AMHC-DTENet: Attention-Based Multipath Hybrid Convolutional Distribution Target Extraction Network for Polarimetric Channel Imbalance Assessment

Manual selection of the distribution target and determination of its spatial location is not only time consuming but also has a serious impact on the polarimetric channel imbalances assessment if the area is incorrect. However, the existing automatic extraction methods are affected by the morphology of the features, and their local feature extraction capability is limited. On the other hand, only local information is considered, while the feature channels and global information are ignored, which in turn leads to the limited application scenarios, especially in some urban areas. The polarimetric channel imbalance estimation accuracy is so low that it exceeds the system tolerance limit. Therefore, in order to more effectively mine the polarimetric features of distributed targets in polarimetric synthetic aperture radar images, we propose an attention-based multipath hybrid convolutional distribution target extraction network for polarimetric channel imbalance assessment. First, in order to develop the local feature extraction capability of the hierarchical network, we design a hybrid convolutional module with adaptive adjustment of the receptive field size. Second, a polarimetric feature channel reconstruction module is constructed in order to utilize the spatial information of the polarimetric feature channels. Then, considering that the polarimetric information of the ground feature is sensitive to the relative geometry of the target attitude to the radar line of sight, the vision transformer architecture is used to capture polarimetric global information and extend it to multipath. Finally, Gaofen-3(GF-3) full-polarimetric data are used for experimental verification. Experimental results demonstrate the effectiveness and reliability of the proposed method.