In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C., Tournaire O. (Eds), IAPRS. Vol. XXXVIII. Part 3A - Saint-Mandé, France. September 1-3. 2010 
120 
A Markov Random Field Model for Individual Tree Detection from 
Airborne Laser Scanning Data 
Junjie Zhang, Gunho Sohn 
GeoICT lab, Earth and Space Science and Engineering Department, York University, 4700 Keele St., Toronto. ON M3.I 1P3, Canada 
- (jason84, gsohn)@yorku.ca 
Commission III, WG 111/2 
KEY WORDS: Airborne Laser Scanning, Tree detection, Markov Random Field, Automation. Segmentation 
ABSTRACT: 
Small-footprint Airborne Laser Scanning (ALS) holds great potential in forest inventory as it surpasses traditional remote sensing 
techniques in terms of rapid acquisition of 3D information of trees directly. With the increasing availability of high density ALS data, 
the derivation of more detailed individual tree information, such as tree position, tree height, crown size and tree species, becomes 
possible from ALS data exclusively. However, single tree detection is a critical procedure for tree-wise analysis in order to retrieve 
more accurate individual-tree-based parameters. The presented research highlights a novel Markov Random Field to model the 
configuration of single trees in ALS data in w'hich a global optimum to isolate individual trees can be achieved and addressed 
difficulties of individual tree detection problem in terms of problem representation and objective function. Firstly, local maxima are 
overpopulated from the CHM recovered from ALS data using a circular type of window filter with variable size. Then trees are 
modelled as objects at the centre of the extracted local maxima and attributed with other features retrieved from CHM image. The 
neighbourhood system is set up by TIN and energy functions are carefully designed to incorporate constraints for penalizing false 
trees and favour true ones. Finally, the optimal tree models are obtained through an energy minimization process. The method is 
applied on ALS data acquired from a coniferous forest and experimental results show a good detection rate. 
1. INTRODUCTION 
Small-footprint Airborne Laser Scanning (ALS), as an active 
remote sensing technology, allows for rapid acquisition of 
accurate 3D information of Earth topography and features in 
large scale. ALS gains popularity in forest survey quickly due 
to its unique capability of penetrating the tree canopy and 
providing relatively direct measurement of 3D structural 
information of trees, as well as the elevation of terrestrial 
surface under the canopy in forested area. This advantage 
makes it an alternative of tradition technology or even preferred 
one in the acquiring some forest parameters. 
Recent development of commercial small-footprint and Hill- 
waveform ALS system make the advantage stand even out. The 
practice of ALS in forestry study has extended from the 
formerly extraction of stand-based forest parameters, to the 
derivation of more detailed individual-tree-based information, 
such as tree position, tree height, crown size and tree species, 
with the increase of point density (Brandtberg, 2007; Hyyppa, 
et al., 2008). The potential of ALS data in forest inventory is 
still being exploited by various researches. 
However, in order to implement tree-wise analysis of forest, it 
is essential to detect individual trees from ALS data first. 
Extensive researches have been done to isolate single trees 
using ALS data and many of them used the methods extended 
from such procedures using aerial photos or satellite images 
(Chen, et ah, 2006). The outer geometry of trees can be directly 
recovered from ALS point clouds and the peaks and valleys on 
the recovered model can be better estimations of treetop 
positions and crown edges, than that obtained from photos or 
images. 
2. RELATED RESEARCH 
In this case, most methods focused on the reconstruction of the 
canopy height model (CHM) and the methods developed for 
optical imagery were transferred to detect trees in CHM. Most 
of those approaches were segmentation-based methods and fall 
in the scope of low-level vision techniques. To the best 
knowledge of the author, those methods include but not limited 
to: seed-based region growing (Solberg, et ah, 2006), valley 
following (Leckie, et ah, 2003), watershed segmentation 
(Pyysalo and Hyyppa, 2002) and its variance marker-controlled 
watershed (Chen, et ah, 2006). One main drawback of those 
methods is that empirical values are often used to set key 
parameters in the solution, which makes them scene sensitive. 
One example is that to increase the detection success rate of 
local maxima, which is helpful in seed-based region growing 
and marker-controlled watershed, in terms of false positives and 
negatives using variable window size, preliminary knowledge 
of tree height and crown size of the study area is needed first, 
whether it is from field survey or experience. Also, in the post 
processing of segmentation results, thresholds were also 
determined subjectively to remove or merge segments with 
certain size or according to other heuristic rules (Chen, et ah, 
2006). 
Recently, some applications utilizing the information contained 
in the 3D ALS point clouds to detect single trees were reported. 
Reitberger et ah (2009) hied to detect tree stems beneath the 
tree canopy using full-waveform ALS data, and combine them 
with local maxima detected on CHM as seeds to segment trees 
in 3D using a graph-cutting algorithm. One possible limit of this 
method is that stem detection may be influenced by ALS 
resolution, canopy closure, undergrowth vegetation and tree 
species, which determine the straightness of the stem. Raham 
and Gorte (2009) attempted to delineate tree crown based on 
densities of high points from high resolution ALS and got