AGarioud commited on
Commit
d01abd8
1 Parent(s): e77d47b

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +52 -21
README.md CHANGED
@@ -10,6 +10,7 @@ size_categories:
10
  # Datset Card for FLAIR land-cover semantic segmentation
11
 
12
  ## Context & Data
 
13
 
14
  The hereby FLAIR (#2) dataset is sampled countrywide and is composed of over 20 billion annotated pixels of very high resolution aerial imagery at 0.2 m spatial resolution, acquired over three years and different months (spatio-temporal domains).
15
  Aerial imagery patches consist of 5 channels (RVB-Near Infrared-Elevation) and have corresponding annotation (with 19 semantic classes or 13 for the baselines).
@@ -168,8 +169,8 @@ The dataset covers 50 spatial domains, encompassing 916 areas spanning 817 km².
168
  <br><br>
169
 
170
  ## Dataset Structure
 
171
 
172
- ### Spatio-Temporal Distribution
173
  The FLAIR dataset consists of 77 762 patches. Each patch includes a high-resolution aerial image (512x512) at 0.2 m, a yearly satellite image time series (40x40 by default by wider areas are provided) with a spatial resolution of 10 m
174
  and associated cloud and snow masks, and pixel-precise elevation and land cover annotations at 0.2 m resolution (512x512).
175
 
@@ -177,10 +178,36 @@ and associated cloud and snow masks, and pixel-precise elevation and land cover
177
 
178
 
179
  ### Band order
180
- Aerial : 1. Red; 2. Green; 3. Blue; 4. NIR; 5. nDSM <br/>
181
- Satellite : 1. Blue (B2 490nm); 2. Green (B3 560nm); 3. Red (B4 665nm); 4. Red-Edge (B5 705nm); 5. Red-Edge2 (B6 470nm);
182
- 6. Red-Edge3 (B7 783nm); 7. NIR (B8 842nm); 8. NIR-Red-Edge (B8a 865nm); 9. SWIR (B11 1610nm); 10. SWIR2 (B12 2190nm)
183
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
184
 
185
  ### Annotations
186
  Each pixel has been manually annotated by photo-interpretation of the 20 cm resolution aerial imagery, carried out by a team supervised by geography experts from the IGN.
@@ -193,7 +220,7 @@ This arrangement ensures a balanced distribution of semantic classes, radiometri
193
  Consequently, every split accurately reflects the landscape diversity inherent to metropolitan France.
194
  It is important to mention that the patches come with meta-data permitting alternative splitting schemes, for example focused on domain shifts.
195
 
196
- Official split: <br/>
197
 
198
  <div style="display: flex; flex-wrap: nowrap; align-items: center">
199
  <div style="flex: 40%;">
@@ -222,6 +249,8 @@ Official split: <br/>
222
  <br><br>
223
 
224
  ## Baseline code
 
 
225
  We propose the U-T&T model, a two-branch architecture that combines spatial and temporal information from very high-resolution aerial images and high-resolution satellite images into a single output. The U-Net architecture is employed for the spatial/texture branch, using a ResNet34 backbone model pre-trained on ImageNet. For the spatio-temporal branch,
226
  the U-TAE architecture incorporates a Temporal self-Attention Encoder (TAE) to explore the spatial and temporal characteristics of the Sentinel-2 time series data,
227
  applying attention masks at different resolutions during decoding. This model allows for the fusion of learned information from both sources,
@@ -231,46 +260,48 @@ U-T&T code repository &#128193; : https://github.com/IGNF/FLAIR-2-AI-Challenge <
231
 
232
  <th><font color="#c7254e"><b>IMPORTANT!</b></font></th> <b>The structure of the current dataset differs from the one that comes with the GitHub repository.</b>
233
  To work with the current dataset, you need to replace the <font color=‘#D7881C’><em>src/load_data.py</em></font> file with the one provided here.
234
- You also need to add the following content to your flair-2-config.yml file under the <em><b>data</b></em> tag: <br>
235
-
236
- <code>
237
-
238
- HF_data_path : " " # Path to unzipped HF dataset
239
-
240
- domains_train : ["D006_2020","D007_2020","D008_2019","D009_2019","D013_2020","D016_2020","D017_2018","D021_2020","D023_2020","D030_2021","D032_2019","D033_2021","D034_2021","D035_2020","D038_2021","D041_2021","D044_2020","D046_2019","D049_2020","D051_2019","D052_2019","D055_2018","D060_2021","D063_2019","D070_2020","D072_2019","D074_2020","D078_2021","D080_2021","D081_2020","D086_2020","D091_2021"]
241
 
