Deep Learning-based Vessel Detection from Very High and Medium Resolution Optical Satellite Images as Component of Maritime Surveillance Systems

Today vessel detection from remote sensing images is increasingly becoming a crucial component in maritime surveillance applications. The increasing number of very high and medium resolution (VHR and MR) optical satellites shortens the revisit time as it was never before. This makes the technology especially attractive for a variety of maritime monitoring tasks. Nevertheless, it is quite a challenge to perform object detection on enormous large satellite images that cover several hundreds of square kilometers and derive results under near real time constraints. This thesis presents an end-to-end multiclass vessel detection method from optical satellite images. The proposed workflow covers the complete processing chain and involves rapid image enhancement techniques, the fusion with automatic identification system (AIS) data, and the detection algorithm based on convolutional neural networks (CNN). To train the CNNs, two versions of training datasets were generated. The VHR training dataset was produced from the set of WorldView-[1-3] and GeoEye-1 images and contains about 40 000 of uniquely annotated vessels divided into 14 different classes. The MR training dataset was generated from the set of Landsat-8 images and contains about 14 000 of uniquely annotated vessels of 7 different classes. The algorithms presented are implemented in the form of independent software processors and integrated in an automated processing chain as part of the Earth Observation Maritime Surveillance System (EO-MARISS). The solution developed from the methods presented has proven its usability within different projects and is used and further developed at the ground station of the German Aerospace Center (DLR) in Neustrelitz.