Redefining Bankfull Stage Using Machine Learning Prediction in a Large Lowland River

Abstract Understanding and predicting bankfull stage is essential for geomorphic analysis, hydraulic modelling and river corridor management, especially regarding floods and connectivity. This study evaluates traditional geomorphic and hydrologic methods and introduces a machine learning (ML) approach to predict bankfull stage along the upper Apalachicola River, Florida, United States. First, we integrated LiDAR point cloud data with a 2010 hydrographic survey through a triangular irregular network (TIN) following coordinate transformation, producing a 1.5‐m resolution digital elevation model (DEM) that captures riverbed elevation. After extracting cross‐sections, we applied two geomorphic methods, Width‐to‐Depth Ratio (WDR) and Hydraulic Mean Depth (HMD), and three (1.1‐, 1.5‐, 2.0‐year) hydrologic return intervals to 344 cross‐sections to determine bankfull stage and assessed their agreement with visual and field‐verified bankfull elevations. A refined HMD method using shape ratio (Rs) and slope inflection corrections improved geomorphic estimates in 131 cross‐sections. Bankfull estimates based on return intervals exhibited substantially lower reliability relative to geomorphic methods, with associated confidence levels falling below 70%. This reduced performance is attributable to stage‐based spatial averaging and limited sensitivity to local topographic variability. By contrast, the geomorphic methods (WDR and HMD) achieved confidence levels exceeding 95%, underscoring their stronger agreement and robustness. To improve accuracy, we developed ML models: Random Forest (RF), Gradient Boosting (GB) and Ensemble model trained on cross‐sectional elevation profiles and engineered features such as cross‐sectional area, top width, maximum depth, symmetry, etc. The GB model outperformed all others ( R 2 = 0.94, MSE = 0.21), with feature importance analysis revealing that elevation at the top bank as well as channel area, top width and maximum depth dominated predictions. While this study advances the integration of ML into fluvial systems and provides a replicable framework for large rivers with limited hydrologic data, we recommend that multiple ML models be evaluated across individual reaches to account for the unique geomorphic and hydraulic characteristics of each river reach.