Tree crown polygons with tree height statistics for Freiburg im Breisgau (Germany)

Description This dataset contains individually segmented tree crown polygons and associated tree heights for the city of Freiburg im Breisgau, Germany. The dataset was generated from summer 2021 RGBI orthophotos (0.20 m spatial resolution; flight date: 1 June 2021; [1]) using the DeepTrees tree crown segmentation framework [2]. Tree height statistics were subsequently derived from a 2021 vegetation-only normalised DSM (1 m spatial resolution; [3]). Details of the processing workflow are provided in the Processing and methodology section below. The following tree height statistics are provided for each tree crown: count – number of nDSM pixels within the tree crown min – minimum tree height max – maximum tree height median – median tree height mean – mean tree height std – standard deviation of tree height Because the nDSM represents vegetation height above ground, the reported height statistics represent the height of vegetation within each detected tree crown. Processing and methodology Tree crown segmentation Tree crowns were segmented from the Freiburg summer 2021 RGBI orthophotos (0.20 m spatial resolution) using DeepTrees v1.6.0 [2]. The source orthophotos were split into 1 km × 1 km tiles prior to processing, with no additional image preprocessing. The initial model was the pre-trained DeepTrees model developed for Halle Unet-resnet18_epochs=209_lr=0.0001_width=224_bs=32_divby=255_custom_color_augs_k=3_jitted.pt provided with the DeepTrees repository [4]. To improve performance for Freiburg, the model was fine-tuned using approximately 30 manually delineated tree crowns located within Freiburg city centre. Tree crowns were digitised manually in QGIS and stored as separate training and validation polygon shapefiles together with their corresponding orthophoto tiles. Training data preparation was performed using the DeepTrees function tcdm.prepare_data() which generated the required segmentation masks, outlines, and distance transforms. Because only two labelled image tiles were available, a 50:50 training/validation split was used. The DeepTrees configuration file used for fine-tuning is included with this dataset as finetune_fbg.yaml. The resulting fine-tuned model was subsequently used to segment all orthophoto tiles covering the study area. Polygon extraction Polygon extraction parameters were iteratively adjusted to maximise mapped tree cover while minimising omission errors during visual inspection. The configuration file is included with this dataset as inference_on_individual_tiles_fbg_v4.yaml. The final extraction parameters were: polygon_extraction: min_dist: 10 mask_exp: 2 outline_multiplier: 5 outline_exp: 1 dist_exp: 0.5 area_min: 3 sigma: 1.5 label_threshold: 0.3 binary_threshold: 0.07 simplify: 0.3 Following segmentation, all detected tree crown polygons were retained. No manual editing or removal of individual polygons was performed after automatic extraction. id – identifier of the detected tree crown within its respective 1 km × 1 km orthophoto tile. Values are unique within each tile but may be repeated between tiles. The dataset index is globally unique. Tree height extraction Tree height attributes were derived from the 2021 vegetation-only normalised Digital Surface Model (nDSM) (1 m spatial resolution) provided by the Landesamt für Geoinformation und Landentwicklung Baden-Württemberg (LGL-BW). The nDSM represents vegetation height above ground level. For each tree crown polygon, every nDSM pixel whose centre point fell within the polygon was assigned to that tree and tree height statistics were calculated, including the number of pixels (count), minimum (min), maximum (max), mean (mean), median (median), and standard deviation (std) of tree height using the zonal_stats() function from the Python package rasterstats [5]. Tree crowns without intersecting nDSM pixels retain missing height attributes. Of the 1,355,118 detected tree crowns, 840,765 (62%) do not contain tree height attributes, primarily because forested areas included in the orthophotos lie outside the spatial extent of the available vegetation-only nDSM. Within the urban area, however, tree crowns with missing height attributes are the minority. Quality control Model fine-tuning and polygon extraction parameters were refined through repeated visual comparison of segmentation results against the source orthophotos. No additional manual correction, editing, or quality filtering of the final tree crown polygons was performed after automated processing. Dataset summary Number of detected tree crowns: 1,355,118 Coordinate reference system: ETRS89 / UTM zone 32N (EPSG:25832) Input orthophotos: RGBI, 0.20 m spatial resolution (2021) Input nDSM: Vegetation-only normalised DSM, 1 m spatial resolution (2021) DeepTrees version: v1.6.0 Output formats: ESRI Shapefile (.shp) and GeoParquet (.parquet) Limitations The following limitations should be considered when using this dataset: Systematic underestimation of tree detection:Despite parameter adjustments, DeepTrees systematically underestimates tree cover in parts of the study area, particularly for trees with lighter reflectance. Fine-tuning the model using a substantially larger and more representative training dataset would likely improve segmentation performance. Orthophoto tile edge effects:Because the orthophotos were processed as individual tiles, trees intersecting tile boundaries can be incorrectly segmented, including cases where trees appear artificially elongated along the tile edge. Processing buffered image tiles would reduce these edge effects. Mismatch between tree crowns and nDSM pixels:The vegetation-only nDSM does not necessarily correspond exactly to the detected tree crown polygons. Consequently, some tree crowns have few or no nDSM pixels assigned to them and therefore have missing or incomplete tree height statistics. Recommended citation Winkler, L. (2026). Tree crown polygons with tree height statistics for Freiburg im Breisgau (Germany). Zenodo. https://doi.org/10.5281/zenodo.21674889 References [1] Stadt Freiburg im Breisgau. RGBI Orthophotos, Freiburg im Breisgau, 2021. Spatial resolution: 0.20 m. Flight date: 1 June 2021. Available from the Freiburg Geodata Catalogue:https://geodaten.freiburg.de/geonetwork/srv/eng/catalog.search#/metadata/69dc1e2f-ef64-4633-a16b-f3c35e27fdc3 [2] Khan, T., Arnold, C., & Grover, H. (2025). DeepTrees: Tree Crown Segmentation and Analysis in Remote Sensing Imagery with PyTorch. Journal of Open Source Software, 10(114), 8056. https://doi.org/10.21105/joss.08056 [3] Landesamt für Geoinformation und Landentwicklung Baden-Württemberg (LGL-BW). (2021). Normalized Digital Surface Model (nDOM). Available from the LGL-BW 3D products catalogue:https://www.lgl-bw.de/Produkte/3D-Produkte/Digitale-Oberflaechenmodelle/nDOM/[4] Khan, T. (2024) DeepTrees Git repository. https://codebase.helmholtz.cloud/taimur.khan/DeepTrees.git [5] Perry, M. (2015). rasterstats: Raster zonal statistics for Python. https://pythonhosted.org/rasterstats/