Day 4

Detailed paper information

Back to list

Paper title Large scale mapping of linear disturbances in forest areas using deep learning and Sentinel-2 data across boreal caribou herd ranges in Alberta, Canada
  1. Ignacio San Miguel Hatfield Consultants Speaker
  2. Olivier Tsui Hatfield Consultants
  3. Jason Duffe Environment and Climate Change Canada
  4. Andy Dean Hatfield Consultants
Form of presentation Poster
  • C1. AI and Data Analytics
    • C1.04 AI4EO applications for Land and Water
Abstract text Large scale mapping of linear disturbances in forest areas using deep learning and Sentinel-2 data across boreal caribou herd ranges in Alberta, Canada

Ignacio San-Miguel1, Olivier Tsui1, Jason Duffe2, Andy Dean1

1 Hatfield Consultants Partnership, 200 – 850 Harbourside Drive, North Vancouver, BC, V7P 0A3, Canada
2Landscape Science and Technology Division, Environment and Climate Change Canada - 1125 Colonel By Drive, Ottawa, ON, K1A 0H3, Canada

In the Canadian boreal forest region habitat fragmentation due to linear disturbances (roads, seismic exploration, pipelines, and energy transmission corridors) is a leading cause for the decline of woodland caribou (Rangifer tarandus) – boreal population; and as a result, a deep understanding of linear disturbances (amount, spatial distribution, dynamics) has become a research and forest management priority in Canada.

Canada imposed regulatory restrictions on the density of forest habitat disturbance in woodland caribou ranges, given the species’ protection under the Species at Risk Act (SARA). To support current regulations, government agencies currently rely on manual digitization of linear disturbances using satellite imagery across very large areas. Examples of these datasets include the Anthropogenic Disturbance Footprint Canada dataset (ADFC) (Pasher et al., 2013) which was derived using visual interpretation of Landsat data to map linear disturbances across more than 51 priority herds covering millions of ha for years 2008-2010 at 30 m and for 2015 at both 30 and 15 m (using the panchromatic band); and the Human Footprint (HF) dataset (ABMI, 2017), a vector polygon layer that captures linear disturbances across a grid of 1,656 3 by 7 km sample sites (~3.5Mha) distributed across the province of Alberta and collected from 1999 to 2017. Such efforts are laudable, yet time consuming and expensive across large areas resulting in incomplete and infrequent coverage. The need for cost-effective methods to map linear disturbances in forest settings is ubiquitous.

Automated methods using machine learning are a desired alternative to enable frequent and consistent mapping of linear disturbances across large areas at a reduced cost. Recent advancements in deep learning (DL) algorithms and cloud computing represent an opportunity to bridge the gap in accuracy between methods using visual interpretation and automated methods relying on machine learning. DL algorithms explicitly account for the spatial context (in case of 2D and 3D convolutional neural networks) and can assemble more complex patterns using local and simpler patterns, which makes them particularly suitable for geometric challenges where the contextual information is relevant, like in linear disturbance detection.

Automatic extraction of roads from satellite imagery using DL is gaining increasing attention, however, to this date, most of the existing methods for the detection of linear features using remote sensing data and DL focus on urban paved roads with no methods focused on linear disturbances in forest areas (e.g., seismic lines, logging roads, pipeline corridors). Linear disturbance extraction in forest areas poses unique challenges compared to the mapping of urban paved roads, which preclude the application of current methods without adaptation. There are several unique challenges, first, the current technology was developed using very high-resolution (VHR) imagery and not high-resolution (HR) imagery like Sentinel-2. Second, linear disturbances in forest areas have very diverse types, each with its particularities, and generally the features are narrower and more irregular than paved roads. Third, linear disturbances in forested areas have different road surface conditions and surrounding vegetation cover, while those in urban settings are more homogenous.

The objective of this research is to develop and evaluate the accuracy of an automated algorithm to extract linear disturbances in forest areas across boreal caribou herd ranges in Alberta, Canada, using DL and 10m spatial resolution Sentinel-2 data. Specifically, this study explores the capacity of various Unet-inspired architectures (Unet, Resnet, Inception, Xception) coupled with transfer learning to perform pixel-level binary classification of linear disturbances.

The HF vector data set was used as training data covering 3.5Mha across Alberta for the year 2017. HF was derived using visual interpretation on SPOT-7 and ortho-imagery, thus capturing some details than are not discernible in the 10m Sentinel-2 data, which introduces some error in the training data.

DL model results are promising, with Intersection over Union (IoU) accuracies ranging from moderate-low to fair (0.3-0.5) for various types of unpaved roads and pipelines, with the finer-scale seismic lines largely undetected (IoU of 0.1). The best performing model used transfer learning using as encoder a Inceptionresnetv2 architecture with weights pre-trained on the Imagenet dataset. The main challenges identified in the accurate prediction of linear disturbances include variability in land cover conditions, occlusion and shadows caused by forested vegetation on adjacent roads, and width of the target linear disturbances, where features < 10m width go largely undetected using Sentinel-2. We discuss the trade-offs challenges and options related to evaluating model accuracy using multiple metrics and DL architectures.

This research demonstrates the potential of a cost-effective method using DL architectures coupled with Sentinel-2 data to maintain current and accurate maps of linear disturbances in highly dynamic forest areas to support caribou conservation efforts. Building upon the standardized methods proposed here, very large areas could be mapped frequently to, potentially, create a comprehensive national linear disturbance database to support decision-making for caribou habitat conservation.

Keywords— deep learning, linear disturbances, Sentinel-2, Unet, Caribou

ABMI Human Footprint Inventory: Wall-to-Wall Human Footprint Inventory. 2017. Edmonton, AB: Alberta Biodiversity Monitoring Institute and Alberta Human Footprint Monitoring Program, May 2019.

Ministry of Forests Lands and Natural Resource, 2020. Digital Road Atlas - Province of British Columbia [WWW Document]. URL (accessed 3.10.20).

Pasher, J., Seed, E., Duffe, J., 2013. Development of boreal ecosystem anthropogenic disturbance layers for Canada based on 2008 to 2010 Landsat imagery. Canadian Journal of Remote Sensing 39, 42–58.

Ronneberger, O., Fischer, P., Brox, T., 2015. U-Net: Convolutional Networks for Biomedical Image Segmentation. arXiv:1505.04597 [cs].

Zhang, Z., Liu, Q., Wang, Y., 2018. Road Extraction by Deep Residual U-Net. IEEE Geosci. Remote Sensing Lett. 15, 749–753.