Improving interpretability of mangrove aboveground biomass modeling via ecological knowledge: machine learning based hierarchical models

Aboveground biomass (AGB) is a key indicator of mangrove ecosystem function, especially in productivity and blue carbon storage. The partial coverage Light Detection and Ranging (LiDAR) data are widely used as a bridge to connect field plots with wall-to-wall satellite imagery for AGB estimation. However, the prediction models lack robustness and interpretability. Inspired by the principle of biomass allometric equations and LiDAR sampling, we propose a hierarchical model with spectral unmixing embedded for AGB estimation, using partial coverage unmanned aerial vehicle (UAV) LiDAR data, satellite imagery (Landsat 8, Sentinel-2, Sentinel-1), and field plots. In this hierarchical model, LiDAR-extracted canopy cover (CC) and canopy height (CH) indicating the volume of trees are first modeled and then extrapolated to the entire study area using satellite imagery. In the second stage, field-extracted AGB is modeled by combining the predicted CH and CC with satellite-imagery-based features. The results show that the proposed method effectively estimates mangrove AGB on Hainan Island with fewer predictors, while improving interpretability and reducing LiDAR processing effort. The optimal hierarchical model integrating UAV-LiDAR and Landsat 8 data outperformed the traditional upscaling method (R² = 0.686 > 0.620, lower RMSE%), providing an efficient and interpretable method for mapping large-scale mangrove AGB.