A Change Detection-Guided Hierarchical Ensemble Framework for Long-Term 30-m Remote- Sensing-Based Ecosystem Type Mapping in the Yangtze River Delta, China
Abstract
Deltaic regions have undergone rapid ecosystem restructuring under urbanization, agricultural intensification, and hydrological regulation. Existing land-cover datasets, however, often lack the thematic resolution and temporal coherence needed to capture ecosystem type units and transitions in these regions. Here, we developed YRD_EMD, a remote-sensing-based ecosystem type dataset for the Yangtze River Delta (YRD), using a continuous change detection (CCD)-guided hierarchical ensemble framework. Specifically, CCD was used to identify stable and potentially changed areas, allowing unchanged samples to be transferred across years. To reduce confusion among spectrally similar ecosystem types and support temporally consistent mapping, hierarchical ensemble classification was applied to update class labels in potentially changed regions. Object-based refinement further delineated waterbody and wetland subclasses using spatial-context information. YRD_EMD maps eight major ecosystem types and 15 subclasses, supporting detailed characterization of landscape heterogeneity and ecological gradients. Validation using visually interpreted samples yielded an overall accuracy (OA) of 86.14% ± 1.45% and a Kappa coefficient of 0.84 ± 0.02 for the 15 subclasses. Independent third-party validation at the major-class level produced an OA of 91.21% ± 6.28% and a Kappa coefficient of 0.81 ± 0.05. Compared with existing products, YRD_EMD more coherently captured wetland–waterbody mosaics, vegetation restructuring areas, and cropland transition zones. Spatiotemporal analysis showed that cropland loss and urban expansion dominated ecosystem transitions over the past 39 years. Irrigated cropland decreased by approximately 38 657.15 km2, whereas urban land increased by approximately 36 836.66 km2. Overall, YRD_EMD provides a regionally adapted data foundation for ecosystem assessment and natural resource management.