Update README.md

README.md

@@ -30,8 +30,7 @@ This dataset includes 135,569 patches, each measuring 50m*50m, covering a cumula
 Each patch represents a monospecific forest, labeled with a single tree species to facilitate classification tasks. 
 The proposed classification features 13 semantic classes, hierarchically grouping 18 tree species from 9 different tree genus.
 
-A reference train/val/test split is provided with class stratification. 
-To account for spatial autocorrelation, each forest exclusively belongs to either the train, validation, or test set.
+A reference train/val/test split is provided.
 
 | Class | Train (%) | Val (%) | Test (%) |
 |-------|------------:|----------:|-----------:|
@@ -64,15 +63,24 @@ Lidar points clouds            |  Aerial imagery
 ## Annotation
 
 Annotation were made at the forest level, and considering only monospecific forests. A semi-automatic approach was adopted in which pure forest polygons
-were selected and then curated by photointerpreters. The annotation polygons came from the [BD Forêt](https://inventaire-forestier.ign.fr/spip.php?article646), 
+were selected and then curated by expert photointerpreters from the IGN. The annotation polygons came from the [BD Forêt](https://inventaire-forestier.ign.fr/spip.php?article646), 
 a forest vector database of tree species occupation in France. Ground truths from the F[rench National Forest Inventory](https://inventaire-forestier.ign.fr/?lang=en) 
-were also used.
+were also used to improve the condidence in the purity of the forests.
 
 ## Data Splits
 
 The polygons were sampled in southern France due to the partial availability of the Lidar data at the time of dataset creation. 
-They are located in 40 distinct French administrative departments. 
+They are located in 40 distinct French administrative departments, covering a large diversity of territories and forests.
+To define a common benchmark, we divided the data into train, validation, and test sets, with a stratification on semantic labels. 
+Annotation polygons are scattered across southern France, leading to a good geographical diversity within each semantic class. 
+To account for the high spatial autocorrelation, the split is performed at the annotation polygon level: 
+each forest exclusively belongs to either the train, validation, or test set.  
+This makes PureForest suitable to evaluate the territorial generalization of classification models. 
+We aimed for a 70%-15%-15% split across the train, validation, and test sets, respectively.
 
+Approximate positions of forests in PureForest
+
+![](./dataset_extent_map.excalidraw.png)
 
 ## Citation
 Please include a citation to the following article if you use the PureForest dataset: