Guglielmo M1, Gardner B2, Gibbs M3, Hughes J4, Petheram C5
1CSIRO, Information Management and Technology, Marsfield, Australia, 2CSIRO, Information Management and Technology, Clayton , Australia, 3CSIRO Environment, Waite, Australia, 4CSIRO Environment, Black Mountain, Australia, 5CSIRO Environment, Sandy Bay , Australia
Biography:
Dr Magda Guglielmo is the Technical Lead for Scientific Computing Applications at CSIRO. She holds a PhD in Astronomy from the University of Sydney, where her research focused on galaxy interactions. Her expertise spans scientific computing, machine learning, computer vision, image processing, and research software engineering. She has worked on projects involving convolutional neural networks, large language models, named-entity recognition, retrieval-augmented generation systems, remote sensing, and software optimisation for scientific applications.
Abstract:
River flow monitoring is necessary for water management, flood assessment, and environmental studies. Satellite video creates new opportunities for estimating river surface velocities using image velocimetry techniques, such as large-scale particle image velocimetry or space-time image velocimetry.
Planet SkySat is one of the few satellite constellations capable of acquiring optical video imagery, providing continuous video sequences of up to 120 seconds over user-defined areas of interest. These observations enable the tracking of river surface features and support the assessment of rapidly evolving hydrological events. However, SkySat videos present several registration challenges due to changes in viewing geometry, scale, and field of view throughout an acquisition. Also, the geolocation derived from the Rational Polynomial Coefficients (RPC) typically exhibits errors of 50-100 m.
This presentation introduces our software pipeline that addresses these challenges and provides accurate stabilisation, registration and geolocation of SkySat imagery. The pipeline combines RPC-based orthorectification, SIFT feature detection and descriptor matching for image registration and georeferencing. Frames are processed in user-defined groups and aligned using both global and local reference frames to accommodate changes in viewing geometry. Finally, feature matching against a georeferenced image is used to derive virtual control points and transfer geolocation across the video sequence. Preliminary results show image alignment ranging from subpixel to a few pixels, with geolocation errors of 2–5 m. The resulting workflow transforms raw video into analysis-ready imagery for river velocity and many other applications.
