Machine Learning-Enabled Multi-Source Remote Sensing Assessment of Wetland Dynamics

Wetlands provide essential hydrological, ecological, biodiversity, and climate-regulation functions, yet their strong seasonal variability and spectral complexity make consistent national-scale monitoring particularly challenging in environmentally heterogeneous and data-sparse regions. This study developed an integrated multi-source remote-sensing framework to characterize wetland distribution and change across Pakistan for the benchmark years 2018, 2020, 2022, and 2024. Sentinel-2 surface-reflectance observations and water-, vegetation-, and moisture-sensitive spectral indices were integrated with topographic and probabilistic land-cover information to discriminate seven classes: non-wetland, tidal flat, mangrove, salt marsh, swamp, marsh, and flooded vegetation. The primary seven-class product was evaluated using an independent stratified reference dataset of 750 observations interpreted without reference to Dynamic World using very-high-resolution imagery and Sentinel-1/-2 temporal information. Independent validation yielded 68.9% unweighted overall accuracy and 84.6% area-weighted overall accuracy, with mangrove showing the highest user’s accuracy (82.0%) and marsh the lowest (52.0%). Spatially blocked cross-validation produced a more conservative national overall accuracy of 68.2% (κ = 0.62), compared with 85.4% under random pixel partitioning. Predictor ablation showed that removing Dynamic World probabilities reduced spatially validated accuracy from 68.2% to 64.9%, indicating a useful but non-dominant contribution. Error-adjusted change analysis reduced the dominant 2018–2024 non-wetland-to-marsh transition from an unadjusted 107,340 km2 to 48,200 km2 (95% CI: 36,800–59,600 km2), while the bias-adjusted national net wetland balance was estimated at +57,300 km2 (95% CI: 41,800–72,800 km2). These estimates represent uncertainty-adjusted differences between benchmark-year classifications and should not be interpreted as direct evidence of permanent ecological wetland formation. The relatively low marsh classification accuracy and the limitations of annual median composites require cautious interpretation of the dominant transition pathway. Direct validation of the 2018–2024 change product yielded 61.4% accuracy, emphasizing that temporal transitions are more uncertain than single-date classifications. The results demonstrate the value of independent validation, spatial generalization testing, and uncertainty adjustment for national-scale wetland monitoring.