Deep Learning in Remote Sensing

This thesis presents a study into Deep Learning with a domain specific lens of Remote Sensing. In particular, the aims are aligned with tasks that are of practical utility for residential solar system (PV) analysis and adaptation, in addition to disaster relief planning and fields in which remote sensing is necessary to map urban landscapes. We present 5 studies which each have their own particular focus. Initially we study the ability by which we may detect and localize PV systems with the help of aerial imagery. We leverage pre-trained models from Germany and explore their out-of-the-box utility for inference in Denmark. The second component of our thesis revolves around the vectorization of building shapes using aerial or satellite imagery. We explore deep generative models in the form of auto-regressive conditional recurrent neural networks and transformer models, and demonstrate state of the art performance on proprietary acquired data for single residential homes as well as benchmark datasets inferring multiple objects from one image. Finally we demonstrate the ability by which we may infer 3D shapes using aerial imagery. By leveraging an auto-regressive deep generative model, we find that we are capable of learning 3D points and surfaces from just a single image per building. We believe our studies in this thesis may provide a strong argument for further study of autoregressive deep generative models in the remote sensing context, and perhaps one day become a natural component in 3D building foundation models.