Mapping post-fire habitat characteristics through the fusion of remote sensing tools

Post-fire snags provide important resources for cavity nesting communities as well as being subject to timber removal through salvage logging practices. Tools that can characterize their distributions along with other features important as wildlife habitat, such as woody shrub cover, would be useful for research and management purposes. Three dimensional lidar data and Landsat time series disturbance products have both shown varying promise in their ability to characterize aspects of dead biomass and understory cover, but studies exploring the combination of the remote sensing datasets calibrated with field data to model difficult to map habitat components are limited. The purpose of this study was to 1) relate lidar and Landsat time series products to field-collected calibration data to produce maps of important post-fire wildlife habitat components including snags of varying sizes and the availability of a woody shrub layer; and 2) compare the individual performance of the Landsat and lidar datasets for predicting the distributions of these difficult to map forest habitat features. Using 164 field calibration plots and remote sensing predictors, we modeled and mapped the distributions of our response variables including snag classes (dbh ≥ 40 cm, ≥ 50 cm, and ≥ 75 cm) and woody shrub cover thresholds (≥ 30 and ≥ 50% cover) at 10 m resolutions. Remote sensing predictors included various lidar structure and topography variables and Landsat time series products representing the pre-fire forest, disturbance magnitude, and current forest condition. A model was chosen for mapping purposes using AIC model selection and then by comparing leave-one-out-cross-validation error matrices to choose among competing models. We were able to predict and map all response variables with moderate accuracies and variable sensitivity (true positive) and specificity (true negative) rates. All snag and shrub models were considered to have “good” predictive performance as indicated by area under the curve values (0.74–0.91), with percent correctly classified values ranging from 69–85% when a probability threshold is chosen that balances false positive and false negative errors. Landsat models performed marginally better than lidar structure models according to classification statistics. Landsat-only models had slightly less accuracy than models that included lidar and Landsat data, but often with greater errors than the combined model. The ability to map the response variables with moderate errors and acceptable accuracies for many applications was through the fusion of these remote sensing datasets.