Improving image derived vegetation maps with regression based distribution modeling

Incorporating ecological information into image-based vegetation mapping remains a challenge. Much attention has been placed on the use of ancillary information layers in image classification (e.g. slope, aspect, elevation) in that they provide indirect links to information that is ecologically relevant to species distributions. The objective of this study was to assess the utility of incorporating regression-based distribution model surfaces with image classification results using a consensus theoretic approach. We used spatially explicit non-parametric regression modeling in order to incorporate ancillary information in the production of an existing vegetation map for the Lake Tahoe basin. Probability surfaces for 19 prevalent species or genera were produced using generalized additive modeling (GAM). Models were fit to plot data obtained from multiple resource agencies using land-form based explanatory variables derived from a digital elevation model. Model evaluation was assessed by examining species response curves and through cross-validation, resulting in a range of accuracies for individual species (ROC values from 0.58 to 0.85). Probability surfaces for the study area were subsequently generated within a GIS. These surfaces were spatially re-sampled and used in conjunction with IKONOS imagery for use in vegetation mapping. The GAM surfaces were combined with maximum likelihood image classification results using consensus theory and a simple iterative weighting scheme. Results from the analysis demonstrate that the inclusion of the GAM surfaces improved individual class accuracies and suggests the need for implementing standardized and objective species modeling techniques for improving vegetation maps.