Automated DEM generation and scientific applications of Planet SkySat triplet stereo and video imagery

Abstract:

The Planet SkySat-C SmallSat constellation can acquire very-high-resolution (0.5 to 0.9 m) triplet and multi-view stereo imagery with sub-daily revisit times, providing an exciting opportunity for global, on-demand 3D mapping of the Earth’s dynamic surface. However, several technical challenges, such as small individual scene footprints and the relatively low default geolocation accuracy, continue to prevent widespread, user-friendly processing and analysis of these valuable data. For example, the default SkySat L1B image geolocation accuracy is ~5 to 30 m, which limits output DEM vertical accuracy and research applications generally requiring sub-pixel accuracy (< 1 m).

We developed an open-source workflow to refine the SkySat camera models and improve absolute image geolocation using external reference elevation data (e.g., DEMs from airborne LiDAR or WorldView stereo, ICESat-2 altimetry), without manual ground control point (GCP) selection. The refined camera models are used to generate accurate and self-consistent SkySat DEMs at 2-m posting and orthoimages at native resolution. We are scaling the workflow on high-performance computing environments and processing SkySat triplet stereo and video imagery over a diverse range of geologically interesting landforms spread across the globe.

In this study, we highlight several quantitative science applications from the derived SkySat DEMs, including: estimating glacier mass balance, seasonal snow depth, landslide deformation, as well as building and canopy height in urban and forested settings, respectively. Our workflow can be extended to produce accurate DEMs and orthoimages from imagery acquired by similar Smallsat/CubeSat constellations (such as PlanetScope), opening up new avenues for geodetic change detection studies.