242
- domains_val : ["D004_2021","D014_2020","D029_2021","D031_2019","D058_2020","D066_2021","D067_2021","D077_2021"]
243
-
244
- domains_test : ["D015_2020","D022_2021","D026_2020","D036_2020","D061_2020","D064_2021","D068_2021","D069_2020","D071_2020","D084_2021"]
245
-
246
- </code>
 
247
 
248
 
249
  <br><br>
250
 
251
 
252
  ## Reference
 
 
253
  Please include a citation to the following article if you use the FLAIR dataset:
254
 
255
  ```
256
- @misc{garioud2023flair,
257
  title={FLAIR: a Country-Scale Land Cover Semantic Segmentation Dataset From Multi-Source Optical Imagery},
258
  author={Anatol Garioud and Nicolas Gonthier and Loic Landrieu and Apolline De Wit and Marion Valette and Marc Poupée and Sébastien Giordano and Boris Wattrelos},
259
  year={2023},
260
- eprint={2310.13336},
261
- archivePrefix={arXiv},
262
- primaryClass={cs.CV}
263
  }
264
  ```
265
 
266
  ## Acknowledgment
 
 
267
  This work was performed using HPC/AI resources from GENCI-IDRIS (Grant 2022-A0131013803). This work was supported by the project "Copernicus / FPCUP” of the European Union, by the French Space Agency (CNES) and by Connect by CNES.<br>
268
 
269
 
 
 
270
 
 
271
 
272
 
273
  ## Dataset license
 
274
 
275
  The "OPEN LICENCE 2.0/LICENCE OUVERTE" is a license created by the French government specifically for the purpose of facilitating the dissemination of open data by public administration.<br/>
276
  This licence is governed by French law.<br/>
 
10
  # Datset Card for FLAIR land-cover semantic segmentation
11
 
12
  ## Context & Data
13
+ <hr style='margin-top:-1em' />
14
 
15
  The hereby FLAIR (#2) dataset is sampled countrywide and is composed of over 20 billion annotated pixels of very high resolution aerial imagery at 0.2 m spatial resolution, acquired over three years and different months (spatio-temporal domains).
16
  Aerial imagery patches consist of 5 channels (RVB-Near Infrared-Elevation) and have corresponding annotation (with 19 semantic classes or 13 for the baselines).
 
169
  <br><br>
170
 
171
  ## Dataset Structure
172
+ <hr style='margin-top:-1em' />
173
 
 
174
  The FLAIR dataset consists of 77 762 patches. Each patch includes a high-resolution aerial image (512x512) at 0.2 m, a yearly satellite image time series (40x40 by default by wider areas are provided) with a spatial resolution of 10 m
175
  and associated cloud and snow masks, and pixel-precise elevation and land cover annotations at 0.2 m resolution (512x512).
176
 
 
178
 
179
 
180
  ### Band order
 
 
 
181
 
182
+ <div style="display: flex;">
183
+ <div style="width: 15%;margin-right: 1;"">
184
+ Aerial
185
+ <ul>
186
+ <li>1. Red</li>
187
+ <li>2. Green</li>
188
+ <li>3. Blue</li>
189
+ <li>4. NIR</li>
190
+ <li>5. nDSM</li>
191
+ </ul>
192
+ </div>
193
+
194
+ <div style="width: 25%;">
195
+ Satellite
196
+ <ul>
197
+ <li>1. Blue (B2 490nm)</li>
198
+ <li>2. Green (B3 560nm)</li>
199
+ <li>3. Red (B4 665nm)</li>
200
+ <li>4. Red-Edge (B5 705nm)</li>
201
+ <li>5. Red-Edge2 (B6 470nm)</li>
202
+ <li>6. Red-Edge3 (B7 783nm)</li>
203
+ <li>7. NIR (B8 842nm)</li>
204
+ <li>8. NIR-Red-Edge (B8a 865nm)</li>
205
+ <li>9. SWIR (B11 1610nm)</li>
206
+ <li>10. SWIR2 (B12 2190nm)</li>
207
+ </ul>
208
+ </div>
209
+
210
+ </div>
211
 
212
  ### Annotations
213
  Each pixel has been manually annotated by photo-interpretation of the 20 cm resolution aerial imagery, carried out by a team supervised by geography experts from the IGN.
 
