Transferability of a Tree-Crown Delineation Approach Using Region-specific Segmentation

Abstract

In this paper we present experiences of a transferability study on single tree delineation by region- specific segmentation of air-borne laser scanning (ALS) data with varying point density. Encouraged by promising results of preceding studies on high-density ALS data (Tiede et al., 2006, Tiede & Hoffmann, 2006) we assessed the transferability to an area, in which (1) lower density ALS data is available obtained by a different flight campaign; (2) the generation of a normalised crown model (nCM) is quite influenced by rough terrain; and (3) optical data (FCIR aerial photographs) differs from ALS data in terms of resolution and time of acquisition. Region-specific segmentation means utilising a-priori knowledge of the respective scale domain of the envisaged target features, while finally representing the scene content in a spatially contiguous one-level- representation (OLR). In total 2,344 single trees taller than 5 m were detected, for almost all of which a tree crown was delineated, even in dense pole forest. That means an average overestimation of 23 % as compared to visual interpretation, mainly due to shady conditions of the aerial photographs. A comparison based on 100 m x 100 m raster cells of manually and automatically extracted tree tops shows high congruence (correlation coefficient: 0.95).