Abstract
Remote sensing is widely used to detect forest disturbances (e.g., wildfires, harvest, or outbreaks of pathogens or insects) over spatiotemporal scales that are infeasible to capture with field surveys. To understand forest ecosystem dynamics and the ecological role of human and natural disturbances, researchers and managers would like to characterize spatiotemporal patterns of several types of disturbance. Recent advances include the development and testing of algorithms to automatically detect and attribute an array of disturbance types across large forest landscapes from remotely sensed imagery. We reviewed the scientific literature for automated disturbance type attribution in forest ecosystems to synthesize current knowledge and inform future work. We created metrics that characterize study contexts, methods, and outcomes to evaluate 34 studies that automated the attribution of several (two or more) forest disturbance types. Reported accuracies of ∼80% for up to eight disturbance types at local (500 km2) to continental (6.5 × 106 km2) spatial extents reaffirm the potential for attributing forest disturbances from remote sensing data at scales relevant to ecological research and forest management. Generally, greater accuracies were reported for attributing disturbance types that affect most of a pixel and exhibit spectral signatures distinct from other disturbances or the prior (undisturbed) condition. Accuracy of attributing disturbance type tends to be notably lower for disturbances with smaller spatial extents or resulting in subtle spectral changes, such as small patches of tree mortality covering <40% of a pixel or areas of tree damage without evident mortality (e.g., defoliator outbreaks). Nearly all (33 of 34) studies used Landsat image time series to attribute disturbance types because the Landsat archive is freely accessible, spans several decades (allowing longer-term change analyses), is calibrated to facilitate change analyses, and has multiple spectral bands that aid algorithms. In our synthesis, most studies focused on specific disturbance types of interest within a particular study area, rather than a comprehensive list of types that would make the algorithm more generally applicable at broader spatial extents. We recommend several future research areas to improve the thematic resolution, accuracy, and potentially the transferability of attribution algorithms. Further development and testing of algorithms will continue to reveal the most effective approaches for attributing disturbance types that have been more difficult to detect with available spectral data to date, such as insect damage or low severity wildfire, and to evaluate model sensitivity as well as accuracy. The lessons that collectively emerge from automated forest disturbance attribution studies have the potential to guide future research and management applications.
[Display omitted]
•We reviewed 34 studies that automated identification of forest disturbance types.•All reviewed studies except two used Landsat image time series as source data.•Studies attributed two to eight disturbance types with ∼22–100% accuracy per type.•Most studies focused on areas <75,000 km2 and separated 3–5 disturbance types.•More work is needed to refine algorithm sensitivity, accuracy, and transferability.