Urban Forest Classification with 12 Tree Species by Fusion of Multi-period LiDAR Data Set and the Vegetation Index of Spectral Satellite Imageries

The importance of forests biomass is being highlighted globally because forests conduct carbon sequestration, which could help to alleviate the adverse effects of climate change. Therefore, many studies have focused on accurately classifying trees and proposed various methods, including classifying using satellite imageries and object-based analysis. Although satellite imageries are widely used for tree species classification, there is a limitation in explaining vertical canopy structures, which could feature each tree species. Recently, with the advance of Light detection and ranging (LiDAR) remote sensing, quantifying vegetation structures in a three-dimensional perspective can be possible. This study aimed to increase the accuracy of the tree species classification by fusing multispectral satellite imageries with Airborne Laser Scanning (ALS) datasets. The study area was Mt. Bongseo in Cheonan city, Republic of Korea, and it consisted of about 12 species. We used Landsat 8 imageries in 2014, Sentinel-2 images in 2016, and ALS datasets from 2014 and 2016 with an average density of 8 points/m2. First, the ALS datasets were tiled into a 10 m by 10 m grain unit, which was the same spatial resolution as satellite imagery. Then, we calculated the canopy height metrics such as maximum heights, mean heights, and relative heights for reflecting the vertical canopy structures. Second, for satellite imageries in 2014 and 2016, cloudless multi-band single imageries were extracted using the Cloud Cover filter of Google Earth Engine (GEE). We derived vegetation-related indices such as NDVI, NDVI, MSAVI, and NDWI using these imageries. Finally, we estimated tree classification using the satellite-derived indices, LiDAR-derived variables, and combined datasets (i.e., using both satellite- and LiDAR-derived variables) using the extreme gradient boosting algorithm (XGBoost). As a result, the overall accuracy (OA) of the combined data was 0.71 and 0.93 in 2014 and 2016, respectively, which were greater than 0.2 and 0.04 when satellite images were used alone. Kappa values showed 0.58 and 0.90, respectively, which were 0.25 and 0.05 greater than when satellite-derived indices were used alone. It was confirmed that the combination of satellite images and LiDAR data increases the accuracy of species classification.