README.md

---
license: mit
task_categories:
- image-to-image
tags:
- remote-sensing
- super-resolution
- satellite-imagery
- earth-observation
- marine
- coral-reef
- seagrass
- mangrove
- landsat
- sentinel-2
- multi-image-super-resolution
pretty_name: MarineMISR
size_categories:
- 1K<n<10K
---

# MarineMISR

MarineMISR is a multi-image super-resolution (MISR) dataset pairing stacks of
Landsat 8/9 scenes (low-resolution, 30 m) with a single co-located Sentinel-2
scene (high-resolution, 10 m) over coastal and marine habitats. Each sample is
a 512x512 pixel patch (5.12 km x 5.12 km) sampled from one of three habitat
types — **coral reef**, **seagrass**, and **mangrove** — so the dataset can be
used to train and evaluate super-resolution models specifically over these
ecologically important, and typically under-represented, marine environments.

The dataset was built with the [MarineSpatialTooling](https://github.com/SpatialDecisionsGroup/MarineSpatialTooling) repository's `superres`
pipeline, which stratified sampling equally across the three habitat classes,
queried Google Earth Engine for the imagery, and postprocessed the rasters
into a pixel-aligned, reflectance-scaled stack.

## Dataset Summary

- **2,001 samples**, split roughly evenly across three habitat classes:

  | Habitat  | Samples |
  |----------|--------:|
  | Coral    |     678 |
  | Mangrove |     672 |
  | Seagrass |     651 |

- Each sample contains **6-8 Landsat 8/9 low-resolution images** (target 8;
  accepted down to 6 when insufficient cloud-free scenes were available in the
  sampling window) and **1 Sentinel-2 high-resolution image**.
- Samples are drawn globally and span **4 seasons** (Winter/Spring/Summer/Autumn,
  based on the Northern Hemisphere calendar) and multiple years, so revisit
  gaps and seasonal illumination/atmospheric conditions vary across samples.
- Coral and seagrass sites are additionally filtered by a Kd490
  (bottom-visibility) threshold so that only optically shallow water — where
  the benthic habitat is actually visible from space — is included. Mangrove
  sites, being emergent, are not filtered this way.
- Maximum cloud cover per scene: 20%.

## Dataset Structure

```text
MarineMISR/
├── raw/                                # As-downloaded rasters (before reflectance scaling)
│   └── sample_000000/
│       ├── landsat/
│       │   ├── landsat_00_2019-03-16.tif   # One file per low-res image, named <index>_<date>
│       │   └── ...
│       ├── sentinel2/
│       │   ├── sentinel2_000000.tif        # High-res target, standardized to the 512x512 grid
│       │   └── raw/
│       │       └── sentinel2_raw_000000.tif  # Unclipped/unstandardized original download
│       └── sample_metadata.json
├── processed/                          # Analysis-ready rasters (recommended for training)
│   └── sample_000000/
│       ├── landsat/                    # Reprojected onto the high-res grid + rescaled to reflectance
│       ├── sentinel2/                  # Rescaled to reflectance (already on the target grid)
│       └── sample_metadata.json
└── metadata/
    ├── dataset_metadata.json           # Full metadata for every sample (JSON)
    ├── dataset_manifest.csv            # Flat manifest, one row per sample
    ├── creation.log                    # Log from the sampling/manifest-creation step
    ├── download.log                    # Log from the download/standardization step
    └── postprocess.log                 # Log from the reflectance-scaling/alignment step
```

`raw/` holds the data as downloaded and grid-standardized: Landsat rasters are
in their native projection/resolution and pixel digital numbers (DN);
Sentinel-2 has both the raw download (`sentinel2/raw/`) and the version
clipped and resampled onto the sample's fixed 512x512 grid.

`processed/` is the analysis-ready version most users want: every low-res
Landsat image has been reprojected onto the same grid as the high-res
Sentinel-2 target (at Landsat's native 30 m resolution, so the two rasters
are pixel-registered but not resampled to a common resolution), and all
optical bands in both `landsat/` and `sentinel2/` have been rescaled from raw
DN to physical surface reflectance. Classification/QA bands (`QA_PIXEL`,
`SCL`) are left unscaled since they aren't reflectance values.

