Detect-Cluster-Reconstruct: Three-Act On-Board UAV Landslide 3D Modeling with Multi-Constraint Clustering

In post-disaster mountainous environments characterized by the simultaneous loss of transportation, communication, and power, the acquisition, transmission, and timely processing of time-critical geospatial data become extremely challenging. These constraints underscore the urgent need for real-time on-board processing of data on lightweight, energy-efficient platforms such as unmanned aerial vehicles (UAVs). However, conventional photogrammetric 3D reconstruction pipelines typically rely on holistic, batch-mode processing and large-scale compute clusters, making them impossible for deployment on low-power embedded systems. To address this limitation, this paper proposes Detect-Cluster-Reconstruct, a three-act on-board 3D modelling framework for real-time reconstruction of typical geological disaster targets, such as landslides, featuring a multi-constraint clustering strategy tailored for resource-limited UAV platforms. Specifically, we first deploy a TensorRT-accelerated YOLOv11n-seg model on an onboard edge computing unit to perform online detection of landslide regions from real-time imagery captured during drone flight, achieving sub-second response times. Second, we employ a multi-constraint cross-view instance clustering strategy that integrates spatial proximity, semantic embedding similarity, and local feature matching consistency to explicitly filter out redundant observations from non-disaster areas and retain only those corresponding to genuine landslide instances. Finally, we conduct localized 3D reconstruction solely over the clustered landslide regions, avoiding unnecessary full-scene modeling and significantly reducing computational and storage overhead, thereby generating high-precision, instance-level 3D models of landslides. Validation on real-world drone datasets from four typical landslide-prone regions in Sichuan Province demonstrates that our framework improves clustering accuracy (measured by the Adjusted Rand Index, ARI) by 67.2% over baseline tracking and traditional photogrammetric methods, while significantly enhancing 3D reconstruction efficiency.