Floodplain wetlands are dynamic and vulnerable ecosystems, requiring accurate and detailed spatial information for sustainable management and conservation. Multi-source and multi-temporal remote sensing data have been commonly used to address spatiotemporal uncertainties, whereas the multivariate dynamic information embedded in these data remains underutilized. Therefore, this study developed three dynamic features derived from temporally dense Sentinel-1/2 imagery to improve floodplain wetland mapping on the Google Earth Engine (GEE) platform. First, hydrological dynamics were captured through the Inundation Frequency (IF) based on the Sentinel-1 Dual-Polarized Water Index (SDWI) time series. Second, vegetation dynamics and combined hydrological-vegetation variations were captured through Phenology Indicators (PIs) and a novel Inundation-Phenology Index (IPI), based on fitted NDVI and NDWI time series using all available Sentinel-2 observations. Finally, these features were introduced to classify a typical floodplain area in the middle Yangtze River, China, using the Random Forest algorithm with superpixels optimized. The results achieved an overall accuracy (OA) of 96.24% and a Kappa coefficient of 0.95, improving by 9.72 percentage points in OA and 0.12 in Kappa over the initial features (median composites of spectral/polarization bands and indices) alone. Comparative experiments, SHAP interpretation, and spatial consistency analysis revealed the following: (1) IF demonstrated notable and stable performance, significantly improving classification in frequently inundated areas; (2) PIs enhanced vegetation prediction, primarily attributed to EOS (end of growing season) and LOS (length of growing season); (3) IPI facilitated the differentiation between open water and transition zones, particularly benefiting complex and highly dynamic transitional areas. This study presents the considerable potential of proposed dynamic features, with results expected to support refined management and ecological conservation.
Xiong et al. (Tue,) studied this question.