Bayesian multimodal fusion for seafloor habitat mapping with autonomous underwater vehicles

Autonomous Underwater Vehicles collect rich multimodal remote sensing data, but robust fusion and classification of heterogeneous sensor streams remains challenging due to noise and habitat variability. We propose a Bayesian multimodal neural network for fusing optical imagery, bathymetry, and side-scan sonar, quantifying epistemic and aleatoric uncertainty to improve classification reliability in complex underwater environments. Our model leverages uncertainty estimates to identify ambiguous predictions and highlight poorly sampled regions, enabling reliable mapping and evaluation of model confidence. We evaluated our model on large-scale Autonomous Underwater Vehicle survey data, achieving 85.5% classification accuracy, surpassing unimodal baselines by ∼ 7%. We provide a preliminary complexity analysis to estimate onboard feasibility, enabling rapid onboard processing for uncertainty-aware habitat classification. We vary the number of Monte Carlo runs to explore the balance between computational cost and uncertainty estimation. We further analyse how spatial context in side-scan and bathymetric imagery affects mapping accuracy, showing that a 30 metre patch size provides the best balance between habitat continuity and boundary delineation. This work demonstrates a scalable, uncertainty-aware multimodal fusion framework, enhancing the reliability and interpretability of underwater classification and mapping. We release the full codebase and trained models via an open-source software library (PyPI: multimodal-auv).