This notebook trains a U-Net convolutional neural network to extract building footprints from aerial images taken in Massachusetts.
The image data consists of 3347 colour raster of dimension 256x256x3 using the NAD83 (EPSG:26986) projection system. Each raster represents an area of 300 square meters. The label data consists of building footprints vectors extracted from OpenStreetMap and converted into a binary raster with the same extent and resolution as the image rasters. The unformatted data was provided by Minh (2013) and can be accessed on his website. Images and labels are randomly partitioned into a training (70%) validation (15%) and test sample (15%).
The U-Net convolutional neural network is coded in Tensorflow. Adam is used as an optimizer and Binary Focal Cross Entropy as the loss function. In Step 8 a spatial dropout layer is applied to enhance performance. The model was trained for 30 epochs with a batch size of 32. It had a loss of .04 and an accuracy of .9
Diagram of the U-Net convolutional network:
Example of model output: