A machine learning framework for modeling and upscaling mangrove carbon productivity (ML-MCP)

Global warming has intensified the pursuit of carbon neutrality, drawing increased attention to mangroves for their critical role in carbon sequestration and CO₂ uptake. Various models have been developed to estimate the Gross Primary Productivity (GPP) of mangrove ecosystems using satellite data; however, uncertainties persist in daily carbon uptake predictions. This study addresses this gap by introducing the Machine Learning-based Mangroves Carbon Production (ML-MCP) model, which integrates process-driven and remote sensing-based approaches. Among the tested machine learning models, XGBoost performed the best, achieving an R² of 0.67 across four sites and 0.78 at the Fujian Zhangjiangkou Mangrove Nature Reserve. However, the complexity of mangrove ecosystems—shaped by environmental gradients and stressors—poses challenges for accurate model predictions. SHapley Additive exPlanations (SHAP) analysis identified the light-limited photosynthetic rate (PI) as the most critical factor influencing GPP, while the soil salinity index (SI) also played a significant role when data was aggregated. Additionally, temperature index (TI) emerged as the second most important factor affecting GPP at the SKR-US and ZJ-CN sites. These findings highlight the necessity of adopting context-specific approaches for effective ecosystem management. While machine learning models show great potential for predicting GPP, the unique challenges of mangrove ecosystems call for enhanced data collection, specialized modeling techniques, and a deeper understanding of mangrove-specific processes to improve model performance in these vital coastal environments.