Moon-Mapping

A project by the STAC core team at IIT Mandi: train super-resolution models to upscale Chandrayaan-2 TMC-2 imagery (~5 m/px) towards 30 cm/px, using LRO NAC and Chandrayaan-2 OHRC imagery as high-resolution ground truth — and from that build a high-resolution lunar atlas.

The approach is to learn the TMC→NAC mapping on aligned image pairs: TMC crops resized to 24×24 px as input, the matching NAC crops at 96×96 px as target. Two model families were tried, an SRGAN and a SwinIR transformer.

Data is not in this repository. Every image, dataset and checkpoint is fetched from its original archive by scripts/download_data.sh. Start with DATA.md.

Quick start

git clone https://github.com/STAC-IITMandi/Moon-Mapping.git
cd Moon-Mapping
pip install -r requirements.txt

scripts/download_data.sh help     # what's available
scripts/download_data.sh all      # dirs + sample NAC images + all public weights

The Chandrayaan-2 half needs a free PRADAN account — see DATA.md.

Repository layout

AI Models/                  the super-resolution models
├── Moon Mapping/             SRGAN: train, generate, cascade to 16x, evaluate
├── Transformer_Model-main/   SwinIR layers written from scratch (RSTB, STL, attention)
├── SwinIR/                   the project's working SwinIR copy + training notebooks
├── SwinIR-0.0/               earlier SwinIR snapshot, kept for comparison
├── Swin-Transformer/         vendored Swin Transformer backbone (Microsoft)
└── REFERENCES.md             papers and code provenance

DataSet/                    footprint geometry and dataset generation
├── Generator.py, Gen2.py     TMC↔OHRC footprint intersection from PDS4 labels
├── tmc_break.py              lunar-ellipsoid lat/lon → pixel mapping
├── crop.py, readOHRC.py      cropping; reading PDS4 OHRC products
├── dataset_prep*.ipynb       dataset assembly notebooks
├── TMC/coordinates.csv       3,246 TMC products, scraped with corner coordinates
└── OHRC_ShapeFiles/          ISRO-supplied OHRC footprint shapefiles

dataset-cleaning-creation-and-analysis/
├── README.md                 how the training pairs were built — read this first
├── scripts/                  NAC calibration (ISIS), coordinate maths, crop generation
└── dataset-preparation/      scraping and intersection notebooks + CSV catalogues

docs/                       ISRO data-product user guides, LTA procedure, presentations
scripts/download_data.sh    fetches all data from PRADAN / LROC
DATA.md                     data setup guide
Mid_ISRO.pdf                mid-project report

How the training data was built

The full account is in dataset-cleaning-creation-and-analysis/README.md. In short:

1. Find NAC images fully contained within an orthorectified TMC image (855 TMC footprints × 14,026 NAC footprints, intersected in the notebooks).
2. Calibrate the NAC EDR with the LROC-recommended ISIS pipeline (nac2png.sh).
3. Crop the TMC image to the NAC extent (NACTMC_coords.py), then match brightness/contrast and rotate so features line up.
4. Cut both into 96×96 px (NAC) and 24×24 px (TMC) pairs, discarding tiles that are mostly no-data (final_gen.py).

One TMC–NAC pair yields roughly 13,000 crops; two pairs were used for training.

OHRC was investigated as ground truth but largely abandoned: of ~90 OHRC images, only ~10 overlapped any TMC image.

Models

SRGAN — AI Models/Moon Mapping/. A 16-block SRResNet generator at ×4, cascaded twice for ×16 (cascade.py). srgan_config.py has a single mode switch (train / generate / evaluate) that selects the paths and hyperparameters for each stage.

SwinIR — AI Models/SwinIR/ and AI Models/Transformer_Model-main/. A shifted-window transformer, trained with L1, SSIM and perceptual losses in separate runs.

Weights and data. scripts/download_data.sh checkpoints fetches the trained SwinIR checkpoints (five files, 777 MB), training-data fetches the crops they were trained on (1.7 GB), and swinir-weights fetches the original SwinIR models the test scripts expect. No trained SRGAN generator is published, so the SRGAN needs training before its generate / evaluate modes are usable — see DATA.md § Model weights and training data.

Data sources

• Chandrayaan-2 TMC-2 and OHRC — ISRO/ISSDC PRADAN
• LRO NAC — LROC archive, Arizona State University

Please observe each archive's access terms and citation requirements. PRADAN in particular enforces session and rate limits.