Table 2. Datasets used to for Deep Neural Network training, cross-validation and selection for detection of aerial calls and ground calls of Leach’s storm-petrels, Hydrobates leucorhoa.
Each dataset consists of 2-second clips that contain either a positive or negative detection.
| Call type | Positive detections | Negative detections | Accuracya (%) | Sensitivityb (%) | Probability thresholdc (%) |
|---|---|---|---|---|---|
| Training dataset | |||||
| Aerial | 6,004 | 3,977 | – | – | – |
| Ground | 6,149 | 20,883 | – | – | – |
| Model selection dataset | |||||
| Aerial | 4,002 | 4,051 | 99.7 | 85.33 | 99 |
| Ground | 4,095 | 15,617 | 85.96 | 52.62 | 50 |
| Randomly sampled test dataset | |||||
| Aerial | 1,357 | 3,414 | 98.35 | 78.92 | 99 |
| Ground | 151 | 4,630 | 75 | 23.84 | 50 |
Notes.
a
Accuracy is calculated as the number of positive detections/(number of positive detections + number of negative detections) above a probability threshold.
b
Sensitivity is calculated as the number of positive detections (above a probability threshold)/number of true positive events in the dataset. The number of true positive events was determined by manual review.
c
The probability threshold is the user selected cutoff above which events are assumed to have the signal of interest.