Accuracy and cost of models predicting bird distribution in agricultural grasslands

• Published in: Agriculture, Ecosystems & Environment
• Main Authors: Barbottin, Aude, Tichit, Muriel, Cadet, Claire, Makowski, David
• Other Authors: Sciences pour l'Action et le Développement : Activités, Produits, Territoires (SADAPT), Institut National de la Recherche Agronomique (INRA)-AgroParisTech, Agronomie
• Format: Article in Journal/Newspaper
• Language: English
• Published: HAL CCSD 2010
• Subjects: BAYESIAN MODEL AVERAGING; BIRD; LIVESTOCK FARMING SYSTEM; LOGISTIC REGRESSION; MODEL SELECTION; SENSITIVITY; SPECIFICITY; COST; [SDV]Life Sciences [q-bio]; Vanellus vanellus
• Online Access: https://hal.archives-ouvertes.fr/hal-01173187; https://doi.org/10.1016/j.agee.2009.10.009

Numerous agro-environmental indicators have been developed to assess the impact of farming systems on biodiversity. They can be combined into logistic models for predicting the presence of species of ecological interest. In general, several models are available for a given species and their practical value depends on their accuracy and the cost of measurement of their input variables. This paper aims to assess the accuracy and cost of implementation of a wide range of models predicting the presence of two grassland bird species, the lapwing Vanellus vanellus and the redshank Tringa totanus. Some of these models were developed using stepwise selection procedures and the others were developed by Bayesian Model Averaging. Sensitivity, specificity, and probability of correctly ranking fields (AUC) were estimated for each model from observational data. The cost of implementation of each model was computed as a function of the number and types of input variables. Results showed that the presence/absence of lapwings can be predicted more accurately than the presence/absence of redshanks, probably due to the stricter ecological requirements of lapwings. For both species, the highest AUC values were obtained with models combining habitat and management variables. The most costly models were not always the most accurate. Full models and models derived by Bayesian Model Averaging were most costly and less accurate than some of the models derived using selection procedures. When large sets of candidate variables were considered, the models selected using the BIC criterion were less costly and sometimes more accurate than the models selected using the AIC criterion.