Evaluation of tropical rainforest crown segmentation and individual tree inventory using airborne LiDAR data

Abstract

Light Detection and Ranging (LiDAR) offers higher accuracy to map forest structure at horizontal and vertical scales and thus improve the forest inventory process especially in the tropical rainforest area in Sarawak. Tree structure attributes such as Canopy Height Model (CHM) or the individual tree-and-area-based information. Tropical forest with dense canopy structures always hampers the accuracy of tree location, while the existing sampling strategy is rather random and insufficient to present the areal coverage. Furthermore, the CHM segmentation has been less tested in tropical regions, and thus, no study to identify the better method to be adapted for the forest in Borneo has been found. The study area is one of the forest plots in the Bukit Hitam Nature Reserves (BHNR) of Limbang, Sarawak covering about 22 hectares. This area was chosen because the forest diversity has been recognised as one of the nature parks in Malaysia. Therefore, this study was aimed to evaluate the existing crown segmentation method for LiDAR so-called the Marker-controlled watershed segmentation on the data measurement in BHNR. The objectives were to develop allometric relation using in-situ data; to estimate the forest crown in raster- and point-based segmentation methods; and to assess the quality of extracted canopy based on comparison with in-situ data. ArcGIS and LiDAR360 were used to perform the crown segmentation. The crown delineation from segmented results were evaluated using univariate statistics and Jaccard Index (JI). Results from ArcGIS showed an under-estimation at JI=0.41 which was less than 50% from inside the reference crown, but LiDAR360 indicated JI=0.61 representing more than half of the matching segmentation. The accuracy of crown diameter produced by ArcGIS and LiDAR360 was 15.7 and 31.7 meters respectively. This study found that LiDAR360 could offer better segmentation in raster and point-based methods and the point-based segmentation has proven that the discrete point is slightly better than the raster-based output.