220
  Consequently, every split accurately reflects the landscape diversity inherent to metropolitan France.
221
  It is important to mention that the patches come with meta-data permitting alternative splitting schemes, for example focused on domain shifts.
222
 
223
+ Official domain split: <br/>
224
 
225
  <div style="display: flex; flex-wrap: nowrap; align-items: center">
226
  <div style="flex: 40%;">
 
249
  <br><br>
250
 
251
  ## Baseline code
252
+ <hr style='margin-top:-1em' />
253
+
254
  We propose the U-T&T model, a two-branch architecture that combines spatial and temporal information from very high-resolution aerial images and high-resolution satellite images into a single output. The U-Net architecture is employed for the spatial/texture branch, using a ResNet34 backbone model pre-trained on ImageNet. For the spatio-temporal branch,
255
  the U-TAE architecture incorporates a Temporal self-Attention Encoder (TAE) to explore the spatial and temporal characteristics of the Sentinel-2 time series data,
256
  applying attention masks at different resolutions during decoding. This model allows for the fusion of learned information from both sources,
 
260
 
261
  <th><font color="#c7254e"><b>IMPORTANT!</b></font></th> <b>The structure of the current dataset differs from the one that comes with the GitHub repository.</b>
262
  To work with the current dataset, you need to replace the <font color=‘#D7881C’><em>src/load_data.py</em></font> file with the one provided here.
263
+ You also need to add the following content to the <font color=‘#D7881C’><em>flair-2-config.yml</em></font> file under the <em><b>data</b></em> tag: <br>
 
 
 
 
 
 
264
 
265
+ ```
266
+ HF_data_path : " " # Path to unzipped HF dataset
267
+ domains_train : ["D006_2020","D007_2020","D008_2019","D009_2019","D013_2020","D016_2020","D017_2018","D021_2020","D023_2020","D030_2021","D032_2019","D033_2021","D034_2021","D035_2020","D038_2021","D041_2021","D044_2020","D046_2019","D049_2020","D051_2019","D052_2019","D055_2018","D060_2021","D063_2019","D070_2020","D072_2019","D074_2020","D078_2021","D080_2021","D081_2020","D086_2020","D091_2021"]
268
+ domains_val : ["D004_2021","D014_2020","D029_2021","D031_2019","D058_2020","D066_2021","D067_2021","D077_2021"]
269
+ domains_test : ["D015_2020","D022_2021","D026_2020","D036_2020","D061_2020","D064_2021","D068_2021","D069_2020","D071_2020","D084_2021"]
270
+ ```
271
 
272
 
273
  <br><br>
274
 
275
 
276
  ## Reference
277
+ <hr style='margin-top:-1em' />
278
+
279
  Please include a citation to the following article if you use the FLAIR dataset:
280
 
281
  ```
282
+ @inproceedings{garioud2023flair,
283
  title={FLAIR: a Country-Scale Land Cover Semantic Segmentation Dataset From Multi-Source Optical Imagery},
284
  author={Anatol Garioud and Nicolas Gonthier and Loic Landrieu and Apolline De Wit and Marion Valette and Marc Poupée and Sébastien Giordano and Boris Wattrelos},
285
  year={2023},
286
+ booktitle={Advances in Neural Information Processing Systems (NeurIPS) 2023},
287
+ doi={https://doi.org/10.48550/arXiv.2310.13336},
 
288
  }
289
  ```
290
 
291
  ## Acknowledgment
292
+ <hr style='margin-top:-1em' />
293
+
294
  This work was performed using HPC/AI resources from GENCI-IDRIS (Grant 2022-A0131013803). This work was supported by the project "Copernicus / FPCUP” of the European Union, by the French Space Agency (CNES) and by Connect by CNES.<br>
295
 
296
 
297
+ ## Contact
298
+ <hr style='margin-top:-1em' />
299
 
300
+ If you have any questions, issues or feedback, you can contact us at: ai-challenge@ign.fr
301
 
302
 
303
  ## Dataset license
304
+ <hr style='margin-top:-1em' />
305
 
306
  The "OPEN LICENCE 2.0/LICENCE OUVERTE" is a license created by the French government specifically for the purpose of facilitating the dissemination of open data by public administration.<br/>
307
  This licence is governed by French law.<br/>