Classification of Wetlands in the Liaohe Estuary Based on MRMR-RF-CV Feature Preference of Multisource Remote Sensing Images

Coastal wetland ecosystems, shaped by the interaction between ocean and land, play a pivotal role in blue carbon sequestration. Understanding the distribution and evolutionary patterns of complex wetland types is essential for advancing wetland conservation, exploring blue carbon potentials, and supporting national carbon neutrality goals. This study utilized the Google Earth Engine cloud platform alongside Sentinel-1 synthetic aperture radar data, Sentinel-2 multispectral data, digital elevation model data, and unmanned aerial vehicle data to develop a feature preference algorithm integrating the minimum redundancy maximum relevance (MRMR) and random forest (RF) methods. Using eCognition, ArcMap, and ENVI platforms, we achieved an object-oriented classification of the Liaohekou Shuangtaizi wetlands. Results revealed that: 1) The MRMR-RF optimized 40 features, ranked by importance as Sentinel-2 spectral > Sentinel-1 index > Sentinel-1 radar > topographic > Sentinel-1 texture; 2) Six sets of comparison schemes were established, and the classification scheme based on the MRMR-RF model achieved the best classification performance, with an overall accuracy of 90.89% and a Kappa coefficient of 0.9; 3) The Liaohekou Estuary wetland was predominantly composed of Phragmites australis (P.australis), shallow sea, and tidal flat, followed by cropland, rivers, breeding pools, Suaeda salsa (S.salsa), reservoirs, puddles, bare soil, and building sites as secondary components; and 4) Between 2000 and 2023, the wetland area of different types in the study area changed significantly, with the changes mainly concentrated in the coastal aquaculture areas, tidal flat areas, and S.salsa growth areas.