A new method for mangrove biomass prediction based on automatic machine learning and multi-source UAV data fusion

Rapid and accurate assessment of mangrove aboveground biomass (AGB) is essential for coastal blue carbon management. However, most existing non-parametric machine learning approaches rely on spectral information and LiDAR point cloud data combined with traditional algorithms, and these approaches often require substantial effort, domain expertise, and repeated manual trial-and-error tuning. To address this issue, this study developed a new method for predicting mangrove AGB using variables derived from UAV LiDAR and multispectral imagery within two automated machine learning (AutoML) frameworks, FLAML and H2O. The results demonstrate the feasibility of applying AutoML to mangrove AGB inversion. Using the coefficient of determination (R2) and root mean square error (RMSE) as evaluation metrics, the XGBoost algorithm achieved R2 values of 0.94 for the training set and 0.89(RMSE1.92) for the test set under the FLAML framework, and 0.79 and 0.76(RMSE1.96), respectively, under the H2O framework. SHAP-based feature-importance analysis showed that percentile-based LiDAR height variables were the most influential predictors. Among spectral variables, the green band, the vegetation factor (Veg), and the Kawashima index exhibited the highest importance. This method provides rapid and accurate estimation of mangrove AGB and enables visual interpretation of contributions from spectral and LiDAR-derived variables, offering data support for local mangrove ecosystem assessment and management.

Actions (login required)

View Item