Geographically local modeling of occurrence, count, and volume of downwood in Northeast China

The Liangshui National Nature Reserve, located in Northeast China, was heavily damaged by severe windstorms in 2008 and 2009, which caused abundant windthrows, especially large trees, and significantly altered the size and structure of the natural forest. A forest survey was conducted to collect data on living trees, downwood on the forest floor, and environmental factors. We were interested in modeling three types of response variables, including the occurrence of downwood (binary), the number of downwood trees (count) and the volume of downwood (continuous). These response variables were regressed to a set of stand and topographic predictors, including the average diameter of living trees, total volume of living trees, elevation, and slope. Both global and local (geographically weighted regression) modeling techniques were utilized to fit the models. Our results show that local models have great advantages over corresponding global models in model fitting and performance, with desirable model residuals. The spatial variations of local model coefficients were visualized in contour maps, which provided detailed information on the relationships between downwood and stand and topographic variables in the local areas. Furthermore, these local models can be readily incorporated into GIS software and combined with statistical graphics and the mapping ability of GIS technology, to become excellent tools for assessing the risk of natural disasters or disturbances for a given local area, predicting damage caused by such disasters, and offering information critical to decision-making and management planning to prevent or reduce the impacts of natural disasters in the future. ► We model the occurrence, count, and volume of downwood, respectively. ► We use global and local logistic, Poisson, and Gaussian regression models. ► Local (GWR) models fit and perform superior to global models. ► Spatial variations of local model coefficients are visualized by contour maps. ► GWR models show details on the effects of predictors in local areas.