Combining multiple feature selection methods and structural equation model for exploring factors affecting stand biomass of natural coniferous-broad leaved mixed forests

Recognition of biotic and abiotic factors affecting biomass of natural mixed forests is of great importance for forest carbon estimation and management. When estimating stand biomass using models, the different variable selection methods often yield inconsistent results, and there is lack of systematic analysis. This study aimed to combine multiple feature selection methods with structural equation modelling (SEM) to identify a set of variables affecting stand biomass more reasonably. Eight methods were applied for feature selection based on data from 286 permanent sample plots in natural coniferous-broad leaved mixed forests in northeast China. These methods included Pearson correlation analysis, two methods derived from principal component analysis (PCA), stepwise regression, redundancy analysis (RDA), generalized additive model (GAM), random forest (RF), and boosted regression tree (BRT). A total of 56 candidate variables were considered, covering stand, biodiversity, climate and soil features. Significant variability was observed in the variables selected, however, there were 6 variables consistently identified across all methods, including species diversity (N_Sp_Div), stand structural diversity (N_ Size_ Div), nearest taxon index (NRI), community weighted mean based on dry matter mass of leaves (CWM.LDMC), soil pH, and degree-days above 18 °C (DD18). Then, these variables were aincluded in the SEM with stand average age and additive stand density index (aSDI) to explore the direction and magnitude of their impacts on stand biomass. The SEM results showed that aSDI and average age had the greatest positive effects on stand biomass, and structural diversity also had a significant positive effect. DD18 affected stand biomass both directly and indirectly, with the total negative effect. Soil pH indirectly affected stand biomass via aSDI. Our findings demonstrate that combining multiple feature selection methods with SEM was an effective approach for understanding multiple factors affecting stand biomass, and provided valuable insights for forest biomass estimation and carbon management.