J. D. Miller, C. N. Skinner, H. D. Safford, E. E. Knapp, and C. M. Ramirez. 2012. Trends and causes of severity, size, and number of fires in northwestern California, USA. Ecology 22:184–203.

Appendix A. Methods used to create a 1987 prefire vegetation map and accuracy assessment results.

To develop a 1987 vegetation map we first developed three cover type maps through classification of Landsat imagery (Fig. A1). The three classifications of cover type were restricted in extent to National Forest and ecoregion boundaries. Ecoregion boundaries were used to minimize confusion between cover types. Each Landsat image was transformed to the Tasseled-Cap brightness, greenness, and wetness features to reduce the dimensionality and image noise of the Landsat data (Crist and Cicone 1984). The brightness, greenness and wetness features have been shown to be correlated to forest structure, stand age and vegetation biomass (Cohen and Spies 1992, Hansen et al. 2001). After transformation, the images were segmented into polygons delineating areas with uniform spectral characteristics using eCognition (Definiens 2007). Image segmentation has replaced traditional pixel based classification algorithms in recent years and has been shown to increase accuracies for stand level classifications by taking advantage of contextual information (Wulder et al. 2004, Meddens et al. 2008). We carried out photo interpretation of cover type categories in 3500 to 6200 randomly chosen polygons per Landsat image in areas undisturbed by fire for training and validation, depending upon size of the study area covered by each Landsat image. We used Random Forests, a variant of Classification and Regression Trees (CART), to develop the three separate classifications of cover type, one for each Landsat image (Breiman 2001, R Development Core Team 2009). Overall classification accuracies for the cover type classifications computed using validation data independent of training data ranged between 69.4 and 89.2% (Tables A1, A2, and A3).

FIG. A1. Map of the overall study area. Northwestern California Forests demarcated in light gray and outlined in black listed clockwise starting from the top left are the Six Rivers, Klamath, Shasta-Trinity, and Mendocino National Forests. Darker gray polygons represent 132 fires that occurred 1987–2008 for which fire severity was analyzed in the current study. Outlined in black are the boundaries of three cover type classifications derived from Landsat images. Classification boundaries are a result of the intersection of National Forest, ecoregion, and image boundaries.

TABLE A1. Confusion matrix of verification points for the cover type classification of Landsat image from Path 46/Row 31.

TABLE A2. Confusion matrix of verification points for the cover type classification of Landsat image from Path 45/Row 32.

TABLE A3. Confusion matrix of verification points for the cover type classification of Landsat image from Path 45/Row 33.

The second step in the development of the 1987 vegetation maps was to carry out a Random Forests classification of vegetation type. Biophysical variables derived from a 30m digital elevation model and climate variables from the DAYMET (http://www.daymet.org/) and PRISM (http://www.prism.oregonstate.edu/) datasets were assigned to each polygon derived in the cover type classification above (Rehfeldt et al. 2006, Evans and Cushman 2009). Vegetation type was trained using the polygons that corresponded to plot locations from the USFS Forest Inventory and Analysis (FIA) program (USDA 1992); due to the small number of FIA plots, all plots were used for training and none were withheld for accuracy assessment. Vegetation types were lumped into broad categories during the classification process based upon the ability of the Random Forests algorithm to separate types into unique classes. The vegetation type classification was applied only to areas mapped as forest (conifer or hardwood, excluding saplings and poles <25.4 cm dbh) in the cover type classification.

