SPUN (Snow Patch U-Net) - model and dataset

This repository contains the trained model weights for SPUN (Snow Patch U-Net), designed for mapping patchy and late-season fractional snow cover in Sentinel-2 satellite imagery. See https://github.com/leamhowe/spun for more guidance on running SPUN with the `SPUN.pt` file. 

Alongside the model weights, this repository provides the complete training and evaluation dataset. The dataset contains 2,455 Sentinel-2 1 km image chips formatted as `.zarr` stores. It utilizes a hybrid approach, combining manually labelled high-quality ground-truth samples with a larger set of pseudo-labelled samples to increase geographic and temporal diversity.

For full details on the model architecture and dataset generation, please refer to the SPUN paper.

 

 

---

The archive is organised to contain both the pre-trained model and the accompanying data stores:

```text

.

├── README.md               # Read this for more information

├── SPUN.pt                 # Pre-trained PyTorch model weights

├── dataset_metadata.csv    # Master list of dataset files, splits, and provenance

└── zarr_files/             # Directory containing all .zarr stores

    ├── original_training_set/   # Contains 100% manually labelled data

    └── [filename].zarr          # pseudo-labelled data.

```

---