by Wesley Wu
The Boreal Disturbance Database (BorealDB) is a spatial database constructed by Ouellette et al (2020) to map wildfire and timber harvesting disturbances within Ontario’s Managed Area (MA). Forest disturbances, such as wildfires and timber harvests, are ecological processes within forested landscapes that are intrinsic to the history and development of boreal forests. BorealDB portrays disturbances as points, arranged in a regular grid-like pattern, derived from individually classified overlapping Landsat scenes. However, due to these overlaps’ classification uncertainty can occur when an individual disturbance point has multiple classifications from each overlapping scene. The overlap produces classification uncertainty within the database by labelling disturbance points with multiple classes that do not always agree.
Ascertaining the uncertainty of disturbance points, requires understanding of the focal neighbourhoods of disturbance points. This study aimed to identified and flag these areas of uncertainty within BorealDB so that they may be further scrutinized to assess classification confidence, the measure of how likely a disturbance point belongs to a disturbance class. To identify areas of uncertainty the focal context, the nearest orthogonal neighbours of a disturbance point, was examined.
Landscapes do not exist in isolation, disturbance points that are closer to one another are likely to have a stronger relationship than points further away. To quantify focal disturbance classes the orthogonal neighbourhood classes of each disturbance point were fed into classification tree (CT) and random forest (RF) models. Classification uncertainty is deemed to exist in areas where BorealDB and the CT and RF predictions disagree. Comparing the BorealDB disturbance classifications with those predicted by the CT and RF algorithms would identify areas of classification uncertainty. The most stable classifier method disturbance clusters were determined by sampling disturbed locations and visually assessed against the original satellite imagery.
When the CT and RF predictor results were compared with the BorealDB classes, the CT predictions had more similarities (85%) than the RF predictions (42%). The high agreement between BorealDB and CT indicates that many point classifications within BorealDB are supported by predictions derived from their orthogonal neighbour classifications indicating a relationship between a disturbance point and its nearest neighbours. The predictions of the CT and RF differed at approximately 50%. Visual assessments found that RF was able to classify fire disturbances better than CT, but overall CT predictions were more consistent. Thus, CT was found to be the most stable classifier method at identifying classification uncertainty within BorealDB.
Going forward, this research opens opportunities for future research. The classifiers were only fed the orthogonal data of BorealDB derived classifications. The thesis serves as a proof of concept demonstrating that predictions derived from the focal scale can reinforce BorealDB’s classification confidence. Future research could expand the focal neighbourhood, while the research was focused solely on the neighbouring disturbance points in the cardinal directions additional research can incorporate the ordinal directions as well, providing more samples to feed the contextual classifiers.
Wesley Wu is a Geography MSc student under the supervision Professor Tarmo Remmel. His research focuses on applying focal contextual methods to a spatial database to assess boreal disturbance mapping uncertainty.
Ouellette, M., Remmel, T.K., Perera, A.H., 2020. A spatial database of historical wildfire and timber harvest in the boreal AOU of Ontario: the methodological framework (No. TR-37), Science and Research Technical Report. Ontario Ministry of Natural Resources and Forestry, Science and Research Branch, Peterborough, ON.