Boreal Forest Tree Species Classification at the Crown Scale Using Fusion of G-LiHT Hyperspectral and Lidar Data

Knowledge of species composition across the boreal region is important for understanding wildfire behavior and patterns of post-fire forest recovery, and can lead to improved estimates of forest carbon stocks and fluxes. In particular, the expansion of deciduous tree coverage in areas previously dominated by black spruce has been hypothesized as a result of post fire succession, but it is not settled as to whether post-fire species shifts represent an ecosystem state change or simply transitional vegetation coverage of smaller forbs or shrubs. As these are moderate to coarse resolution optical remote sensing products, the commonly used MODIS and Landsat imagery do not permit analysis of these forest processes that occur at a fine spatial scale and require information on tree and terrain structure. More detailed analyses of forest structure and composition requires both high spatial resolution data and the structural information to monitor species at a scale appropriate for ecosystem processes and stakeholder (e.g., USDA Forest Service) inventory needs.

In this paper, we fused Goddard Lidar Hyperspectral Thermal (G-LiHT) hyperspectral imagery and LiDAR data at multiple object scales to map tree species distributions across the Tanana Forest Inventory Unit in boreal Alaska. Employing a multiscale segmentation process, we extracted visible to near infrared (VNIR) spectral information at the crown level, canopy height characteristics at the stand level, and elevation, slope, and aspect data from larger terrain features using the superpixel (SLIC) segmentation of the gridded LiDAR data. The resulting combination of hyperspectral bands, canopy characteristics, and terrain-derived factors were used to train a random forest algorithm to classify tree species and leaf type at the individual crown level. Validation data came in the form of crowns manually delineated from G-LiHT data, Forest Service inventory plots, and existing high resolution UAV imagery. Preliminary results from calibration/validation datasets indicate an above 90% overall accuracy for leaf type. We expect that the accuracy of species classification may be improved from the addition of canopy and terrain information. This work also represents the first step towards fine scale species discrimation across spatially extensive boreal regions.