This land-use and land-cover classification dataset is derived from 2017 Sentinel-2 multi-spectral imagery for the Athabasca Oil Sands surface mineable area in northeastern Alberta. The oil sands surface mineable area, Township 89 to 103, Range 3 to 14, west of the 4th Meridian, falls within the Lower Athabasca Regional Plan (LARP) area. As part of Alberta’s Land-use Framework, LARP was developed in 2012 to set the stage for robust growth, vibrant communities, and a healthy environment within the region. One of its implementation objectives is to balance the economic development of oil sands and impacts on ecosystems and the environment. This is to be achieved through enhanced science-based monitoring to improve the characterization of landuse over time and understand cumulative effects on the environment.
This land classification dataset contains 14 classes: 1 - Water, 2 - Exposed - Barren Land, 3 - Bryoids, 4 - Shrubland, 5 - Wetland, 6 - Wetland - Treed, 7 - Herbs, 8 - Coniferous, 9 - Deciduous, 10 - Mixed Wood, 11 - Burned Areas - Little Biomass, 12 - Oil Sands Mining, 13 - Tailings Ponds, and 14 - Developed Footprints. These categories can be used as baseline data for planning, managing, and monitoring surface infrastructure needs and impacts.
The Overall accuracy is the total number of correctly classified pixels divided by the total number of pixels in the error matrix N.
The Producer accuracy is an indication of the probability of a reference pixel being correctly classified. It is calculated by dividing the number of correctly classified pixels in a class by the total number of pixels in that class.
The User accuracy is a ratio of the total number of correctly classified pixels in a class divided by the total number of pixels that were classified in that class. It indicates the probability that a pixel classified on the map represents that class on the ground.
The Kappa coefficient indicates if the accuracy level is significantly better than a random result, providing a better comparison of different classifications.
Accuracy assessment was conducted based on the High-resolution Forest Land Cover for Canada (2015), Alberta Ground Cover Classification (AGCC), and Digital Integrated Dispositions (DIDs) data. For accuracy asssment purpose, Burned Areas, Oil Sands Mining, and Tailings Ponds classes were fused with the classes they were derived from, i.e., Exposed/Barren Land, Developed, and Water, respectively.
LULC Classification result of 2017
Overall Accuracy = (2930810/3300000) 88.81%
Kappa Coefficient = 0.88
Class|Producer Accuracy (%)|User Accuracy (%)|Producer Accuracy (Pixel Ratio)|User Accuracy (Pixel Ratio)
Water|98.60|98.10|295797/300000|295797/301511
Exposed - Barren Land|99.45|98.73|298364/300000|298364/302210
Bryoids|66.61|88.05|199820/300000|199820/226927
Shrubland|96.95|87.32|290845/300000|290845/333091
Wetland|96.27|88.57|288802/300000|288802/326071
Wetland - Treed|84.21|97.69|252636/300000|252636/258605
Herbs|81.85|86.34|245552/300000|245552/284416
Coniferous|93.85|92.51|281537/300000|281537/304334
Deciduous|68.26|96.95|204779/300000|204779/211219
Mixed Wood|95.35|71.05|286058/300000|286058/402592
Developed|95.54|82.12|286620/300000|286620/349024
Process steps were performed in ENVI 5.5 software to produce the classification map:
1. A Sentinel-2 cloud-free composite dataset (10 m spatial resolution) was produced based on 2017 mid-summer imagery using the Google Earth Engine as an input to the land-use/land-cover (LULC) classification process.
2. Ground-reference datasets were created from the High-resolution Forest Land Cover for Canada (HFLCC), Alberta Ground Cover Classification (AGCC) and Digital Integrated Dispositions (DIDs) data.
3. A Maximum Likelihood classification algorithm was applied to the LBAPC data to produce LULC classifications for the Water, Exposed - Barren Land, Bryoids, Shrubland, Wetland, Wetland - Treed, Herbs, Coniferous, Deciduous, and Mixed Wood classes.
4. The Developed Footprint class was produced using the Constraint Energy Minimization and Spectral Angle Mapper partial unmixing method followed by a K-means clustering. Post-classification techniques (i.e., majority filter, high-pass filter, clump, and sieve) were applied to all of the classes to refine the result and to minimize false detections. False detections (e.g., sandbars and fire scars misclassified as developed footprint) were further reduced by matching the DIDs areas that fall within this class.
5. The Oil Sands Mining class was derived from Developed class by spatial matching with DIDs data. Similarily, the Tailings Ponds class was derived from the Developed class by spatial matching with DIDs data. In addition, the Burned Areas class was derived from the Exposed/Barren Land class by spatial matching with historical wildfire data sourced from the Government of Alberta.
Part of the process steps were adapted from our previous work published in the following journal paper.
Chowdhury, S., Chao, D.K., Shipman, T.C. and Wulder, M.A. (2017): Utilization of Landsat data to quantify land-use and land-cover changes related to oil and gas activities in west-central Alberta from 2005 to 2013: GIScience and Remote Sensing, p. 700-720, at http://dx.doi.org/10.1080/15481603.2017.1317453.