Role of environmental and stand factors on forest dieback: An approach using structural equation modelling and machine learning

Abstract

The Zagros forest region in western Iran is facing dieback processes, impacting both the environment and the local economy. Understanding the cause of forest dieback remains a challenge for forest experts due to the wide set of factors usually involved in the dieback phenomenon. To achieve this goal, we measured woody species dieback intensity (DI) using 124 plots distributed in oak forests across a wide range of environmental conditions. On these plots, we measured stand and vegetation factors (e.g. woody species richness, WSR), various soil properties and topographic indices (in particular the SAGA wetness index) and climatic factors. We then modeled dieback intensity using structural equation modeling (SEM) and machine learning (ML). We found a significant positive correlation between WSR and DI (r=0.48), and very low to low but still significant negative correlations with the SAGA wetness index (r=-0.10) and soil clay content (r=-0.16), respectively. Other soil properties and climate factors showed no significant influence on DI. After six rounds of respecification, the SEM fitting gave satisfactory results based on the validation map of DI indices (Goodness-of-fit Index = 0.948 and Comparative fit index =0.958). The DI map was generated using a random forest (RF) machine learning model to achieve an accuracy of 75% and a kappa index of 0.45, and then categorized into four classes. The map showed that the intensity of dieback was contrasting in our area: 41.1% of plots had no dieback, 58.4% had low DI, and less than 1% had moderate or high DI. We also found that soil clay content and woody species richness were crucial factors, explaining around 50% of the variance in DI. In conclusion, combining the methods from SEM and RF models provides an efficient approach to understand and identify key factors in the complex process of dieback.