Supercomputing to go back in time: Reconstruction of the history of Canada’s forests
Figure: Multi-year proxy BAP time series animation
Canada is a vast nation, dominated by forested ecosystems. From a total area approaching 1 billion hectares, Canada has about 350 million hectares of land with trees.
At the Canadian Forest Service (CFS), Senior Research Scientist Mike Wulder and his colleagues Joanne White and Geordie Hobart have been working with imagery from the Landsat series of satellites for decades. Previously they were limited by the capacity for data transfer, computing software and hardware, and the degree of automation possible due to the state of the science and the nature of the data.
In addition to improvements in computing capacity, government agencies worldwide are increasingly offering free and open access to analysis-ready data products. This has been a game changer in the remote sensing community. However, if data is not provided in an analysis-ready form, users are still required to build and implement custom approaches and tools prior to any value-added processing.
In 2008 the United States Geological Survey made all archival and new data collected by the Landsat series of satellites free and open access, as well as analysis ready. From a forest monitoring point-of-view, data from the Landsat series of satellites have many qualities that make them a data source of choice, the most important of which is a spatial resolution (30 m) that can capture human impacts over large areas.
Given the CFS mandate to monitor and report on Canada’s forested ecosystems, Wulder was interested in how best to use the vast data archive of Landsat to re-create the recent history of Canada’s forests.
Wulder and his colleagues had a goal. But how were they going to get there? Issues to address included the impacts of cloud, shadow, smoke, haze, seasonality, satellite revisit, on the data.
When problems are big, it is often fruitful to create a team with complementary skills. From a government perspective, universities offer additional brainpower to help address scientific and technical questions, such as how to deal with billions of pixels over space and time.
In the case of this research, Wulder’s team partnered with the Integrated Remote Sensing Studio from the Faculty of Forestry at the University of British Columbia, led by Professor Nicholas Coops, a Canada Research Chair in Remote Sensing. Coops brought his insights and team to help address many of the obstacles that stood in the project’s way. Critically, Dr. Txomin Hermosilla joined the team as a Post-Doctoral Fellow and has been instrumental in leading the analysis and providing scientific insight. Through this collaboration, the team developed an approach to generate consistent, time series informed, image composites, which captured forest dynamics on the landscape.
Now, the missing part was the capacity to consistently, transparently, and with some flexibility, process the data. Enter WestGrid, a regional partner of Compute Canada.
“We have been able to port our processing algorithms to WestGrid, enabling more rapid processing: scripts that used to take weeks to complete now take days or hours,” says Wulder. “We are able to collaborate in a shared environment, working from the same base data and ensuring consistent outcomes. The gains in computing power offered by WestGrid has allowed for increasingly ambitious analyses.”
Using WestGrid/Compute Canada computing resources, they have been able to undertake research to capture and label three decades of forest change over Canada’s forests (see http://forests.foundryspatial.com/) and are now generating annual maps of land cover for the same 30-year period. The expert computing support offered by one of WestGrid’s support team analysts, Belaid Moa, has been crucial to Wulder’s team’s success.
By way of scope, recall that Canada is approaching 1 billion hectares in size, which is in turn represented with about 10 billion Landsat pixels. This spatial coverage of ~10 billion pixels is populated with a number of spectral channels and indices on an annual basis. These annual coverages are replicated for each year in the time series.
“WestGrid offered the capacity for dealing with such spatially large and temporally dense data sets,” says Wulder. “No longer limited by computing considerations, the algorithms we are developing are increasingly sophisticated, providing new and otherwise unavailable information on Canada’s forested land cover and the disturbances and recovery occurring over the forested land base. Canadian science, monitoring, and reporting activities are strengthened by the computing capacity offered by WestGrid, as well as by the cross-institutional collaborative work environment.”
In addition to ensuring changes in forest cover are considered in the land cover mapping processes, this research will improve the CFS’s ability to refine the assignment of land cover by leveraging knowledge of forest successional processes with the time series of data. The long baseline offered by the science outcomes to date are especially informative to investigations related to climate change.
For more information on Wulder’s team’s work, please see: http://bit.ly/2e1FvBk.