Significant Topographic Changes in the United States
Filtering of Elevation Difference Mask to Identify True Topographic Changes
The output of the differencing and thresholding procedures described above was a set of pixels with elevation differences large enough to be judged significant. Ideally, these differences would represent the set of true topographic changes that could then be used for further analysis. In reality, this was not the case, as the selected differences included areas that clearly did not reflect topographic changes but were related to characteristics of the input elevation datasets. Even though the NED and SRTM data were processed to register them as precisely as possible, there remained residual offsets in some areas that were large enough to cause vertical differences that exceeded the significance thresholds.
This condition was not consistent across the study area. In some areas the registration of the NED and SRTM data was excellent, whereas in other areas the residual offset was very noticeable. The cause of the misregistration is not known; however, it may be caused by variations in the internal geometry of the SRTM and possibly the NED data. The false vertical differences due to the residual horizontal offsets were most prevalent in the higher relief areas in the desert southwest of the United States as seen in the figure below. In these higher relief areas, the misregistration may be slight but can still lead to significant vertical differences.
Another primary cause for differences being falsely included in the change mask is the first return nature of SRTM elevation data. Even though the accuracy-based threshold accounts for variable accuracy by land cover type, some areas of vegetation and built structures are the sole cause of SRTM – NED differences large enough to exceed the threshold. Even though the geographic distribution of the reference control point dataset used for accuracy assessment was extensive, it may not have been truly representative of the full range of land cover conditions. Thus, the accuracy-based thresholds do not capture all the variability in vertical accuracy due to land cover. The figure below presents an example of the presence of a vegetation canopy falsely being detected as topographic change.
Extensive viewing and examination of the results from the thresholding operations made it clear that some filtering and refinement of the change mask was necessary to limit it to areas of true topographic change. The threshold results were overlaid on shaded relief images from the NED and SRTM data, Landsat satellite images, aerial photography, and various topographic maps to visually check the quality of the results in different terrain settings. This visual examination was critical for developing a set of criteria with which to filter the selected elevation differences. Visual inspection of areas contained in the change mask overlaid on reference data is an effective method of verifying if real topographic changes exist; however, such a manual approach is impractical for the large study area used in this investigation. Therefore, a set of empirical decision rules based on observed conditions was developed and applied systematically to the significant elevation differences across the conterminous United States. Previous studies using difference grids derived from multitemporal elevation data and a thresholding approach to detect topographic changes also included a post-threshold filtering step to eliminate false detections.
The filtering criteria are based on specific characteristics of the difference areas, and these characteristics are calculated directly from the input elevation data and land cover datasets. Although performed on a pixel-by-pixel basis, the threshold procedure results in groups of connected pixels that form separate entities. Each of these groups of contiguous pixels was treated separately and attributed as a distinct feature. The following attributes were assigned to each feature:
The changes in elevation, relief, slope, and aspect were calculated by differencing those terrain parameters as derived from the NED and SRTM data. The majority land cover class was assigned from the NLCD. The volume was calculated from the difference grid. Proximity to known mining locations was determined from the NLCD.
Proximity to SRTM data voids was calculated from the SRTM data. Voids in SRTM data occur where reliable elevations could not be calculated from the radar data because of layover, shadowing, or low interferometric correlation. They occur infrequently, but where they do exist they often fall on steep slopes. The figure below shows an example of SRTM data voids. It is important to note their location for filtering criteria because significant elevation differences that occur immediately adjacent to the voids are often false changes that are selected only because of slight misregistration of the NED and SRTM data in steep terrain.
The NED source DEM production method code is assigned from the NED spatially referenced metadata. This is an important consideration for the filtering criteria because some areas in the NED are derived from older USGS 7.5-minute DEMs made with obsolete photogrammetric methods. These methods often left artifacts in the data, and this noise can create areas of erroneous differences when the NED is subtracted from the SRTM data (see figure below).
A test group of 30 1x1-degree tiles was examined intensively to develop the set of decision rules to filter and refine the change mask. The test group in the figure below necessarily included a wide range of terrain, land cover, and data quality conditions, and it represents a sample of about 3 percent of the 934 tiles used as data processing units.
Interactive analysis was conducted by overlaying the areas selected as significant differences from the thresholding procedure on ancillary reference data. This visual check allowed for determination of which features in the change mask represented actual topographic modifications and for the recording of the characteristic range of attribute values for those features. The range of attribute values for the real change features formed the basis for the filtering decision rules. Whereas the differencing, thresholding, and assignment of terrain and land cover attributes had to be performed on gridded (raster) data, the distinct groups of connected pixels of significant differences were then converted to polygon (vector) feature data, which facilitated the filtering by allowing the polygon attributes to be queried and analyzed in a relational database environment.
