Memory-Efficient Georeferencing of Segmented Satellite Imagery
TL;DR
Memory-efficient georeferencing of segmented satellite imagery, an automatic georeferencing pipeline that uses land–sea–cloud segmentation to reliably identify ground control points, and a subpixel accuracy demonstrates that high-precision georeferencing is attainable despite inherent resource constraints and large initial telemetry errors.
Abstract
Accurate georeferencing is essential for the prompt and rapid use of Earth Observation data. Direct georeferencing algorithms, using satellite orbital position and attitude to compute geolocation pixels, are commonly used. However, small satellites employing commercial off-the-shelf components can experience significant uncertainties in attitude and position telemetry data. This article introduces memory-efficient georeferencing of segmented satellite imagery, an automatic georeferencing pipeline that uses land–sea–cloud segmentation to reliably identify ground control points. These are used to correct the satellite attitude for georeferencing by minimizing the reprojection error. The corrected attitude state is then used to reestimate each pixel geolocation through direct georeferencing. The proposed pipeline is scalable and designed for future implementation onboard, and it achieves a geolocation root-mean-square error of 0.81 pixel. This subpixel accuracy demonstrates that high-precision georeferencing is attainable despite inherent resource constraints and large initial telemetry errors. The method is validated and demonstrated using the HYPSO-1 and HYPSO-2 satellites, which each carry a hyperspectral pushbroom sensor.