Delineation of Individual Tree Crowns from ALS and Hyperspectral data: a comparison among four methods

Forest structural properties are traditionally acquired during extensive fieldwork campaigns. A great potential for time saving is given by remote sensing assisted inventories. Recently great attention has been devoted to individual tree crowns (ITC) level forest inventories. In ITC inventories a key step is the delineation of the tree crowns. Thus, in this study we will compare four methods for the identification of individual tree crowns (ITC) based on high density airborne laser scanning (ALS) and hyperspectral data. The study area is an alpine forest located in Lavarone at 1400 m (Trento Province, Italy) above sea level. 600 trees were inventoried in a plot of 4800 m2, of which 58% Silver Fir (total basal area: 37 m2/ha), 23% European beech (total basal area: 2 m2/ha), 19% Norway Spruce (total basal area: 22 m2/ha). ALS data were acquired by an Optech ALTM 3100EA sensor, with a mean density of 8.6 points/m2 for the first return (laser pulse wavelength 1064 nm, laser repetition rate 100 kHz) and with up to four recorded returns for each laser pulse. Methods 1 and 2 exploit both a CHM in raster and point cloud formats. The main difference among them is that method 1 uses a watershed segmentation to delineate the ITC, while method 2 uses a region growing algorithm. In greater detail the two methods can be summarized as follows: i) a raster CHM is created from point cloud; ii) the CHM is properly filtered to avoid inclusion of non-vegetated objects; iii) a watershed segmentation (method 1) or a region growing (method 2) is applied; iv) ITCs are reshaped using a morphological filter and their raw ALS point cloud distribution. Method 3 is based on raw ALS cloud and focuses on the delineation of intermediate and suppressed trees. In greater detail the method can be summarized as follows: i) the point cloud is divided into horizontal layers to which a 3D K-means clustering is applied; ii) K-means clusters are grouped using a prolate ellipsoid shape along all the layers; iii) the distribution of points in the clusters is estimated along x and y axes and the uneven distributed clusters are separated into two new 3D clusters; iv) clusters are eventually merged along all layers, in 2D space, and grouped into final 3D clusters representing ITCs. Method 4 is based on hyperspectral data. ITCs are delineated on a single raster band (band at 810 nm). In greater detail the method can be summarized as follows: i) the raster image is filtered with a low pass filter, and with a thresholding filter in order to highlight only the tree crowns; ii) a watershed segmentation algorithm is applied; iii) a morphological filter is used in order to reshape the final ITCs. The above described methods will be compared with the field inventory data for the identification of individual trees and their canopy sizes.