Deep Learning for Building Density Estimation in Remotely Sensed Imagery

This paper is about point-wise estimation of building density from remote sensing optical imagery using deep learning methods. Convolutional neural network (CNN) based deep learning approaches are used for this work. Pre-trained VGG-16 and FCN-8s deep architectures are adapted to the problem and line-tuned with additional training. Estimated values are used to generate building heat maps in urban areas. Comparative simulation results of the two architectures reveal that accurate density estimation is possible without the need for detailed maps of building locations during supervised training.