Update README.md

README.md

@@ -31,17 +31,20 @@ The dataset is designed to do binary semantic segmentation of burned vs unburned
 
 ## Dataset Structure
 
+We opted to use HDF5 to grant better portability and lower file size than GeoTIFF.
+
 ### Dataset opening
 
-Dataset was compressed using `h5py` and BZip2 from `hdf5plugin`. **WARNING: `hdf5plugin` is necessary to extract data**
+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.
+Pre-patched version with matrices of size 512x512xC version are provided, too. In this case only mask with at least one positive pixel are present.
 
 ### Data Fields
 
-In each HDF5 file, you can find post-fire, pre-fire images and binary masks. The file is structured in this way:
+In each standard HDF5 file, you can find post-fire, pre-fire images and binary masks. The file is structured in this way:
 
 ```bash
 ├── foldn
@@ -66,6 +69,20 @@ In each HDF5 file, you can find post-fire, pre-fire images and binary masks. The
 
 where `foldn` and `foldm` are fold names and `uidn` is a unique identifier for the wilfire.
 
+For the pre-patched version, the structure is:
+```bash
+root
+|
+|-- uid0_x: {post_fire, pre_fire, mask}
+|
+|-- uid0_y: {post_fire, pre_fire, mask}
+|
+|-- uid1_x: {post_fire, mask}
+|
+...
+```
+the fold name is stored as attribute.
+
 ### Data Splits
 
 There are 5 random splits whose names are: 0, 1, 2, 3 and 4.