Enhanced Multi-View Subspace Clustering via Unified Anchor Graph and Random Projection for Hyperspectral and LiDAR Data

In recent years, multi-view subspace clustering methods have demonstrated significant progress in multimodal remote sensing data clustering. However, traditional shallow models are limited in handling complex nonlinear features and fail to fully exploit the deep semantic information in remote sensing data. To address this challenge, we propose enhanced multi-view subspace clustering via unified anchor and random projection (SCUARP), which couples a parameter-shared nonlinear autoencoder, modality-dependent anchors generated by full-sample random projection, and a single cross-modal anchor-representation matrix within one training objective. This shared representation couples the structures of co-registered hyperspectral (HS) and LiDAR samples while avoiding costly full graph modeling across all samples. Experimental results on three large-scale multimodal remote sensing datasets demonstrate that our method achieves superior performance compared to existing approaches, with improvements in overall accuracy (OA) by up to 8.87 percentage points on the MUUFL Gulfport dataset and 9.67 percentage points on the Houston dataset, respectively, while offering a favorable practical trade-off between clustering accuracy and computational cost. This study contributes to the field of remote sensing and machine learning by providing an effective and scalable solution for multimodal remote sensing data clustering. Our code can be obtained at https://github.com/kcarol1/SCUARP