Machine learning algorithms for operational ship detection in synthetic aperture radar images

The project will develop new and improved algorithms for automatic ship detection and extraction of information about sea vessels and their movement from satellite images captured by radar sensors. Ship detection is the main focus of the project, but extraction of higher-level information such as ship type, size and heading is a natural next task. The project will utilize modern machine learning methodology which has revolutionized image analysis in recent years, known as deep learning with artificial neural networks. Deep learning is particularly well suited to exploit the vast amounts of training data available through the existing processing environment at Kongsberg Satellite Services (KSAT) for operational ocean monitoring, which includes both multimission satellite images and ship information from the Automatic Identification System tracking vessels at sea. This methodology can efficiently learn how to distinguish between ship targets and their look-alikes based on patterns in the data that can be discovered by artificial intelligence. The aim of the work is to develop methods that can be applied to both existing satellites and upcoming missions, including new and innovative radar imaging modes, such as those employed by the planned constellation of Norwegian MicroSAR satellites. A subtask is to find strategies for efficient adaptation of developed algorithms to new sensors and radar imaging modes. This can be done by transfer learning, which includes judicious reuse of certain parts of the neural network architectures, while modifying and retraining other parts. The project will utilize KSAT's unique access to an operational processing chain for satellite images, large data amounts, application knowledge, and experience with customer requirements to processing speed, detection performance, reliability and robustness.

The project will develop new and improved algorithms for operational inference of information about the presence and properties of sea vessels based on deep learning in SAR images. Ship detection is the main focus of the project, but extraction of higher-level information, such as ship type, size and heading, is a natural next task. The work will both target existing satellites and prepare for upcoming missions with innovative SAR modes. A subtask is to find strategies for efficient adaptation of developed algorithms to new sensors and SAR modes. This can be done by transfer learning, which includes judicious reuse of certain parts of the deep architectures, while modifying and retraining other parts. Deep learning methods can efficiently exploit contextual information to identify the characteristics of vessels that extend over multiple pixels and reject the typical spatial patterns of range and azimuth ambiguities. Another motivating property is their ability to learn features implicitly from training data, thus avoiding explicit feature extraction and encoding of prior knowledge. Operational ship detection introduces challenges that can be efficiently dealt with by deep learning from example data, including rejection of ambiguities and target look-alikes. The literature holds some examples of successful application of deep learning to ship detection, but much work is still needed to transfer research into an operational systems and obtain the performance required in an environment governed by customer specifications and near real-time constraints, concerning processing speed, detection performance, reliability and robustness. The validation process will utilize the operational processing environment at KSAT with access to large amounts of training data. The project will profit on unique access to multimission SAR data, ground truth data from the Automatic Identification System network, and knowledge about user requirements.