## Data Fields

Each sample directory contains a `sample_metadata.json` with fields including:

| Field | Description |
| --- | --- |
| `location_id` | Integer ID of the sample (matches the `sample_NNNNNN` directory name) |
| `latitude`, `longitude` | Center coordinates of the sampled patch (WGS84) |
| `season_id` | 0=Winter, 1=Spring, 2=Summer, 3=Autumn |
| `habitat_class` | `coral`, `seagrass`, or `mangrove` |
| `depth_m` | Bathymetric depth at the site (from GEBCO), meters |
| `date_range` | Search window used to find cloud-free imagery for this sample |
| `lowres_satellite` / `highres_satellite` | `landsat` / `sentinel2` |
| `lowres_images` | List of Landsat scenes used, each with GEE asset ID, acquisition date, and cloud cover |
| `highres_images` | Same, for the Sentinel-2 scene |
| `alignment_crs` | UTM CRS the sample was reprojected into |
| `patch_size_pixels` / `patch_size_meters` | 512 / 5120 |
| `target_origin_x` / `target_origin_y` | Grid origin in `alignment_crs` |
| `highres_aoi_geojson` | Polygon footprint of the sampled area |
| `lowres_count` | Number of low-res images actually included (6-8) |

`metadata/dataset_manifest.csv` contains the same information flattened to
one row per sample (used internally to drive/resume downloading).

## Data Specifications

| | Sentinel-2 (high-res) | Landsat 8/9 (low-res) |
| --- | --- | --- |
| Resolution | 10 m | 30 m |
| Bands | 13: `B1,B2,B3,B4,B5,B6,B7,B8,B8A,B11,B12,SCL,AOT` | 8: `SR_B1..SR_B7, QA_PIXEL` (Collection 2, Level-2 surface reflectance) |
| Collection | `COPERNICUS/S2_SR_HARMONIZED` | `LANDSAT/LC08/C02/T1_L2` + `LANDSAT/LC09/C02/T1_L2` (merged for faster revisit) |
| Images per sample | 1 | 6-8 |
| Patch size | 512x512 px (5.12 km x 5.12 km) | native resolution, reprojected to the same footprint |

Reflectance scaling (`processed/` only): `reflectance = DN * scale + offset`,
using each satellite's documented Collection-2/L2A constants (Sentinel-2
optical bands and AOT: scale `0.0001`/`0.001`, offset `0`; Landsat SR bands:
scale `0.0000275`, offset `-0.2`). `QA_PIXEL` and `SCL` are unscaled masks.

## Loading a Sample

```python
import json
import rasterio

sample_dir = "processed/sample_000000"

with open(f"{sample_dir}/sample_metadata.json") as f:
    meta = json.load(f)

with rasterio.open(f"{sample_dir}/sentinel2/sentinel2_000000.tif") as ds:
    highres = ds.read()  # (13, 512, 512), surface reflectance

with rasterio.open(f"{sample_dir}/landsat/landsat_00_2019-03-16.tif") as ds:
    lowres = ds.read()  # (8, H, W), surface reflectance, same footprint as highres
```

## Provenance

This dataset was generated end-to-end with the `superres` pipeline in the
`MarineSpatialTooling` repository: samples were drawn evenly from global
coral, seagrass, and mangrove habitat polygons (UNEP-WCMC / Global Mangrove
Watch extents), imagery was queried from Google Earth Engine, downloaded and
standardized to a fixed grid, then reprojected/rescaled in the postprocessing
step described above. See that repository for the full sampling methodology,
CLI tooling, and dataset-integrity/analysis scripts used to validate this
release.

## License

MIT. Note that the underlying Landsat and Sentinel-2 imagery is subject to
the respective open-data terms of the USGS/NASA and Copernicus programmes.