Overlapping Landsat scene classifications and focal context identify boreal disturbance mapping uncertainty
Mon statut pour la session
Despite the vastness of the boreal and the wealth of information about this ecosystem, there is inconsistency in how large disturbances are mapped and attributed (Remmel and Perera, 2017). The BorealDB dataset (Ouellette et al., 2020) implements a consistent methodology for mapping historical fire and timber harvesting disturbances for Ontario as point data derived from a mosaic of independently classified Landsat scenes. This study assessed the confidence of BorealDB classifications within overlapping scene margins since common locations can have multiple disturbance classifications allowing for potential uncertainty. We identify all areas within overlapping margins that express disagreement and scrutinize these locations for classification uncertainty. For each point of interest in BorealDB, the disturbance state of its four nearest spatial orthogonal neighbours were extracted and used to produce classification tree (CT) and random forest (RF) predictions of the class at the focal point.
Uncertainty was assessed as being greatest when neighbouring locations and overlapping disturbance classes disagree or, where BorealDB and predicted CT or RF classes disagree. Sampled uncertain locations were additionally compared with original Landsat scenes and visually assessed to scrutinize the confidence of BorealDB’s classifications against CT and RF predictions. We conclude that the degree of uncertainty varied between contextual classifiers in predicting focal disturbance classes, with RF returning worse results than CT (58% disagreement versus 15% respectively). We also conclude that timber harvest classifications were more poorly classified by contextual classifiers than fire disturbances. This is reflected in BorealDB and RF class comparisons which found harvest disagreement to be greater for harvesting sites than for fire (97.93% versus 14.92% respectively). To a lesser degree, BorealDB and CT class comparisons performed almost equally with disagreement for harvesting and fire being 19.92% and 19.25% respectively. Visual assessments corroborate these conclusions, suggesting that effectively predicting focal class memberships with nearest orthogonal neighbour states is both classifier and class dependent.
Agreement between observed and CT predicted classes in any of the four cardinal directions were always high, regardless of whether looking at fire or harvesting disturbance classes. This differed substantially for RF, where agreements were very high for the fire class, but exceptionally poor for the harvest class. Overall, the uncertainty analysis indicates that fire classification in BorealDB is significantly more certain than for harvesting within areas where multiple overlapping classifications are available. Further, when seeking to estimate missing data or to assess neighbourhood influences on BorealDB’s disturbance classes, we recommend using CT rather than RF predictions.