• Converted the points to 1" cells in the space of the Global DEM. As the GDEM cell centers fall on the edges of Hydrosheds cells, there is a uniform 0.5" shift, but this does not worry us. In fact it partly compensates for the 1.5" NE shift of the HydroSHEDS data.
• Calculated the locus of 1" grid cells closest to each streampoint, clipping to an arbitrary buffer of .05 degrees. These regions, called Thiessen polygons or Voronoi cells before clipping, were computing with the eucallocation function in ARC/INFO.
• Found the lowest elevation within each region. (zonalmin)
• Assigned that elevation back to the point.
• For diagnostic purposes, those lowest cells were converted to points and are displayed below as stars. In the case of ties, more than one point may be shown.

	The Wang Chhu is the major river of Bhutan. It continues across the India before joining with other rivers, entering Bangladesh, and flowing toward the Brahmaputra.
Here are three versions of the profile of the Wang Chhu. The hydroSHEDS profile looks well behaved, though the flat regions are unlikely to be real. The SRTM DEM is shows some spikiness. It would have looked better if we had been able to obtain a 3" DEM generated as the minimum of the 9 component 1" values. The ASTER GDEM is more spiky, and it shows some low spots that we would like to take a look at.

Above is the SRTM image of an area in the second box of the context map. The graph on the left is a detail of the full river profile. This area concerns us. If the GDEM contains bogus pits, our data filtering becomes much more difficult. The image on the right shows the 3" hydroSHEDS DEM, the profile created from it, and red points selected where the GDEM elevations were selected. The two big blue points are the troublesome pit on the graph.

This is a [15m lossy] Landsat image of the same area. It is flipped upside down for those of us who cannot interpret aerial/satellite photos illuminated from the bottom.

This is the 1" Global DEM with a special color stretch. The dip in our river profile is clearly visible as a feature of the river channel, comprising many data points and centered on the channel. This leads us to conclude that these are legitimate elevations, and that apparently higher areas downstream are artifacts of tree cover. The two pixels have 9 and 10 components values (The minimum value from 9 ASTER scenes was used.). This is typical for the area, though it seems that high points on the profile have lower data quality. We will look at this statistically later.

Here is one artifact of my clever method. Profile points has a small search area on the inside of curves. If the actual channel lies to the inside of the curve, that particular HydroSHEDS point is inhibited from searching the channel for a data point. The result is a spike in the profile. This is not a big problem, as we expect upward spikes and remove them.
Here is a similar artifact. Because of a poor match between the profile and the actual channel, a profile point has serched across an oxbow to find a data point. Such an error could result in a downward spike, but here it merely produces an upward spike.
Here we see our methodology start to fall apart as the Wang Chhu becomes a braided river near the Indian border. The river has moved considerably between the Landsat (~2000), SRTM (Feb, 2000?), and multiple ASTER passes.
This is the first place where I suspect that a floodplain is lower than the surface of the river. This is way upriver at about 2000 meters.

Now we have revisit our smoothing algorithm. For now we will assume that all bumps are artifacts and that no low points will be discarded. As our data is in integer meters, we cannot (except where points are discarded) represent non-zero slopes less than one meter over thirty meters. This is unacceptable even in Bhutan.

We have run this procedure on most of the Brahmaputra basin, and are starting to evaluate the results.

GDEM2