CE394K GIS in Water Resources Fall 2002

Drainage Analysis Using DEM from Different Sources

Term project report

Larry W. Teng

http://www.ae.utexas.edu/~mave/ce394k/

Summary

Drainage analysis is one of the fundamental applications of digital elevation models (DEM). By demonstrating the impact on drainage network, this project introduces different DEM products existing nowadays, and discusses the advantages and disadvantages of each product. Also, a solution to integrate different DEMs is presented.

The following sections present some facts of different DEMs and show the strengths and weaknesses of each of them. Finally, a comparison in drainage networks using these DEMs will be discussed.

Topographic Maps and DEMs

To generating a topographic map can be a real task. It can be drawn by survey, or by extracting the depth from comparing the stereo aerial photos of the same location. Either way, it requires tremendous human efforts to delineate the shape of the terrain surface. Figure 1 shows an example of the topographic map. To read a topographic map is not intuitive for general public, and even for experienced user, erroneous interpretation of a series of contours happens occasionally. Also, its inaccuracy of vertical resolution can be another problem in applications that need precise positioning.

DEM is more visualizable as the terrain height can be represented using color wheels like gray level. The most conventional way to generate a DEM is to digitize the topographic map from either digitizing table or scanned topographic map. The topographic contours are digitized by tracing each of them manually, and stored as lines with attributes and locations. Using the numeric methods, the digitized contours can be interpolated so that smoothly varied surface results and the value of each point on this surface represents the mean height of the resolution cell.

· Figure 1 USGS topographic map over Austin area. Each of the brown contours represents the equal terrain heights. (Source: TerraServer, http://terraserver.homeadvisor.msn.com/default.aspx)

Figure 2 displays a DEM from National Elevation Datasets (NED), which is available from the Internet. This product is an improved version of the digitized DEM using some sophisticate algorithm to mosaic individual digitized topographic maps. Gray level is used to represent the heights, so the brightness is proportional to the terrain heights. Comparing to Figure 1, it is natural to interpret the terrain shape qualitatively, and the quantitative measurements are available at each resolution cell. The terrain surface is smooth because of heavy interpolation. It is suitable for stream network delineation. However, the error to the true terrain height is large, because this method is not a direct measurement of each resolution cell but the result of interpolation of neighboring cells where contours exist. It is insensitive to the variation of true heights, especially for flat terrains where the contours are widely spread. For the area in the gap between the contours, the readings are just a rough approximation.

· Figure 2 Digital elevation model of National Elevation Datasets around downtown area of Austin. The resolution shown is 30 meters.

InSAR DEM

Digitizing is a tremendous and tedious task. With the development of radar remote sensing, researchers discover the technique of radar interferometry (InSAR) to extract the height information from differencing the phase angles of two radar measurements for the same location but with slightly different look angle. The idea is quite analogous to the stereo aerial photos that technicians used to generate the topographic contours in previous section. Please refer to [1] and [2] for more information.

Radar matches the echoes of its own transmitted microwave to detect the existence of target and measure the distance between the sensors. In other words, the distances from sensor to targets are determined by the elapsed time after the transmission of the microwave burst. Another fact is that the radar emits the energy as a fan. Although the radar beamwidth is shaped to be narrow, emitting from the high-altitude platform causes a one-degree beam resulting in tens kilometers coverage. For these reasons, to avoid ambiguity, the imaging radar uses so-called “side-looking” geometry to view the terrain from the platform. Figure 3 depicts the viewing geometry of an imaging radar.

· Figure 3 Radar looking geometry. (Source: Alaska SAR Facilities. http://www.asf.alaska.edu)

Under this viewing geometry, several distortion patterns are caused in a radar map, and they may affect the outcome of the InSAR-derived DEM. Figure 4 illustrates these distortions, and the most unrecoverable one is the shadowing effect. The shadowing effect causes nulls at these locations, and there is no way to recover it except using another radar map observed from the other side, or the assistance from other data source. In principle, a DEM product that the user received should be corrected to eliminate these distortions as much as possible. However, the nature of this viewing geometry should not be ignored.

Another phenomenon that degrades the quality of an InSAR-derived DEM is the existence of noises, or “speckles,” in radar jargon. When radar listens the echoes from target, it is not preventable to record some other signals that do not belong to the specific target. Nevertheless, a lot of filtering techniques are applied on the InSAR DEM, but the user should be aware of the existence of this kind of error source.

InSAR DEM takes advantage of two similar radar maps to extract the phase information and derive it into height measurement. If at some locations of the two radar maps are not comparable, like water surface where fluctuations of waves continue happening, there is no way to calculate the phase and it is called “decorrelated.”

