TreePseCo: Scaling Individual Tree Crown Segmentation using Large Vision Models

Abstract. Forest monitoring through individual tree crown delineation is essential for sustainable management and carbon cycle assessment. This paper presents TreePseCo, an adaptation of the PseCo framework leveraging foundation models for automated tree crown segmentation in aerial imagery. Our approach implements a three-stage pipeline: (1) tree center detection using a modified Segment Anything Model (SAM) decoder that generates probability heatmaps, (2) instance mask generation through prompt-guided segmentation utilizing SAM's visual features, and (3) boundary refinement via specialized classification to eliminate false positives. We validate our method on two datasets: the extensive NEON dataset covering diverse U.S. forest ecosystems and the Valle d’Aosta dataset (VdA), a custom set of high-resolution RGB aerial images from northwestern Italian forests. Experiments against the popular DeepForest demonstrate that, while the baseline maintains excellent performance on its native NEON dataset, TreePseCo exhibits superior generalization capabilities when applied to new geographical contexts, achieving higher mAP scores on the VdA dataset. Our approach shows strength in detecting trees in densely clustered formations and identifying smaller tree instances, areas where existing methods often struggle. Overall, results suggest TreePseCo provides a robust foundation for comprehensive forest inventory applications across diverse environments.