This project was designed to study techniques and prototype methodologies for mapping meaningful vegetation parameters that can be used as input variables to fire management systems, such as FARSITE. Satellite remote sensing data may be effectively used to model and map the natural vegetation types and vegetation structure classes that are essential elements of fire fuels characterization. The k-Nearest Neighbor (k-NN) classification algorithm was evaluated for several parameters to yield the most useful data for fuels mapping.

When classifying pixel "x", the k-NN algorithm finds the training plots that are most similar to x and chooses the class prevailing among these plots; where k is the number of nearest training plots. The algorithm can utilize any continuous variable for training and classification, and output continuous or class values. Previous studies using a k-NN classifier, usually to characterize land cover, have not evaluated several key input parameters for best results, especially for fire fuels mapping. This study was conducted with the k-NN algorithm to evaluate the size of the sample to use (number of plots), the number of neighbors (k), and the number and combination of spectral bands and dates of satellite data, from three seasons of imagery. Several additional input variables, including topography (elevation, slope, aspect, topographic position) and soils (available water content, organic carbon, and a quality index) were examined.

The k-NN algorithm and the results of the parameter evaluation were used to map the spatial distribution of several key vegetation variables by integrating U.S. Forest Service Forest Inventory and Analysis (FIA) field survey plot data with Landsat-7 ETM+ remote sensing imagery. Mapping results were evaluated using several methods, including a field evaluation of the classifications. The results displayed in the graphs and the table are from the analysis of a Chesapeake Bay test site. The map examples are from a southern Utah test site. Forest type classes were successfully mapped, along with estimates of basal area of coniferous and deciduous forest, total above ground biomass, crown cover, tree height and forest size class (saplings, pole, and saw timber). The research results will be useful for science and land management agencies to map fire fuels with remotely sensed data for predicting fire behavior using models such as FARSITE.