Tree species classification with color-coded images generated from laser scanner point clouds

Abstract This paper investigates the use of color-coded images generated from laser scanner point clouds in deep learning networks for tree species classification. We introduce a novel approach that encodes 3D structural tree information contained in the laser scanned point cloud into the RGB channels of standard 2D images, preserving three key spatial attributes: tree depth (how much the tree extends inward within a pixel), tree fill (how much the tree occupies spaces within a pixel), and distance (minimum distance of the points in each pixel to the projection plane). Tree classification is based on multiple images (various viewpoints, stem-restricted views, vertical sections) together with metadata on visible dimensions, maximum depth, viewing angle, and image type. Low-resolution images (224 × 448 px) were classified using shallow convolutional neural networks, and the final tree-level prediction was obtained by aggregating classifications from multiple images. The proposed classifier, 3DTreeView, was evaluated using the FOR-species20K benchmark dataset, as well as across different biogeographic regions separately. Its performance was assessed in relation to scanner types, biogeographic regions, and tree height categories. The classifier achieved an accuracy of 91% for the 7 species in the boreal biogeographic region, 80% for the 22 species in the temperate region, and 68% for the full dataset comprising 33 species. Overall, the results demonstrate that colored images of the trees obtained from projecting the 3D point clouds combined with multi-view deep learning provide an effective and practical approach for tree species recognition across diverse forest conditions and biogeographic regions, ranging from small to large numbers of trees.