Single Tree Detection from Airborne Laser Scanning Data: A Stochastic Approach
MetadataShow full item record
Characterizing and monitoring forests are of great scientific and managerial interests, such as understanding the global carbon circle, biodiversity conservation and management of natural resources. As an alternative or compliment to traditional remote sensing techniques, airborne laser scanning (ALS) has been placed in a very advantageous position in forest studies, for its unique ability to directly measure the distribution of vegetation materials in the vertical direction, as well as the terrain beneath the forest canopy. Serving as basis for tree-wise forest biophysical parameter and species information retrieval, single tree detection is a very motivating research topic in forest inventory. The objective of the study is to develop a method from the perspective of computer vision to detect single trees automatically from ALS data. For this purpose, this study explored different aspects of the problem. It starts from an improved pipeline for canopy height model (CHM) generation, which alleviates the distortion of tree crown shapes presented on CHMs resulted from conventional procedures due to the shadow effects of ALS data and produces pit-free CHM. The single tree detection method consists of a hybrid framework which integrates low-level image processing techniques, i.e. local maxima filtering (LM) and marker-controlled watershed segmentation (MCWS), into a high-level probabilistic model. In the proposed approach, tree crowns in the forest plot are modelled as a configuration of circular objects. The configuration containing the best possible set of detected tree objects is estimated by a global optimization solver in a probabilistic framework. The model features an accelerated optimization process compared with classical stochastic models, e.g. marked point processes. The parameter estimation is another issue: the study investigated both a reference-based supervised and an Expectation-Maximization (EM) based unsupervised method to estimate the parameters in the model. The model was tested in a temperate mature coniferous forest in Ontario, Canada, as well as simulated coniferous forest plots with various degrees of crown overlap. The experimental results showed the effectiveness of our proposed method, which was capable of reducing the commission errors produced by local maxima filtering based methods, thus increasing the overall detection accuracy by approximately 10% on all of the datasets.