Modelling Wetland Vegetation Trajectories using Machine Learning and Multispectral Sentinel-2 Imagery

Wetlands are important to water quality, nutrient cycling, and biodiversity, and have been identified as a management priority across Canada. Reclaiming wetlands that have been impacted can be a long-term, costly process depending on the type of wetland, location, and potential access constraints. One challenge with wetland reclamation is that the work may cause additional or new wetland impacts during access and execution of the work. Concurrent with wetland reclamation priorities, there is a focus on restoration of caribou habitat and wetlands are prevalent in planned caribou restoration areas. Improvements in wetland reclamation will also benefit caribou habitat restoration efforts.

Advances in remote sensing and machine learning are enabling the completion of more complex, local-scale monitoring programs and the development of cost-effective monitoring strategies that integrate natural variation.  Compared with traditional wetland monitoring, these new approaches are expected to reduce the cost of environmental monitoring and reclamation while increasing the defensibility of monitoring efforts allowing operators to identify disturbance effects in wetlands.

Connacher Oil and Gas Limited and Matrix Solutions Incorporated are exploring the use of machine learning approaches to create a low-impact and cost-effective approach to assessment of wetland vegetation trajectories by linking field observations to multispectral Sentinel-2 imagery. The study area is an oil sands lease situated within the Boreal Forest Natural Region of Alberta, Canada, with a mosaic of upland forests and extensive areas of peatland ecosystems. Much of the lease has been impacted by forest fire; most notably the 1995 Mariana Lake fire, as well as older fires that occurred in 1981 and 1982. A set of vegetation metrics that define the condition and recovery of wetland communities and ecological sensitivity to disturbance were chosen to be modeled.

The Random Forest algorithm was used to train predictive models. Training data consisted of 72  homogeneous wetland and upland vegetation communities that were spatially delineated in 2021 and 2022 using a custom data collection application by direct visual assessment from a helicopter, supplemented by high resolution base imagery. Estimates of vegetation cover were used to calculate a set of proportional cover metrics tailored to the study area and research objectives. Percentage of cover metrics had to account for masking of lower vegetation strata by upper ones and the effect of solar zenith angle on the reflected spectral signature received by the satellite sensor. Explanatory variables included individual Sentinel-2 bands, spectral indices, and indices of seasonal change from summer to fall. Three key vegetation metrics were expressed as visible proportions in each 10×10m pixel of the study area for each year of the Sentinel-2 program and converted to trajectories.

Training data collected as polygons were found to be robust against normal variation in spectral response that occurs at a pixel scale and permitted validation at the same community scale as the model’s inferential units. Discretized, raster-derived geospatial layers enable vegetation ecologists to interpret the patterns of presence and change over time for jack pine, tamarack, and herbaceous cover to inform future management decisions in the study area.high-resolution