Ecohydrology‐Informed Machine Learning for Floodplain Vegetation Classification to Support Riverine Nature‐Based Solutions

ABSTRACT Floodplain restorations are an increasingly common nature‐based solution (NbS) to support flood risk reduction and enhance ecosystem services while protecting biodiversity. Monitoring vegetation structure and habitats in restored floodplains is critically important to support evidence‐based NbS implementation and biodiversity conservation interventions. Remote sensing (RS)‐based machine learning (ML) methods for floodplain vegetation mapping are complicated by hydrology driven, spatiotemporally dynamic habitat mosaics, high heterogeneity, spatial interspersion, and spectral mixing. Leveraging prevailing theory of floodplain ecology, ecohydrology‐informed ML modeling could improve habitat classifications. We tested the effectiveness of a novel RS and hydrological data fusion to classify vegetation and land cover types in floodplains along the lower Missouri River using supervised ML algorithms. Extreme Gradient Boosting model results showed higher classification accuracy from RS hydrological data fusion (overall accuracy, OA = 85.8%, kappa = 0.84) over multisensor (optical radar) RS data fusion (OA = 80%, Kappa = 0.77). Certain land cover classes, particularly those with structurally similar vegetative communities but ecologically important differences in hydrology (e.g., soil moisture, subcanopy inundation), showed greater classification uncertainty. The user's and producer's accuracies for palustrine emergent and upland herbaceous, two dominant land cover types associated with high uncertainty, increased by ≥ 10%, an improvement from 55% to 60% in base model to 70%–75% in ecohydrology‐informed model. Post hoc SHapley Additive exPlanations (SHAP) analysis further highlighted the importance of hydrological features in model predictions and distinguishing spectrally mixed hard‐to‐separate classes. Our results demonstrate how incorporating ecohydrological knowledge into ML modeling improves vegetation classification in heterogeneous landscapes such as large river floodplains. This methodological framework can support Natural Flood Management and NbS implementation monitoring, but further validation across river systems and years will be needed to assess broader transferability. Future work utilizing new and additional high‐resolution RS products for canopy height and subcanopy water could also make great contributions to habitat classification in floodplains.