HyperR3SNet: Leveraging Hyperbolic Space and Vision Foundation Models for Remote Sensing Semantic Segmentation

Remote sensing semantic segmentation is driven by land-use monitoring, urban planning, and ecological assessment, yet progress is hampered by scarce pixel-level labels. To address this issue, we present HyperR3SNet, which is an efficient framework for remote sensing semantic segmentation that tackles data scarcity and scale variations in overhead imagery. HyperR3SNet transfers self-supervised vision foundation models (VFMs) to remote sensing, providing strong feature generalization with minimal labeled data. Building on this, a multiscale cross-attention (MSCA) module is inserted into the backbone, Vision Transformer layer, enabling the network to extract richer features across widely varying object scales. To keep the model lightweight, a matrix-factorized task head is employed, sharply reducing parameters and computation while sustaining accuracy. HyperR3SNet is the first model to integrate a hyperbolic pixel-level loss with VFM adaptation for cross-domain and low-annotation remote sensing segmentation, leveraging the exponential expansion property of hyperbolic geometry to capture latent interclass relations and preserve structural consistency under weak supervision. Evaluated on the widely adopted remote sensing segmentation datasets (e.g., iSAID, LoveDA, Potsdam, and Vaihingen), HyperR3SNet achieves mean intersection over union (mIoU) of 67.60%, 55.86%, 80.07%, and 96.46%, respectively, and on average surpasses a broad range of state-of-the-art (SOTA) methods.