[Paper Review] OAM-TCD: A globally diverse dataset of high-resolution tree cover maps

Josh Veitch-Michaelis, Andrew Cottam+6 more|arXiv (Cornell University)|Jun 1, 2024

Remote Sensing in Agriculture5 citations

TL;DR

The paper introduces OAM-TCD, a large open-access high-resolution tree crown delineation dataset with over 280k individual trees and 56k groups, along with baseline models and an open training/inference pipeline.

ABSTRACT

Accurately quantifying tree cover is an important metric for ecosystem monitoring and for assessing progress in restored sites. Recent works have shown that deep learning-based segmentation algorithms are capable of accurately mapping trees at country and continental scales using high-resolution aerial and satellite imagery. Mapping at high (ideally sub-meter) resolution is necessary to identify individual trees, however there are few open-access datasets containing instance level annotations and those that exist are small or not geographically diverse. We present a novel open-access dataset for individual tree crown delineation (TCD) in high-resolution aerial imagery sourced from OpenAerialMap (OAM). Our dataset, OAM-TCD, comprises 5072 2048x2048 px images at 10 cm/px resolution with associated human-labeled instance masks for over 280k individual and 56k groups of trees. By sampling imagery from around the world, we are able to better capture the diversity and morphology of trees in different terrestrial biomes and in both urban and natural environments. Using our dataset, we train reference instance and semantic segmentation models that compare favorably to existing state-of-the-art models. We assess performance through k-fold cross-validation and comparison with existing datasets; additionally we demonstrate compelling results on independent aerial imagery captured over Switzerland and compare to municipal tree inventories and LIDAR-derived canopy maps in the city of Zurich. Our dataset, models and training/benchmark code are publicly released under permissive open-source licenses: Creative Commons (majority CC BY 4.0), and Apache 2.0 respectively.

Motivation & Objective

• Provide a globally diverse, high-resolution (10 cm/px) dataset for individual tree crown delineation (TCD) in RGB aerial imagery.
• Enable benchmarking of semantic and instance segmentation models for tree mapping at sub-meter scales.
• Assess model generalization across biomes and urban/natural environments using biome-stratified cross-validation.
• Release open-source tooling for labeling, training, and evaluating TCD models on large orthomosaic imagery.

Proposed method

• Assemble a global 10 cm/px RGB dataset from OpenAerialMap with 2048x2048 tiles.
• Provide semantic masks and MS-COCO style instance polygons for trees and canopy groups.
• Train baseline semantic segmentation models (UNet with ResNet backbones; SegFormer variants) and instance segmentation models (Mask-RCNN) on the dataset.
• Evaluate models with k-fold biome-stratified cross-validation and a holdout split using IoU, accuracy, and F1 metrics.
• Compare performance to existing datasets and test on independent data such as Zurich urban canopies and LIDAR-derived canopies.
• Release training/prediction pipeline under Apache 2.0; host dataset on HuggingFace and MS-COCO formats.

Experimental results

Research questions

• RQ1 Can a globally diverse, high-resolution RGB dataset enable accurate instance-level tree crown delineation at sub-meter scales?
• RQ2 How do semantic and instance segmentation models perform on this dataset across biome-stratified splits and holdout data?
• RQ3 What is the generalization of trained models to independent urban and non-urban imagery, and how do results compare to ground-truth inventories or canopy maps?

Key findings

• OAM-TCD comprises 5072 tiles of 2048x2048 px at 10 cm/px with annotations for over 280k individual trees and 56k canopy groups.
• SegFormer-based models achieve the highest IoU and F1 scores in cross-validation and holdout tests among the tested architectures.
• Instance segmentation with Mask-RCNN yields competitive mAP50 scores (about 41.8–43.2 on holdout) relative to COCO benchmarks.
• Models show strong agreement with Zurich LIDAR-derived canopy maps for urban canopies (IoU up to 0.791, accuracy 0.922, F1 0.883 for tcd-segformer-mit-b5).
• The dataset and baseline models are openly released (CC BY 4.0 / Apache 2.0 licenses), enabling zero-shot evaluation of foundational EO models on high-resolution tree mapping.