Understanding sediment redistribution in shallow coastal lagoons remains challenging due to strong nonlinear interactions between hydrodynamics, sediment supply, and basin morphology. Although satellite observations enable detailed mapping of bathymetric change and suspended sediment dynamics, most studies remain descriptive and lack physically interpretable modeling frameworks. This study presents a spatially aware and explainable deep-learning framework for investigating pixel-scale morphodynamics in Chilika Lagoon, India, a monsoon-dominated coastal lagoon characterized by highly heterogeneous erosion-deposition patterns. Satellite-derived bathymetry, suspended sediment concentration (TSS), plume indicators, and hydrodynamic forcing variables were integrated within a convolutional neural-network framework using spatial patches. Morphodynamic states (erosion, stability, and sedimentation) were classified across more than 4.6 million valid pixels. Spatial block cross-validation demonstrated robust predictive performance, with the primary physics-based model achieving 71.6% overall accuracy and nearly 80% spatial pixel agreement despite excluding explicit geographic coordinates. Comparative ablation experiments revealed that removing row and column coordinates produced only negligible performance reductions, confirming that predictions were governed primarily by physically meaningful hydro-morphodynamic controls rather than spatial memorization. SHAP-based explainable AI identified wave forcing, distance-to-inlet connectivity, bathymetric depth, and tidal forcing as the dominant predictors controlling erosion-deposition organization. Threshold sensitivity analysis further demonstrated that morphodynamic separability strongly depends on the Δ D e p t h classification strategy. The proposed framework provides a spatially robust and physically interpretable approach for lagoon-scale morphodynamic analysis and environmental management.
Mre et al. (Mon,) studied this question.