Disturbance regimes are a critical component of ecosystems, and vary across ecosystem types. However, methods for characterizing regimes across scales and disturbance types remain elusive. Here, we investigate disturbance regimes over seven Landsat scenes located across the contiguous United States. Specifically, we locate and attribute forest disturbances using a new method, disturbr, and then determine disturbance severity, extent, and spatial heterogeneity by calculating NDVI declines and disturbance patch structure. We ask the question, how does ecoregion influence disturbance regime characteristics?
We applied the disturbr algorithm to 30m Landsat NDVI from 2000-2016. This algorithm is novel in that it uses both multi-pixel (spatial) and single-pixel (temporal) variables to summarize time series characteristics (e.g., magnitude NDVI decline), and because it applies a machine learning approach to both disturbance detection and attribution. We trained the disturbr algorithm with >600 pixels per scene and validated the results using time series plots, satellite images, and historical imagery. We then calculated severity as disturbance magnitude, extent as disturbance patch area, and spatial heterogeneity using several patch-based metrics.
Results/Conclusions
The disturbr algorithm resulted in a mean kappa detection accuracy of 75% among all seven Landsat scenes (59-97%), where disturbances were attributed to insect outbreaks, fire, clearcut harvest, and even changes in river path over time. Following our test of the disturbr algorithm, we characterized disturbance regimes among different ecosystems in terms of disturbance severity, extent, and spatial heterogeneity. Preliminary results show substantial differences among ecosystems and disturbance types, underscoring the variation in subsequent disturbance impacts among ecosystems and the importance of a full understanding of disturbance patterns across the landscape. For example, mean severity in Southwestern Oregon (fire = -0.14, harvest = -0.12, insects = -0.05) differed considerably from mean severity in Southeastern South Carolina (fire = -0.17, harvest = -0.06, anthropogenic land use change (LUC) = -0.15). Mean disturbance patch extent (connected pixels) also demonstrated differences among these scenes (Oregon: fire = 8888 m2, harvest = 15272 m2, insects = 1087 m2; South Carolina: fire = 1307 m2, harvest = 8833 m2, LUC = 5679 m2). This approach lays the groundwork for a more comprehensive quantitative approach to characterizing and studying disturbance regimes at broad spatial scales, with the potential to inform analyses of fundamental ecosystem processes.