.png){kind=icon}
.png){kind=icon}
Return
Balancing segmentation accuracy and hydrological plausibility in multi-source remote sensing flood inundation mapping with adaptive terrain-constrained learning
.png){kind=icon}
PRE
AI
Y
G
A
B
J
Y
Q
DOI：10.1016/j.jhydrol.2026.136209.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Remote sensing-based flood inundation mapping is commonly optimized as a semantic segmentation task using pixel-overlap objectives. However, flood extent is also a terrain-controlled spatial process, and accuracy-oriented models may generate locally implausible patterns, including downstream discontinuities, false positives on steep slopes, and incomplete detection of low-lying inundation. This study frames flood inundation mapping as a multi-objective problem that requires balancing image-based segmentation accuracy with terrain-hydrological plausibility. We propose an adaptive terrain-constrained learning framework that dynamically coordinates pixel supervision, regional overlap, boundary refinement, and DEM-derived terrain-consistency within a unified U-Net architecture. The terrain constraints are not intended to reproduce hydrodynamic processes; instead, they serve as lightweight soft constraints for reducing terrain-implausible patterns in discrete flood probability maps. Cross-regional 10-fold validation on the Sen1Floods11 dataset shows that reliable multi-source observability is a prerequisite for terrain-constrained optimization. Under the full SAR-optical-terrain input setting, fixed terrain-constrained losses reduced flow-direction violations but degraded F1-score and IoU, revealing an accuracy-plausibility trade-off. In contrast, the proposed Adaptive-Dynamic strategy achieved the best overall segmentation performance, with an F1-score of 0.864 and an IoU of 0.761, while maintaining low high-slope false positives and good lowland flood recovery. These results indicate that terrain-derived hydrological priors are most effective when introduced as adaptive soft constraints that regularize image-derived flood predictions without dominating image-based flood evidence, thereby reducing local prediction patterns that contradict basic terrain logic rather than simulating hydrological processes.
Keywords:
Flood inundation mapping
Terrain-constrained learning
Hydrological plausibility
Multi-objective optimization
Multi-source remote sensing
Adaptive loss weighting
Journal
IF:
6.3
Papers:
2.3W
Citations:
9.8W
Organization
Cited Papers