Individual tree identification using different LIDAR and optical imagery data processing methods

Abstract

The most important part in forest inventory based on remote sensing data is individual tree identification, because only when the tree is identified, we can try to determine its characteristic features. The objective of research is to explore remote sensing methods to determine individual tree position using LIDAR and digital aerial photography in Latvian forest conditions. The study site is a forest in the middle of Latvia at Jelgava district (56º39' N, 23º47' E). Aerial photography camera (ADS 40) and laser scanner (ALS 50 II) were used to capture the data. A LIDAR data is 1.4 to 9 p/m2 depending on the altitude. Image data is RGB (Red, Green, and Blue), NIR (Near Infrared) and PAN (Panchromatic) spectrum with 20 to 50 cm pixel resolution depending on the altitude. Image processing was made using Fourier transform and RGB colour segmentation. LIDAR data are processed with DBSCAN algorithm, global maximum algorithm, and local maximum algorithm. Field measurement's parameters were tree coordinates, species, height, diameter at breast height, crown width, and length. Best results on both ALS and ADS data were achieved using local maximum methods. 1. Introduction The most responsible and important part in forest inventory based on remote sensing data is individual tree identification, because only when the tree is identified, we can try to determine its characteristic features, like tree species, tree height, diameter at breast height, volume, and biomass (Secord et al., 2006; Edson and Wing, 2011). In the studies of forest inventory using remote sensing sensors, one of the main problems that the authors mentioned is tree identification and tree location accurate determination (Hyyppä et al., 2008; Kane et al., 2010), especially in Middle Europe (Diedershagen et al., 2006), since there is a mixture of different deciduous and coniferous trees. As a result, the identification is more difficult. Many authors in their conclusions highlight that the usage of LIDAR and airphoto methods to determine forest inventory parameters will never be one hundred per cent correct (Onge et al., 2004; Rombouts, 2006), especially applying automated tracking methods (Hyyppä et al., 2004; Junttila et al., 2010). Practically for all researchers, so far it has been difficult to identify small trees (Pitkänen, 2001; Pouliot and King, 2005) and closely growing trees (Pouliot and King, 2005; Koch et al., 2006), as well as high density hardwood stands with homogeneous crown (Koch et al., 2006; Rahman and Gorte, 2008). Automated tree identification and tree location accurate determination are still problematic (Popescu et al., 2002; Junttila et al., 2010), even in cases where different types of data (Vauhkonen et al., 2008) are available. This is mainly by the fact that trees vary in crown size (Tokola et al., 2008), shape and optical properties (Tokola et al., 2008; Vauhkonen et al., 2008). For example,