Process-informed deep learning for interpretable forested wetland mapping under dense canopy

Canopy-obscured hydrologic signals and gradual wetland–upland transitions hinder the delineation of forested wetlands, particularly when heterogeneous Earth-observation evidence is fused as an unstructured feature stack. We developed a process-informed multi-stream convolutional neural network (CNN) for mapping forested wetlands on a 10-m modelling grid in the Changbai Mountains, China. The model used 85 covariates derived from Sentinel-1, Sentinel-2, ALOS-2 PALSAR-2, and Copernicus DEM GLO-30. These covariates were assigned a priori to five predictor streams and processed by separate CNN branches before fusion. Relative to a capacity-matched single-stream baseline, the grouped architecture increased forested-wetland F1 by 3.42 percentage points across five validation folds, whereas the gain on the fixed held-out test set was smaller (+1.03 percentage points). For map production, the probability-averaged five-fold ensemble achieved an overall accuracy of 0.873, a macro-F1 of 0.882, and a forested-wetland F1 of 0.836 on the fixed held-out test set. To interpret correlated predictors, group-aggregated SHapley Additive exPlanations (SHAP) and k-nearest-neighbor conditional replacement were used to characterize global and conditional model reliance. Among correctly classified forested-wetland samples, global reliance was greatest for temporal features and phenology, whereas conditional replacement indicated the greatest residual model reliance for hydrology, followed by topography. Direct pixel counting yielded a map-derived forested-wetland area of 686.6 km 2 , concentrated mainly in peripheral plains and hills. This framework combines process-informed feature organization with dependence-aware interpretation to support interpretable regional mapping of canopy-obscured forested wetlands.