Paper: Fusing DEMs

Notes from two papers discussing similar methods for merging elevation rasters.

1. Distance-based blending of SRTM and ASTER

Robinson, N., Regetz, J., & Guralnick, R. P. (2014). EarthEnv-DEM90: A nearly-global, void-free, multi-scale smoothed, 90m digital elevation model from fused ASTER and SRTM data. ISPRS Journal of Photogrammetry and Remote Sensing, 87, 57-67.

  • Merged dataset from ASTER and CIGAR SRTM
  • Highlights importance off well-documented methodology
  • They shifted ASTER by a 1/2 pixel to align cell edges (instead off centres) with exact degree lines.
  • Voids filled with Delta Surface Fill (DSF) using an ancillary dataset.
    • Extrapolates dH over the hole, then adds that to the ancillary dataset.
    • Which is nice in that it works for datasets that have a regular bias compared to the size of the hole, like a different datum.
    • Grid search to choose the best interpolation smoothing and tension parameters, taking training data from shifting actual holes.
  • SRTM was also shifted by a half pixel.
  • Both DEMs were smoothed to reduce noise. Smoothing factor was chosen based on an estimate of noise.
    • This probably not a bad idea for SRTM and ASTER (which are very noisy) but I’m hoping won’t be needed for gpxz)
  • They used weighted average blending with weight ~ exp(x^2).
    • This resulted in equal weights at a distance of 2km (about 200 pixels) and 1% at 6km.

2. Error-based blending

Petrasova, A., Mitasova, H., Petras, V., & Jeziorska, J. (2017). Fusion of high-resolution DEMs for water flow modeling. Open Geospatial Data, Software and Standards, 2(1), 1-8.

  • Distance-weighted blending with spatially varying blending zone width based on elevation difference.
  • DEMs must have common CRS and resolution (though I just thing for their discretisation and notation, the theory should hold otherwise.)
  • Large elevation differences along the edge of the hires raster (typically representing vegetation or built structures) should be removed first.
  • DEMs are averaged in a blend zone by weight that varies from 0 to 1 across the zone. The weight can be a function of distance into the blend zone: most simply directly proportional, or could be logistic. (Something like logistic would be less smooth, but would make the result closer to the either input DEM.)
  • They suggest a blend width increasing with height difference decreased by transition angle (which you specify).
  • dH is smoothed with a 5 pixel radius.