Physics-Constrained Deep Learning for Mangrove Distribution and Wind Modeling

This study introduces a physics-informed neural network (PINN) framework that integrates remote sensing with physical principles for multi-scale coastal ecosystem monitoring. A hybrid PINN-CNN-LSTM processes Sentinel-1/2 imagery and ERA5 data while enforcing conservation laws to predict mangrove distribution and coastal wind dynamics. Applied to East Luwu, South Sulawesi, Indonesia, the model attains 93.6% accuracy for mangrove classification (RMSE = 0.14731) and strong wind prediction performance (R2 = 0.91, RMSE = 0.048 m/s). Physical validation confirms drag coefficients (Cd = 0.8–1.4) and 97.9% energy closure, consistent with canopy flow theory. Cross-site validation across three mangrove systems shows transferability with <15% RMSE increase. The framework supports regional-to-global monitoring, offering mechanistic insights into carbon sequestration with an estimated 198,535 tons C/month capacity under physically constrained uncertainty. These results demonstrate the framework’s potential for credible blue carbon assessment and ESG-aligned coastal management.