Mapping wetland vegetation associations in the Yellow River Delta: a phenology-based segmentation approach using dense Sentinel-2 time series

Accurate spatial distribution information of wetland vegetation serves as the foundation for assessing the health and diversity of wetland ecosystems and is essential for monitoring environmental changes and predicting future ecological trends. Focusing on the complex wetlands of the Yellow River Delta, this study proposes a novel, robust classification framework to achieve the first-ever association-level vegetation map for the region. Our methodology leverages dense Sentinel-2 NDVI time series data (December 2022 to January 2024). The data were rigorously pre-processed, including quantitative optimization of the Savitzky-Golay (S-G) filter parameters to ensure high fidelity of the phenological trajectories. Employing the phenology-based unsupervised segmentation approach, we completed the first wetland vegetation mapping at the association level in the Yellow River Delta region. We developed a new framework based on the principle of temporal similarity among vegetation covers of the same type, segmenting NDVI dense time series data using the X-means algorithm. By correlating survey data with patch categories and actual vegetation association types, we ultimately producing maps for 9 wetland vegetation association types. Within this framework, we addressed the separability of vegetation association structure in the study area. Results indicate that unsupervised segmentation algorithms effectively exploit latent information in NDVI dense time series data, achieving an overall classification accuracy of 85.00%, Kappa of 82.4%, and Macro-F1 of 85.22%, all higher than those of the random forest classifier. The resulting association-level map accurately reveals the spatial distribution characteristics and relationships among different vegetation communities in the Yellow River Delta, providing critical data support for the refined ecological management and effective conservation of this vital coastal wetland. • Leveraging temporal signatures rather than spectral features alone in wetland vegetation classification. • Novel application of unsupervised learning for phenological pattern recognition. • First association-level vegetation classification map developed for the Yellow River Delta region.