Skip to content

MarsRecon

A geospatial dataset manager and deep-learning training framework for NASA's HiRISE Mars imagery.

MarsRecon wraps the HiRISE Reduced Data Records (RDR) and Digital Terrain Model (DTM) collections as TorchGeo datasets, and provides a flow-matching monocular depth pipeline (DepthFM) and a tri-modal CLIP model (MarsCLIP) trained on top of them.

  • Dataset layer

TorchGeo GeoDataset wrappers for HiRISE RDR and DTM stereo pairs, with spatiotemporal indexing, async PDS downloading, and radiometric calibration.

Architecture

  • DepthFM training

Flow-matching monocular depth estimation adapted for Mars DTMs, on multi-GPU PyTorch Lightning.

DepthFM pipeline

  • MarsCLIP

Tri-modal CLIP (image + elevation + text) for Mars imagery, with MAE pretraining.

MarsCLIP

  • Get started

Install, run the dataset pipeline, and launch a training job.

Quickstart

Why MarsRecon?

HiRISE produces the highest-resolution images of Mars available (~25 cm/pixel RED, ~50 cm/pixel colour), but the raw archive is awkward for deep learning: thin rotated parallelogram strips, JP2/IMG formats, per-observation projections, and a >10 TB total volume. MarsRecon handles all of that plumbing so you can focus on the model.

  • Strip-aware sampling — HiRISEGeoSampler pre-grids valid patch centres inside actual strip polygons, avoiding the 60–90 % empty-pixel patches you'd get from bounding-box sampling.
  • Single CRS hub — all observations are reprojected on-the-fly into a common IAU 2000 Mars geographic CRS.
  • Fast I/O — cog_conversion pre-converts JP2 → Cloud-Optimized GeoTIFF; the LitData streaming path serves training at full GPU saturation.
  • DepthFM training — flow-matching elevation prediction with photometric (Lunar-Lambert) consistency, normals, multi-scale gradients, and ordinal-ranking losses.

Repository layout (at a glance)

src/
  dataset/         # MarsHiRISE / MarsHiRISEDTM / HiRISEGeoSampler
  depth_fm/        # DepthFM model, Lightning training, losses, viz
  clip/            # MarsCLIP tri-modal model and MAE pretraining
configs/           # OmegaConf YAML configs
scripts/           # entry points (training, inference, viz, ablations)
tests/             # pytest suite
docs/              # this site

See Architecture · Overview for the full source-tree map.

Live API reference

Every module under src/ has an auto-generated reference page driven by mkdocstrings — new public symbols appear automatically on the next build.

Browse the API reference