Unlocking coastal wetland dynamics: a 40-year evolutionary analysis using HASX-Net and time series insights from Yancheng, China

Coastal wetlands face increasing pressures from climate change and human activities, requiring breakthrough technological approaches to obtain high-precision spatiotemporal dynamic information. This study proposes the Hybrid Attention Swin Xception Network (HASX-Net), applying it to the coastal wetland dataset of Yancheng, and bridging the research gap in deep learning for long-term and multi-temporal wetland extraction. To address the limitations of traditional wetland classification in capturing tidal dynamic changes, this study develops a novel time-series curve fitting method based on the Water and Vegetation Index. It systematically analyzes the spatiotemporal evolution patterns of coastal wetlands over the past 40 years, incorporating human activity-driven dynamics. The results demonstrate that (i) HASX-Net achieved outstanding performance in wetland extraction, with mean precision, recall, and F1 scores of 93.54%, 93.24%, and 0.9339, respectively. (ii) Over the past 40 years, the total wetland area declined sharply from 2586.91 km2 to 1883.10 km2, with salt marshes shrinking to just 269.19 km2. This study presents an innovative approach that integrates deep learning with time-series dynamic analysis, offering a robust solution for long-term monitoring of coastal wetlands. It establishes a new paradigm for high-precision, multi-temporal wetland extraction and provides actionable insights to advance global wetland conservation and sustainable management practices.