· Figure 4 Geometric distortion caused by side looking of imaging radar. (Source: Alaska SAR Facilities. http://www.asf.alaska.edu

The significance of the height derived from radar interferometry is its direct measurement of the geographic heights. Unlike the interpolated digitized DEM, the terrain heights are determined on pixel-by-pixel basis. A quick implication is that the terrain surface will not be as smooth as the DEM interpolated from contours. Figure 5 is an example of Shuttle Radar Topographic Mission (SRTM). It is apparent that the buildings appear on the center right of this picture but not in the contrast in Figure 1. The InSAR DEM captures a lot of useful details that represent the real-world terrain height. However, those “details” can also be caused by noises in the acquisition. Another fact to be noticed is the black areas shown on the upper portion of Figure 5, and they are the nulls caused by decorrelation of water surface.

· Figure 5 Digital elevation model from Shuttle Radar Topographic Mission around downtown area of Austin. The resolution shown is 30 meters.

LIDAR DEM

The application of laser is hot in a lot of research areas, and so is the survey science. The Laser altimetry or LIDAR, stands for light detection and ranging, forms another branch to measure the geographic heights from remote sensors. Analogous to radar, by measuring the two-way distance between the sensor and the target using laser, lidar gives a direct measurement of distance from the sensor to the target. Comparing to radar, which has a fan-like radiation pattern, though narrow beamwidth is used, the lidar is a sharp beam with high energy and scans right beneath the sensor, so the spatial resolution of lidar can be as high as one meter for airborne sensors. However, the downside is to handle the massive data points to cover the same area as conventional DEM.

By using different frequency, laser can penetrate different media. The most significant example of intermediate media is the vegetation. Figure 6 and Figure 7 show the lidar DEM over Austin area that detects the top of trees and the ground, respectively. Note that the buildings in Figure 7 at downtown are removed as the ground heights of these locations are not measurable.

(For more information about lidar, please refer to website like http://www.optech.on.ca/ or read Pierre Gueudet’s report.)

· Figure 6 Lidar DEM from ALTM of the top of the ground truth around downtown area of Austin. The resolution shown is 5 meters.

· Figure 7 Lidar DEM from ALTM of the ground height (bottom) around downtown area of Austin. The resolution shown is 5 meters.

An important feature of lidar DEM that is different from conventional DEM is that the grid of measurements can be very sparse, based on the viewing geometry of lidar. Therefore, it is generic to use triangulated irregular network (TIN) to represent the terrain surface. However, in practice, we usually interpolate the grids of lidar measurements on a regularly spaced grid like ordinary DEM.

Data Fusion

Data fusion is an active area in information science as well as in remote sensing. Data fusion is a formal framework in which are expressed the means and tools for the alliance of data originating from different sources. The basic idea is to merge useful information from different sources and result an optimized version combining desired features in each source. A multiscale approach has been developed to incorporate spatial data with different resolution to generate an optimized merge at desired resolution.

· Figure 8 Illustration of multi-scale algorithm.

The basic idea of the multi-scale algorithm is that the detailed information can be regarded as the corresponding coarse information plus individual noises at each resolution cell. The methods to control the noises in details can be categorized into two major groups: deterministic and stochastic. An example of deterministic approach is known as wavelet analysis [3], as the multi-scale Kalman filter is an example of stochastic way to merging data at different layers. Figure 8 illustrates the construction of the multi-scale algorithms.

Let us use a problem to demonstrate how the algorithm works. Suppose the bottom layer in Figure 8 with finest resolution is, say, the 5-m lidar DEM, which has some artifact at some regions. The intermediate layer can be inserted by 10-m TOPSAR InSAR DEM, which does not have artifact but consists of no-data area due to the shadowing effect somewhere. The final result from the multi-scale algorithm is desired as a map without artifact in lidar DEM and has no-data areas filled in InSAR DEM. Figure 9 shows the result before and after data fusion. It can be seen that most artifacts in the original DEMs, either in lidar or in InSAR, are attenuated in the fused result. Quantitatively, the uncertainties of estimates shown in the lower right say that the average errors in height are less than one meter. It is a great improvement, especially for InSAR sensor whose generic error is around five meter.

· Figure 9 Data fusion result associated with the uncertainty of estimates. The upper raw displays the original DEM.

The idea of data fusion can also be extended to combine data in time series. Suppose the optimal information has existed for a while and new technology is introduced to result in better quality but with some limitation that can be compensated by the previously optimized data. Figure 10 illustrate the updating experiment using three TOPSAR DEM and a coarse ERS DEM. Suppose the fist pyramid is optimized, another two decent measurements are received later. The multi-scale algorithm can be repeated to update existing optimized result. Figure 11 shows the result of Figure 10 and the significance is the attenuation of the uncertainties after three runs of the algorithm.

· Figure 10 Updating scheme using multi-scale algorithm.

