Estimation of Tree Species Proportions of Forest Stands Using Laser Scanning

Tree species proportions of forest stands were estimated using data measured by Toposys laser scanner. Employed estimation method was multilayer perceptron neural network with error back-propagation training algorithm. Measurements were divided into 2*2 m squares and then those measurement which were higher than determined DTM were used to compute features. The purpose of extracted features were to determine and characterize the vertical distribution of measurements within squares. Four different featuresets were computed in order to test and compare different strategies to extract features. The estimation errors for those stands selected for training were really small but the estimation errors for other stands were considerably worse. When stands were classified according to dominant tree species, classification error for stands selected for training were nonexistent or small, but the classification errors for other stands were larger. Bad results may be due to small number of stand (79), but it also indicates that only 3D-coordinates are not enough to estimate stand tree species proportions. The alternatives to get better results are either classify single trees and compute tree species proportions from these or use also intensity information by using aerial images or laser which can measure intensity.