Stefano Puliti
Research Scientist
Authors
Xinlian Liang Yinrui Wang Jun Pan Janne Heiskanen Ningning Wang Siyu Wu Ilja Vuorinne Jiaojiao Tian Jonas Troles Myriam Cloutier Stefano Puliti Aishwarya Chandrasekaran James Ball Xiangcheng Mi Guochun Shen Kun Song Guofan Shao Rasmus Astrup Yunsheng Wang Petri Pellikka Mi Wang Jianya GongAbstract
Accurate individual tree delineation (ITD) is essential for forest monitoring, biodiversity assessment, and ecological modeling. While remote sensing (RS) has significantly advanced forest ITD, challenges persist, especially in complex forest environments. The use of imagery data is compelling given the rapid increase in available high-resolution aerial and satellite imagery data, the increasing need for image-based analysis where reliable 3D data are unavailable, the widening gap between data supply and processing capabilities, and the limited validation of state-of-the-art (SOTA) methods across diverse real-world conditions. This study aims to advance ITD research by evaluating SOTA instance segmentation approaches, including both recently developed and established methods. The analysis evaluates ITD algorithm performance using the largest forest instance-segmentation imagery dataset to date and standardized evaluation protocols. This study identifies key factors affecting accuracy, reveals remaining challenges, and outlines future research directions. Findings in this study reveal that ITD accuracy is heavily influenced by image resolution, forest structure, and method design. Findings also reveal that, while algorithm innovations remain important, robustness and transferability that ensure generalization across diverse environments are what differentiate method performances. In addition, this study highlights that commonly used evaluation metrics may fail to adequately capture precise performance in specific applications, e.g., individual-tree-crown segmentation in this study. Assessment reliability can be strengthened through the adoption of stricter criteria. Future research should focus on expanding datasets, refining evaluation protocols, and developing adaptive models capable of handling varying canopy structures. These advancements will enhance ITD scalability and reliability, contributing to more effective forest research and management at a global scale.
Authors
Adrian Straker Paul Magdon Marco Zullich Maximilian Freudenberg Christoph Kleinn Johannes Breidenbach Stefano Puliti Nils NölkeAbstract
Deep learning-based tree species classification models applied to terrestrial laser scanning data achieve high accuracies, yet their classification decisions remain poorly understood. Addressing this gap, we present a novel explanation framework that systematically links salient regions of Finer-CAM saliency maps—which highlight image features contributing to the classification of a target species while suppressing features shared with similar species—to structural tree features in 2D side-view images of terrestrial laser scanning (TLS) point clouds. We applied this framework using YOLOv8-based tree species classification models trained on a subset of the FOR-species20K benchmark data set. Faithfulness analysis confirmed that Finer-CAM reliably identifies the most discriminative image regions for model predictions, validating its suitability for enhancing model interpretability. Analysis of 630 saliency maps indicate that the models primarily rely on image regions associated with tree crowns across most species, particularly for Silver Birch, European Beech, English Oak, and Norway Spruce, while stem regions contribute more strongly to the differentiation of European Ash, Scots Pine, and Douglas-fir. Image perturbation experiments further demonstrated that the visibility of tree structure—such as branching patterns—in the 2D side-view images enhances classification performance, indicating YOLOv8’s capabilities to leverage detailed structural information encoded in high-density TLS point clouds. Using a cross-validation approach our trained models achieved a mean overall accuracy of 96% (SD = 0.24%) on a test data set unseen during training. Beyond assessing the contribution of image regions associated with structural tree features to tree species differentiation, our methodological framework can facilitate a better understanding of model limitations, the detection of data set biases and shortcut learning, and contributes in building confidence in deep learning-based tree species classifiers.
Authors
Michele Torresani Vítězslav Moudrý Christopher R. Hakkenberg Vojtěch Barták Duccio Rocchini Stefano Puliti Paweł Hawryło Krzysztof Stereńczak Alexander Cotrina-Sanchez Gabriele Giuseppe Antonio Satta Luca da Ros Patrick Kacic Roberto TognettiAbstract
Monitoring forests globally through the assessment of structural characteristics is indispensable in times of increasing disturbances and biodiversity loss. The recent development of a 1-meter resolution Global Canopy Height Map by Meta and the World Resources Institute (Meta/WRI CHM) offers new opportunities for large-scale forest structure analysis. However, its reliability for estimating key forest structural metrics in selected forest sites of the Italian Alps remains largely untested. In this study, we compared estimates of canopy cover, mean and maximum canopy height, tree count, and crown size computed from the Meta/WRI CHM with the corresponding metrics computed from airborne laser scanning (ALS)-based CHMs across five forested sites in Alpine ecosystems representing diverse forest structures, species compositions, and management practices. Our results show that the Meta/WRI CHM provides reliable estimates of canopy cover (R = 0.82–0.92, RMSE = 8%–15%) and, to some extent, mean canopy height (R = 0.69–0.85, RMSE = 2.2–2.9 m). However, it substantially underestimates maximum canopy height and fails to reliably estimate tree count, positions, and crown size, overestimating the number of trees by 200–750 per hectare. In addition, the quality of the evaluated metrics estimates varied with respect to topographic gradients (i.e., slope, aspect, and altitude). Our findings define clear application boundaries for the Meta/WRI CHM in Alpine forest environments: while the product can support stand-level estimates of canopy cover and mean canopy height, it remains unsuitable for estimating maximum canopy height and individual-tree-level metrics such as tree count, tree position, and crown size. Future efforts should focus on refining global CHMs to improve accuracy and expand their applicability for forest monitoring.
Division of Forest and Forest Resources
A Decision Support System for emerging forest management alternatives
This project aims to develop advanced tree growth models using LiDAR-derived, high-density point cloud data to improve the simulation of forest dynamics under close-to-nature silvicultural practices. By modeling tree-level growth in structurally complex and heterogeneous stands, these models will support more accurate, spatially explicit forest simulations and inform sustainable and diversified forest management decisions.
Division of Forest and Forest Resources
SFI SmartForest: Bringing Industry 4.0 to the Norwegian forest sector
SmartForest will position the Norwegian forest sector at the forefront of digitalization resulting in large efficiency gains in the forest sector, increased production, reduced environmental impacts, and significant climate benefits. SmartForest will result in a series of innovations and be the catalyst for an internationally competitive forest-tech sector in Norway. The fundamental components for achieving this are in place; a unified and committed forest sector, a leading R&D environment, and a series of progressive data and technology companies.