CharlesGaydon
commited on
Update README.md
Browse files
README.md
CHANGED
@@ -32,10 +32,6 @@ tags:
|
|
32 |
Lidar and imagery data were acquired over several years in distinct programs, and up to 3 years might separate them. The years of acquisition are given as metadata.
|
33 |
|
34 |
|
35 |
-
VHR Aerial imagery (ORTHO HR) | ALS points clouds (Lidar HD)
|
36 |
-
:-------------------------:|:-------------------------:
|
37 |
-
![](./imagery_18_classes.png) | ![](./lidar_18_classes.png)
|
38 |
-
|
39 |
## Dataset content
|
40 |
<hr style='margin-top:-1em; margin-bottom:0' />
|
41 |
The PureForest dataset consists of a total of 135,569 patches: 69111 in the train set, 13523 in the val set, and 52935 in the test set.
|
@@ -43,6 +39,10 @@ Each patch includes a high-resolution aerial image (250 pixels x 250 pixels) at
|
|
43 |
Band order is near-infrared, red, greeb, blue. For convenience, the Lidar point clouds are vertically colorized with the aerial images.
|
44 |
|
45 |
|
|
|
|
|
|
|
|
|
46 |
### Annotations
|
47 |
<hr style='margin-top:-1em; margin-bottom:0' />
|
48 |
Annotations were made at the forest level, and considering only monospecific forests. A semi-automatic approach was adopted in which forest polygons
|
@@ -68,15 +68,12 @@ were also used to improve the condidence in the purity of the forests.
|
|
68 |
|
69 |
### Dataset extent and train/val/test split
|
70 |
<hr style='margin-top:-1em; margin-bottom:0' />
|
71 |
-
The polygons were sampled in southern France due to the partial availability of the Lidar data at the time of dataset creation.
|
72 |
-
They are
|
73 |
-
To define a common benchmark, we divided the data into train, val, and test sets, with a stratification on semantic labels.
|
74 |
-
Annotation polygons are scattered across southern France, leading to a good geographical diversity within each semantic class.
|
75 |
-
To account for the high spatial autocorrelation, the 70%-15%-15% split is performed at the annotation polygon level:
|
76 |
-
each forest exclusively belongs to either the train, val, or test set.
|
77 |
-
This makes PureForest suitable to evaluate the territorial generalization of classification models.
|
78 |
|
79 |
-
|
|
|
|
|
80 |
|
81 |
![](./dataset_extent_map.excalidraw.png)
|
82 |
|
@@ -85,14 +82,14 @@ Approximate positions of forests in PureForest:
|
|
85 |
Please include a citation to the following Data Paper if PureForest was useful to your research:
|
86 |
|
87 |
```
|
88 |
-
@
|
89 |
title={PureForest: A Large-scale Aerial Lidar and Aerial Imagery Dataset for Tree Species Classification in Monospecific Forests},
|
90 |
-
author={Gaydon
|
91 |
year={2024},
|
92 |
-
|
|
|
93 |
url={https://arxiv.org/abs/2404.12064}
|
94 |
-
|
95 |
-
language = {en},
|
96 |
}
|
97 |
```
|
98 |
|
|
|
32 |
Lidar and imagery data were acquired over several years in distinct programs, and up to 3 years might separate them. The years of acquisition are given as metadata.
|
33 |
|
34 |
|
|
|
|
|
|
|
|
|
35 |
## Dataset content
|
36 |
<hr style='margin-top:-1em; margin-bottom:0' />
|
37 |
The PureForest dataset consists of a total of 135,569 patches: 69111 in the train set, 13523 in the val set, and 52935 in the test set.
|
|
|
39 |
Band order is near-infrared, red, greeb, blue. For convenience, the Lidar point clouds are vertically colorized with the aerial images.
|
40 |
|
41 |
|
42 |
+
VHR Aerial images (Near-Infrared, Red, Green) [ORTHO HR] | ALS points clouds [Lidar HD]
|
43 |
+
:-------------------------:|:-------------------------:
|
44 |
+
![](./imagery_18_classes.png) | ![](./lidar_18_classes.png)
|
45 |
+
|
46 |
### Annotations
|
47 |
<hr style='margin-top:-1em; margin-bottom:0' />
|
48 |
Annotations were made at the forest level, and considering only monospecific forests. A semi-automatic approach was adopted in which forest polygons
|
|
|
68 |
|
69 |
### Dataset extent and train/val/test split
|
70 |
<hr style='margin-top:-1em; margin-bottom:0' />
|
71 |
+
The annotation polygons were mostly sampled in the southern half of metropolitan France due to the partial availability of the Lidar HD data at the time of dataset creation.
|
72 |
+
They are scattered in 40 distinct French administrative departments and span a large diversity of territories and forests within each semantic class.
|
|
|
|
|
|
|
|
|
|
|
73 |
|
74 |
+
To define a common benchmark, we split the data into train, val, and test sets (70%-15%-15%) with stratification on semantic labels.
|
75 |
+
We address the high spatial autocorrelation inherent to geographic data by splitting at the annotation polygon level:
|
76 |
+
each forest exclusively belongs to either the train, val, or test set.
|
77 |
|
78 |
![](./dataset_extent_map.excalidraw.png)
|
79 |
|
|
|
82 |
Please include a citation to the following Data Paper if PureForest was useful to your research:
|
83 |
|
84 |
```
|
85 |
+
@misc{gaydon2024pureforest,
|
86 |
title={PureForest: A Large-scale Aerial Lidar and Aerial Imagery Dataset for Tree Species Classification in Monospecific Forests},
|
87 |
+
author={Charles Gaydon and Floryne Roche},
|
88 |
year={2024},
|
89 |
+
eprint={2404.12064},
|
90 |
+
archivePrefix={arXiv},
|
91 |
url={https://arxiv.org/abs/2404.12064}
|
92 |
+
primaryClass={cs.CV}
|
|
|
93 |
}
|
94 |
```
|
95 |
|