In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C... Tournaire O. (Eds), 1APRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France. September 1-3. 2010
125
-30:- ~r~ T' - - “T—' -7 ~
-40 -
g; -50
s
fi - 60 '
E - 70 ■
O
O -80
-90 -
- ,0 °0 20 40 60 80 100 120 140
Iteration
Figure 5: Global Energy Curve
Figure 5 shows the corresponding global energy curve during
the searching for optimal configuration. As can be seen from
those figures, at the beginning, the global energy decreases
quickly with the removal of false trees, and reaches the minima
of global energy get at 42 iterations. Later, the global energy
start to increase when trees are over-pruned, which indicates the
effectiveness of the designed energy functions in our model.
Finally, 127 trees arc detected from the CHM data. What we
can interpret as well is that in the optimal configuration, the
“false” treetops located on the crown edge and trees with over
lapping crowns are removed at a high accuracy, when compared
with the initial configuration. At the meantime, trees with big
crown are kept even with some extend of overlaps.
6. CONCLUSION AND FUTURE WORKS
We have presented in this paper detecting individual trees from
ALS data. The innovation of the method is formulating the tree
detection in the data as a high-level MRF labeling problem, and
highlights the problem representation and energy function de
sign in the Markov Random Field model. In this approach, trees
are modeled as objects with treetops, crown radius and some
other features extracted from the data and the data is regarded
as configurations of those objects. Then, neighborhood system
is proposed to introduce relationships between the objects and
energy functions are carefully designed to corporate the con
straints in model. Finally, the optimal configuration is found
through an energy minimization process. The experimental re
sult shows a good detection rate of single trees in the data.
The advantage of the method lies in that low level vision me
thod is first used to extract priori information from the data, and
trees are abstracted from that information in a high level. Then
the problem is formulated using a Markov Random Field model.
In such a way, the size of configuration space is greatly reduced
and much less computation will be needed in the searching of
optimal solution. Furthermore, under such a mathematic
framework, other features or constraints extracted from data or
even other sources, which help in the detection of trees, e.g.
stems detected underneath the canopy cover, can be easily add
ed and integrated into the current model without having to alter
the structure of algorithm. However, there are still some issues
to be studied in the future in order to improve the method. The
first one will be model optimization. A RJMCMC embedded
simulated annealing is suggested to be introduced for searching
the configuration space more thoroughly to get the optimal con
figuration. Secondly, it will also be interesting to explore some
algorithm can be employed to help find best weighting coeffi
cients automatically. Finally, more investigation will be imple
mented to find out which scoring functions designed play more
significant roles in the penalization of false trees and detection
of true trees. Also, this method will be applied to more datasets
to test its feasibility to forests of different types or structures.
ACKNOWLEDGEMENTS
This research was supported by a grant for a project entitled
"Automated Change Detection of 3D Landscape Objects for
Powerline Corridor Mapping by Integrating Airborne LiDAR
and Multiple-viewing Digital Cameras” funded by Ontario
Centres of Excellence and GeoDigitial International Inc.
REFERENCES
Besag, J. (1974). Spatial interaction and the statistic analysis of
lattice systems. J. Royal. Statist. Soc. B, 36(2), 192-236.
Brandtberg, T. (2007). Classifying individual tree speices under
leaf-off and leaf-on conditions using airborne lidar. ISPRS
Journay of Photogrammetiy and Remote Sensing, 61(5), 325-
340.
Chen, Q., Baldocchi, D., Gong, P., & Kelly, M. (2006).
Isolating Individual Trees in a Savanna Woodland Using Small
Footprint Lidar Data. Photogrammetric Engineering <£ Remote
Sensing, 72(8), 923-932.
Descombes, X., & Pechersky. E. (2006). Tree Crown
Extraction using a Three State Markov Random Field.
Technical Report n°5982, INR1A.
Genian, S., & Geman. D. (1984). Stochastic relaxation, Gibbs
distributions and the Bayesian restoration of images. IEEE
Trans, on Patt. Anal and Mach. Intell.. 6(6), 721-741.
Hyyppa, J., Hyyppa, H., Leckie, D., Gougeon, F., Yu, X., &
Maltamo, M. (2008). Review of methods of small-footprint
airborne laser scanning for extracting forest inventory data in
boreal forests. Ini. J. Remote Sens., 29(5), 1339-1366.
Leckie, D., Gougeon, F., Hill, D., Quinn, R., Amstrong, L., &
Shreenan, R. (2003). Combined high-density lidar and
multispectral imagery for individual tree crown analysis.
Canadian Journal of Remote Sensing, 29(5), 1-17.
Li, S. Z. (1994). Markov Random Field Models in Computer
Vision Lecture Notes in Computer Science (Vol. 801, pp. 361-
370). Berlin/Heidelberg: Springer-Verlag.
Perrin, G., Descombes, X., & Zerubia, J. (2005). Point
processes in forestry : an application to tree crown detection.
Techical Report n°5544, INRIA.
Pyysalo, U., & Hyyppa. H. (2002). Reconstructing Tree Crowns
from Laser Scanner Data for Feature Extraction. IAPRSSIS,
Vol. 34 (Part 3B), Graz, Austria, 218-221.
Rahman, M., & Gorte, B. (2009). Tree Crown Delineation
From High Resolution Airborne. Lidar Based on Densities of
High Points. IAPRSSIS, Vol. 38 (Part 3/W8), Paris, France,
123-128.
Reitberger, J., Schnorr, Cl, Krzystek, P„ & Stilla, U. (2009). 3D
segmentation of single trees exploiting full waveform LIDAR
data. ISPRS Journal of Photogrammetiy and Remote Sensing,
64(6), 561-574.
Solberg, S.. Ntesset, E., & Bollandsas, O. (2006). Single tree
segmentation using airborne laser scanner data in a structurally
heterogeneous Spruce forest. Photogrammetric Engineering &
Remote Sensing, 72(12), 1369-1378.
