[Data & Code] A drone imagery dataset for semantic segmentation of urban garden ground covers in biodiversity studies

Yasamin Afrasiabian, Chenghao Lu, Dr. Anirudh Belwalkar, Hany El Sharawy, Xiaoxin Song, Ying Yuan, Fei Wu, Xiang Su, Elisa Van Cleemput, Monika Egerer, Prof. Dr. Kang Yu

Last updated on Jun 17, 2026 Data

Visual comparison of ground-cover semantic segmentation on test-set orthomosaics: RGB tile, eight-class label mask, and the DeepLabV3+ prediction.

This dataset accompanies the Data Descriptor published in Scientific Data by Afrasiabian et al. (2026), providing a high-resolution benchmark for mapping the fine-scale ground-cover heterogeneity of urban community gardens.

Urban gardens support city biodiversity through diverse ground covers that provide habitat, pollination, pest control, and soil functions — but their spatial heterogeneity has been poorly mapped due to a lack of high-resolution imagery. This dataset addresses that gap.

What’s in the dataset

24 RGB orthomosaics (GeoTIFF) and their 24 pixel-level label masks, collected in 2021–2022 at five community gardens in Munich, Germany
Organised into Train / Val / Test splits (14 / 5 / 5 orthomosaics and masks)
Spatial resolution 3.2–7.9 mm per pixel (18.9–146.4 megapixels; EPSG:25832)
Eight ground-cover classes: grass, herb, litter, soil, stone, straw, wood, and woodchip
A patch-based version (cropped image/mask patches) ready for deep-learning pipelines
A metadata CSV documenting per-image dimensions, resolution, CRS, and Agisoft processing/flight properties

Benchmarks

The descriptor benchmarks both deep-learning (UNet, DeepLabV3+) and traditional machine-learning (Random Forest, XGBoost, Maximum Likelihood) classifiers. DeepLabV3+ achieved the best performance, with an overall accuracy of ≈93.2% and an Intersection-over-Union of 69.4.

The dataset is intended to support research on urban biodiversity, habitat modelling, garden management, and remote sensing, and can be combined with other fine-scale datasets to advance sustainable urban green planning.

📦 Dataset (Zenodo): https://doi.org/10.5281/zenodo.18757882

💻 Code (GitHub): https://github.com/paglab/ugc-mapping

📄 Citation:

Afrasiabian, Y., Lu, C., Belwalkar, A., Elsharawy, H., Song, X., Yuan, Y., Wu, F., Su, X., Van Cleemput, E., Egerer, M., & Yu, K. (2026). A drone imagery dataset for semantic segmentation of urban garden ground covers in biodiversity studies. Scientific Data, 13, 590. https://doi.org/10.1038/s41597-026-07152-z

Drone RGB Imagery Orthomosaics Semantic Segmentation Urban Biodiversity Ground Cover Deep Learning Urban Gardens

Chenghao Lu

Research assistant & Joint Phd student

My research interests are Agricultural Digital Twin and Smart Breeding. My project focuses on the construction of a Digital Twin system for maize, aiming to fuse Genetic x Environmental x Management data to construct a high-precision and dynamically updated maize growth virtual twin. By introducing a DNA language model with a multimodal fusion approach, I hope to achieve accurate simulation of maize phenotypes and breeding to support sustainable agricultural development.