Enhanced aboveground biomass density estimation in Central Vietnamese forests

Accurate estimation of spatially explicit forest aboveground biomass density (AGBD) is essential for supporting climate change mitigation strategies. Recent studies have demonstrated the predictive effectiveness of the random forest (RF) algorithm in forest AGBD estimation utilizing multi-source remote sensing (RS) data. However, the RF-based estimates may be further enhanced by integrating RF with kriging techniques that account for spatial autocorrelation in model residuals. Therefore, we investigated the performance of random forest ordinary kriging (RFOK) and random forest co-kriging (RFCK) for estimating AGBD in Central Vietnamese forests using Advanced Land Observing Satellite-2 Phased Array L-band Synthetic Aperture Radar-2 (ALOS-2 PALSAR-2), Sentinel-1 (S1), and Sentinel-2 (S2) imageries. 277 predictors, including spectral bands, radar backscatter coefficients, vegetation indices, biophysical variables, and texture metrics, were derived from these RS datasets and statistically linked to field measurements from 104 geo-referenced forest inventory plots. The results showed that textures, modified chlorophyll absorption ratio index (MCARI), and radar backscatters were key contributors to AGBD variability. The fusion of ALOS-2 PALSAR-2 and S2 data yielded the highest RF performance, with coefficient of determination (R2), root mean square error (RMSE), and mean absolute error (MAE) achieving 0.75, 39.15 t.ha-1, and 32.20 t.ha-1, respectively. Incorporating interpolated residuals by ordinary kriging and co-kriging into RF predictions enhanced estimation accuracy, with relative improvements of 5.74–7.04 % in R2, 8.73–10.91 % in RMSE, and 13.62–15.27 % in MAE, yet these gains remained limited. Although RFOK achieved marginally better accuracy (R2 = 0.80, RMSE = 34.88 t.ha-1, MAE = 27.28 t.ha-1) compared to RFCK (R2 = 0.79, RMSE = 35.73 t.ha-1, MAE = 27.81 t.ha-1), the latter reduced estimation bias more effectively, likely due to the inclusion of elevation as a covariate in the co-kriging process. These findings underscore the potential of the hybrid RF-kriging frameworks for improving spatial AGBD estimation, offering a robust approach for carbon accounting in tropical ecosystems.