Fusing radar and optical remote sensing for biomass prediction in mountainous tropical forests

Field measured estimates of aboveground biomass (AGB) in the mountainous region of Bwindi Impenetrable National Park (`Bwindi´), Uganda were used to train remote sensing models in order to estimate AGB within the park. AGB estimates were extrapolated using dual-polarization radar satellite data from ALOS PALSAR, optical imagery from Landsat 7 and a fusion of both, and compared to field estimates as indicators of the model prediction strength. Significant geolocation errors existed in the radar data due to the extreme terrain. Fusing the radar and optical data using the non-parametric algorithm Random Forest (RF) in R, provided lower error than using either radar or optical data alone (RMSE ~120 Mg ha^-1), however, saturation at higher biomass levels was evident. The AGB in Bwindi was estimated at 8.91 Tg ± 0.39 Tg (260.9 Mg ha^-1 ± 11.4 Mg ha^-1).