Capabilities and limitations of Landsat and land cover data for aboveground woody biomass estimation of Uganda

Aboveground woody biomass for circa-2000 is mapped at national scale in Uganda at 30-m spatial resolution on the basis of Landsat ETM + images, a National land cover dataset and field data using an object-oriented approach. A regression tree-based model (Random Forest) produces good results (cross-validated R² 0.81, RMSE 13 T/ha) when trained with a sufficient number of field plots representative of the vegetation variability at national scale. The Random Forest model captures non-linear relationships between satellite data and biomass density, and is able to use categorical data (land cover) in the regression to improve the results. Biomass estimates were strongly correlated (r = 0.90 and r = 0.83) with independent LiDAR measurements. In this study, we demonstrate that in certain contexts Landsat data provide the capability to spatialize field biomass measurements and produce accurate and detailed estimates of biomass distribution at national scale. We also investigate limitations of this approach, which tend to provide conservative biomass estimates. Specific limitations are mainly related to saturation of the optical signal at high biomass density and cloud cover, which hinders the compilation of a radiometrically consistent multi-temporal dataset. As a result, a Landsat mosaic created for Uganda with images acquired in the dry season during 1999–2003 does not contain phenological information useful for discriminating some vegetation types, such as deciduous formations. The addition of land cover data increases the model performance because it provides information on vegetation phenology. We note that Landsat data present higher spatial and thematic resolution compared to land cover and allow detailed and spatially continuous biomass estimates to be mapped. Fusion of satellite and ancillary data may improve biomass predictions but, to avoid error propagation, accurate, detailed and up-to-date land cover or other ancillary data are necessary.