Aboveground Biomass Inversion of Mangrove in Dongzhai Harbor Based on Semi-Supervised Coregression

Mangroves are a vital component of coastal blue carbon ecosystems. The aboveground biomass (AGB) of mangroves must be accurately estimated for meeting carbon neutrality objectives. However, traditional field biomass measures face many constraints in estimating the AGB of mangroves because of the special ecological status of mangroves as nationally protected plants, challenging tidal environments, and high costs associated with field sampling. As such, the limited number of field-measurement sample plots available for mangrove AGB inversion imposes serious challenges, which we addressed by developing a novel semi-supervised coregression (COREG) approach. We developed and validated the semi-supervised COREG regression method based on Sentinel-2 optical, GEDI vegetation canopy height, and field survey data from 120 standard plots (10m×10m) established in Dongzhai Harbor Mangrove National Nature Reserve, Hainan, China. Firstly, in the study area the method evaluated six classical AGB regression models, from which the two base models—the random forest and XGBoost algorithms— with the highest accuracy were selected. Subsequently, collaborative feature importance assessment was used to select seven key feature variables (MTVI2, MSRren, NDVIre, CIre, MDI2, GEDI canopy height, TNDVI), and then dual-regressor collaborative training was applied. Then the method successfully expanded the original 96 training samples to 413 samples via incorporating 317 high-confidence pseudo-labeled samples from 1,000 unlabeled data points through confidence-based pseudo-label generation with a threshold of 0.9. By leveraging a limited number of real samples, the selected effective features, and the expanded pseudo-samples, our method enables more effective AGB inversion. Based on the test sample set, the results of the developed model were more accurate than those of traditional supervised methods, with semi-supervised COREG achieving a coefficient of determination (R2) of 0.85, significantly outperforming supervised random forest regression (R2=0.77) and XGBoost regression (R2 = 0.74). The method ranked first on all accuracy metrics, with a mean absolute error (MAE) of 3.04 t/hm2, mean relative error (MRE) of 2.75%, and root mean square error (RMSE) of 4.705 t/hm2. The total AGB of the Dongzhai Harbor mangroves was 201,389 tons with an average density of 112.86 t/hm2, ranging from 68.93 to 132.86 t/hm2, in 2022. Spatial analysis revealed that the AGB of mangroves was higher in southern than in northern regions, which was consistent with field observations showing predominantly arboreal species (Sonneratia apetala and Kandelia obovata) in the south and shrub species (Rhizophora stylosa and Aegiceras corniculatum) in the north. The minimum sample size analysis indicated that the method maintained acceptable accuracy (R2> 0.75) with as few as 70 training samples, representing an approximately 27% reduction in the field sampling requirement compared with that of traditional supervised approaches. However, the performance of the method is sensitive to the spatial distribution of field samples, particularly struggling with high-density regions (≥125 t/hm2) when training samples were insufficient in these areas. This study provides a practical solution for large-scale mangrove biomass monitoring in data-scarce scenarios, and can be used for guiding blue carbon ecosystem assessment and coastal wetland management to meet carbon-neutrality initiatives.