Systematic refinement of HRSC level 4 and level 5 DTMs using deep learning

Large area three dimensional mapping at metre scale resolution is critical for investigations of the origin and modification of the Martian surface and for the design and operation of landed robotic missions and forthcoming human exploration. The High Resolution Stereo Camera (HRSC) on ESA Mars Express has provided global coverage, but the standard photogrammetric production of level 4 single strip digital terrain models (DTMs) at 50 m/pixel and level 5 quadrangle mosaics remains labour intensive and computationally demanding. Processing a few hundred orbital strips can occupy expert teams for years (Jaumann et al., 2007; Gwinner et al., 2016).

 

We present an automated refinement pipeline that upgrades the complete HRSC level 4 archive together with the MC quadrangle level 5 mosaics from their native 50 m/pixel resolution to 12.5 m/pixel without manual intervention. The method is implemented within the Free University Berlin Mars Monocular Image to Surface Topography Toolbox (MISToolbox) which will become publicly available in 2025 (Tao et al., in prep.).

 

The core of the pipeline is based on a u-net structure using multiscale vision transformer (MViT) encoder followed by an interactive fusion decoder. MViT is able to generate feature maps at multiple resolutions to ensure that both fine details and broader contextual information are preserved. This design of the encoder captures global context across the entire image, which is crucial for understanding the low-frequency and high-frequency spatial relationships in large planetary images. The decoder of the u-net then iteratively fuses feature maps from different resolutions, leading to more accurate height/elevation predictions. The decoder avoids the common pitfall of losing local information, which can occur with direct upsampling methods.

 

Network training employed a combination of HiRISE and HRSC photogrammetric DTMs that were pre filtered to suppress artefacts. Evaluation against independent MOLA reference DTMs and against existing photogrammetric products shows superior accuracy as well as an effective increase in spatial resolution for the refined HRSC products. Effective resolution rises by 3.5 times when assessed with gradient based metrics and craters with diameters down to 300 m are reliably reconstructed (see figure 1).

 

Throughput on a single NVIDIA RTX3090 GPU is approximately 10 minutes per strip, enabling global reprocessing of the full HRSC catalogue in less than two months. All refined single strip DTMs and quadrangle mosaics together with derivative hillshades will be released through the FUB HRSC repository and the inhouse WebGIS portal by EPSC 2025. We anticipate that these higher resolution HRSC DTM products will become a standard base layer for future geologic mapping and/or landing site certification and planning on Mars.

Figure 1. An overview of the refined MC13E HRSC DTM mosaic (left) and zoom in views of the original DTM shaded relief (middle) and refined DTM shaded relief images (right).