· Figure 11 Result of the updating scheme. The upper row shows the estimate of height, and the lower row shows the uncertainty of the height estimates.

Comparison of drainage networks from these DEMs

Applying the terrain processing package in ArcHydro [4], the drainage network can be delineated from each DEM and the impact of each technology is discussed in this section.

The DEMs used in this project are listed in Table 1. Except for the fused DEM, references of the rest can be found on the Internet. If interested in the data fusion, please contact me. Two kinds of the reference stream networks are used and they are National Hydrological Data (NHD) and the local stream network by City of Austin GIS data. For more information about these two datasets, please refer to their websites.

DEM Name	Generating method	Resolution
NED	Digitized	Resampled to 30-m (originally 1-arcsec)
SRTM DEM	InSAR	Resampled to 30-m (originally 1-arcsec)
TOPSAR DEM	InSAR	10-m
ALTM LIDAR DEM	Lidar	Resampled to 5-m (originally 1.25-m)
Fused DEM	Data fusion	5-m basis

· Table 1 DEMs used in this project.

The DEMs used in this project are stacked and shown in Figure 12 to show the extent of each DEM. The very bottom one is NED and SRTM; both are processed at 30-meter resolution. The extent bounded by green box covers TOPSAR InSAR DEM, resolution 10 meters. The purple box represents the lidar data resampled to five meters. Finally, the red box area represents the fused DEM at five meters, too. Note that the actually used lidar data is as large as the fused DEM due to the limitation of processing.

· Figure 12 DEM stack used in this project.

Following are a series of results showing the stream network delineated using different DEMs. Figure 13 shows the streams from NED (green) and SRTM (blue), both at 30-m resolution. An interesting part in this figure is at the center where the stream defined by SRTM offsets a significant amount from both NHD and NED stream. The reason causing this problem is because the water in the lake generates few backscattering to the radar and most of this area is defined as no-data. As a result, the error at this region is large and incorrect stream delineation results.

Figure 14 shows the stream network using TOPSAR. Comparing to the 30-m SRTM result, this shows a lot of improvement and it captures several branches that are not shown in the 30-m data. Figure 15 is using lidar (blue) and fused DEM using TOPSAR and lidar (red), and they found some smaller streams that did not find in 10-m TOPSAR. However, some places in the network seem unnatural and it is caused by the interference by too many details (or noise) that confuse the direction of stream flow.

· Figure 13 The stream network delineation result using NED (green) and SRTM (blue) with City of Austin stream network. The yellow network shown here is NHD.

· Figure 14 The stream network delineation result using TOPSAR InSAR DEM (orange) with City of Austin stream network. The yellow network shown here is NHD.

· Figure 15 The stream network delineation result using lidar (blue) and the fused DEM (red) with City of Austin stream network. The yellow network shown here is NHD.

· Figure 16 Comparison to the same DEM (fused DEM using multiscale algorithm) with different reference stream network. Yellow lines are NHD, and green are City of Austin stream network. The results in red are using NHD and that in blue are using City of Austin stream network.

Figure 16 shows the comparison between using NHD and City of Austin stream network as reference to process. City of Austin stream network is a more detailed data than NHD. As a result, the stream network generated by using City of Austin stream network is closer to the truth, as it generates a better AgreeDEM for following processing steps in terrain processing package. Another fact can be observed in this figure is the stream lines near Lake Austin (upper left corner of the fused DEM.) With different reference streams, the results are quite different.

Conclusion

The factors affect the result of stream network delineation using DEM are as following:

1) Resolution is important and needs to be considered with the scale of the resulting output. The DEM at finer resolution contains massive data and the load of processing is tremendous.

2) Reference stream network is crucial to generate desired result.

3) Radar data is efficient to generate stream network but the existence of noise may ruin the result.

4) Processing lidar data over large area needs a lot of computation.

5) Either lidar or radar DEM needs to be filtered to get good result. Therefore, heavy filter can be used in this application, as the focus in this application is not the terrain surface but the derived stream network.

References

[1] Zebker, H.A., “Studying the Earth With Interferometric Radar,” Computing in Science and Engineering, IEEE, 2000.

[2] Rosen, P.A., Hensley, S., Joughin, I.R., Li, F.K., Madsen, S.N., Rodriguez, E., Goldstein, R.M., “Synthetic Aperture Radar Interferometry,” Proc. of the IEEE, IEEE, 2000.

[3] Slatton, K.C., Teng, L., and Crawford, M.M., “Multiscale Fusion of INSAR Data for Hydrological Applications,” 2002 AIRSAR Earth Science and Application Workshop, 2002. http://airsar.jpl.nasa.gov/documents/workshop2002/papers/S8.pdf

[4] Maidment, D.R., “Arc Hydro: GIS for Water Resources,” ESRI Press, 2002.