Deep Convolutional Neural Networks for Modeling Patterns of Spaceborne Interferometric SAR Systems Signals

摘要

Spaceborne radar systems (SAR) are nowadays considered as very beneficial schemes towards a successful implementation of many engineering applications such as surveillance, maritime traffic management, reconnaissance, etc. Among others, modeling and prediction of ionospheric disturbances are considered as crucial towards successful SAR-based engineering applications. However, modeling ionospheric disturbances behavior is a very challenging research task due to high non-linearities involved in the mature of the data and their dynamics. For this reason, previous research efforts have been concentrated on the use of adaptable neural networks models and echo state machines that enable the effective modeling and prediction of the ionospheric disturbances. In this paper, we investigate the use of deep machine learning algorithms and particularly of Deep Convolutional Neural Networks (CNNs). Deep machine learning paradigm has been introduced in the last decade as a new advanced tool for modeling complex dynamic processes. Deep learning better emulates human brain operation, and makes effective processing of large amounts of unlabeled training data for extracting structures and internal representations from the raw sensory inputs. In this paper, CNNs are used to identify patterns in Spaceborne Interferometric SAR (InSAR) systems signals. CNNs can be controlled by varying their depth and breadth, and they also make strong and mostly correct assumptions about the nature of data taking into account local dependencies and varying statistics. SAR systems signals derived from real interferograms produced by earthquakes occurred in Greece the last fifteen years from the Dionysos Satellite Observatory of the National Technical University of Athens (NTUA) in Greece and objective criteria, such as false positives and negatives, are used to evaluate the efficiency of the proposed schemes.