Tree Crown Segmentation | LINKS

Introduction

Forest monitoring through precise crown delineation is a fundamental component for sustainable management, carbon cycle assessment, and biodiversity monitoring. However, traditional methods often struggle in dense, heterogeneous forest environments where canopy structures are complex and overlapping.

TreePseCo is our advanced AI solution designed to address these challenges through automated segmentation of individual tree crowns in aerial imagery. The system adapts the PseCo framework by leveraging the powerful feature extraction capabilities of the Segment Anything Model (SAM). By combining point-based detection with prompt-guided segmentation, the model demonstrates superior generalization capabilities, particularly when applied to new geographical contexts and challenging canopy environments compared to existing baselines.

Supporting image

(Lungo Vaschetti et al., 2025)

Key Features

Foundation Model Integration: Leverages the Segment Anything Model (SAM) to generate robust visual features, enabling high adaptability to diverse forest types.
Three-Stage Pipeline: Implements a distinct workflow consisting of tree center detection via heatmaps, instance mask generation through prompts, and boundary refinement to eliminate false positives.
High Generalization: Demonstrates superior performance in new geographical contexts, without requiring extensive retraining.
Dense Canopy Detection: Specifically optimized to detect trees in densely clustered formations and identify smaller tree instances where other methods typically fail.

Technologies Used

Computer Vision: Incorporates a ResNet50-FPN backbone within a modified Faster R-CNN architecture for the final classification and bounding box refinement stage.
Foundation Models: Utilizes a modified Segment Anything Model (SAM) (ViT-H encoder) for feature extraction and mask generation.
Remote Sensing: Optimized for High-Resolution RGB aerial imagery.

Use Cases

Forest Inventory & Monitoring: Enables accurate counting and delineation of individual trees to support large-scale forest mapping and inventory updates.
Biomass Estimation: Provides precise crown dimensions which correlate strongly with above-ground biomass, aiding in carbon stock quantification.
Ecosystem Management: Facilitates species classification, forest health assessment, and growth monitoring in diverse environments, from temperate deciduous forests to complex Alpine terrain.

Live Demo

Explore the practical results of our AI solution through this interactive demo. The application displays precomputed inference results on high-resolution aerial imagery capturing Alpine woods in the Valle d’Aosta region.

Try it out

Try out the solution in real time on Hugging Face Spaces:

👉 Launch Demo

If you cannot access the demo, please contact us. Due to privacy considerations, certain images cannot be displayed publicly.

Integration

The TreePseCo model weights, including the point decoder and classification head, are available for download at the github repo to facilitate integration into custom forestry pipelines. Users can finetune the model on their own datasets following the COCO format to adapt the solution to specific local forest characteristics.

Publications

Lungo Vaschetti, J., Arnaudo, E., & Rossi, C. (2025). TreePseCo: Scaling Individual Tree Crown Segmentation using Large Vision Models. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLVIII-M-7-2025, 275–282.