Research Of Polarimetric SAR Detection And Classification Based On Features In Rotation Domain And Deep CNN

Polarimetric synthetic aperture radar(SAR)has the ability to work day and night under almost all weather conditions and is also able to acquire the fully polarimetric information of targets.In the field of polarimetric SAR target scattering mechanism interpretation,due to the target orientation diversity effect,our research team have proposed the rotation domain along the radar line of sight to mine characterization information of target hidden in its orientation diversity.And a series of hidden features can be derived in the rotation domain.Moreover,the advanced convolutional neural network(CNN)has the great potential in the field of polarimetric SAR image processing.In this vein,this paper proposes the idea of combining the selected hidden features in rotation domain with some classical features.Some conventional classifiers or deep CNN classifier in the field of machine learning are used to develop the novel methods in the applications of polarimetric SAR land cover classification and ship detection.Specifically,the main works of this paper are as follow:(1)The polarimetric SAR land cover classification method based on selected polarimetric features and conventional classifier is proposed.This method uses the land cover discrimination ability complements of the selected hidden features in rotation domain and the classical roll-invariant features.Two conventional classifiers are used to obtain the final classification results.Comparison experiments based on AIRSAR and multi-temporal UAVSAR data have verified the better performance and robustness of this classification method.(2)The polarimetric SAR land cover classification method based on selected polarimetric features and deep CNN classifier,named as the polarimetric-feature-driven deep CNN classifier is proposed.And the SAR image classification performaces of the deep CNN classifier under different polarization modes are compared with each other.Comparison experiments based on AIRSAR and three-temporal UAVSAR data using two CNN classifiers with different structures have verified the better performance and generalization of this classification method.(3)The polarimetric SAR ship detection method based on selected polarimetric features and deep CNN is proposed.Based on the result of the sea-land segmentation processing to the full scene,the well-trained deep CNN is adopted to do the ergodic detection processing using moving window with various sizes.Comparison experiments based on RadarSat-2 and GF-3 data have verified the better performance and transfer ability of this detection method.