An Enhanced Remote Sensing Method for Agent Attribution of Forest Disturbance

Abstract:

Forest disturbance patterns in the Sierra Nevada of California are shifting as a result of changing climate and land use. The new dynamics have induced high rates of tree mortality and may also yield species replacement and biodiversity loss. Advancing understanding of the spatio-temporal patterns and consequences of forest disturbance remains a research priority. Of particular urgency is the need for a monitoring system that can both detect disturbances and attribute them to specific agents. Building such a system will aid forest management and conservation, deepen scientific understanding of forest carbon and hydrological cycles, and potentially improve forests’ resilience to changing climatic conditions. This study proposed and evaluated an enhanced method for disturbance agent attribution. A Random Forest model was developed to identify the location, timing, and agent of disturbance occurring in Stanislaus National Forest, California, U.S.A., between 1999 and 2015. The model was trained on a predictor set that accounted for both canopy spectral responses and topographic characteristics that regulate vegetation and disturbance dynamics. Predictors included measures of spectral behavior acquired through temporal segmentation of the Landsat time series; measures of patch geometry; and elevation, slope, and aspect. As a novel contribution, the model also exploited metrics of textural pattern in disturbed and undisturbed patches, using the Grey-Level Co-Occurrence Matrix (GLCM). Model results showed that approximately 48% of the National Forest experienced disturbance over the study period; 18% of the Forest was burned, 24% was harvested or thinned, and 6% experienced drought or biological stress. Overall model accuracy was 72.2%, and per-agent accuracy ranged between 71.0% and 91.9%, evaluated with out-of-bag (OOB) observations. Altogether, the method yielded adequate accuracy for disturbance identification and attribution to multiple agents. Quantifying and mapping agent-explicit disturbance patterns in this way offers actionable data for ecological and hydrological modelers, as well as forest managers, who can use the information to adapt resource management strategies to new forest dynamics. The methods may also enhance development of a complete forest disturbance monitoring system.