Building Section Instance Segmentation From Satellite Images Using Deep Learning Networks

Kurzfassung

We present an end-to-end deep learning framework for building section instance segmentation. With the combined use of learning based approaches and classical image processing we were able to fulfil the task on World-View4 high resolution imagery and reach high quality results. We show that two well known but different deep learning models can tackle the issue with different architectures and inputs comparably. A ground truth raster image with pixel value 1 for buildings and 2 for their touching borders was generated to train the models to predict both as classes on a semantic output. Most developed frameworks present building segmentation on a semantic level only, which can be crucial when the exact number and boundaries of individual buildings is needed. In our work we post process the semantic outputs with the help of watershed labelling to generate segmentation on the instance level. The approach reaches F1-scores of up to 91.48% for buildings and 43.58% for touching borders.