Geometry-Aware Supervision for Connectivity-Preserving Road Extraction via SAM2

Road extraction from remote sensing imagery is a structured problem: a mask with high region overlap may still contain breaks that make the road network unusable, and the difficulty grows when annotated sources are small, heterogeneous, and unequally represented during joint training. We revisit this problem on a SAM2 encoder while keeping all added complexity in training rather than in the deployed network. First, we do not use SAM2 as a black box: its top-down pathway refines only the coarsest levels, leaving the quarter-resolution features our decoder needs semantically shallow, so a feature-enhancement neck restores the missing top-down path and a directional attention module matches the elongated geometry of roads. Second, we decompose road structure into five complementary geometric views—skeleton, direction, width, neighbourhood affinity, and endpoints—used as a shared training-time prior; each is a deterministic function of the segmentation mask, so the prior needs no extra annotation and no learned generator, and its heads are removed after training, leaving the inference graph free of auxiliary computation; the feature-enhancement neck and directional attention module remain in the deployed model, as they serve the segmentation head. A soft-clDice term supervises centrelines, and a square-root sampler rebalances the 5.2:1 update-frequency skew across sources. On Massachusetts, DeepGlobe, and CHN6-CUG, the model reaches an aggregate IoU of 0.7403 and CC-agreement of 0.6874, improving CC-agreement by 44% over D-LinkNet and 10% over the strongest same-backbone adaptation. We also make evaluation part of the contribution: we measure an empirical uncertainty scale for topology metrics and hold every claim to exceed it. Held to this standard, the prior’s CC-agreement gain clears the scale only in the scarce-annotation regime that limits operational mapping yet costs nothing to retain elsewhere. Finally, region overlap and CC-agreement rank models differently, so both metric families should be reported against a stated measurement scale.

Authors

Institutions

Publication Details

Journal
Remote Sensing
Published
2026-11-18
DOI
https://doi.org/10.3390/rs18223966
Primary Topic
Automated Road and Building Extraction
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Geometry-Aware Supervision for Connectivity-Preserving Road Extraction via SAM2

Zhangjia Jiang, 谢利明 Xie Liming, Kai Yang, Jinlong Li et al.
3 links
Remote Sensing
Automated Road and Building Extraction
article

Geometry-Aware Supervision for Connectivity-Preserving Road Extraction via SAM2

Zhangjia Jiang, 谢利明 Xie Liming, Kai Yang, Jinlong Li, Yu Zhang
article en

Abstract

Road extraction from remote sensing imagery is a structured problem: a mask with high region overlap may still contain breaks that make the road network unusable, and the difficulty grows when annotated sources are small, heterogeneous, and unequally represented during joint training. We revisit this problem on a SAM2 encoder while keeping all added complexity in training rather than in the deployed network. First, we do not use SAM2 as a black box: its top-down pathway refines only the coarsest levels, leaving the quarter-resolution features our decoder needs semantically shallow, so a feature-enhancement neck restores the missing top-down path and a directional attention module matches the elongated geometry of roads. Second, we decompose road structure into five complementary geometric views—skeleton, direction, width, neighbourhood affinity, and endpoints—used as a shared training-time prior; each is a deterministic function of the segmentation mask, so the prior needs no extra annotation and no learned generator, and its heads are removed after training, leaving the inference graph free of auxiliary computation; the feature-enhancement neck and directional attention module remain in the deployed model, as they serve the segmentation head. A soft-clDice term supervises centrelines, and a square-root sampler rebalances the 5.2:1 update-frequency skew across sources. On Massachusetts, DeepGlobe, and CHN6-CUG, the model reaches an aggregate IoU of 0.7403 and CC-agreement of 0.6874, improving CC-agreement by 44% over D-LinkNet and 10% over the strongest same-backbone adaptation. We also make evaluation part of the contribution: we measure an empirical uncertainty scale for topology metrics and hold every claim to exceed it. Held to this standard, the prior’s CC-agreement gain clears the scale only in the scarce-annotation regime that limits operational mapping yet costs nothing to retain elsewhere. Finally, region overlap and CC-agreement rank models differently, so both metric families should be reported against a stated measurement scale.

Remote SensingVol. 18(22)
Southwest Jiaotong University (CN)
National Key Research and Development Program of China
Industry, innovation and infrastructure
Automated Road and Building Extraction
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon