Authors:
- Gabriel da Silva Zech (GabZech)
- Julian Kath (juka19)
- Krishnamoorthy Manohara (KrishnaM313)
- Florian Winkler (f-winkler)
- Nassim Zoueini (nassimzoueini)
This tutorial provides an end-to-end workflow of image segmentation based on aerial images. It introduces a U-net convolutional neural network approach to segmenting buildings from aerial imagery as a specific application of deep learning in a public policy context. Built in a PyTorch environment, the tutorial provides users step-by-step explanations of image segmentation and an example of reproducible, working code in a self-contained notebook. Users will benefit from a structured and practical overview of how to collect and pre-process aerial image data, how to create a custom dataset that annotates aerial images using building footprints, and how to train and fine-tune an image segmentation model on aerial imagery. The tutorial can be extended to further projects that involve a similar approach to aerial image segmentation, such as segmenting roads or crop fields.
The self-contained tutorial notebook can be found in this repository under Tutorial_notebook.ipynb.
© Header image was created by ourselves in Python using raw images and building masks from GEOportal.NRW