Plot data from the United States Department of Agriculture Forest Service (USFS) Forest Inventory and Analysis (FIA) program (USDA 1992) were used to train a classification of vegetation type. FIA protocols call for a grid sampling that results in approximately one plot per 6000 ac, but the US Forest Service, Pacific Southwest Region, Remote Sensing Lab (RSL) enhances the number of FIA plots with additional field sampling. After including RSL enhanced plots and excluding plots that had been sampled in plantations and fires that had occurred since 1987, we ended up with 750 conifer and 53 hardwood plots. Vegetation type was assigned to plots based upon tree species in each plot using CALVEG classification rules used by the RSL. We grouped the CALVEG dominance types measured in the plots into 7 types that could exhibit unique fire regime characteristics (Table A4). We did not withhold any plots for accuracy assessment due to the sparse sample of plots. Independent variables used in a Random Forests classification consisted of variables computed from a 30m digital elevation model and climate variables from the DAYMET (http://www.daymet.org/) and PRISM (http://www.prism.oregonstate.edu/) datasets (Table A5). Separate classifications were performed for conifer and hardwood life forms to minimize confusion between dominance types. Elevation was the most important factor in the classification of conifer types (Table A6). Although we did not withhold plots for a true accuracy assessment we did compute confusion matrices to get a sense for how well the classifications performed. The overall accuracies for the conifer and hardwood classifications were 71.7 and 64.2% respectively (Tables A7 and A8). The vegetation type classifications were applied only to the areas mapped as having either conifer or hardwood cover types, excluding sapling/poles, produced in the cover type classification described above.

TABLE A4. Crosswalk between forest vegetation types for the 1987 vegetation classification to CALVEG types for the FIA plots used for training of the classification.

TABLE A5. Independent variables used in the vegetation dominance type classification.

TABLE A6. Importance variables from random forests classification of vegetation type.

TABLE A7. Conifer vegetation types confusion matrix.

TABLE A8. Hardwood vegetation types confusion matrix.

LITERATURE CITED

Beven, K. J., and M. J. Kirby. 1979. A physically based, variable contributing area model of basin hydrology. Hydrological Sciences Bulletin 24:43–69.

Breiman, L. 2001. Random Forests. Machine Learning 45:5–32.

Cohen, W. B., and T. A. Spies. 1992. Estimating structural attributes of Douglas-fir/Western hemlock forest stands from Landsat and SPOT imagery. Remote Sensing of Environment 41:1–17.

Crist, E. P., and R. C. Cicone. 1984. A Physically-Based Transformation of Thematic Mapper Data - the TM Tasseled Cap. IEEE Transactions on Geoscience and Remote Sensing 22:256–263.

Daly, C., R. P. Neilson, and D. L. Phillips. 1994. A statistical-topographic model for mapping climatological precipitation over mountainous terrain. Journal of Applied Meteorology 33:14–158.

Definiens. 2007. Definiens Professional 7. Munchen, Germany.

Evans, J. S., and S. A. Cushman. 2009. Gradient modeling of conifer species using random forests. Landscape Ecology 24:673–683.

Hansen, M. J., S. E. Franklin, C. Woudsma, and M. Peterson. 2001. Forest structure classification in the North Columbia mountains using the Landsat TM Tasseled Cap wetness component. Canadian Journal of Remote Sensing 27:20–32.

Jenness, J. 2006. Topgraphic Position Index extension for ArcView. Jenness Enterprises, Flagstaff, AZ.

Meddens, A. J. H., A. T. Hudak, J. S. Evans, W. A. Gould, and G. Gonzalez. 2008. Characterizing Forest Fragments in Boreal, Temperate, and Tropical Ecosystems. Ambio 37:569–576.

R Developement Core Team. 2009. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria.

Rehfeldt, G. E., N. L. Crookston, M. V. Warwell, and J. S. Evans. 2006. Empirical analyses of plant-climate relationships for the Western United States. International Journal of Plant Sciences 167:1123–1150.

Thornton, P. E., S. W. Running, and M. A. White. 1997. Generating surfaces of daily meteorological variables over large regions of complex terrain. Journal of Hydrology 190:214.

USDA. 1992. Forest Service Resource Inventories: An Overview. USDA Forest Service: Forest Inventory, Economics, and Recreation Research, Washington, DC.

Wulder, M. A., R. S. Skakun, W. A. Kurz, and J. C. White. 2004. Estimating time since forest harvest using segmented Landsat ETM+ imagery. Remote Sensing of Environment 93:179–187.