kipoju commited on
Commit
2d6cf8d
·
verified ·
1 Parent(s): 9f6b4ac

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +1 -1
README.md CHANGED
@@ -16,7 +16,7 @@ size_categories:
16
 
17
  ## WorldFloods Phisat-2 dataset
18
 
19
- The **WorldFloods Phisat-2 dataset** is an adaptation of the [WorldFloodsv2 dataset](https://huggingface.co/datasets/isp-uv-es/WorldFloodsv2) created with the purpose of simulating a Phisat-2 like dataset for
20
  Phisat-2 like images were simulated using the code in the [OrbitalAI Challenge](https://github.com/AI4EO/orbitalAI). For storing the dataset, we followed the [PhiSatNet repo](https://github.com/sirbastiano/PhiSatNet) Data Specification Format.
21
  Since the WorldFloodsv2 dataset contains full flood scenes rather than chips, we created non-overlapping Train and Validation patches of size 512 from the simulated images, and only stored patches with less than 20% cloud cover, as we do on the fly to train the flood segmentation models described in the paper [Global flood extent segmentation in optical satellite images](https://www.nature.com/articles/s41598-023-47595-7).
22
  For Test samples, however, we kept the full scenes as this is a good practice when evaluating flood detection models.
 
16
 
17
  ## WorldFloods Phisat-2 dataset
18
 
19
+ The **WorldFloods Phisat-2 dataset** is an adaptation of the [WorldFloodsv2 dataset](https://huggingface.co/datasets/isp-uv-es/WorldFloodsv2) created with the purpose of simulating a Phisat-2 like dataset for traning Foundation Models for this mission.
20
  Phisat-2 like images were simulated using the code in the [OrbitalAI Challenge](https://github.com/AI4EO/orbitalAI). For storing the dataset, we followed the [PhiSatNet repo](https://github.com/sirbastiano/PhiSatNet) Data Specification Format.
21
  Since the WorldFloodsv2 dataset contains full flood scenes rather than chips, we created non-overlapping Train and Validation patches of size 512 from the simulated images, and only stored patches with less than 20% cloud cover, as we do on the fly to train the flood segmentation models described in the paper [Global flood extent segmentation in optical satellite images](https://www.nature.com/articles/s41598-023-47595-7).
22
  For Test samples, however, we kept the full scenes as this is a good practice when evaluating flood detection models.