What question did this study set out to answer?

The aim is to develop a deep learning workflow to map urban tree canopy using accessible NAIP imagery.

June 11, 2026Open Access

An Open and Transferable Deep Learning Framework for Mapping Urban Tree Canopy Using NAIP Imagery

JYJooyoung YooEmory University YQYi QiUniversity of Southern California IAIsaac Ashe-McNalleyUniversity of Southern California

Key Points

The aim is to develop a deep learning workflow to map urban tree canopy using accessible NAIP imagery.
Developed U-Net model for canopy segmentation and YOLOv9e for individual tree detection.
Conducted experiments across two neighborhoods, annotating 17,466 trees in a structured workflow.
Assessed model performance and requirements for data sufficiency in training and validation.
U-Net achieved a Dice coefficient of 0.824 for canopy segmentation.
YOLOv9e attained an F1-score of 0.687 for individual tree detection.
Model performance stabilizes with around 130 annotated trees per 320x320 pixel tile.

Abstract

The urban tree canopy is an important resource that spans public and private property and whose form and quantity change over short distances. Although remote sensing and deep learning approaches have been used to map urban tree canopy, the high cost of commercial imagery and the technical complexity of model development have limited their adoption by urban forestry practitioners. We developed a structured and reproducible deep learning workflow optimized for freely available USDA National Agriculture Imagery Program (NAIP) imagery. The workflow incorporates a reproducible U-Net segmentation model for canopy delineation and a YOLOv9e object detection model for individual tree identification, enabling complementary estimation of the canopy extent and individual tree locations. Across two neighborhoods in Los Angeles, the optimized U-Net achieved a Dice coefficient of 0.824 for canopy segmentation, while YOLOv9e reached an F1-score of 0.687 for individual tree detection on a held-out test set with 17,466 annotated trees. A data sufficiency experiment showed that model performance stabilizes when approximately 130 trees are annotated per 320 × 320 pixel (px) tile, corresponding to about 25,379 training and 2641 validation labels, providing a practical target for annotation effort. Additional experiments demonstrate a structured workflow for spatial sampling, training data requirements, and the use of model inferences to estimate tree canopy extent and individual tree locations. The workflow also shows encouraging evidence of transferability to previously unseen urban areas without retraining. By relying solely on NAIP-optimized approaches, this new workflow bridges the gap between complex deep learning techniques and the practical needs of urban foresters; empowers local stakeholders to create accurate, affordable, and timely urban tree inventories; and fosters data-driven decision-making for the sustainable management of urban green infrastructure.

Ask AI

Helpful

Bookmark

View Full Paper