Low-rank acoustic variation adapters with spectro-modulation prototype selector for incremental underwater acoustic target recognition

Underwater acoustic target recognition (UATR) in long-term marine deployments necessitates continuous adaptation to newly encountered vessel categories. However, updating models in resource-constrained environments poses a fundamental tension between computational feasibility and knowledge retention, making compact yet effective adaptation essential. To this end, we analyze cross-session acoustic variations in the frozen Transformer feature space and observe that the dominant covariance-difference energy is concentrated in a small number of singular directions. This empirical low-rank property suggests that session-dependent acoustic shifts can be effectively modeled by compact residual corrections rather than full-dimensional model updates. Based on this observation, we propose LAVA-SMPS, a lightweight session-routed incremental learning framework for UATR. The proposed framework integrates Low-Rank Acoustic Variation Adapters (LAVA) with Spectro-Modulation Prototype Selector (SMPS). Specifically, LAVA introduces session-specific low-rank residual adapters into a shared pre-trained Transformer backbone, enabling the model to capture dominant session-dependent acoustic variations while preserving previously learned knowledge. To handle the absence of session identity during inference, SMPS performs training-free session selection by fusing broadband Log-Mel prototype distances with auxiliary DEMON-based modulation cues. Experiments on public underwater acoustic datasets demonstrate that LAVA-SMPS effectively mitigates catastrophic forgetting and achieves superior incremental recognition performance compared with representative baselines.