Answers are grounded in this paper and cite their sources.
Wetlands represent vital ecosystems requiring precise monitoring for conservation, with hyperspectral imaging serving as a crucial remote sensing technology. However, significant domain shifts caused by hydrological dynamics (e.g., tidal fluctuations and seasonal water variations) and imaging environment disparities critically degrade model generalization in unseen wetland environments. To overcome this limitation, we propose HydroMetaNet—a hierarchical learning framework integrating hydrological prior knowledge with deep domain generalization, including first, a flexible wetland spectral–spatial adaptation network (F-WSDN) that dynamically encodes state-specific features through adaptive instance-batch normalization hybrid and spectral–spatial attention. Second, wetland multiscale feature pyramid (WMFP) capturing hierarchical patterns from textures to large-scale distribution via parallel dilated convolutions; and third, hydro-feature enhancer disentangling water-impacted components through water-index-inspired saliency gating. These representations are synergistically optimized by invariant contrastive learning and adaptive meta-learning prototypes, establishing dynamic decision boundaries that absorb cross-scene hydrological differences. Evaluated extensively across three cross-domain wetland benchmarks in China, HydroMetaNet demonstrated overall accuracy improvements of 7.19% in cross region, 1.62% in cross-seasonal, and 0.79% in cross-sensor generalization tasks compared to the best baselines. The framework exhibits exceptional capability in minority-class discrimination while maintaining leading-average accuracy and kappa metrics across all scenarios.
No takes yet. Share an insight, caveat, or question.
Zhao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: