HyperBridge: A multisensor hyperspectral benchmark with geospatial embedding anchors for cross-sensor representation learning

Multi-sensor hyperspectral representation learning requires models that can compare observations acquired by different imaging spectrometers while preserving the narrow-band spectral structures that make imaging spectroscopy physically informative. Existing hyperspectral benchmarks and foundation models have advanced large-scale pretraining, but they are often organized around single-sensor archives or harmonized spectral interfaces and therefore provide limited support for studying dense semantic alignment under sensor-native heterogeneity. This paper introduces HyperBridge, a multi-sensor hyperspectral-geospatial embedding benchmark containing 342,616 quality-controlled paired samples from EMIT, EnMAP, PRISMA, and GF-5. Each sample preserves the native wavelength sampling, bandwidth characteristics, and valid-band mask of the source sensor, and is paired with a co-registered 64-dimensional AlphaEarth Foundations (AEF) geospatial embedding reference. Based on HyperBridge, we propose HyperAnchor, an AEF-anchored hyperspectral representation framework that combines a continuous spectral tokenizer, a dual-branch semantic?spectral encoder, gated spatial?spectral fusion, and spectral-fidelity supervision. The objective is to align sensor-native hyperspectral observations with a shared geospatial semantic space without suppressing recoverable spectral structure. Experiments on the geographically disjoint HyperBridge test split evaluate geospatial semantic retrieval, leave-one-sensor-out generalization, sensor-native spectral reconstruction, progressive ablation, and frozen downstream dense interpretation. HyperAnchor improves full-gallery AEF retrieval from the best reference R@10 of 6.17% to 31.16%, achieves the strongest average R@10 under strict leave-one-sensor-out evaluation, improves sensor-native reconstruction fidelity over interpolation-based autoencoder baselines, and increases average frozen-transfer mIoU from 19.22% to 30.83% under the same low-label protocol. These results support HyperBridge as a reusable benchmark for multi-sensor hyperspectral?geospatial alignment and show that effective AEF semantic anchoring should be coupled with explicit preservation of sensor-native spectral information. The HyperBridge dataset, benchmark protocols, and source code will be made publicly available at https://github.com/HIT-SIRS/HyperBridge.