Planting contexts affect urban tree species classification using airborne hyperspectral and LiDAR imagery
Dengkai Chi, Jingli Yan, Kang Yu, Felix Morsdorf, Ben Somers
Abstract
• Intraspecific variations were found in trees in different urban planting contexts. • These variations can affect remote sensing-based tree species classification. • We proposed a planting context-specific modelling approach. • The proposed approach significantly improved tree species classification accuracy. • The approach will facilitate more effective urban forest management. Different urban planting contexts, such as streets and parks, can lead to significant intraspecific biochemical and structural variations in trees. These variations present challenges for remote sensing-based tree species classification and effective urban forest management. However, few studies have explored how planting contexts influence the accuracy of remote sensing-based tree species identification in urban environments. This study introduced a planting context-specific modelling approach (i.e., models trained for street trees and park trees separately) for classifying seven dominant broadleaved tree species in the Brussels Capital Region, Belgium using airborne hyperspectral and leaf-on LiDAR data. This approach was compared to a traditional general modelling approach. Linear discriminant analysis with principal component analysis was employed to classify tree species at the individual tree level using different feature sets. Our results showed that a planting context-specific modelling approach with combined hyperspectral and LiDAR features achieved an overall accuracy (OA) of 84.2%. It improved the OA of LiDAR-based classifications by 7.6 and 8.9 percentage points for street trees and park trees respectively and of hyperspectral-based street tree species classification by 4.2 percentage points. The decreased discriminatory power of features in general models can be partly attributed to their sensitivity to planting context. We concluded that a planting context-specific modeling approach can enhance urban tree species classification, ultimately supporting improved urban forest management.