Less is more: two UAV-LiDAR metrics outperform complex machine learning algorithms for aboveground carbon estimation in hurricane-disturbed mangroves

Mangrove forests sequester disproportionate amounts of carbon relative to their global extent, yet their structural monitoring at fine spatial scales remains logistically challenging, particularly following disturbance events. Whilst UAV-LiDAR provides centimetre-scale canopy data, its capacity to estimate aboveground carbon (AGC) in hurricane-disturbed mangroves – where vertical structure is partially decoupled from biomass – remains insufficiently understood. In this study, we developed and validated a UAV-LiDAR processing pipeline to estimate AGC across 26 field plots in the mangroves of Marismas Nacionales, Nayarit, Mexico, following Category 3 Hurricane Roslyn (October 2022). Using a DJI Zenmuse L1 sensor across four seasonal campaigns (2023–2024), we extracted 20 structural metrics and evaluated six modelling approaches, ranging from parsimonious linear regression to machine learning algorithms. Contrary to the prevailing trend towards algorithmic complexity, a parsimonious linear model using only the 99th percentile (hp99) and the Rumple Index (RI) outperformed complex architecture (R2 = 0.54, RMSE = 11.41 Mg C ha−1). Plot-level cross-validation confirmed that this performance was not inflated by repeated seasonal measurements (R2 = 0.52, RMSE = 11.67 Mg C ha−1). Model performance decreased with increasing complexity under limited sample size conditions (n = 48). Notably, even competitive configurations such as deep neural networks(R2 = 0.49) failed to surpass the linear model, whilst gradient-based methods showed poor generalization. These results indicate that canopy height and structural rugosity provide sufficient information to estimate AGC in disturbed mangroves, and that increasing model complexity does not improve predictive performance under structural decoupling and small-data constraints. Parsimony, rather than algorithmic sophistication, emerges as the operative principle for UAV-LiDAR carbon monitoring in disturbance-driven ecosystems, with model reliability highest in the most severely damaged stands – precisely where accurate carbon accounting is most critical for blue carbon MRV applications.