The Canadian Forest Service of Natural Resources Canada, partnered with the University of British Columbia, with support from the Canadian Space Agency, has been developing the science and methods to track and characterize the history of Canada’s forests. For this research we use Landsat imagery to detect changes, identify the year in which the changes occur, and estimate a change type (such as harvest or wildfire). A change detection approach was applied to an annual time series of Landsat data, enabling the detection of abrupt (e.g. fire) and gradual (e.g. drought) changes. From the time series we can also assess the recovery of forests after disturbances by wildfire and harvest. In this research, we have developed an explore and discovery tool to portray and communicate forest change over Canada. The map shows points representing locations where forest change has been observed between 1985 and 2011. The source Landsat imagery is a continuous surface of pixels, each representing a 30m x 30m square area. For visualization purposes the pixels have been converted to points, and generalized at different zoom levels.
The map makes use of modern web-map technologies, and performs optimally on newer computer systems running Google Chrome, Firefox, or Safari internet browsers. The main map component of the interface shows the location and type of observed forest change. The mapped presence of change is more reliable than the change type labels. Untyped pixels are present when the algorithm used was not able to unambiguously assign a pixel to a given category (that is, 2 categories had very similar assignment probabilities). At closest zoom levels, each point represents one 30m x 30m pixel. At further zoom levels, each point represents progressively larger areas, allowing for the character of change to be preserved when viewing at regional, provincial or national scales. The stacked column charts at the bottom of the interface show relative amounts of change which occurred in individual years, with the colors representing the types of change. The column chart can be used to restrict the points shown on the map, by clicking and dragging on the chart to select years of interest. On the right side of the map interface, four buttons identify the types of change observed, Fire, Harvest, Infrastructure and Undifferentiated, with colors corresponding to the points on the map and the components of the column chart. On each button, the area associated with each change type currently shown on the map is displayed. Clicking each button toggles the points displayed on the map and in the column chart on and off.
The accuracy of the change products was evaluated using independent validation data. Overall, change events were detected with a 90% accuracy. Fires were detected with a user's accuracy of 98%, while harvesting was detected with a user's accuracy of 88%. These results indicate that the automated change detection and attribution algorithms are robust, but errors will exist. Web tools such as this aid in visualizing and sharing results, but they are also designed to enable feedback that can further support refinement of the change detection and attribution algorithms.