Accurate tree species classification plays a central role in forest management, biodiversity monitoring, and ecological research. In boreal environments, the relatively low species diversity and simple canopy structure offer favorable conditions for species-level analysis using remote sensing. However, classification remains challenging due to structural heterogeneity and the scattered distribution of ecologically significant broadleaved species such as European aspen. This dissertation evaluates the potential of uncrewed aerial vehicles (UAVs) equipped with RGB, multispectral (MSP), and LiDAR sensors to classify tree species and detect key biodiversity indicators in boreal forests of Finland.
The research is based on four sub-studies conducted in boreal conditions using various UAV platforms and sensor configurations, including helicopter-based RGB imagery, photogrammetric point clouds, and UAV-mounted LiDAR. Tree species included in classifications were Scots pine (Pinus sylvestris L.), Norway spruce (Picea abies (L.) Karst.), birches (Betula pendula Roth, B. pubescens Ehrh.), and European aspen (Populus tremula L.). Classification methods included linear discriminant analysis (LDA), support vector machines (SVM), and random forest (RF), using features derived from spectral, structural, texture, and shape attributes.
The highest classification accuracy (OA = 95%) was achieved using early-season multispectral imagery combined with manually delineated crown segments. However, automatic segmentation using RGB-derived point clouds also performed strongly, achieving 92% overall accuracy and a kappa coefficient of 0.90. Integrating structural and spectral information, particularly when acquired simultaneously, consistently improved classification outcomes. The methods also proved effective in identifying ecologically important elements such as European aspen and standing dead trees, with F1-scores for aspen reaching up to 97%.
The results further show that seasonal conditions significantly influence classification accuracy, with early phenological stages providing optimal spectral separability. Feature importance analyses underscored the value of combining spectral, structural, and textural information to maximize classification performance. While multi-sensor setups offered the highest accuracies, well-executed single-sensor approaches still produced reliable results under favorable conditions.
This thesis demonstrates that UAV-based remote sensing offers a cost-efficient, high-resolution alternative to field inventories in tree species classification and biodiversity assessment. The findings support the development of flexible, operational workflows tailored to boreal forest conditions and biodiversity monitoring needs, advancing the role of UAV technologies in sustainable forest management.