Scaling a Fully Automatic CTX-to-HRSC Coregistration Pipeline using Phase-Correlation and (A)KAZE Feature Matching

### Introduction

 

The Context Camera (CTX) on Mars Reconnaissance Orbiter provides near-global coverage of Mars at ~6 m/pixel and underpins countless geomorphological, stratigraphic, and time-series studies. Its absolute pointing, reconstructed from spacecraft trajectory and attitude, is typically uncertain at the level of dozen to hundreds of metres — the limiting factor for change detection, mosaicking, and joint analysis with HiRISE.

The HRSC instrument on Mars Express delivers photogrammetrically controlled orthorectified products with metre-to-decametre geodetic accuracy and global-scale internal consistency, making them an ideal reference frame for tying CTX into a homogeneous geodesy.

 

### Pipeline overview

 

We present a fully automatic two-regime coregistration pipeline, applying similar but extended techniques as reported in [1] and [2].

Where HRSC orthomap coverage is available, we derive dense sub-pixel shift fields between CTX and HRSC via phase-correlation using the AROSICS Python library.

Where HRSC coverage is absent, we match features between overlapping CTX frames themselves, anchored to neighbouring HRSC-tied images, we start using the KAZE and AKAZE detectors inside ISIS' `findfeatures`, and refine it further using the AROSICS detector.

Both regimes feed a single ISIS `jigsaw` bundle adjustment, producing a unified, quality-controlled control network.

 

### AROSICS regime

 

Each CTX cube is map-projected with `cam2map` using the corresponding HRSC product as map template, preserving projection, pixel resolution and `CenterLongitude` to avoid resampling artefacts.

AROSICS searches for tie points on a pre-set regular grid, producing quality factors that can be filtered on; the grid is converted to an ISIS control network for `jigsaw`.

 

### `findfeatures` / AKAZE regime and scaling

 

On the ISIS side, we replaced `autoseed`+`pointreg` with ISIS `findfeatures` using the KAZE and AKAZE detectors.

These coarser scale feature identifications are useful because they are able to shift even strongly offset CTX images without former determination of a search radius.

An overlap-aware greedy set-cover algorithm selects reference images for fromlist matching, and the resulting per-reference networks are merged via `cnetmerge`.

After successful first alignment, several extra rounds using AROSICS are added to improve the internal coregistration before the merge with the on-ortho CTX coregistered images is performed.

 

### Validation and metrics

 

We consistently report both σ₀ (a tail-sensitive RMS) and median residuals: aggressive `pointreg` settings on smooth Martian dust plains can fabricate correlation peaks that inflate σ₀ by an order of magnitude while leaving the median sub-pixel.

Per-pair versus all-at-once bundle adjustments converge currently to comparable σ₀ (0.448 vs 0.438 px) under conservative parameters, which supports pre-coregistration work with final merges, because all-at-once tie-point generations can run into memory issues using the findfeatures tool.

 

### Outlook

 

Work in progress includes (i) anchoring the gap bundle to background MOLA/THEMIS basemaps to improve coregistration from the beginning towards absolute geodetic control, (ii) closing the loop across full HRSC tiles, and (iii) extending the pipeline application to HiRISE data for precise change monitoring analyses.

 

### References

[1] Robbins, S. J., Kirchoff, M. R., & Hoover, R. H. (2020). Fully Controlled 6 Meters per Pixel Mosaic of Mars's South Polar Region. Earth and Space Science, 7, e2019EA001054. doi:10.1029/2019EA001054

[2] Robbins, S. J., Kirchoff, M. R., & Hoover, R. H. (2023). Fully Controlled 6 Meters per Pixel Equatorial Mosaic of Mars From MRO CTX Images, Version 1. Earth and Space Science, 10, e2022EA002443. doi:10.1029/2022EA002443