Ecological Archives C006-057-A4
J. Ruffault and F. Mouillot. 2015. How a new fire-suppression policy can abruptly reshape the fire-weather relationship. Ecosphere 6:199. http://dx.doi.org/10.1890/es15-00182.1
Appendix D. Temporal variations in the performance of boosted regression trees (BRT) models in predicting the probability of fire start and spread.
Fig. D1. Temporal variation in the performance of boosted regression trees (BRT) models predicting the probability of fire start (fires > 0 ha) or the probability of fire spread (fires > 15 ha) on a 5-year moving average window. Mean (curve) and standard deviation (colored area) of an ensemble of 25 models are reported. The gray shaded areas indicate the beginning and the end of the 5-year time period during which the new fire policy was introduced. Changes in model performance during this period were investigated using two statistical methods: a sequential F test (maximum F statistic over the period: FS ) and a OLS-based CUSUM test (statistic: S0 ; ns: p > 0.05, *p < 0.05, **p < 0.01, ***p < 0.001). AUC is the Area under the Receiving operator curve (ROC) curve. The commission error (false positives) is the percentage of fire events misclassified as absences, whereas the omission error is the percentage of non-fire events misclassified as presences (false negatives). The probability threshold is minimized according to the sum of these two values.