Abstract
Mangrove forests in southern China’s Gaoqiao Mangrove National Nature Reserve (Guangdong–Guangxi border) have undergone significant decline followed by partial recovery, driven by human activities and conservation efforts. Traditional monitoring methods struggle to capture their complex spatiotemporal dynamics. This study develops a practical two-stage deep learning framework: an enhanced U-Net with Squeeze-and-Excitation (SE) and Convolutional Block Attention Module (CBAM) first extracts high-quality annual mangrove masks from multi-temporal Landsat imagery (1993–2023), achieving IoU = 0.815 and F1-score = 0.928. These masks are then used for spatiotemporal forecasting, with U-Net–ConvLSTM recommended as the primary architecture due to its excellent balance of accuracy, simplicity, and computational efficiency. An optional asymmetric Ecological Constraint Loss (ECOLOSS) can be added to form the ConvLSTM+ECOLOSS variant, providing marginal additional accuracy (IoU = 0.793 vs. 0.787, MAE = 6.70% vs. 6.83%) on the test period (2019–2023) by acting mainly as an ecological safeguard against unrealistic long-term runaway trends. Forecasts for 2024–2026 indicate continued slow recovery under current management. The U-Net–ConvLSTM pipeline offers a transparent and efficient tool for operational mangrove monitoring and conservation planning in subtropical China.