Habitat suitability quantifies the relative similarity of a pixel to environmental conditions that support the ecological needs of a conservation target. Values range from 0 to 100, with increasing values representing greater similarity.
TerrAdapt’s species-based habitat suitability models are empirically derived using species distribution modeling methods like Maxent and Random Forests that use machine learning algorithms to relate species occurrence to environmental covariates. These methods follow a ‘use-availability’ study design where environmental conditions at ‘use’ locations (i.e., where the species was observed to be present) are contrasted with conditions at ‘available’ locations distributed broadly across the study area.
Species occurrence data is collected from GPS or radio telemetry collars fitted to animals or remote camera stations. Typically many hundreds if not thousands of observations are collected from regional partners. The environmental predictors of occurrence vary from species to species but typically include covariates derived from Landsat imagery, landcover, topography, climate, disturbance, energy and transportation infrastructure, and soils.
We validate our species-based habitat suitability models in two ways. First, we perform k-fold cross-validation where a fraction (typically 10%) of the training data is withheld from model training and then evaluated to determine how well the fitted model predicts species presence in the withheld observations. Second, we typically withhold an independent occurrence dataset from a different effort than was used for model training and assess the predictive accuracy of the model on the independent data. In both approaches, we assess standard model fit metrics for categorical models including area under the receiver-operator curve (AUC), sensitivity, specificity, kappa statistics, etc. For the five species models we have developed, all have an AUC > 0.85 indicating strong model fit.