Alberta Terrestrial Imaging Centre
July 6, 2017
Executive Summary
Assessing condition of reclaimed wellsites in wetlands usually involves intensive field visits that are often constrained by the human and economic resources available. Field surveys are maybe one of the most accurate way to characterize these sites but they are mainly effective in small areas and are not manageable when regular monitoring is required at a large scale. The scope of this work is to assess the value of hyperspectral remote sensing technologies for mapping vegetation condition in a number of reclaimed or like-reclaimed wetlands in Alberta in support of field-based assessments. A set of methods to derive information related to vegetation composition and health from remote sensing data is investigated. Furthermore, Hyperspectral technologies were also assessed against spaceborne multispectral remote sensing data.
Two flight lines of the AISA airborne hyperspectral data acquired at 2-meter spatial resolution in August 2013 over the Stoney Long Lake area forms the foundation of this study. Field data related to vegetation composition and fractional cover were collected in July/August of 2014 and 2015 in a set of wellsites and adjacent areas. Furthermore, Sentinel-2 multispectral data acquired at a 10/20-meter spatial resolution were collected over the Stoney Long Lake area in August 2016. Assessment of Sentinel-2 data was also conducted in the MATRIX study area, where ground measurement were collected across a range of natural and project disturbances related to seismic lines and oil sands exploration pads.
The K-means unsupervised classification was applied to AISA and Sentinel-2 data to map landcover types in the selected study areas. A set of landcover classes were determined based on visual interpretation of remote sensing data, Google Earth and StreetView. Furthermore, the Multiple Endmember Spectral Mixture Analysis (MESMA) was applied to the AISA data to determine fractional cover at a sub-pixel level for a set of vegetation and background targets. Accuracy assessment was conducted for Sentinel-2 and AISA landcover maps using ground data over the MATRIX site, and a set of validation data extracted from orthophos for the Stoney Long Lake due to a limited size of ground measurement over the latter site. In addition, fractional cover per landcover type in wellsites and control adjacent areas was extracted for each of AISA and Sentinel-2 in the Stoney Long lake area and compared to ground data.
Hypespectral narrow-band and multispectral broad-band indices such as NDVI, MCARI2, ZM, and IRECI were derived using AISA and Sentinel-2 data to assess vegetation condition within wellsite and adjacent areas in the Stoney Long Lake area.
Landcover classification using K-means was achieved with moderate to high accuracies using the 2013 AISA data, except for grass/herbaceous and black spruce classes depending on the AISA flight lines assessed. Different types of misclassification errors due to confusion between classes were observed. Shaded areas in wellsites and adjacent areas tends to be classified as water or conifers. Open areas dominated by wetland background tend to be classified as black spruce. Some confusion was also observed between trees and shrubs. Moderate to high accuracies were also achieved using Sentinel-2 data over the Stoney Long Lake and MATRIX study areas. The classes having the lowest accuracies differs between the two study areas. Low accuracies were observed in white spruce, deciduous and regeneration classes. In the Stoney Long Lake, the average fractional cover for the 45 wellsites in wetland areas was found about 0.5 for grass/herbaceous with an average fractional cover of 0.5. The average fractional cover for shrub and black spruce was 0.13 while it equals 0.1 for bareground/builtup. In adjacent areas, vegetation composition included black spruce, deciduous, shrub, and grass/herbaceous with average fractional cover values of 0.5, 0.15, 0.10, and 0.10 respectively.
Some of the causes behind classification errors include class labeling in the K-means classification process using visual interpretation, and assigning a landcover class to test sites based on the interpretation of ground measurements. The K-means classification seems to identify the same landcover classes observed in the field. However, quantifying the agreement between the measured and observed fractional cover was not achievable due to conceptual differences in the ground- and remote-sensing-based methods used. In addition, errors in wellsite boundaries delineation and discrepancies between the areas sampled on the ground and the one assessed in the remote sensing products all prevent a quantitative measure of the agreement between mapped and measured fractional cover.
Compared to K-means, MESMA classification revealed a number of issues such as unclassified pixels scattered throughout the mapped area, and larger classification errors. Deciduous are overestimated at the expense of shrubs and up to 30% of wellsite and control areas were unclassified. The image-based selection of endmember is possible source of errors due to the lack of pure pixels and/or the unrepresentativeness of the endmember spectral variability.
Assessment of vegetation condition using vegetation indices was conducted for both AISA and Sentinel-2 data based on a spatial subset of the Stoney long lake that was not affected by the 2016 wildfire. NDVI, MCARI2 and ZM vegetation indices were calculated using the AISA hyperspectral data. MCARI2 is more sensitive than NDVI to vegetation biomass and tend to show a wider variation range suggesting MCARI2 is a better indicator of revegetation progress. For a similar MCARI2 magnitude, the ZM index was found to be different between vegetation communities, which could be due to inherent variations in chlorophyll content between these communities. Differences in the ZM and IRECI indices were observed within and between wellsites and could be attributed to the chlorophyll variation between different vegetation communities and/or other factors such as senescence or stress.
Difference in spatial resolution between AISA and Sentinel-2 resulted in a high proportion of shaded pixels in the former, and mixed pixels in the latter, both causing misclassification errors. Furthermore, landcover type with low fractional cover tends to be underestimated in Sentinel-2 data. Assessment of chlorophyll-related indices also suggested the Sentinel-2 IRECI vegetation index to be less sensitive than ZM. Possible causes of the differences observed between AISA and Sentinel-2 could be a reduction in vegetation productivity due to senescence, disturbance, vegetation stress and/or inconsistency between AISA and Sentinel-2 data due to differences in their spectral sampling and resolution.
# 16-ERPC-09