Aboveground biomass estimation of individual trees with airborne Lidar data

Abstract

This thesis presents a framework of the aboveground biomass (AGB) estimation for individual trees from airborne lidar data. To reduce the impact of topography on object points, a voxel-based multiscale morphological airborne lidar filtering algorithm is proposed to distinguish ground points from object points and normalise the height values of identified object points. Because of the presence of pits demonstrating abnormally lower elevation values than the surroundings in the canopy height model (CHM), a multiscale morphological algorithm is proposed to rectify the pits and improve the accuracy of the CHM. Once the pits in the CHM are rectified, a hybrid method is proposed to segment lidar points into individual trees. A modified CHM-based tree segmentation algorithm is used to identify highly accurate tree tops, which are served as seeds in a point-based profile analysis to segment individual trees meanwhile recognising understory trees. Due to the lack of points representing breast height of individual trees, diameter at breast height (DBH) cannot be extracted from segmented trees directly. Hence, a DBH regression model is generated based on extracted tree height and crown width. The principle component analysis and the ridge regression are conducted to investigate if multicollinearity exists in the input variables for AGB regression model generation and remove the irrelevant variables. Three existing generalised AGB allometric models are exploited to compute the AGB estimates as the reference, respectively since the field samples of the AGB estimates are not available. Once the input variables are set, the AGB regression models are generated by various machine learning techniques to examine which technique has the best performance in regression model generation, including random forest, support vector regression, multilayer perceptron and radial basis function. The qualities of the various AGB regression models are assessed by model efficiency index, adjusted coefficient of determination, leave-one-out cross validation, Akaike information criterion, and normalised-mean-square-errors of the AGB estimates. According to the results, the best AGB regression model can explain up to 99% of the variation of the AGB estimates. In conclusion, random forest yields the most accurate AGB regression models and is most robust in regression model generation.