Modelling future shrinkage of the East Kolkata Wetlands for sustainable planning through a hybrid machine learning integrated CA-ANN framework

The East Kolkata Wetlands (EKW), a Ramsar site in India, is facing accelerating degradation from rapid urbanisation. Yet, existing land-use change models offer limited capacity for spatially targeted, evidence-informed wetland management. This study integrates ensemble machine learning (ML) and Cellular Automata-Artificial Neural Network (CA-ANN) modelling to assess historical wetland shrinkage, identify its drivers, map spatial vulnerability, and project future land-use trajectories for the EKW. Multi-temporal Landsat analysis (1991 to 2025) documents a decline in wetland extent from 91.2 km² to 33.4 km², a 63.4% loss, primarily through conversion to agricultural land (44.67 km²) and built-up areas (5.87 km²). Four ML algorithms (RF, LGBM, XGB, and SVM) were applied to quantify driver importance; tree-based ensemble models achieved R² values of 0.943–0.946, while SVM recorded R² = 0.855. SHAP-based attribution identified distance from built-up areas as the dominant shrinkage driver (average contribution 19.63%; 32.82% in RF), followed by agricultural proximity, dumping ground proximity, hydrological connectivity, and land surface temperature. The ensemble vulnerability assessment identified 47.83% of the remaining wetland area as High or Very High risk, concentrated in the western and north-eastern sectors, while the south-eastern sector exhibited comparatively low vulnerability. The CA-ANN framework, validated at 84% overall spatial accuracy against an independently classified 2025 reference map, indicates continued wetland contraction to 24.9 km² by 2050 under a business-as-usual scenario. These outputs represent indicative spatial trajectories rather than deterministic forecasts. The findings provide spatially explicit, evidence-based insights into wetland vulnerability and shrinkage dynamics in the EKW, with potential relevance for spatial prioritisation in wetland planning and management. • Wetlands shrank by 63.4% between 1991 and 2023. • Wetland loss largely converted to agriculture and built-up land along urban corridors. • Ensemble ML models explained shrinkage with high accuracy (R² ≈ 0.94). • SHAP ranked proximity to built-up areas as the strongest driver of wetland loss. • Projections indicate wetland extent may decline to 24.9 km² by 2050.