ReportsNet: Rotation-Enhanced Polar-Radial Topology Synergy Network for Hyperspectral and LiDAR-Derived DSM Joint Classification

Abstract

Joint classification of hyperspectral (HS) imagery and light detection and ranging (LiDAR)-derived digital surface models (DSMs) commonly represents local neighborhoods on Cartesian grids, where changes in object orientation alter the spatial arrangement presented to the fusion network. Existing fusion architectures do not explicitly encode center-relative radial distances, cyclic angular relationships, or geometrically corresponding cross-modal locations. To address this gap, the rotation-enhanced polar-radial topology synergy network (ReportsNet) is proposed, which locally reorganizes co-registered HS and DSM patches into target-centered polar-radial representations and models complementary distance-based and relative-angle-aware context for multimodal fusion. On the MUUFL, Trento, and Houston datasets, ReportsNet obtains OAs of 91.99%, 99.45%, and 90.69%, respectively. It achieves the highest OA on all three datasets, exceeding the next-best compared methods by 1.44, 0.44, and 1.28 percentage points, respectively. In the Houston rotation experiment from 0° to 90°, ReportsNet yields an OA standard deviation of 0.06%, indicating stable performance under the evaluated perturbations. The code is available at https://github.com/LuyangLyu/radial-flow