A Bag-of-Meaningful-Words Integrating Scattering and Structure Information for PolSAR Image Classification

We propose a feature descriptor that integrates both the physical scattering properties of local targets and the contextual structure of HR PolSAR image patches. Firstly, the physical scattering properties are integrated by learning a meaningful scattering vocabulary, composed of the physically optimized entropy\anisotropy\alpha (H\A\alpha) classification labels, computed on small windows (e.g., of 5 x 5 pixels). Then, in a Bag-of-Meaningful-Words (BoMW) fashion, we compute the histograms of these labels over the learned vocabulary on larger windows (e.g., of 64 x 64 pixels), thus capturing the structure of image patches. We validated the quality of the obtained feature descriptors by computing K-Nearest-Neighbor (KNN) classifications of two HR, L-band, airborne polarimetric SAR images of the F-SAR and UAVSAR sensors, showing significant classification improvements compared with other state-of-the-art methods.