The automated recognition of buildings is the first step towards an automated assessment of building damage on the basis of satellite images in the event of a disaster, which in turn offers a time advantage for the emergency forces on site. For the creation of such a system, a corresponding dataset was provided by the xView2 Challenge.
This Repo provides a training pipeline for a U-Net for building footprint recognition on sattelite imagery. 2 U-Netversions with similar F1 scores are provided within this repo. The U-Nets generalise on different regions/ building types in the world. Testing scripts to evaluate F1 scores are also provided.
How to use:
- To Use the U-Nets for classification you can load them from the .pth files in the U-Net folder like so ''' torch.load("U-Net/U-Net2.0/unet2.0_disaster_vision_BCEWLL_.pth") '''
- Make sure to give 1024 x 1024 as input images
- specifiy the folder of the input images in config.py under TEST_PATHS
- For U-Net2.1 adjust the thershhold in config.py to 0.85 For U-Net1.1 adjust the thershhold in config.py to 0.31
- create a folder called f"predictionT{threshhold * 100}" with the subfolders
- predictions
- gt-targets
- visible-predictions
Note that there will be damage predictions saved in the folder, you can ignore them. They contain the same pictures as the localization predictions and are only needed to satisfy the scoring_xView2.py script
Score: U-Net2.1 and U-Net 1.1 with their specified threshholds reach an F1 score of 0.62