A Validation-Controlled Label-Efficient Framework for Coastal Wetland Habitat Mapping Using Multi-Season Sentinel-1 and Sentinel-2 Data

Reliable coastal wetland habitat mapping is often constrained by the scarcity and the cost of reliable reference data, especially in data-limited coastal environments. We propose a validation-controlled, label-efficient framework pairing multi-season Sentinel-1 and Sentinel-2 predictors with a CatBoost teacher and a lightweight MLP student. A candidate is pseudo-labeled only when both separately calibrated models agree and exceed class-specific thresholds; accepted labels are class-balanced and down-weighted. The framework was evaluated at the Sidi Moussa–Oualidia wetland complex and Merja Zerga lagoon in Morocco. At Sidi Moussa–Oualidia, 62 configurations were compared through nested polygon-grouped validation and then frozen before a five-seed held-out evaluation. The supervised MLP and Agreement-augmented MLP achieved mean Macro-F1 values of 0.9518±0.0044 and 0.9509±0.0062, indicating that augmentation did not materially change the already strong full-data baseline. Under a stricter budget of 30 training and 20 validation observations per class, Agreement yielded a mean Macro-F1 of 0.9092±0.0102 compared with 0.9023±0.0093 for the supervised baseline and produced pseudo-labels in all five seeds. A spatial-range sensitivity analysis further showed that both models retained Macro-F1 values of 0.9391 and 0.9403 for test observations located beyond the largest estimated within-class autocorrelation range. At Merja Zerga, the native six-class supervised MLP achieved 0.9456±0.0050, compared with 0.9401±0.0047 after Agreement augmentation. Spatially blocked four-class experiments nevertheless showed that 20 to 30 local training labels per class recovered approximately 96–98% of the corresponding full-data performance. The framework therefore supplies an operational criterion for using unlabeled observations: augmentation is adopted only where calibrated filtering yields adequate class coverage, and validation confirms a downstream effect; otherwise the supervised model is retained. For the strict Sidi Moussa–Oualidia reduced-label experiment, the reported development budgets count every site-specific label used for fitting, early stopping, and calibration. The Merja Zerga blocked experiments separately quantify training-label sensitivity while retaining their blocked validation resources.