Topology-preserving river vectorization via HydroSAM and differentiable graph optimization under coarse labels

Abstract

Accurate river-network extraction supports global hydrological management and environmental monitoring. Automated extraction from remote-sensing imagery remains challenging because river morphology is highly variable, fine tributaries have weak visual signals, canopy and infrastructure can obscure channels, and dense annotations are scarce or spatially incomplete. Conventional raster predictions can also be fragmented and lack structural continuity. We present HydroSAM, a three-stage framework for raster extraction and topology-aware vectorization. A frozen SAM 2 Hiera encoder is adapted with QV-LoRA and the proposed Directional Curvilinear Prior Injection (DCPI), while a gated Digital Elevation Model (DEM) branch contributes coarse-scale topographic context. Segmentation is produced in a zero-prompt setting by the native SAM 2 mask decoder. A graph-guided differentiable vectorization stage then converts probability maps into connected parametric river networks. We also present M2Hydro, a benchmark containing 7,947 high-resolution 4096 × 4096 source image-annotation pairs across China. In single-run tests, HydroSAM-MM attains an F1 of 0.6995 and clDice of 0.7753 on M2Hydro. The evaluations indicate that HydroSAM supports parameter-efficient adaptation and is more resilient to the evaluated cross-regional domain shift than the convolutional and Transformer baselines.