This project is part of the master thesis Possibilities to improve the performance and robustness of U-Nets for image segmentation in satellite images with a focus on attention mechanisms and transfer learning with the Department of Data Science and Knowledge Engineering at Maastricht University.

• Thesis
• Presentation

This master thesis aims to perform semantic segmentation of buildings on satellite images from the SpaceNet challenge 1 dataset using the U-Net architecture. To enable a more focused use of feature-maps, soft attention mechanisms are integrated into the U-Net and examined. Furthermore, possibilities of transfer learning are investigated by using convolutional neural networks pre-trained on ImageNet as encoders for the U-Net. Finally, the performance and robustness for the segmentation task is evaluated for both approaches.

It is assumed that you have anaconda Python installed. You can create a Python environment with the required dependencies by following:

conda create -n thesis_env python=3.7
conda activate thesis_env
conda install numpy pandas rasterio shapely geopandas matplotlib imageio scikit-image
conda install pytorch torchvision torchaudio cudatoolkit=11.0 -c pytorch
conda install tifffile

If your CUDA versions differ please refer to Start Locally | PyTorch.

The model weights can be found at:

• U-Net
• Attention U-Net
• CBAM U-Net
• Residual Attention U-Net
• scAG U-Net
• DenseNet121 U-Net
• MobileNetV2 U-Net
• ResNet34 U-Net
• VGG11 U-Net

• Philip Rinkwitz

This project is licensed under the MIT License - see the LICENSE.md file for details. The helper methods in Utils/SolarisHelpers.py are taken from the Solaris project that is licensed under the Apache-2.0 License. Please refer to that license if you plan on reusing that code.