An Explainable CausalFormer for Multimodal Reservoir Shoreline Extraction

Source abstract

Reservoir shoreline extraction is complicated by water-level fluctuation, cloud-water confusion, exposed sediment, and asynchronous sensor evidence. This study introduces CausalFormer, an explainable spatiotemporal Transformer that separates shoreline-relevant information from acquisition-specific nuisance variation through matched counterfactual interventions. Six optical bands, dual-polarization synthetic-aperture-radar backscatter, terrain derivatives, and daily reservoir levels are encoded as aligned multiscale tokens. A causal gate estimates modality relevance under cloud, speckle, and stage interventions, while region, signed-distance contour, and evidential uncertainty heads jointly produce a shoreline and confidence layer. Evaluation used 4,860 patches from six reservoirs, including 960 test patches distributed across dry, normal, and flood stages. CausalFormer obtained 91.8% mean intersection-over-union, 89.6% boundary F1, a 1.74 m shoreline-displacement error, and an expected calibration error of 0.031. Relative to the strongest multimodal baseline, these results correspond to gains of 3.7 and 4.4 percentage points in region and boundary accuracy and a 0.83 m reduction in displacement. Five-seed evaluation and paired reservoir-date bootstrap analysis were used to separate model gains from initialization and spatial dependence. Under 40% cloud occlusion, mean intersection-over-union remained 86.9%, and uncertainty-based screening raised accepted-segment boundary F1 while directing ambiguous arcs to review. The vector workflow retains shoreline geometry, confidence, intervention provenance, and review status for subsequent quality control. The results show that matched intervention evidence can support accurate, calibrated, and auditable reservoir shoreline products when its causal interpretation is restricted to the stated nuisance assumptions.