Cross-sensor continual maritime object recognition via spectral shift modulation in heterogeneous infrared–visible UAV imagery

Multi-modal maritime object recognition is pivotal for UAV-assisted search and rescue. However, large-scale multi-platform UAV deployments inevitably encounter heterogeneous infrared sensors with diverse spectral response functions, leading to non-stationary multi-modal distributions and severe dynamic spectral shifts across successive multi-platform UAV deployment phases. Existing continual learning paradigms typically focus on mitigating catastrophic forgetting via memory replay, but overlook the sensor-specific biases inherent in spectral variations. Such biases induce negative transfer and performance degradation, as the model fails to decouple stable object features from dynamic spectral fluctuations. To address these challenges, we propose C3SM, a cross-sensor spectral shift modulation framework. The core of C3SM is a design-inspired spectral shift parsing prior that decouples sensor-specific spectral responses from sensor-invariant structural cues. By anchoring the fusion features to a structure-consistent reference space within the spatial-frequency domain, the proposed prior quantifies the cross-sensor discrepancy into a spectral shift index. Leveraging this proxy index, C3SM integrates a prior-guided cross-sensor memory modulation module. This module stabilizes the continual learning trajectory by adaptively suppressing unreliable high-shift memory experiences. Furthermore, a shift-robust contrastive learning strategy enforces feature alignment across spectral shift clusters, ensuring the capture of robust cross-sensor correlations. Extensive evaluations on the UAVSensorCL dataset across four heterogeneous sensors illustrate that C3SM achieves a 48.88% mAP at 7.15 FPS, outperforming state-of-the-art CODA-Prompt-PL (45.92% mAP) and ER-PL (45.09% mAP). By mitigating spectral interference, C3SM provides a robust baseline for maritime object recognition across dynamic environments. The dataset and code are available at https://github.com/Dingyuan-Chen/C3SM .