Tree crown segmentation based on a geometric tree crown model for prediction of forest variables

A new algorithm for tree crown segmentation from airborne laser scanning data was validated at a test site in southern Sweden (lat. 58 degrees N, long. 13 degrees E). The tree crown segmentation algorithm used a correlation surface created by fitting a geometric tree crown model and was also controlled using an a priori probability function. If the model fit alone was used, 69% of the field-measured trees were detected but when a priori information was used, the proportion of detected trees increased to 75%. The proportion of detected trees represented 95% of the total stem volume for all field measured living trees. The tree crown segments, with zero, one, or several trees, were used as input to an imputation algorithm for prediction of forest variables, which yielded relative root mean square errors of 8.9% for stem volume, 7.2% for basal area, 3.8% for mean tree height, 6.3% for mean stem diameter, and 15% for stem density, after aggregation to plot level for cross-validation. Thus, automatic tree crown delineation using the segmentation algorithm could be used for imputation of tree stems to obtain high accuracy predictions of several forest variables.