Dataset DOI: 10.5061/dryad.n8pk0p39w

Description of the data and file structure

This dataset contains 1,643 validated multi-linestring geometries representing surfaced upland hill tracks across Great Britain, manually digitised from aerial imagery. The data constitute the ground truth reference dataset used for training and validation of a convolutional neural network (CNN) segmentation model developed to automate the detection of surfaced hill tracks in UK upland moorland. Track lines were originally digitised by Nottingham Trent University (2020) and subsequently reviewed and edited by RSPB. The dataset does not represent model outputs; predicted track outputs are described in the associated publication.

Files and variables

File: ValidatedSurfacedTracks_EPSG27700_20260603.gpkg

Description: The dataset is stored as a GeoPackage in British National Grid (EPSG:27700). Each feature includes the year of aerial imagery used for digitising (Year_1), OS BNG 10 km square reference (BNG_square), track type classification (track_type), access track category (access_tra), country (country), track length in kilometres (Length_km), and data provenance (Created). An RSPB editorial review flag (RSPB_edi) indicates features that have undergone additional quality checking. The dataset supports analyses of track overlap with environmentally sensitive features including deep peat soils and statutory protected areas, and investigation of track density variation across dominant upland land uses and management types across Scotland, England, and Wales.

Access information

Data was derived from the following sources:

• Clutterbuck, B., Burton, W., Smith, C., & Yarnell, R. W. (2020). Vehicular tracks and the influence of land use and habitat protection in the British uplands. Science of the Total Environment, 737, 140243.

File: Joint_model_fit.R

Description: Joint_model_fit.R contains the R code used to fit a Bayesian joint model examining the association between upland land use/management and predicted hill track density across moorland in the UK. The script implements a hurdle-gamma outcome model (predicting track density in km per km² of moorland) and a categorical treatment model (predicting land management class), fitted jointly in a single brms multivariate model to account for potential confounding by land management assignment.

Input data

UK_AOI_output_polygons_with_length_and_Density_and_covars_2025.gpkg

File: UKbatch1_.ipynb

Description: Jupyter Notebook implementing the inference pipeline for predicting surfaced upland hill tracks across Great Britain using the trained U-Net model. Requires the pre-trained weights produced by UnetSegmentation_surfaced_tracks_training_pipeline.ipynb. Full methodological detail is provided in the paper. Loads pre-trained U-Net weights and runs batch prediction on UK imagery tiles, producing georeferenced softmax probability rasters (2-band GeoTIFF). Tiles are then grouped by source image, mosaicked using max-value overlap, and thresholded to binary track/no-track rasters.

File:UnetSegmentation_surfaced_tracks_training_pipeline.ipynb

Description: Jupyter Notebook containing the full U-Net training pipeline for binary semantic segmentation of surfaced upland hill tracks from RGB imagery. Full methodological detail is provided in the paper.

Defines and trains a U-Net convolutional neural network, including: dataset loading and augmentation (segDataset); train/test split (90/10, seed 101); U-Net architecture (3-channel input, 2-class output, bilinear upsampling); mean IoU loss function; Adam optimiser (lr = 1e-3) with StepLR decay; and model checkpointing. The trained weights are saved to saved_models/unet_batch32_epoch_17_0.17201.pt, which is required as input to UKbatch1_.ipynb.

Python dependencies

torch, torchvision, numpy, opencv-python (cv2), albumentations, Pillow, scipy, scikit-learn, matplotlib

File: UK_AOI_output_polygons_with_length_and_Density_and_covars_20250620RevG.gpkg

Description:GeoPackage of 1 km² polygon grid cells covering upland moorland areas of Great Britain, used as the unit of analysis in the joint Bayesian model examining the association between land use/management and predicted surfaced upland hill track density. Each polygon contains predicted track density, land management class, and model covariates. Land management classes were assigned to grid cells based on criteria detailed in the paper. Full methodological detail is provided in the paper.

File: dataset6.zip

Description: Training dataset for the U-Net semantic segmentation model, consisting of 6,036 paired image and binary mask tiles covering a GB-wide sample of upland moorland. Images are RGB tiles derived from Getmapping PLC aerial photography (25 cm pixel resolution). Binary masks were generated by rasterising the validated surfaced track polygon dataset (ValidatedSurfacedTracks_EPSG27700_20260603.gpkg, deposited separately on Dryad). Full methodological detail is provided in the paper.
Structure
dataset6/
├── images/ # 6,036 RGB image tiles (.tif)
└── masks/ # 6,036 corresponding binary mask tiles (.tif)
This dataset is the input to UnetSegmentation_surfaced_tracks_full_pipeline.ipynb. The root variable in that notebook should be set to the path of the dataset6/ directory after download.