Predicting chronic wasting disease in white-tailed deer at the county scale using machine learning.

Continued spread of chronic wasting disease (CWD) through wild cervid herds negatively impacts populations, erodes wildlife conservation, drains resource dollars, and challenges wildlife management agencies. Risk factors for CWD have been investigated at state scales, but a regional model to predict locations of new infections can guide increasingly efficient surveillance efforts. We predicted CWD incidence by county using CWD surveillance data depicting white-tailed deer (Odocoileus virginianus) in 16 eastern and midwestern US states. We predicted the binary outcome of CWD-status using four machine learning models, utilized five-fold cross-validation and grid search to pinpoint the best model, then compared model predictions against the subsequent year of surveillance data. Cross validation revealed that the Light Boosting Gradient model was the most reliable predictor given the regional data. The predictive model could be helpful for surveillance planning. Predictions of false positives emphasize areas that warrant targeted CWD surveillance because of similar conditions with counties known to harbor CWD. However, disagreements in positives and negatives between the CWD Prediction Web App predictions and the on-the-ground surveillance data one year later underscore the need for state wildlife agency professionals to use a layered modeling approach to ensure robust surveillance planning. The CWD Prediction Web App is at https://cwd-predict.streamlit.app/ .

Using landscape habitat associations to prioritize areas of conservation action for terrestrial birds.

Predicting species distributions has long been a valuable tool to plan and focus efforts for biodiversity conservation, particularly because such an approach allows researchers and managers to evaluate species distribution changes in response to various threats. Utilizing data from a long-term monitoring program and land cover data sets, we modeled the probability of occupancy and colonization for 38 bird Species of Greatest Conservation Need (SGCN) in the robust design occupancy modeling framework, and used results from the best models to predict occupancy and colonization on the Iowa landscape. Bird surveys were conducted at 292 properties from April to October, 2006-2014. We calculated landscape habitat characteristics at multiple spatial scales surrounding each of our surveyed properties to be used in our models and then used kriging in ArcGIS to create predictive maps of species distributions. We validated models with data from 2013 using the area under the receiver operating characteristic curve (AUC). Probability of occupancy ranged from 0.001 (SE < 0.001) to 0.995 (SE = 0.004) for all species and probability of colonization ranged from 0.001 (SE < 0.001) to 0.999 (SE < 0.001) for all species. AUC values for predictive models ranged from 0.525-0.924 for all species, with 17 species having predictive models considered useful (AUC > 0.70). The most important predictor for occupancy of grassland birds was percentage of the landscape in grassland habitat, and the most important predictor for woodland birds was percentage of the landscape in woodland habitat. This emphasizes the need for managers to restore specific habitats on the landscape. In an era during which funding continues to decrease for conservation agencies, our approach aids in determining where to focus limited resources to best conserve bird species of conservation concern.