Abstract
Background: Understanding the relationships between forest structure, in particular attainable height, and the environment is important for sustainable forest management. Similarly, modeling structural attributes improve our understanding of forest growth dynamics and may identify key drivers of long-term changes in the forest ecosystem. Due to the inherent complexity of these relationships, quantification of some drivers of forest growth is often not available, resulting in spatially auto-correlated errors of the regression model. Methods: To explore the tree height-environment relationships of oriental beech we compared the performance of a standard regression model (multiple linear regression, MLR) to those accommodating a spatial correlation structure, specifically a Generalized Least Squares model with exponential correlation structure (GLS) and three variations of the Simultaneous Autoregressive Model (SAR): the spatial lag model (SLM), the spatial Durbin model (SDM) and the spatial error model (SEM). Across 127 0.1 ha circular sample plots in the primeval World Heritage Hyrcanian Forests of Iran, we collected data on tree height and edaphic and topographic. Within each plot, the height of all trees with DBH ≥ 7 cm was measured. Results: The results showed that SAR and GLS models reduced spatial autocorrelation of model residuals and improved model fit, with both SDM and SEM slightly superior to the SLM in removing spatial autocorrelation in the model residuals. SDM performs better than SEM in terms of RMSE and adjusted R2. Conclusions: Although SAR-based models performed marginally better than GLS, we still recommend GLS for spatial analyses due to their easier implementation and ease-of-use compared to SAR models. However, when the computation time is a concern, SAR-based models can be more useful because of faster execution. Keywords: spatial autocorrelation; Hyrcanian forests; multiple linear regression model; simultaneous autoregressive model; generalized least squares