OptiSAR-Net: A Cross-Domain Ship Detection Method for Multisource Remote Sensing Data
Jun Dong, Jiewen Feng, Xiaoyu Tang
Abstract
Optical and synthetic aperture radar (SAR) remote sensing are crucial for ship detection. Integrating SAR’s all-weather imaging with optical data’s shape recognition enhances downstream applications. However, current cross-domain methods often use unsupervised or semi-supervised techniques for single-source detection, limiting their practical use in cross-domain ship detection. Inspired by human visual cortex mechanisms, this article proposes OptiSAR-Net, an end-to-end cross-domain multisource ship detection network. Specifically, OptiSAR-Net features dual adaptive attention (DAA) for extracting standard features from SAR and optical images, and bilevel routing deformable spatial pyramid pooling-fast (BSPPF) for adapting to multiscale changes. To mitigate SAR noise, we employ VoV-GSCSP with spatial shuffling attention (VSSA) in the neck. OptiSAR-Net achieved state-of-the-art average precisions (APs) of 88.6% and 91.3% on the optical datasets DOTA and HRSC2016, respectively, and showed strong performance on the SAR datasets HRSID and SSDD. On the cross-domain heterogeneous dataset (CDHD), OptiSAR-Net differentiated ship targets effectively with only 2.7 million parameters and 11.7 GFLOPs, achieving an inference speed of 89 FPS on an NVIDIA RTX 3090. These results demonstrate that cross-domain multisource detection significantly enhances performance and application potential compared to single-source detection. Code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/SCNU-RISLAB/OptiSAR-Net</uri>.