Testing the efficacy of tree-ring methods for detecting past disturbances

The retrospective study of abrupt and sustained increases in the radial growth of trees (hereinafter 'releases') by tree-ring analysis is an approach widely used for reconstructing past forest disturbances. Despite the range of dendrochronological methods used for release-detection, a lack of in-depth comparison between them can lead researchers to question which method to use and, potentially, increases the uncertainties of disturbance histories derived with different methods. Here, we investigate the efficacy and sensitivity of four widely used release detection methods using tree-ring width series and complete long-term inventories of forest stands with known disturbances. We used support vector machine (SVM) analysis trained on long-term forest census data to estimate the likelihood that Acer rubrum trees experiencing reductions in competition show releases in their tree-ring widths. We compare methods performance at the tree and stand level, followed by evaluation of method sensitivity to changes in their parameters and settings. Disturbance detection methods agreed with 60-76% of the SVM-identified growth releases under high canopy disturbance and 80-94% in a forest with canopy disturbance of low severity and frequency. The median competition index change (CIC) of trees identified as being released differed more than two-fold between methods, from - 0.33 (radial-growth averaging) to - 0.68 (time-series). False positives (type I error) were more common in forests with low severity disturbance, whereas false negatives (type II error) occurred more often in forests with high severity disturbance. Sensitivity analysis indicated that reductions of the detection threshold and the length of the time window significantly increased detected stand-level disturbance severity across all methods. Radial-growth averaging and absolute-increase methods had lower levels of type I and II error in detecting disturbance events with our datasets. Parameter settings play a key role in the accuracy of reconstructing disturbance history regardless of the method. Time-series and radial-growth averaging methods require the least amount of a priori information, but only the time-series method quantified the subsequent growth increment related to a reduction in competition. Finally, we recommend yearly binning of releases using a kernel density estimation function to identify local maxima indicating disturbance. Kernel density estimation improves reconstructions of forest history and, thus, will further our understanding of past forest dynamics.