Abstract Above-Ground Biomass (AGB) assessment using remote sensing has been an active area of research since the 1970s. However, improvements in the reported accuracy of wide scale studies remain relatively small. Therefore, there is a need to improve error analysis to answer the question: Why is AGB assessment accuracy still under doubt? This project aimed to develop and implement a systematic quantitative methodology to analyse the uncertainty of remotely sensed AGB, including all perceptible error types and reducing the associated costs and computational effort required in comparison to conventional methods. An accuracy prediction tool was designed based on previous study inputs and their outcome accuracy. The methodology used included training a neural network tool to emulate human decision making for the optimal trade-off between cost and accuracy for forest biomass surveys. The training samples were based on outputs from a number of previous biomass surveys, including 64 optical data based studies, 62 Lidar data based studies, 100 Radar data based studies, and 50 combined data studies. The tool showed promising convergent results of medium production ability. However, it might take many years until enough studies will be published to provide sufficient samples for accurate predictions. To provide field data for the next steps, 38 plots within six sites were scanned with a Leica ScanStation P20 terrestrial laser scanner. The Terrestrial Laser Scanning (TLS) data analysis used existing techniques such as 3D voxels and applied allometric equations, alongside exploring new features such as non-plane voxel layers, parent-child relationships between layers and skeletonising tree branches to speed up the overall processing time. The results were two maps for each plot, a tree trunk map and branch map. An error analysis tool was designed to work on three stages. Stage 1 uses a Taylor method to propagate errors from remote sensing data for the products that were used as direct inputs to the biomass assessment process. Stage 2 applies a Monte Carlo method to propagate errors from the direct remote sensing and field inputs to the mathematical model. Stage 3 includes generating an error estimation model that is trained based on the error behaviour of the training samples. The tool was applied to four biomass assessment scenarios, and the results show that the relative error of AGB represented by the RMSE of the model fitting was high (20-35% of the AGB) in spite of the relatively high correlation coefficients. About 65% of the RMSE is due to the remote sensing and field data errors, with the remaining 35% due to the ill-defined relationship between the remote sensing data and AGB. The error component that has the largest influence was the remote sensing error (50-60% of the propagated error), with both the spatial and spectral error components having a clear influence on the total error. The influence of field data errors was close to the remote sensing data errors (40-50% of the propagated error) and its spatial and non-spatial Overall, the study successfully traced the errors and applied certainty-scenarios using the software tool designed for this purpose. The applied novel approach allowed for a relatively fast solution when mapping errors outside the fieldwork areas.

Document Type


Publication Date