territories,Canada 3. Netherlands 12. Andes Mountains and 4. Madagascar Peru-Chile Trench 5. Haiti 13. Brazil Cerrados 6. U.S.A. 14. Australia 7. Greenland 15. Japan 8. Africa 16. ETOPO5 bathymetry 9. Europe 17. FNOC topography The next step in model development was testing for significant errors and artifacts. A number of data spikes were identified in this process and methods were developed for correcting them. Several dozen spikes were corrected using an automated method that replaced each erroneous cell with the mean of the neighboring eight cells. Several other spikes were corrected by modifying obvious errors such as replacing missing negative signs or replacing missing zeros. Contact NGDC for a listing of corrected cells. Reducing Boundary Discontinuities One artifact of producing a global model using the method described above is vertical misalignment between the borders of regional models. Vertical offsets which break the continuity of the terrain surface represented by the model have been detected along many of the boundaries where regional models adjoin. Fortunately, nearly all of the offsets are small and are well within the vertical tolerance of the 5-minute global model. The only region where the offsets are significant are in South America along the boundaries of the Brazil Cerrados DTM and in Antarctica at the southern edge of the ETOPO5 bathymetry model. Offsets along these boundaries range from tens to several hundreds of meters. In South America, the main cause of the offsets is the lower data accuracy in the areas where the input models adjoin. Offsets are most prominent in the Amazon Basin, Paraguay, and Southern Brazil, where the quality of the source models appears relatively poor. In these areas, the density of source data used to generate the regional models are low, and consequently the quality of the models in these areas is poor. The low density of source data manifests itself as areas of low grid detail in the FNOC model and as broad undulating spurious features in the Brazil model. Where the two models adjoin, the data show practically no agreement between each model. Since there is a lack of good comparative data for the area, it is not yet possible to determine which model (FNOC, Andes, or Brazil) is more accurate along the boundaries. A few different methods were tested to reduce the discontinuities between data sets in South America. The only method that worked satisfactorily was to smooth the data along the boundaries using a matrix filter. A 5x5 filter was applied to an area that extended 5 cell widths from each boundary. The boundaries where the filter was applied are shown in Figure 5.10. The filter has the effect of replacing each cell with the non-weighted mean of itself and the surrounding 24 neighboring cells to create a highly smoothed output grid. Figure 5.11 shows the before and after appearance of a portion of the smoothed boundary which illustrates the abrupt offset being converted to a smoothly tapered offset. Such a modification is merely cosmetic, since filtering does not improve the reliability of the data. Boundary offsets are undesirable from a scientific standpoint whether they are smoothed or not. The only desirable option for correcting these offsets is to replace the data along the boundaries with more accurate values. Unfortunately, such a correction is not possible at this time due to the lack of reliable data for the area. Hopefully, better quality data can be obtained to alleviate this problem in future releases of TerrainBase. Discontinuities near Antarctica were not smoothed due to the nature of the offsets in this area. Below 78 S, where the ETOPO5 bathymetry model terminates, there are no other bathymetric data available. Ocean cells south of 78 S were filled with numeric zeros which denote null values. This produces a broad, flat, zero-elevation terrace extending from 78 S to the Antarctic coast. At 78 S, a sharp drop-off occurs where the shelf meets the edge of the bathymetry model. Smoothing this boundary would have little purpose since it would entail modifying the bathymetry data, in part, to conform with the null data. In addition, smoothing would also have minimal impact on the cosmetic appearance of this discontinuity. This problem only affects the southernmost coastal areas of the Ross and Weddell Seas. *TERRAINBASE ANCILLARY ENVIRONMENTAL DATA TerrainBase Digital Elevation Model #\data\ncillary\tbase.img