UAV-Mapping of Aboveground Biomass in Arid Mangrove Forests: A Crown-to-Grid Machine Learning Approach

Mangrove forests are among the most carbon-rich coastal ecosystems, yet their aboveground biomass (AGB) remains poorly quantified in arid regions where structural complexity, closed canopies, and logistical constraints limit conventional field surveys. Improving AGB estimation in these understudied ecosystems is essential for advancing blue-carbon inventories, understanding ecological functioning under extreme environmental conditions, and supporting conservation and restoration initiatives. To address this gap, we present a UAV-based framework designed to generate high-resolution, non-destructive AGB estimates for Avicennia marina mangroves along the Saudi Arabian Red Sea coast, where data on AGB and carbon stocks remain scarce. The proposed approach implements a crown-to-grid framework that simulates quadrat-based AGB sampling at the site-scale using UAV-LiDAR and multispectral data. Field-measured trees are used exclusively to provide reference AGB values derived from an existing allometric relationship for Middle Eastern Avicennia marina. For model training, the crowns of these reference trees are manually delineated and partitioned into 1 m × 1 m grid cells; to augment the training dataset and reduce sensitivity to grid placement, each crown is sampled using 10 shifted grid configurations generated by systematically offsetting the grid origin. Tree-level AGB is then distributed across the cells using the canopy height model as a structural weighting function, generating a physically consistent, cell-level AGB reference while conserving total tree biomass. Spectral, structural, and index-based features extracted at the cell-level are used to train a Random Forest regression model. Model performance is evaluated using leave-one-tree-out cross-validation by aggregating predicted cell-level AGB back to the tree-scale and comparing it against field-derived AGB reference values. Once trained, the model is applied to a continuous 1 m × 1 m grid across the entire UAV-covered area, enabling spatially explicit AGB mapping without requiring individual-tree delineation. In addition to the methodological contributions, our results provide quantitative insights into AGB distribution in arid mangrove ecosystems. Mean site-level AGB densities ranged from ~25 to 31 Mg ha⁻¹, with localized hotspots associated with denser or taller vegetation. By resolving sub-canopy variability and integrating structural and spectral information, the framework improves our ability to characterize vegetation patterns that influence ecosystem function, productivity, and resilience, which are key components of blue-carbon dynamics in extreme environments. Finally, the approach establishes a pathway for upscaling UAV-derived AGB estimates to broader coastal regions, offering a critical bridge between field observations, high-resolution remote sensing, and satellite-based AGB products. Such scalable, non-destructive methods are essential for developing robust blue-carbon inventories, improving carbon accounting in regions where destructive sampling is limited, and supporting management and restoration strategies under accelerating climate and anthropogenic pressures.