AirNet: A Deep Learning Approach to Extracting Building Information from Remotely Sensed Imagery

Abstract

The rapid development of deep learning has given rise to new and improved ways of performing image segmentation. However, there are not much work done specifically for the task of segmenting remotely sensed imagery. The model called AirNet is, therefore, presented in this thesis. This model tests how a segmentation network performs the task of extracting building information from remotely sensed imagery. The architecture is based on the popular network called SegNet (Badrinarayanan et al., 2015), but with some modifications. The model is implemented using the open source library TensorFlow (Abadi et al., 2016).

The model is tested using two datasets that are constructed from vector data and aerial images of Norway. They contain images of 0.1-meter resolution one with RGB (red, green, blue) images, and one with IR (infrared) images.

Three versions of the model are tested; AirNet-basic, AirNet-extended, and AirNet- dropout. Results show that the best accuracy is achieved through pretraining the AirNet-extended model, and then finetuning it with the IR dataset. This gives an accuracy of 83.28 mIoU. By looking examples of the performed segmentation, it can be observed that the network manages to segment the images into areas that reflect the shape of the buildings well. Almost every building has at least some correctly classified pixels, and few areas are falsely classified as buildings.