Work

Unified knowledge transfer boosts individual tree crown segmentation without scene-specific labels

Year: 2026

Type: article

Abstract: Accurate delineation of individual tree crowns is essential for forest inventory, ecosystem monitoring, and climate-informed management. However, most existing deep learning approaches rely on scene-specific annotations that are costly to produce and difficult to generalize across heterogeneous regions and cross-resolution scenarios, particularly when transferring from labeled high-resolution source imagery to unlabeled low-resolution target imagery. To tackle these challenges, we propose UniTree, a unified framework that enables cross-region and cross-resolution (from high-resolution source imagery to unlabeled low-resolution target imagery) individual tree crown segmentation without scene-specific annotations. UniTree integrates three key components: (1) domain knowledge transfer to mitigate geographic heterogeneity through adversarial feature alignment; (2) resolution knowledge transfer, in which a high-resolution teacher distills semantic-, geometric-, and feature-level knowledge to a low-resolution student; and (3) a saliency-to-structure training paradigm that guides learning from salient to structural crown features. Evaluations across multiple test areas in Yosemite National Park, California, demonstrate that UniTree outperforms state-of-the-art (SOTA) methods (DeepLabv3, U-Net, attention-enhanced U-Net, DeepForest, and Detectree2), achieving gains of 1–26 percentage points in mean Intersection over Union and gains of 2–60 percentage points in F1 score. Moreover, analysis across crown-size bins and spectral band combinations reveals that increasing network depth or adding more bands does not necessarily improve segmentation performance. By eliminating the need for scene-specific annotations while maintaining high accuracy and generalization, UniTree offers a scalable solution for automated tree crown delineation, enabling large-scale digital forest inventories and ecological assessments across diverse landscapes.

Source: Remote Sensing of Environment

DOI: 10.1016/j.rse.2026.115646

Authors: Jiaqi Yang, Kyle Kabasares, Ming Liu, Taejin Park, Min Chen

Institutions: University of Wisconsin–Madison, Ames Research Center, Bay Area Environmental Research Institute

Language: English

Topic: Remote Sensing and LiDAR Applications

Subfield: Environmental Engineering

Field: Environmental Science

Domain: Physical Sciences

SDG: Life on land

Open Access status: hybrid

Estimated APC paid: $4,450