Abstract
Mangrove carbon assessments commonly emphasize woody biomass and sediments, whereas species-level variation in leaf carbon traits remains insufficiently quantified at landscape scales. This study has developed a field-calibrated Earth observation–machine-learning (EO-ML) framework to quantify and map area-based leaf organic carbon (LOCa) across the western Sundarbans, India. Field sampling at 100 georeferenced locations during 2024 represented seven dominant mangrove species. Carbon mass fraction was measured by dry combustion, leaf mass per area was derived from dry mass and projected one-sided leaf area, and LOCa. Landsat-8 spectral indices, canopy and species layers, terrain attributes, and hydroclimatic covariates were integrated through LASSO-based feature selection and comparative ML modeling. Nested five-fold cross-validation identified random forest (RF) and gradient boosting regression (GBR) as the strongest models (R2 = 0.87 and 0.86, respectively; normalized RMSE = 0.03), substantially outperforming support vector regression (SVR) (R2 = 0.45). Species-level high-class LOCa ranged from 58.13 g C m-2 in Acanthus ilicifolius to 102.02 g C m-2 in Bruguiera gymnorhiza; Avicennia officinalis and Aegialitis rotundifolia also exhibited comparatively high values. Spatial predictions revealed species-specific hotspots and marked within-species heterogeneity, while bivariate analysis identified discrete sectors where elevated LOCa coincided with high mangrove richness. Secondary standing-leaf carbon and stoichiometric CO2-equivalent products were retained as static, assumption-bound sensitivity representations rather than annual sequestration, total ecosystem carbon, or verified offsets. Research establishes a robust trait-to-landscape approach for species-resolved assessment of mangrove leaf carbon density, providing an important regional baseline for ecological monitoring and future investigations of mangrove carbon dynamics under changing environmental conditions.
[Display omitted]
•Field-calibrated EO–ML framework maps species-level mangrove leaf organic carbon.•Random Forest achieved robust LOCaprediction with R2= 0.87 across the study area.•Species-resolved maps reveal pronounced spatial heterogeneity in leaf carbon density.•Bivariate hotspots identify priority zones of high LOCaand species richness.•Trait-to-landscape framework supports long-term mangrove ecological monitoring.