Single-image estimation of tree volume via pixel-mapped 3D reconstruction: A low-cost solution using deep learning...

Although LiDAR is widely used for tree measurement, its high cost and operational complexity remain significant barriers to widespread adoption. With advances in photogrammetry and deep learning, efficient and accurate alternatives have become increasingly important for forest resource surveys. Accordingly, we propose an automated trunk-parameter measurement framework that maps image pixels to physical units. The framework integrates the SegFormer deep-learning model, trunk-skeleton extraction, an adaptive curvature-segmentation algorithm, and segment-wise 3-D reconstruction, thereby enabling image segmentation, curvature analysis, three-dimensional reconstruction, and measurement. To validate its practical value, we collected images of 3013 trees across four species in the Beijing region. Additionally, we acquired point-cloud data and conducted destructive measurements on 141 trees of various species for comparative evaluation. Experimental results indicate that the stem segmentation algorithm effectively extracts trunk regions in images, and the adaptive segmentation method substantially improves trunk volume estimation accuracy. The approach achieves only 2.01 %-7.68 % error in single-tree volume and height measurements-primarily due to segment-height inaccuracies-and offers an approximately 6.9-fold improvement in efficiency compared with the existing HMLS method. In summary, this method provides an efficient, low-cost solution for forestry surveys and shows great potential for monitoring tasks that require high accuracy under resource constraints. This innovative method is expected to further advance forest resource assessment.