Fully Automated DEM Generation using Satellite Imagery

Status:
Accepted

High resolution and accurate Digital Elevation Model (DEM) generation from satellite imagery is a challenging problem. In this work, a multiview stereo reconstruction framework is proposed that is applicable to non-stereoscopic satellite images which may have been captured by different satellites. Given a cross-platform satellite image archive, the images are first geolocation corrected with respect to each other using a fully-automated processing pipeline that applies sparse feature matching and bias correction of Rational Polynomial Coefficient (RPC) camera models. The images are then randomly paired to create affine rectified patches via local RPCs and matched using a standard stereo technique. The resulting disparity maps are converted to orthographic height maps with 1 meter Ground Sampling Distance (GSD). The height maps from different pairs covering the same geographic region are in excellent pixel-to-pixel alignment thanks to relative geo-registration and thus it becomes possible to combine them via a simple median filtering technique. The DEM given by a single pair may contain gaps or areas with erroneous surface geometry due to collection time differences, shadows, appearance changes and/or unfavorable relative viewing geometry of the two images; however, when sufficient number of pairs contribute to a combined DEM, the quality is much approved and all the gaps are filled. The quality of the high resolution DEMs generated by the proposed pipeline is evaluated using publicly available coastal LIDAR data from South East United States. The root mean square error of resultant elevation values is measured to be 1.54 meters relative to LIDAR data with 15 cm accuracy. It is also shown via stabilization of the variance in error that the selection of the images for pairing doesn’t matter after a minimum of 10 images are selected and all the 90 pairs given by them are used. The quality of the combined DEM is attesting to the power of harnessing big data in Remote Sensing. An interesting application of the proposed fully automated DEM generation pipeline is detection of structural changes for which various examples are shown from a 5 square kilometer urban region in Sydney, Australia.

Slides

Session details
Schedule info
Session Time Slot(s):
Wednesday, May 4, 2016 - 17:20 to 17:35

Comments

we mainly used following open source tools:
VXL (an open source C++ libraries for Computer Vision, including some geospatial capability), openCL, gdal and libgeotiff.
Plus, OSM data and LiDAR are served as ground truth for accuracy assessment

Public comment