Local neural-network-weighted models for occurrence and number of down wood in natural forest ecosystem

The natural forest ecosystem has been affected by wind storms for years, which have caused several down wood (DW) and dramatically modified the fabric and size. Therefore, it is very important to explain the forest system by quantifying the spatial relationship between DW and environmental parameters. However, the spatial non-stationary characteristics caused by the terrain and stand environmental changes with distinct gradients may lead to an incomplete description of DW, the local neural-network-weighted models of geographically neural-network-weighted (GNNWR) models are introduced here. To verify the validity of models, our DW and environmental factors were applied to investigate of occurrence of DW and number of DW to establish the generalized linear (logistic and Poisson) models, geographically weighted regression (GWLR and GWPR) models and GNNWR (GNNWLR and GNNWPR) models. The results show that the GNNWR models show great advantages in the model-fitting performance, prediction performance, and the spatial Moran's I of model residuals. In addition, GNNWR models can combine the geographic information system technology for accurately expressing the spatial distribution of DW relevant information to provide the key technology that can be used as the basis for human decision-making and management planning.