Long-Term Coastal Wetland Mapping Using SCE-TS: A Spatiotemporal-Context-Enhanced Multi-Source Remote Sensing Time-Series Approach

The invasion and subsequent removal of Spartina alterniflora in recent years have induced pronounced changes in landscape patterns and ecological processes of China’s coastal wetlands. Long-term and continuous remote sensing monitoring of its dynamics is therefore of considerable scientific and management importance. However, long-term wetland monitoring based on multi-source remote sensing data commonly faces several challenges, including inconsistencies in temporal information caused by asynchronous image acquisition, limited and unstable training samples, and pronounced spatial noise resulting from highly fragmented wetland landscapes. These issues constrain the stability and reliability of long-term classification results. To address these challenges, a spatiotemporal-context-enhanced multi-source remote sensing time-series (SCE-TS) is proposed. First, phenological dynamics from different sensors are preserved through feature-level joint modeling, avoiding phenological distortions introduced by forced temporal alignment. Subsequently, representative and stable temporal prototypes are extracted through repeated feature selection and clustering of local temporal features, and combined with feature enhancement strategies to improve the representation of class-specific temporal characteristics under limited sample conditions. Furthermore, spatial neighborhood convolution is incorporated during feature construction to integrate temporal information from the central pixel and its surrounding neighbors, thereby mitigating the effects of mixed pixels and pixel-level temporal instability. Finally, an improved cascade forest model is employed for classification and mapping. The Yellow River Delta (YRD) and Yancheng wetlands (YC), characterized by distinct geographic settings and landscape structures, were selected as study areas. Using 602 Sentinel-1 and Sentinel-2 images, wetland classification map was generated for the period from 2016 to 2025. Experimental results show that the proposed method achieves overall accuracies of 95.02% in the YRD and 94.48% in the YC, while maintaining stable classification performance across multiple years. Long-term monitoring results further indicate that the removal of Spartina alterniflora has promoted the recovery of wetland vegetation structure; however, post-removal wetland ecosystems still exhibit complex dynamics in terms of wetland area and carbon storage. Overall, the proposed method provides a robust framework for evaluating invasive species management and supporting sustainable wetland ecosystem monitoring.