FINE ACQUISITION OF VESSEL TRAINING DATA FOR MACHINE LEARNING FROM SENTINEL-1 SAR IMAGES ACCOMPANIED BY AIS IMFORMATION

Abstract

Ship detection in coastal regions accompanied by machine learning could be effective in economic and martial issues. However, conventional researches on ship detection mainly focused on modifying the training model itself instead of obtaining qualified training data. In order to ameliorate the training data in the aspect of quality and quantity, this research aims to directly obtain training data from AIS information. From discrete MS information, interpolation on SAR acquisition time was conducted, followed by adjusting the Doppler frequency shift caused by each vessel's velocity. Training data was constructed from the adjusted position using internal location of AIS sensor on each type of vessel. Extracted training data by the proposed algorithm from Sentinel-1 images was tested with CNN model. As the detection performance of the extracted training data exceeded that from visual interpretation, this study concluded that qualified training data could be extracted from the proposed algorithm.