AerialToken: Geometry-Guided Domain Generalization for Remote Sensing Semantic Segmentation
Abstract
Domain-generalized semantic segmentation (DGSS) is critical for remote sensing and Earth observation imagery, where models trained on annotated satellite or unmanned aerial vehicle (UAV) imagery from a source region often suffer substantial performance degradation when deployed to unseen sensors, spatial resolutions, seasonal conditions, and geographic areas. However, acquiring dense pixel-level annotations for every new remote sensing scene remains costly and difficult to scale. Recent adapters built upon frozen visual foundation models (VFMs) have improved source-only generalization, but they still rely primarily on RGB appearance, whose spectral and textural statistics vary considerably across remote sensing domains. Although monocular depth provides structural cues, dense depth predictions in overhead imagery are sensitive to flight altitude, ground sampling distance, sensor intrinsics, and scene layout. We propose AerialToken, a geometry-aware framework that uses a frozen depth foundation model as an intermediate geometry token generator rather than a final depth prior. Predecoder representations from Depth Anything V2 are placed in parameter-free block-wise correspondence with a frozen visual encoder through geometry representation alignment, injected through layerwise dual route token calibration, and aggregated into geometry-guided decoder queries for the query-conditioned structural decoder. Experiments on two single-source cross-domain remote sensing protocols spanning satellite and UAV benchmarks show consistent improvements on unseen datasets without target-scene annotations. The results indicate that intermediate geometry representations provide more stable height-like structural cues than final dense depth maps where the domain shift alters depth scale, and that the two representations become comparable where it does not. The code is available at https://github.com/SY-Ch/AerialToken