Domain Adaptive Target Detection with Optimal Transportation for Different Satellite SAR Images

Satellite-borne SAR image target detection is widely used across various fields. However, there are different data distributions among SAR images from different satellite sensors and different imaging conditions. This makes it challenging for deep learning based detectors to achieve performance generalization for different data domains with distributional discrepancies. In this paper, an Optimal Transport based domain Adaptive target Detection (OTAD) method is proposed to improve target detection performance under the data distributional discrepancies in the realistic scenarios. OTAD utilizes geometry-aware clustering assignment to estimate prototypes corresponding to different geometries. Subsequently, unbalanced optimal transport is employed to map target features to source prototypes corresponding to their respective geometries. In this way, OTAD accomplishes feature adaptation between source and target domains, reducing feature distributional discrepancies while preserving intra-domain structures. Two cross-domain target detection tasks are constructed using Gaofen-3 and Sentinel-1 SAR data in the experiments. The cross-domain detection results demonstrate that OTAD effectively reduces feature distributional discrepancies and improves target detection performance for different satellite SAR images.