Helipad detection from satellite imagery using convolutional neural networks

Location data about U.S. heliports is often inaccurate or nonexistent in the FAA\u27s databases, which leaves pilots and air ambulance operators with inaccurate information about where to find safe landing zones. In the 2018 FAA Reauthorization Act, Congress required the FAA to collect better information from the helicopter industry under part 157, which covers the construction, alteration, activation and deactivation of airports and heliports. At the same time, there is no requirement to report private helipads to the FAA when constructed or removed, and some public heliports do not have up to date records. This thesis proposes an autonomous system that can authenticate the coordinates in the FAA master database, as well as search for helipads in a designated large area. The proposed system is based on a convolutional neural network model that learns optimal helipad features from the data. We used the FAA\u27s 5010 database and others to construct a benchmark database of rotorcraft landing sites. The database consists of 9,324 aerial images, containing helipads, helistops, helidecks, and helicopter runways in rural and urban areas, as well as negative examples, such as rooftop buildings and fields. The dataset was then used to train various state-of-the-art convolutional neural network models. The outperforming model, EfficientNet-b0, achieved nearly 95% accuracy on the validation set. We subsequently implemented a gradient-based explainability map depicting the most salient pixels that influenced the network\u27s prediction