Clustering Versus Regression Trees for Determining Ecological Land Units in the Southern California Mountains and Foothills

Abstract

Abstract A landscape stratification was required for simulation modeling of fire disturbance and succession in the mountains and foothills region of the Peninsular Ranges within San Diego County, California. Two quantitative approaches to mapping ecological land units (ELUs) were compared for a 3,878 km2 study area. These were: (a) clustering of climate overlain with key terrain variables; and (b) regression tree modeling of climate, geology, and terrain variables using a normalized difference vegetation index (NDVI), derived from a Landsat Thematic Mapper image, as the dependent variable. Terrain variables derived from a digital elevation model included slope gradient, cosine transformed slope aspect, potential solar insolation, and a topographic moisture index. For the simulation model ELUs were required that would stratify the landscape according to biomass (fuel) accumulation dynamics, related to site productivity, and probabilities of plant species establishment. Therefore, ELUs were defined using abiotic variables, and the resulting stratifications were evaluated by their ability to reduce within-class variance in the NDVI (as an index of biological productivity), and by comparing them to a map of existing vegetation. While the regression tree method resulted in classes that explained more variance in NDVI than classes resulting from unsupervised clustering, the difference was not large. In contrast, the unsupervised approach resulted in ecological land classes that were more strongly related to existing vegetation patterns. FOR. SCI. 49 (3):354–368.