Spatially adaptive water occurrence mapping from Sentinel-1 imagery using long-term backscatter statistics and behaviour-based zone stratification

Wetland surface-water mapping with synthetic aperture radar (SAR) remains challenging because low backscatter is not unique to open water: in deltaic wetlands, dry sandy substrates, dunes, smooth bare soils, and other low-return surfaces can produce water-like responses, while inundated vegetation may not exhibit the low-backscatter response typical of smooth open water. This study develops a spatially adaptive Sentinel-1 water-occurrence mapping framework for the Romanian Danube Delta using the full IW GRDH VV archive from 2015 to 2024. Two long-term per-pixel backscatter descriptors, the temporal median and tenth percentile (P10), were derived to characterize contrasting water-detection behaviour. Stable not-water and permanent-water reference pixels from the JRC Global Surface Water Occurrence product were used to compare threshold-selection criteria, with the Youden J statistic selected as the operational criterion. The resulting thresholds were −15.74 dB for the median descriptor and −19.33 dB for the P10 descriptor. A P10-based zoning mask separated water and not-water areas, after which the median-derived threshold was applied in water-prone zones and the P10-derived threshold in not-water zones. The occurrence map covered 3,588 km2 of valid area. Most of the area was never classified as water (2,323.5 km2; 64.8%), while 781.4 km2 (21.8%) showed episodic water detection between 1% and 10%. Permanent or near-permanent water detection covered 378.0 km2 (10.5%), including 271.3 km2 (7.6%) above 90% occurrence. Validation against JRC Monthly Water History produced pooled accuracy of 0.950, F1-score of 0.837, MCC of 0.811, and balanced accuracy of 0.930. Near-synchronous comparison with Dynamic World yielded accuracy of 0.940, F1-score of 0.794, MCC of 0.766, and balanced accuracy of 0.848. Zone-stratified and beach-ridge plains analyses showed that the adaptive strategy reduced false water detections over sandy low-backscatter areas while preserving high recall in water-prone zones.