Wetlands reduce disaster risks, regulate climate, support biodiversity and provide essential services to urban communities. Yet climate change and rapid urbanization have driven wetland loss and degradation since 1970. Existing global products often map broad land-cover categories, use incompatible resolutions or omit seasonal dynamics. They rarely follow the Ramsar classification system across an entire city, limiting their value for accreditation assessments. Pixel-based approaches can fragment small urban wetlands, while rule-based classifications struggle with similar-looking rivers, lakes, ponds and aquaculture areas. Based on these challenges, in-depth research is needed on standardized, fine-scale and temporally consistent wetland mapping for Ramsar Wetland Cities.
Led by Beijing Normal University, researchers from seven additional Chinese institutions published (DOI: 10.34133/remotesensing.1065) the study on July 7, 2026, in the Journal of Remote Sensing. The team developed the Object-Knowledge-based Hierarchical Optimization Cascade (OKHOC) method to address a practical monitoring gap: cities need whole-area maps that align with Ramsar definitions, distinguish seasonal and human-made wetlands, preserve fragmented urban features and remain consistent across accreditation cycles. The resulting product covers 43 Ramsar Wetland Cities (RWCs) in 17 countries and five benchmark years.
OKHOC combines object-based analysis, ecological knowledge and machine learning in a three-stage spectral, geometric and seasonal cascade. Unlike products built from land-cover taxonomies, it generates the Global Wetland City Fine Classification System at 10-meter resolution (GWC FCS10), separating 18 Ramsar wetland classes and six nonwetland classes. A Random Forest (RF) first establishes wetland boundaries. Spectral clustering and Extreme Gradient Boosting (XGBoost) resolve nonlinear differences among geometrically similar water bodies, while inundation rules distinguish permanent, seasonal and floodplain wetlands. A geometric-similarity inheritance mechanism reduces unrealistic year-to-year type switching. The product captured fragmented small water bodies and microscale marshes more effectively than seven comparison datasets.
The researchers processed 27,879 Sentinel-1 scenes and 16,104 cloud-filtered Sentinel-2 scenes for 2016, 2018, 2020, 2022 and 2024. They built 49 spectral, texture, polarization, topographic, geometric and seasonal features, then optimized city-specific feature sets and model parameters. Validation used 54,186 high-accuracy samples covering 18 wetland types across all 43 cities. The fine-scale product achieved a five-year mean overall accuracy of 94.69% and a Kappa coefficient of 0.925; the primary-stage classification reached 94.98% accuracy. By 2024, mapped wetlands totaled 3,063,592.66 hectares. Across the 43 cities, wetland area increased by a net 145,781.41 hectares from 2016 to 2024. Seventeen cities showed annual increases of at least 1%, while six declined by at least 1%. Coastal classes performed best, whereas seasonal rivers, seasonal lakes and floodplains remained more difficult because of changing water levels and vegetation. At the national level, accredited cities in 11 countries recorded net gains, whereas those in five countries experienced slight losses.
“The framework converts satellite observations into a consistent, standards-aligned picture of urban wetlands,” the researchers said. “It can help cities document restoration gains, identify losses and prepare evidence for accreditation. With improved data fusion and learning models, future versions could become more reliable in cloudy tropical regions and fragmented landscapes.”
The team combined Sentinel-1 radar and Sentinel-2 optical time series with elevation, bathymetry and existing wetland datasets. Processing was conducted through Google Earth Engine (GEE), ArcGIS Pro, Google Earth Pro and Collect Earth. Simple Non-Iterative Clustering created image objects, while Recursive Feature Elimination with Cross-Validation selected informative variables. Random Forest separated primary classes; spectral clustering and XGBoost subdivided similar water bodies. Intra-annual Water Inundation Frequency rules classified permanent, seasonal and floodplain wetlands, and an 80% geometric-similarity threshold guided interannual inheritance.
GWC FCS10 could support Ramsar accreditation and renewal, ecosystem-service assessment, restoration planning and wetland inventories. Its standardized classes may help organizations compare progress and direct conservation resources toward emerging losses. The team plans to extend coverage from 43 cities to all 74 Ramsar Wetland Cities and additions. Spatiotemporal data fusion could improve imagery in cloud-prone regions, while graph neural network–Transformer models may better capture small, seasonal and complex wetlands. These advances could make urban wetland governance more timely, transparent and comparable.
###
References
DOI
10.34133/remotesensing.1065
Original Source URL
https://spj.science.org/doi/10.34133/remotesensing.1065
Funding Information
National Key Research and Development Program of China, Grant Nos. 2023YFF0807204 and 2024YFF1306105; National Natural Science Foundation of China, Grant Nos. U21A2022 and 41571077; and the Open Fund of the State Key Laboratory of Remote Sensing Science and Digital Earth and Beijing Engineering Research Center for Global Land Remote Sensing Products, Grant Nos. OF202509 and OF202508.
About Journal of Remote Sensing
The Journal of Remote Sensing, an online-only Open Access journal published in association with AIR-CAS, promotes the theory, science, and technology of remote sensing, as well as interdisciplinary research within earth and information science.