ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision", Graz, 2002 
  
BAYESIAN OBJECT RECOGNITION FOR THE ANALYSIS OF COMPLEX FOREST 
SCENES IN AIRBORNE LASER SCANNER DATA 
Hans-Erik Andersen * *, Stephen E. Reutebuch 5 Gerard F. Schreuder*? 
* University of Washington, College of Forest Resources, Seattle, WA, 98195 USA - 
(hanserik, gsch)@u.washington.edu 
® USDA Forest Service, Pacific Northwest Research Station, Seattle, WA, 98195 USA - sreutebuch@fs.fed.us 
Commission III, WG III/3 
KEY WORDS: Forestry, laser scanning, LIDAR, object, recognition, remote sensing, statistics 
ABSTRACT: 
Bayesian object recognition is applied to the analysis of complex forest object configurations measured in high-density airborne laser 
scanning (LIDAR) data. "With the emergence of high-resolution active remote sensing technologies, highly detailed, spatially 
explicit forest measurement information can be extracted through the application of statistical object recognition algorithms. A 
Bayesian approach to object recognition incorporates a probabilistic model of the active sensing process and places a prior 
probability model on object configurations. LIDAR sensing geometry is explicitly modelled in the domain of scan space, a three- 
dimensional analogue to two-dimensional image space. Prior models for object configurations take the form of Markov marked point 
processes, where pair-wise object interactions depend upon object attributes. Inferences are based upon the posterior distribution of 
the object configuration given the observed LIDAR. Given the complexity of the posterior distribution, inferences are based upon 
dependent samples generated via Markov chain Monte Carlo simulation. This algorithm was applied to a 0.21 ha area within Capitol 
State Forest, WA, USA. Algorithm-based estimates are compared to photogrammetric crown measurements and field inventory data. 
1. INTRODUCTION 
1.1 Automated forest inventory 
While national and local inventories often utilize remotely 
sensed data for stratified sampling and classification of general 
forest type, most of these programs remain heavily reliant upon 
expensive field data for individual tree-level information. At the 
national level in the United States, individual tree inventory 
information is collected at considerable expense. It is significant 
that senior researchers within the USDA Forest Service Forest 
Inventory and Analysis (FIA) program have recognized the 
need for the development of automated forest interpretation and 
measurement algorithms to reduce human intervention and 
labor costs (Gulden, 2000). 
1.2 The LIDAR technology 
LIDAR (LIght Detection And Ranging) is an operationally 
mature remote sensing technology that can provide highly 
accurate measurements of both forest canopy and ground 
surface. While specifications vary among systems, LIDAR 
systems emit from 5,000 - 100,000 pulses per second. In 
forested areas, individual LIDAR pulses can penetrate the forest 
canopy through gaps, and can therefore acquire information 
relating to three-dimensional forest structure as well the 
underlying terrain surface. 
1.3 LIDAR analysis for forest measurement applications 
In recent years, there has been increasing interest in the use of 
LIDAR for automated detection and measurement of forest 
features. Research efforts in the last fifteen years were focused 
  
* Corresponding author. 
on the use of small footprint (0-1 m pulse diameter) LIDAR 
systems to estimate forest stand level parameters (Nelson et al., 
1988; Means ef al, 2000). Researchers in Canada used a 
model-based approach to recover tree heights from LIDAR 
canopy height measurements (Magnussen et al., 1999). Three- 
dimensional mathematical morphology has been applied to a 
high-resolution LIDAR-based canopy surface model to extract 
individual tree measurements (Andersen ef al., 2001). 
1.4 Automated individual tree crown recognition through 
template matching 
With the recognition that high-resolution remotely sensed 
spatial data can support more intensive forest management 
practices, there has been increasing interest in recent years in 
the development of algorithms for automated identification and 
measurement of individual trees using high-resolution, two 
dimensional digital imagery. Several studies have used a model- 
based approach to locate individual trees using tree crown 
template models (Pollock, 1996; Larsen, 1998; Sheng et al, 
2001). 
Researchers in Scandinavia have attempted to model the 
relationship between the spatial distribution of individual trees 
and the position of spectral maxima in a digital image (Dralle 
and Rudemo, 1997). Another Scandinavian study has used 
deterministic parameter search methods for maximum 
likelihood estimation on a spatial point process model to infer 
the parameters of a disturbance model that relates the true 
position of tree-tops to those observed on an aerial photograph 
(Lund and Rudemo, 2000).