Logo image
Back
Navigating Hierarchies in Hyperbolic Space: A Knowledge-Driven Graph Network for Semantic Segmentation of Remote Sensing Imagery
Journal article   Peer reviewed

Navigating Hierarchies in Hyperbolic Space: A Knowledge-Driven Graph Network for Semantic Segmentation of Remote Sensing Imagery

Wenqi Cui, Peng Luo, Weisong Li, Xing Xu, Jiale Chen and Chuan Chen
International journal of applied earth observation and geoinformation, Vol.152, p.105435
08/2026
DOI: 10.1016/j.jag.2026.105435
url
https://doi.org/10.1016/j.jag.2026.105435View
Published (Version of record) Open Access

Abstract

•Introduce a knowledge-driven framework based on fully hyperbolic neural operators.•Propose a Fully Hyperbolic Graph Attention Network for heterogeneous RS graphs.•The model integrates multi-scale relational knowledge in a bottom-up manner.•It improves object-level accuracy and MIoU by 7.2% and 7.6% over Euclidean baselines.•FHGAT outperforms the tangent-space approximation model by 5.4% in accuracy. Semantic segmentation of multi-scale remote sensing objects faces two key challenges: representing complex relationships and effectively integrating geographic knowledge. These relationships generate hierarchical patterns embedded with eigen properties, forming a nested system that exhibits exponential growth, a characteristic poorly captured by Euclidean-based deep learning models limited to polynomial expansion. Hyperbolic space offers a geometrically faithful alternative, but current hyperbolic neural networks rely on tangent space projections for computation, introducing significant approximation errors that degrade network performance. To address these challenges, we introduce a knowledge-driven framework founded on fully hyperbolic neural network operators. These operators are designed to be manifold-preserving, enabling all computations to be performed directly within hyperbolic space and thereby eliminating tangent space approximation errors. Building on this foundation, we propose a novel Fully Hyperbolic Graph Attention Network (FHGAT) that operates on a multi-relational heterogeneous graph representation of remote sensing objects. Our model effectively mines and integrates embedded knowledge from complex, multi-scale relationships in a bottom-up fashion. Extensive experiments validate its superiority; compared to a standard Euclidean baseline, our model increases object-level accuracy and Mean Intersection over Union (MIoU) by 7.2% and 7.6%, respectively. Notably, it also surpasses a state-of-the-art tangent-space approximation model by 5.4% in accuracy, confirming the benefits of its end-to-end hyperbolic design for resolving spectral confusion. This research marks a significant step in advancing remote sensing analysis from a data-driven to a more robust and interpretable knowledge-driven paradigm.
 Heterogeneous Graph Space Hyperbolic Space Multi-Relation Remote Sensing Semantic Segmentation

Details

Metrics

1 Record Views
1 Times Cited - Web of Science
Logo image