Multi-View-Stereo DSM from Planetscope satellites

Master Thesis at ifp - Valentin Schmitt

Valentin Schmitt

Multi-View-Stereo DSM from Planetscope satellites

Duration: 6 months
Completition: February 2022
Supervisor: Kelsey Jordahl (Planet Labs Inc), Prof. Dr.-Ing. Norbert Haala
Examiner: Prof. Dr.-Ing. Norbert Haala


 

Planetscope

Since 2015, Planet Labs has operated Cubesat 3U satellites for Earth observation and remote sensing monitoring. The Planetscope constellation contains around 130 Doves which are frame based sensors and they reached the daily time resolution in 2018. The whole setting is close to airborne photogrammetric acquisitions that allow digital surface modelling computation with Structure from Motion pipelines. Unlike low flight, Doves perch on orbits around 475 km high with a small field view angle acquiring 24 km x 16 km scenes and thus they hold a weak intersection geometry at the ground. Nevertheless, they provide a very large redundancy. Product knowledge of overall organisation, hardware and software description as well as internal process specification outlines first hassles and sets a coarse methodology  forth.

Figure 1: Dove overlaps, left: in-track, middle: cross track, right: cross date

Development challenges

This study aims to design an automated process chain for 3D reconstruction enhanced with information redundancy. Former studies about Planetscope matching built object oriented methodology which returned promising results. However, this research focuses on an image oriented approach which is closer to low flight custom. The research part sets the whole chain up which is divided into scene block creation, bundle adjustment and reconstruction sections. From the entire Planetscope database, the scene selection retrieves the best ones and sets a working directory in. Then, the adjustment section presents the strength and weakness of incoming Dove’s location models and it figures out a correcting procedure making use of the SRTM only. The image matching provides stereo point clouds with a large redundancy and the reconstruction part ends with a multi-view stereo method merging elevations and returns a standard DSM and its computation accuracy. External data (SRTM) is only involved during the bundle adjustment and reconstruction results are compared to ground truth without further transformation.

Figure 2: Integer disparity in epipolar geometry, one scene sliding over the other and displaying the radiometric difference, white: no difference, top-right: stereopair footprint

Results

Finally, an assessment runs  over test sites (Providence Mountain, Stuttgart and Mount Saint Helens) and the last part outlines process issues and summarises the achievable accuracy. The final accuracy reaches 8 m in vertical direction (2 pixels) along a 4 m horizontal grid. All in all, this research sets the basis of Planetscope reconstruction with image approach, even though some computation limits remain and it covers only the case of Dove-Classic. 

Figure 3: Providence mountain elevation result (band 1)
Figure 4: Providence mountain accuracy result (band 2)

Further work

Hence, the work could be extended during a Planet partnership. It remains process improvements and extension turning the current work stable. Then, the following satellite generations shall be included improving outcomes by redundancy. The ultimate goal is a stable and automatic algorithm which can be implemented on Planet customer portal.

References

S. Aati, J.-P. Avouac, 2020, Optimization of Optical Image Geometric Modelling, Application to Topography Extraction and Topographic Change Measurements Using PlanetScope and SkySat Imagery, Remote Sensing, December 2020.

S. Bhushan, D. Shean, O. Alexandrov, S. Henderson, 2021, Automated Digital Elevation Model (Dem) Generation From Very-High-Resolution Planet Skysat Triplet Stereo And Video Imagery, ISPRS Journal of Photogrammetry and Remote Sensing, January 2021.

S. Ghuffar, 2018, DEM Generation from Multi Satellite PlanetScope Imagery, Remote Sensing, September 2018.

Ansprechpartner

Dieses Bild zeigt Norbert Haala

Norbert Haala

apl. Prof. Dr.-Ing.

Stellvertretender Institutsleiter

Zum Seitenanfang