CaBuAr P1

# California Burned Areas Dataset

This is the first part of the dataset.

### Dataset Summary

This dataset contains images from Sentinel-2 satellites taken before and after a wildfire.
The ground truth masks are provided by the California Department of Forestry and Fire Protection and they are mapped on the images.

### Supported Tasks

The dataset is designed to do binary semantic segmentation of burned vs unburned areas.

## Dataset Structure

### Dataset opening

Dataset was compressed using `h5py` and BZip2 from `hdf5plugin`. **WARNING: `hdf5plugin` is necessary to extract data**

### Data Instances

Each matrix has a shape of 5490x5490xC, where C is 12 for pre-fire and post-fire images, while it is 0 for binary masks.

### Data Fields

In each HDF5 file, you can find post-fire, pre-fire images and binary masks. The file is structured in this way:

```bash
├── foldn
│ ├── uid0
│ │ ├── pre_fire
│ │ ├── post_fire
│ │ ├── mask
│ ├── uid1
│ ├── post_fire
│ ├── mask
│
├── foldm
├── uid2
│ ├── post_fire
│ ├── mask
├── uid3
├── pre_fire
├── post_fire
├── mask
...
```

where `foldn` and `foldm` are fold names and `uidn` is a unique identifier for the wilfire.

### Data Splits

There are 5 random splits whose names are: 0, 1, 2, 3 and 4.

## Dataset Creation

### Source Data

#### Initial Data Collection and Normalization

Data are collected directly from Copernicus Open Access Hub through the API. The band files are aggregated into one single matrix.