We use a convolutional neural network to detect buildings in satellite data. We have tried several models and acheive >0.6 intersection over union. We use our trained model on images of areas affected by natural or manmade disasters to estimate the number of buildings distroyed.
For details of the ideas and models used in this project please refer to the full project report.
A result example from our neural net model can be seen below
The number of buildings in these threshold images are counted using by detecting the contected blobs in the
images
Finally we use the trained model to count the number of destroyed buildings in before and after images of natural or manmade disasers