SAR-AIS anomaly detection via cross-modal consistency modeling and reliability-aware fusion

SAR and AIS provide complementary information and constitute important data sources for maritime target perception. However, errors in AIS static attributes and mismatches in SAR-AIS association reduce the reliability of cross-modal fusion, making consistency assessment between SAR observations and AIS attributes essential. To address this issue, this paper proposes a Reliability-aware Cross-modal Fusion Network (RCF-Net) for SAR-AIS anomaly detection. RCF-Net consists of a SAR image branch and an AIS attribute branch to encode SAR observations and static AIS attributes, respectively. A cross-modal interaction module constructs difference and product features, from which a sample-specific gating weight is learned to regulate the contribution of AIS features during fusion. Using the FUSAR-Ship dataset, three controlled anomaly scenarios involving semantic, geometric, and association inconsistencies are constructed for evaluation. Experimental results show that RCF-Net improves detection performance over Direct Fusion under the MMSI-level grouped protocol, thereby validating the effectiveness of the proposed method for controlled SAR-AIS cross-modal anomaly detection. However, scene-level evaluation indicates that generalization to unseen SAR acquisition scenes remains limited.