Buildings in NRW¶
Image data source: NRW OpenGeodata
In our third case study we want to detect buildings in remotely sensed image data.
Research objective¶
Our aim is to classify an image (orthophoto) into building- and no-building pixels. To this end, we want to compare the results of a simple pixel based model (Random forest) with a more complex deep learning model (UNet).
As the state of North Rhine-Westphalia (NRW) provides high quality orthophoto and real estate data on its NRW OpenGeodata platform, we will use these data sets for model training and validation:
Orthophoto |
Real estate data |
|---|---|
|
|
Task¶
Develop a conceptual approach for this task and discuss it in your group.
How would you implement the processing chain for a Random forest application?
Try to find information on UNets and how they handle data.
How could such a model be applied in our case?
What could be possible difficulties?
How do we have to adapt a “traditional” UNet model for our purpose?