PSFNet: A Feature-Fusion Framework for Persistent Scatterer Selection in Multitemporal InSAR

In the field of multitemporal interferometric synthetic aperture radar (MT-InSAR), the selection of persistent scatterer (PS) is crucial for acquiring ground deformation product. To obtain precise ground deformation, pixels with as high signal-to-noise ratio (SNR) as possible should be selected, while pixels with low SNR should be avoided. To this end, we propose a novel framework, referred to as the PS feature-fusion network (PSFNet), for efficient PS selection. Specifically, we propose a data-driven two-branch network consisting of a ResUNet with spatial and channel attention, as well as a TANet with 3-D convolutional layers and a time-step attention block (T-Attention block), which can use not only spatial features of SAR image but also time-series phase features when selecting PS pixels. In particular, a time-step attention mechanism is proposed for accommodating to interferometric pairs with different SNRs to enhance the feature representation ability of the network. The proposed method was tested using the Sentinel-1 images, showing that it can select more PSs with higher quality compared with StaMPS. In addition, the prediction time of PSFNet requires only 0.26% of the running time of StaMPS, which greatly improves the efficiency of PSFNet for practical applications.