Joint Learning of Underwater Terrain Matching and Suitability Analysis via SO(2) Elevation Transformer

Abstract

Underwater navigation reliability heavily depends on accurate terrain evaluation. While conventional approaches treat terrain matching and suitability assessment as separate processes, this separation often leads to inconsistent and imprecise terrain evaluations. To address these limitations, we propose an integrated framework that employs a unified terrain map encoder to simultaneously handle both matching and suitability analysis tasks. At the core of our framework lies the SO(2) Elevation Embedding Transformer (SEET), which combines rotation-equivariant CNN with elevation embeddings. The SEET encoder is pre-trained through self-supervised contrastive learning on underwater elevation data, eliminating the need for manual labeling. Our extensive experimental validation demonstrates the framework’s effectiveness, showing superior matching accuracy with minimal navigation deviation. The distinct performance gap observed between suitable and unsuitable regions further validates the effectiveness of our suitability analysis approach.

Keywords

• Rotation invariance
• suitability analysis
• SO(2) transformer
• terrain evaluation
• underwater navigation