Supplementary: Mapping Trails and Tracks in the Boreal Forest using LiDAR and Convolutional Neural Networks

Dataset Background

This dataset consists of various geospatial layers and raster files used in the study of mapping trails and tracks in the boreal forest using LiDAR data and convolutional neural networks (CNNs). The layers include manually labeled training data, test polygons for accuracy assessment, trail density map, example of digital terrain model, original pixel-wise trail and track maps, and refined vectorized trail and track maps generated from both airborne and drone-based LiDAR data. Additionally, the dataset contains ancillary data such as industrial disturbance footprints, land cover maps, and random points for model performance evaluation.

These data were collected and processed to develop and validate models capable of automatically mapping trails and tracks across diverse land-cover types. This dataset supports remote sensing and ecological research by providing detailed spatial data for automated trail and track mapping, enhancing understanding of human and wildlife movement patterns in the boreal forest.

The dataset accompanies a research paper titled Mapping Trails and Tracks in the Boreal Forest using LiDAR and Convolutional Neural Networks, authored by Greg McDermid, Irina Terenteva, and Xue Yan Chan, which will be submitted to PeerJ. We are part of the Applied Geospatial Research Group (University of Calgary, Canada) and the Boreal Ecosystem Recovery and Assessment (BERA) project, a multi-sectoral research partnership involving academic institutions, private-sector companies, public-sector divisions, and a not-for-profit organization. The BERA project aims to understand the effects of industrial disturbances on natural ecosystem dynamics in Alberta's boreal forest and develop restoration strategies.

Our GitHub page, featuring tools and code for the trails and tracks mapping project, can be found here:

https://github.com/appliedgrg/trails-tracks-mapper

Dataset Contents

• VisualDelineation_for_1stCNN.gpkg

This dataset contains manually labeled training data for an initial U-Net model. Visual interpretation of aerial LiDAR data was used to map prominent trails and tracks, primarily on seismic lines, which were then buffered to create pixel-wise labels for the first CNN training.

• AApolygons_VisualDelineation_trail_attributes.gpkg

This dataset includes 11 test polygons, each 50x50 meters, used for accuracy assessment. Each polygon contains visually delineated trails and tracks with information on their length, area (assuming a width of 30 cm), and percent cover.

• AApolygons_VisualDelineation_strong_trails.gpkg

This dataset catalogs "strong" trails and tracks within the 11 test polygons, identified through visual interpretation of ultra-high resolution (0.5 cm) drone orthomosaics (see visual interpretation key).

• AApolygons_VisualDelineation_weak_trails.gpkg

This dataset catalogs "weak" trails and tracks within the 11 test polygons, identified through visual interpretation of ultra-high resolution (0.5 cm) drone orthomosaics (see visual interpretation key).

• 2ndCNN512_aerialDTM50cm_AA.tif

This dataset contains CNN-predicted trails and tracks within the 11 test polygons using a 50-cm airborne LiDAR data provided by Alberta-Pacific Forest Industries, Inc.

• 2ndCNN512_droneDTM10cm_AA.tif

This dataset contains CNN-predicted trails and tracks within the 11 test polygons using a 10-cm drone LiDAR data collected with a Zenmuse L1 sensor.

• Kirby_FLAIM_industrial.gpkg

This dataset includes visually delineated large industrial disturbances within the study area, such as roads, pipelines, and powerlines. These areas were masked out and not included in final calculations of trail lengths and areas.

• 2ndCNN_Trails10cm_postprocessed.gpkg

This dataset contains a refined map of trails and tracks across the study area, generated using a CNN model. The initial output was a pixel-wise probability map indicating the likelihood of each pixel being part of a trail or track. To ensure a conservative estimate, a score threshold of 40 was applied. The resulting raster map was then transformed into centerlines, producing a vectorized representation of the trails and tracks.

• drone_lidarDTM_microtopography10cm_KirbySouth.tif

This dataset represents a microtopography map of the test area with a 10-cm resolution drone-based LiDAR data.

• 2ndCNN512_5m_density.tif

This dataset is a density map of CNN-predicted trails and tracks, calculated for a 5-meter grid by summing trail and track areas within the grid. High values indicate areas with abundant trails and tracks, while lower values indicate no-trail areas.

• keras_unet.zip

This file contains the implementation of the Keras U-Net used to train the CNN models. The implementation details can also be found at this GitHub repository.

• 2ndCNN_droneDTM10cm_KirbySouth_scorethreshold10.tif

This dataset represents a map of trails and tracks produced by applying a CNN model to drone-based 10-cm LiDAR data within the test area. The output is unprocessed pixel-wise probability map indicating the likelihood of each pixel being part of a trail or track.

• Kirby_FLAIM_seismic_line_footprint.gpkg

This dataset includes polygons of seismic line footprints developed using the AI-augmented Forest Line Mapper tool, available at this GitHub repository.

• AApolygons_randompoints.gpkg

This dataset contains random points used to assess model performance within the test area.

• Landcover_BigKirby.tif

This dataset is a land cover map for the study area, created in Google Earth Engine using a combination of satellite imagery and machine-learning classification techniques.

• Landcover_BigKirby.qml

This file provides the layer style for the land cover dataset.

• Training_random_points_for_2ndCNN.gpkg

This dataset contains random points used to create training patches for the CNN.

• CNN_AA_drone_vs_aerial_DTM.csv

This spreadsheet contains raw results of accuracy assessment for drone and aerial DTM datasets.

• AA_testpoints_aerial_DTM10cm_scorethre15.gpkg

This dataset contains accuracy assessment results for the CNN-predicted trail and track map using drone-based LiDAR data with a 10 cm resolution and a score threshold of 15. The file includes classifications of true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN).

• AA_testpoints_drone_DTM10cm_scorethre15.gpkg

This dataset contains accuracy assessment results for the CNN-predicted trail and track map using airborne LiDAR data with a 10 cm resolution and a score threshold of 15. The file includes classifications of true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN).

Acknowledgement

This research is part of the Boreal Ecosystem Recovery and Assessment (BERA) project (www.bera-project.org), and was supported by a Natural Sciences and Engineering Research Council of Canada Alliance Grant (ALLRP 548285-19) in conjunction with Alberta-Pacific Forest Industries, Alberta Biodiversity Monitoring Institute, Alberta Environment and Protected Areas, Canadian Natural Resources Ltd, Cenovus Energy, ConocoPhillips Canada, Imperial Oil Ltd, and Natural Resources Canada.