The first artificial intelligence algorithm for identification of bat species in Uruguay

Abstract Acoustic bat identification is a complementary method to traditional mist-netting for chiropteran surveys. In this work, an algorithm based on artificial intelligence techniques was developed for bat species identification in Uruguay. An acoustic library of 662 search phase pulses of the 10 most common bat species in Uruguay was obtained. Random Forests, Support Vector Machines and Artificial Neural Networks algorithms were trained to predict bat species from acoustic variables. Variable selection was performed in an independent subset (one third) of the dataset, using the function varImpPlot of randomForest R package. Model performance was evaluated by means of the test error with the remaining two thirds of the data. To do this, data were split randomly into a training and a test set, then the model was trained with the training sample and its performance was assessed using the test sample. The procedure was performed 100 times and the test errors were averaged to have an unbiased measure of the performance of the models. The best predictor was a Random Forests classifier that considered 12 predictor variables. The achieved accuracy was comparable to other international published products. Additionally, a threshold value for classification probability was optimized to define an “unclassified” class that allows using this algorithm even when the training sample does not represent an exhaustive sample of local richness. A web application returning the predicted class and a confidence measure for a given observation was developed permitting the use of this tool by a broad spectrum of users, from biologist to technicians.