Remote sensing and artificial intelligence integration for seabed seagrass distribution and probabilistic seagrass biomass carbon potential

The Seto Inland Sea (SIS), Japan's largest semi-enclosed coastal sea, hosts seagrass meadows that form an important and quantifiable reservoir of biomass-based carbon. Reliable mapping of these meadows is therefore essential for ecosystem assessment and coastal management. This study integrates a multiband depth-invariant bottom index (BI) derived from Sentinel-2 imagery with machine learning and deep learning classifiers to map seabed seagrass distribution and estimate the probabilistic seagrass biomass carbon potential of the inner SIS between 2019 and 2025. Sentinel-2 imagery was processed in Google Earth Engine, and Bands 2 (blue), 3 (green), and 4 (red) were combined into three BI configurations ( 𝐵 ⁢ 𝐼 2 − 3 , 𝐵 ⁢ 𝐼 3 − 4 , and 𝐵 ⁢ 𝐼 2 − 4 ). Of nine classifiers tested, Random Forest (RF) performed best, reaching 92.31% accuracy and 𝜅 = 0.81 under a single stratified split. More conservative validation across 50 repeated random splits gave 87.4 ± 9.5% accuracy and 𝜅 = 0.686 ± 0.225, while spatial leave-one-site-out (LOSO) validation gave 77.4% accuracy and 𝜅 = 0.50, indicating reduced transferability to unsurveyed sites. Among the three band-pair configurations, 𝐵 ⁢ 𝐼 2 − 3 was the dominant predictor (variable importance = 0.93). Building on this classification, seagrass biomass carbon potential was estimated by combining satellite-derived seagrass probabilities with field-based biomass distributions through Monte Carlo uncertainty propagation. The resulting estimate reflects standing seagrass biomass carbon only, excluding sediment organic carbon, burial, decomposition, and lateral carbon export, and should be read as a biomass carbon potential rather than a total blue carbon stock. Monthly carbon potential followed a clear seasonal cycle, rising from winter into spring and peaking in May–June. Together, these results show that combining a multiband BI with artificial intelligence (AI)-based classification enables site-specific seagrass mapping and biomass-based carbon assessment in optically complex coastal waters, while underscoring the need for spatially independent validation before extrapolating beyond surveyed sites.