Conference Agenda

The development of high-performance, affordable UAVs has transformed vegetation monitoring, enabling observation of forest canopies at an unprecedented level of detail. UAV-derived datasets now provide high-fidelity structural and physiological information at the individual tree level across entire forest stands, offering novel insights into forest dynamics. In the context of increasing tree mortality, such data are becoming essential for understanding forest resilience and adaptation. However, exploiting this data requires effective individual tree crown segmentation algorithms (ITCS) at the forest scale, capable of tackling large-scale data and variability introduced by the environment. In this paper, we developed a new workflow designed to process UAV hyperspectral imagery at the forest scale, enabling automated ITCS and analysis. Our pipeline integrates hyperspectral-to-RGB conversion, ITCS, and centroid-based mask fusion. To assess the performance of our pipeline, we evaluated the model on two replicated white spruce common gardens in Canada, each comprising approximately 6,000 trees of similar age and structure. The experiments rely on a large multi-temporal dataset of hyperspectral imagery acquired during 60 UAV missions between 2022 and 2024, allowing us to evaluate the robustness of the proposed pipeline across a wide range of seasonal and acquisition conditions. Results show that the proposed pipeline achieves a mean segmentation performance of 0.536 mAP (0.885 mAP50) on the annotated dataset. At the forest scale, the system demonstrates strong detection capability with F1-scores of 0.948 at the Pintendre site and 0.863 at the Pickering site, successfully detecting most trees while maintaining stable performance across varying environmental conditions.

Individual tree detection and crown delineation (ITDCD) in dehesa landscapes is complicated by geometric distortions from steep terrain, varying tree densities, and the partly multi-crown 'broccoli-like' structure of holm and cork oaks. This study evaluates the usability of a modified StarDist deep learning model, which has recently shown effectiveness for ITDCD in Canadian forests. Moreover, this study develops a workflow transforming the original StarDist, designed for microscopy images, into an ITDCD solution, taking the georeferencing of geospatial data into account. The tile-wise organized ground truth dataset is created with the pretrained Tree Segmentation model available in the ArcGIS Living Atlas, combined with manual revision. Several augmentation methods are applied, resulting in 960 images, which are split into 85 % for training and 15 % for validation. Following the approach of the Canadian forest study, the StarDist implementation is modified by introducing a constraint to the probability loss function. Rather than computing loss across all pixels, the modified loss function considers only pixels explicitly annotated as objects, while background pixels are excluded. An additional dataset of 1,200 trees serves as ground truth for testing the prediction across the entire study area. Using an Intersection over Union of 0.5, this test demonstrates good performance (Accuracy: 87.50 %; F1-score: 0.85). The accuracy varies with tree density: in areas with sparse tree cover, nearly all tree crowns are detected; in moderately dense areas, a number of tree crowns are missed; whereas in very dense tree layers, the frequency of missed detections increases.

Individual tree crowns detection and delineation from airborne LiDAR data is essential for forest inventory, carbon stock estimation, and ecosystem monitoring. In mixed-wood forests, however, this task remains difficult due to high stand density, multi-layered canopy structure, and the wide variation in crown size and shape across coniferous and deciduous species.

This study addresses two core limitations of existing deep learning methods for individual tree crown delineation. Standard instance segmentation models rely on blind anchor-based proposals that frequently miss small understorey trees in dense canopies, and their pixel-based mask representations struggle to accurately capture crown boundaries for small or irregular crowns.

We propose a multi-task learning framework that jointly trains a structure-aware treetop detection head and a crown segmentation head on a shared backbone network. The treetop detection head generates spatially precise crown seeds guided by canopy height and allometric relationships, replacing blind anchor proposals with data-driven initialisation. Two segmentation strategies are evaluated within this framework: a Mask R-CNN pixel-based approach and a StarDist contour-based approach.

Experiments are conducted on a high-density airborne LiDAR dataset acquired over a mixed-wood forest in Ontario, Canada, comprising 4,417 manually delineated reference crowns. Results demonstrate improved detection completeness for small crowns and more accurate boundary delineation for overlapping larger crowns compared to single-task baselines.

This study examines the use of multi-temporal UAV hyperspectral and LiDAR data to identify tree species in a mixed deciduous forest in southern Ontario, Canada. Weekly UAV flights were conducted from summer through spring to capture structural and spectral changes associated with leaf development, senescence, and leaf drop. Field measurements were collected to provide species labels and biometric information for individual trees.

LiDAR data are processed to delineate individual tree crowns and to derive structural metrics such as crown height, width, density, and vertical canopy profile. Hyperspectral imagery, consisting of more than 300 bands, is co-registered with the LiDAR-derived crowns to extract spectral signatures and compute vegetation indices. These data support the development of a spectral library for the main species in the study area.

The multi-temporal dataset allows evaluation of how phenological changes influence separability among species. Early leaf loss in autumn and differences in budburst timing in spring are expected to produce temporary structural and spectral contrasts that aid classification. Machine learning models, including random forest and neural networks, are applied to assess the contribution of structural, spectral, and seasonal features to species discrimination.

Accurate 3D reconstruction and instance segmentation of soybean seedlings are crucial for early phenotyping and precision agriculture. This study presents a UAV-based sparse-view 3D reconstruction and plant-level segmentation framework that integrates 3D Gaussian Splatting (3DGS) with Mobile-SAM, enabling efficient and high-fidelity modeling under routine field conditions. Traditional LiDAR and MVS approaches, while detailed, are constrained by cost, acquisition density, and computational complexity. By contrast, 3DGS offers explicit Gaussian primitives for fast rendering and direct geometric access but often fails under sparse-view UAV imagery due to weak multi-view constraints and repetitive canopy structures.

To overcome these limitations, the proposed method introduces a mask–geometry co-optimization mechanism: YOLO-generated bounding-box prompts guide Mobile-SAM to produce accurate single-view plant masks, which serve as semantic priors to associate 2D observations with 3D Gaussian primitives. Iterative refinement aligns rendered and observed masks, ensuring spatial consistency and coherent 3D plant boundaries.

Field experiments on a soybean plot demonstrated the method’s effectiveness, achieving high reconstruction quality and visually precise seedling segmentation. The resulting 3D models capture fine structural details and distinct plant instances even under sparse-view UAV data.

This work highlights the potential of combining explicit geometric modeling and lightweight semantic segmentation to achieve robust, scalable, and field-deployable 3D crop reconstruction, offering a promising pathway for high-throughput plant phenotyping and yield estimation in real-world agricultural applications.

In this study, we propose an upscaling approach based on 8-band PlanetScope SuperDove imagery (Coastal Blue, Blue, Green I, Green, Yellow, Red, Red Edge, NIR) combined with UAV data. We employed an evidential Dirichlet neural network to estimate the fractional cover of 13 herbaceous and shrub species typical of Mediterranean coastal dunes, previously mapped at 3 cm using a traditional Random Forest classifier trained on UAV multispectral samples. The overall goal is to enable large-scale mapping of coastal vegetation using high-resolution satellite imagery.