A new approach to developing a logistic regression model variables to predict tree mortality, based on tree-ring growth dynamics

Abstract

The annual tree increment is one of the integral indicators of abiotic and biotic processes occurring in the forest ecosystem. The use of logistic regression models based on annual tree-ring growth data is a promising approach to studying tree mortality. The diversity of logistic variables in scientific research is a result of various choices of statistics (average, median, growth trend, etc.) and their score in the time-window for the past N (5, 10, ..., 40) years. We propose a new scheme for the formation of logistic variables that involves fixing the statistics for calculating the average and choosing two non-intersecting time-windows based on measurements of the annual tree-rings growth. The choice of non-overlapping «windows» enables setting the ratio of the average growth of annual rings of trees between the windows for different periods of time. We examined the past 41 years of tree growth. Logistic regression models are constructed on a set of pairs of non-intersecting «windows» with a limit on the values of the sensitivity and specification of at least 1.6. The calculation of the percentage prediction if a tree is living or dying was done based on the contingency table in the logistic regression model. The logistic regression models were visualized using ROC curves. The models were compared on an expert scale based on the calculated area under the ROC curves. The obtained logistic regression model was verified by the bootstrap method. The calculations were carried out for the Siberian stone pine Pinus sibirica du Tour growing in the Baikal region (the Khamar-Daban Ridge) using the R programming language. The computed logistic regression model helped us predict live and dead trees in more than 80 % of cases.

Keywords

• dendrochronology
• annual increment
• siberian stone pine pinus sibirica du tour
• khamar-daban