Iterative bundle adjustment without ground control points on SPOT-7 stereogrammetry and its effect on DEM geometric accuracy

Abstract

Generating Digital Elevation Models (DEMs) from stereo optical satellite data has been a well-established practice for many years. The typical workflow involves performing Bundle Adjustment (BA) to align the stereo imagery, often supplemented with Ground Control Point (GCP) data for accurate vertical values. However, acquiring high-precision GCPs using geodetic GPS can be costly and time-consuming. In this study, we utilize a GCP-less approach that combines iterative bundle adjustment with an existing DEM for SPOT-7 tri-stereo imagery. The Semi-global matching algorithm is employed as the stereo correlator for all three stereo combinations (Forward-Nadir, Nadir-Backward, Forward-Backward). We also explore four alternative approaches: single BA without GCPs, single BA with 9 GCPs, 9 GCPs only, and single BA without GCPs but with DEM co-registration. To build the DEM, we utilize the SPOT-7 panchromatic band (1.5m) and upscale it by a factor of two to achieve a Ground Sampling Distance of 3 meters. We evaluate the horizontal shift in the x and y directions of the produced DEMs using DEMNAS as the reference. Additionally, the vertical accuracy is assessed using the Root Mean Square Error (RMSE) by comparing the results to a combination of geodetic Independent Control Points (ICPs), Unmanned Aerial Vehicle (UAV) Digital Surface Models (DSMs), and filtered ICESat-2 ATL-08 points as the reference data. Preliminary findings indicate that the GCP-less iterative BA approach outperforms all but one other method on average. The iterative BA method yields average x and y shifts of 1.72 meters and 0.95 meters, respectively. These values are lower than those obtained using single BA (13.01 and 3.27), single BA with 9 GCPs (2.85 and 2.62), and 9 GCPs only (3.94 and 1.00) approaches. The only approach that produces lower shifts is the single BA with DEM coregistration, which results in 0.60m and 0.72m for x and y shifts, respectively.