The primary objective of the filtering process was to reduce the errors of commission as much as possible, even at the expense of omitting areas that appeared visually to have undergone topographic change but their attributes did not meet the filtering criteria. This goal is consistent with the thresholding procedure in which areas had to pass strict statistical criteria to be considered as candidate topographic change areas, tests that undoubtedly eliminated some areas of actual change. If the area of interest for a topographic change study is small, then the set of elevation differences that represent actual topographic changes could be selected manually with a very high degree of accuracy, and the statistically based thresholding and filtering processes could be avoided altogether. However, when the study area is large and data quality is variable, as is the case for this study, automated procedures must be used that result in a manageable set of features. Thus, a conservative approach to labeling polygons as areas of true topographic surface change is the method that has been implemented for this study.
The primary criteria for acceptance of a polygon into the set of significant topographic surface changes involve the degree to which major terrain parameters as measured from the NED and SRTM data have changed. The parameters of interest are mean elevation, relief, slope, and aspect. If a combination of these parameters have changed enough, then the polygon under consideration is deemed acceptable as a representation of significant geomorphic change. As a matter of practicality, before the terrain parameters are checked, polygons of very small size are eliminated from further consideration. Polygons comprising areas of less than 50 contiguous pixels, which equates to an area of about 0.045 square kilometers, were deleted from the change mask. The designation of this size threshold was based on the experience of visual examination, and in practice this rule was quite effective at removing a lot of background noise from the change mask, which greatly aided interpretability of the remaining polygons.
Based on visual analysis of the test tiles, it was observed that the following conditions were generally true for polygons that delineated areas of actual topographic change:
The attributes of each candidate polygon were checked to see how many of these four conditions were met. Those polygons with three or four of the conditions being met were retained in the change mask, while those with only one or two conditions being met were eliminated. This filter greatly reduced the number of false detections of change caused by residual misregistration in higher relief areas or the presence of non-ground-level surfaces (forest canopy or built structures). Subsequent visual analysis indicated that the change mask still had some problems with delineations of the areas of actual topographic change, so a series of additional criteria was tested and implemented as sequential filter steps to address specific conditions.
The table below outlines the filter steps applied to obtain the final change mask. As listed in the table, the sequential filter steps make extensive use of the terrain parameters measured from the elevation datasets as well as land cover information from the NLCD.
The input change mask to the filtering and refinement process consisted of all the polygons from the individual 1x1-degree processing units merged into a feature dataset covering the entire study area. The change mask contained 219,156 distinct polygons outlining areas of significant elevation differences from the thresholding process (and after the minimum size criterion of 50 contiguous cells had been applied). The primary filter that checks for at least three of four terrain parameter change criteria being met reduced the number of candidate polygons in the change mask by more than 88 percent to 26,064 polygons. The subsequent filter steps (steps 3 through 10 in the table above) further reduced the number to 5,263 polygons, or 2.4 percent of the prefiltering total. These polygons represent the final set of features that delineate significant topographic surface changes across the conterminous United States.
An example of an area in West Virginia helps to illustrate the effects of the filtering and refinement applied to polygons outlining significant elevation differences. For the 1x1-degree tile extending between 80 and 81 degrees west longitude and between 38 and 39 degrees north latitude, 8,252 polygons that outlined areas of significant elevation differences originally passed the elevation thresholds. Most of these polygons were very small and were eliminated by the first filter step that checks for a minimum size of 0.045 square kilometers. After the size criterion was applied, 293 polygons remained, and after all the filter steps were applied a final total of 41 polygons were included as topographic change polygons.
This figure displays a small area within the example tile that shows some polygons that were eliminated and some that survived the filtering process. In this case, even the eliminated polygons appear to outline areas that have experienced topographic change. However, these polygons were eliminated because they did not meet the requirement for change in terrain parameters (filter step 2).
As part of the filtering and refinement process, polygons delineating areas that were entirely flat in the SRTM data were extracted and treated separately. These areas form a special case of surface change because most of them represent reservoirs or water bodies that have changed in water level. Detection of these features was aided by the fact that the SRTM data, as part of its production process, had water bodies and shorelines edited to reflect conditions present during the February 2000 data collection period. Interferometric radar data generally have little useful information about the elevation of water bodies because of the weak returns of the radar signal from water, so an extensive editing process was performed on the SRTM data to ensure correct representation of water features. Flat areas in the SRTM data that coincided with significant elevation increases often indicate a new reservoir (if the underlying NED data are not flat) or a significantly increased water level in an existing reservoir or lake (if the underlying NED data are flat and the NLCD indicates a cover type of water). It was also observed that some of the flat features in the SRTM data with a corresponding significant increase in elevation represent features associated with surface mining operations, such as settling ponds and graded spoil banks. A total of 364 polygons were extracted by segmenting flat areas with significantly increased elevations from the SRTM data, and each of these polygons was labeled "new," "existing," or "mining" depending on the relief (or the lack thereof) in the NED and the NLCD class designation for the area.