First Generation Fisher Habitat Model, West Coast

Predicted fisher habitat for the west coast, classified into 3 bins: selected for, intermediate, and selected against.

This layer is derived by an expert opinion model for the northeastern region created by Katherine Fitzgerald (USFWS) and a Maxent (version 3.3.3k) model created by CBI with 456 verified fisher detection localities provided by multiple sources (fisher_data_summary_april_2013.xlsx), and an array of 22 environmental data layers (updated_fisher_ppo_supporting_info.docx; for supporting documents, see
http://databasin.org/datasets/c81d82489d1e4be2a852968f4d8eaf7a).

The study area was subdivided based on ecoregional subsection divisions into 6 overlapping model regions. Three regions currently occupied by fishers were used for calibrating models, which were then projected to the 3 unoccupied regions (updated_fisher_ppo_supporting_info.docx). Projections were omitted for the northeastern portion of the study area, where results of an expert opinion model are used instead.

Before modeling, fisher detection points were filtered by removing non-verified detections (no physical evidence to verify fisher identification), detections prior to 1970, detections of reintroduced animals, and telemetry detections. Remaining localities were further filtered to ensure spatial independence by using a nearest-neighbor distance of 5km. If two or more detections were within 5km of one another, the most reliable and recent was retained, or in case of a tie, by random selection.

Potential environmental predictors included vegetation, climate, elevation, terrain, and Landsat-derived reflectance variables at 30-m and 1-km resolutions. Environmental variables were averaged over a 10-km2 moving window and then resampled to 90 m. Urban and open water land covers were masked out.
Maxent was run separately on the three calibration regions, using 10-fold cross validation, initially using all 22 environmental predictors. Next, correlated variables(r > 0.7) were eliminated by retaining the one that yielded the maximum decrease in training gain when excluded from the model. Next, variables that provided the minimum decrease in training gain when excluded were systematically removed using a stepwise procedure until obtaining a model with the fewest predictors having an average training gain not significantly different than the full model. Significance was defined as lack of overlap between 95% confidence intervals for training gain averages.

The final models were projected onto adjacent unoccupied regions. The regional models were combined in areas of overlap using distance-weighted averaging, classified using a strength of selection analysis, and then mosaicked together by Dave LaPlante.

THESE ARE NOT FINAL RESULTS. They have not yet been peer reviewed or evaluated with independent test data. These “first generation” results are provided for discussion and analysis by US Fish & Wildlife Service, and should not be further distributed or used without permission.