Fig 1.
Detection pipeline for search-phase bat echolocation calls.
(a) Raw audio files are converted into a spectrogram using a Fast Fourier Transform (b). Files are de-noised (c), and a sliding window Convolutional Neural Network (CNN) classifier (d, yellow box) produces a probability for each time step. Individual call detection probabilities using non-maximum suppression are produced (e, green boxes), and the time in file of each prediction along with the classifier probability are exported as text files.
Fig 2.
Spatial distribution of the BatDetect CNNs training and testing datasets.
(a) Location of training data for all experiments and one test dataset in Romania and Bulgaria (2006–2011) from time-expanded (TE) data recorded along road transects by the Indicator Bats Programme (iBats) [7], where red and black points represent training and test data, respectively. (b) Locations of additional test datasets from TE data recorded as part of iBats car transects in the UK (2005–2011), and from real-time recordings from static recorders from the Norfolk Bat Survey from 2015 (inset). Points represent the start location of each snapshot recording for each iBats transect or locations of static detectors for the Norfolk Bat Survey.
Fig 3.
Precision-recall curves for bat search-phase call detection algorithms across three testing datasets; (a) iBats Romania and Bulgaria; (b) iBats UK; and (c) Norfolk Bat Survey. Curves were obtained by sweeping the output probability for a given detector algorithm and computing the precision and recall at each threshold. The commercial systems or algorithms that did not return a continuous output or probability (SCAN’R, Segment, and Kaleidoscope) were depicted as a single point.
Table 1.
Average precision and recall results for bat search-phase call detection algorithms across three different test sets iBats Romania and Bulgaria; iBats UK; and Norfolk Bat Survey.
Fig 4.
Comparison of the predicted bat detections (calls and passes) for two different acoustic systems using monitoring data collected from Jersey, UK.
Acoustic systems used were SonoBat (version 3.1.7p) [43] using analysis in [49], and BatDetect CNNFAST using a probability threshold of 0.90. Detections are shown within each box plot, where the black line represents the mean across all transect sampling events from 2011–2015, boxes represent the middle 50% of the data, whiskers represent variability outside the upper and lower quartiles, with outliers plotted as individual points. See text for definition of a bat pass.