Alignment

Alignment#

Even a clean stereo DEM usually sits slightly off true ground, because small camera-model errors translate into a shift of the whole surface. pc_align removes that shift by registering the DEM to a trusted reference.

The reference dataset#

ICESat-2 ATL06-SR for Earth, MOLA for Mars, LOLA for the Moon, or any prior high-quality DEM (3DEP, ArcticDEM, regional lidar). A good reference does not need to be dense; it needs to be accurate and to sample stable terrain. A few thousand well-placed altimetry points constrain a rigid shift better than a dense but biased DEM.

What pc_align does#

The algorithm is iterative closest point (ICP): pair each point of your DEM with the nearest reference point, find the rigid transformation (rotation and translation; a translation-only or scale-solving variant if you ask for one) that minimizes those distances, apply it, and repeat until the fit stops improving. The output is your DEM moved onto the reference.

pc_align translates the DEM so it passes through the reference altimetry points

When it helps and when it doesn’t#

pc_align fixes what a rigid move can fix: near-uniform bias and small global tilt. It cannot fix localized blunders, matching noise, or non-rigid warp; those need attention upstream (better bundle adjustment, better stereo parameters, or better input imagery). The difference-to-reference maps below, both from tutorial runs, show the two cases:

Difference maps: a near-uniform offset alignment can fix, and spatially varying error it cannot

If the difference map against the reference still shows structure after alignment, the problem is not alignment.

Reading the alignment report#

pc_align prints percentile breakdowns of the point-to-reference distances before and after alignment, plus the transformation it applied. From a run of the WorldView tutorial’s DEM against its COP30 reference:

Input:  error percentile of smallest errors (meters): 16%: 3.87315,  50%: 7.13745, 84%: 11.2421
Output: error percentile of smallest errors (meters): 16%: 0.574244, 50%: 2.08287, 84%: 5.95369
Translation vector (North-East-Down, meters): Vector3(2.1610355,3.4450347,7.0166972)
Translation vector magnitude (meters): 8.1100172

Three things to read off:

  • The 50% lines are the headline metric: the median point-to-reference distance dropped from about 7 m to about 2 m.

  • The translation vector says where the DEM moved: here mostly down by 7 m, with a couple of meters horizontally. It should be a plausible magnitude for your sensor.

  • A median that barely drops, or a huge translation, means the DEM and reference disagree in a way ICP cannot reconcile; check the vertical datums and the overlap area first.

Alignment inside asp-plot#

The asp_plot CLI runs pc_align by default (the --pc_align option, on unless altimetry plotting is disabled) against the appropriate altimetry reference for the body (ICESat-2 for Earth, MOLA for Mars, LOLA for the Moon) and appends alignment-report pages to the PDF. This is how the WorldView tutorial aligns its DEM: its report command runs pc_align without any extra flag.

Where to read more#