LM Modelling workflow

In the following exercises, we want to use a deep neural network, specifically U-Net, to predict the presence of buildings in the southern part of Marburg. For this, we will use a DOP of Marburg as well as a vector file containing the outlines of the buildings for the extent of the DOP. The DOP file can be downloaded here.

Image: An exemplary workflow for creating a U-Net model for buildings in Marburg.

In the first exercise, we will create masks and split the data. Then, we will prepare the data for the actual U-Net and complete a step called augmentation. In the next session we will apply the actual U-Net model and calculate performance metrics. Finally, we will use the U-Net to make spatial predictions that we can then compare with the random forest models from the previous unit.

This entire example is based on the tutorial Introduction to Deep Learning in R for the Analysis of UAV-based Remote Sensing Data [CC BY 3.0 DE] by Christian Knoth. The scripts from this tutorial are available here.