Monitoring the composition of forest stands using airborne laser scanning and aerial photography data

Reliable identification of the species composition of forest stands is required for forest inventory, environmental monitoring, and assessment of forest resources. Airborne laser scanning provides a three-dimensional description of the forest canopy, including height, point density, crown structure, and reflected signal characteristics. However, the geometric structure of a point cloud alone is often insufficient for accurate tree species classification. Species with similar canopy morphology may have close geometric characteristics, especially in mixed forest areas and under conditions of uneven point cloud density.

Purpose – The purpose of this study is to modernize the RandLA-Net neural network architecture for tree species classification based on the joint processing of airborne laser scanning point clouds, aerial photography data, and contextual characteristics of a forest plot.

A modified neural network architecture, MMF-RandLA-Net, is proposed. In this architecture, the three-dimensional branch processes the geometry of the airborne laser scanning point cloud, the optical branch extracts image features from aerial photography data, and a separate contextual block takes into account generalized characteristics of the forest plot. Multilevel feature fusion is used to combine these representations. During the experimental study, the modified neural network model was compared with the baseline RandLA-Net architecture using several classification quality metrics. The obtained results show that the proposed model improves classification quality by 1.5–3.7 percentage points depending on the metric.

The proposed neural network architecture can be used for preparing thematic maps of tree species composition, refining forest inventory materials, monitoring the condition of forest stands, and processing airborne observation data.