ISPRS Commission III, Vol.34, Part 3A ,Photogrammetric Computer Vision“, Graz, 2002 
  
  
  
  
  
72690x10? L1 pvr bran eel er et | Lid edly 
  
0° **'2.57130x10° 2871405105 7150105 E ORI Trond 
Easting (m) 
Figure 6. Planimetric view of MAP estimate of crown 
configuration (black circles), photogrammetric 
crown measurements (short dashes) and 0.081 ha 
circular inventory plot boundary (long dashes). 
S. DISCUSSION 
Results indicate that the algorithm is generally successful in 
identifying structures associated with individual tree crowns 
within this forest area. The MAP estimate of the crown 
configuration generated by the algorithm closely matches the 
spatial patterns evident in the LIDAR data (Figure 5). The 
algorithm appears to be very sensitive to the data, and in some 
areas added spurious small crowns to increase the likelihood of 
the data. 
In general, the MAP estimate of crown locations corresponds to 
the photogrammetric crown measurements (see Figure 6). It 
should be noted that accurate recognition and delineation of 
overlapping tree crowns is difficult even in high-resolution 
aerial imagery. In this case, there is a systematic discrepancy of 
1-4 meters in the north-south direction between algorithm-based 
crown locations and photo-based crown locations. This offset is 
probably due to the effect of crown layover and/or 
misregistration of the aerial photography. 
Field data was available for a 0.081 ha circular inventory plot 
located within the study area (see Figure 6). Interestingly, the 
number of codominant (overstory) trees found within the plot in 
the field (14) matches the number found by the algorithm and 
measured in the photographs. 
6. CONCLUSIONS 
Bayesian object recognition provides a promising framework 
for the analysis of complex forest scenes using high-density, 
three-dimensional LIDAR data. It is clear that modelling 
assumptions will have a strong influence on the results; for 
example, it is apparent that crowns with an asymmetrical, 
irregular shape will be difficult to detect given the constraints of 
the generalized ellipsoidal crown model used here. The use of 
more complex crown models may improve recognition of 
irregularly shaped crowns. 
Future research will focus on comparing algorithm results to 
field-based measurements and assessing the influence of 
automated measurement error on  stand-level parameter 
estimates. In addition, Bayesian object recognition offers a 
flexible modelling approach that allows for fusing the 
information content from multiple sources of data. Such 
multiple data sources are becoming more available as vendors 
offer simultaneous acquisition of georeferenced imagery and 
LIDAR data. As the data enter the model only through the 
likelihood function in Bayesian object recognition, other types 
of remotely sensed data (including aerial photography and high 
resolution satellite imagery) can be easily incorporated into the 
model through adjustment of the likelihood function. 
7. REFERENCES 
Andersen, H., S. Reutebuch, and G. Schreuder, 2001. 
Automated individual tree measurement through morphological 
analysis of a LIDAR-based canopy surface model. In: 
Proceedings of the First International Precision Forestry 
Symposium, Seattle, WA, USA. 
Baddeley, A., and M. van Lieshout, 1993. Stochastic geometry 
models in high-level vision. In: Mardia, K.V. and G.K. Kanjii, 
eds. Advances in Applied Statistics 1. Carfax, Abingdon, 
Oxfordshire, pp. 231-256. 
Besag, J., 1993. Towards Bayesian image analysis. In: Mardia, 
K.V. and GK. Kanjii, eds. Advances in Applied Statistics 1. 
Carfax, Abingdon, Oxfordshire, pp. 107-119. 
Dralle, K., and M. Rudemo, 1997. Automatic estimation of 
individual tree positions from aerial photos. Canadian Journal 
of Forest Research, 27, pp. 1728-1736. 
Goudriaan, J., 1988. The bare bones of leaf-angle distribution in 
radiation models for canopy photosynthesis and energy 
exchange. Agricultural and Forest Meteorology 43, pp. 155- 
169. 
Green, P., 1995. Reversible jump Markov chain Monte Carlo 
computation and Bayesian model determination. Biometrika, 
82(4), pp. 711-732. 
Gulden, R., 2000. Remote sensing at the dawn of a new 
millennium: A Washington DC perspective. In: Proceedings of 
the Eighth Biennial Remote Sensing Applications Conference, 
Albuquerque, NM, USA. 
Kirkpatrick, S., C.D. Gelatt, Jr., and M.P. Vecchi, 1983. 
Optimization by simulated annealing. Science, 220(4598), pp. 
671-680. 
Larsen, M., 1998. Finding an optimal match window for spruce 
top detection based upon an optical tree model. In: Proceedings 
of the International Forum on Automated Interpretation of 
High Spatial Resolution Digital Imagery for Forestry, Victoria, 
BC, Canada. 
Lund, J. and M. Rudemo, 2000. Models for point processes 
observed with noise. Biometrika, 87(2), pp. 235-249. 
Magnussen, S., P. Eggermont, and V.N. LaRiccia, 1999. 
Recovering tree heights from airborne laser scanner data. Forest 
Science, 45(3), pp. 407-422. 
Mea 
Hen 
airb 
Rem 
Nel: 
bion 
of th 
Poll 
in a 
ima; 
Unr 
Ripl 
Ripl 
Ant 
Leci 
Stat 
Rue 
Bio! 
Rue 
with 
30, 
She 
Crov 
ima, 
67(¢ 
Sun 
fore 
Sen. 
Van 
arch 
99- 
van 
neig 
recc 
van 
ana 
Net 
Wel 
of 
Phy 
Ack 
The 
Wa: 
Sen 
Sup] 
to 
Res