STAH-Net: A Spatio-Temporal Attention Hybrid Network for Coastal Wetland Vegetation Classification Using Multi-Source Sentinel-1/2 Time Series

Accurate mapping of coastal wetland vegetation is essential for ecological conservation and blue carbon assessment, yet it remains challenging due to fragmented boundaries, spectral similarity among species, pronounced phenological variability, and limited generalization under small-sample conditions. We propose a spatio-temporal attention hybrid network (STAH-Net) for multi-source time-series classification. A 29-band feature cube is constructed from Sentinel-1 and Sentinel-2 data, integrating monthly Normalized Difference Vegetation Index (NDVI) and vertical–vertical (VV) polarization time series with static spectral and polarimetric indices. An improved U-Net backbone with dual-branch 3D convolutions extracts optical and Synthetic Aperture Radar (SAR) temporal features in parallel, incorporating a Temporal Enhancement Attention Module (TEAM) for phenological weighting, a Local Channel Attention (LCA) module for channel recalibration, and Atrous Spatial Pyramid Pooling (ASPP) for multi-scale context. An edge-weighted hybrid loss combining edge-weighted cross-entropy and Dice loss mitigates class imbalance and boundary blurring. Experiments in Hangzhou Bay, Liaohe Estuary, and Minjiang Estuary achieves overall accuracy (OA) values of 98.16%, 97.23%, and 95.76%, respectively, with a maximum Kappa of 97.49%, outperforming ConvLSTM, U-Net, SegFormer, SITSMamba, and GLMambaNet under identical settings. Ablation and visual analyses confirm the effectiveness of the proposed modules, with NDVI phenology providing primary discrimination and VV offering complementary structural information. Cross-year and cross-region tests show promising transferability, with OAs of 97.21% and 94.75% in Liaohe Estuary for 2024 and 2020, respectively, and 93.83% in Shanghai. These results demonstrate that STAH-Net provides an effective and transferable workflow for coastal wetland vegetation mapping.