Attribution of disturbance change agent from Landsat time-series in support of habitat monitoring in the Puget Sound region, USA

To understand causes and consequences of landscape change, it is often not enough to simply detect change. Rather, the agent causing the change must also be determined. Here, we describe and test a method of change agent attribution built on four tenets: agents operate on patches rather than pixels; temporal context can provide insight into the agent of change; human interpretation is critical because agent labels are inherently human-defined; and statistical modeling must be flexible and non-parametric. In the Puget Sound, USA, we used LandTrendr Landsat time-series-based algorithms to identify abrupt disturbances, and then applied spatial rules to aggregate these to patches. We then derived a suite of spectral, patch-shape, and landscape position variables for each patch. These were then linked to patch-level training labels determined by interpreters at 1198 training patches, and modeled statistically using the Random Forest machine-learning algorithm. Labeled agents of change included urbanization, forest management, and natural change (largely fire), as well as labels associated with spectral change that was non-informative (false change). The success of the method was evaluated using both out-of-bag (OOB) error and a small, fully-independent validation interpretation dataset. Overall OOB accuracy was above 80%, but most successful in the numerically well-represented forest management class. Validation with the independent data was generally lower than that estimated with the OOB approach, but comparable when either first or second voting scores were used for prediction. Spatial and temporal patterns within the study area followed expectations well, with most urbanization occurring in the lower elevation regions around Seattle Tacoma, most forest management occurring in mid-slope managed forests, and most natural disturbance occurring in protected areas. Temporal patterns of change agent aggregated to the watershed level suggest substantial year-over-year variability that could be used to examine year-over-year variability in fish species populations. (C) 2015 Elsevier Inc. All rights reserved.