Investigation of CNN-based models for Frog Calling Activity Detection

Frogs plays an important role in monitoring and evaluating the health of wetland. Globally, many frog species are threatened. Therefore, it is valuable to first detect frog activity, then classify species, and lastly predict frog population. In this study, we used a CNN-based model to detect frog calls. Specifically, continuous field recordings were divided into fixed duration segments. Then, different CNN architectures and various time-frequency representations of frog calls were investigated. The different CNN models were evaluated using different training and test datasets to assess model robustness. Experimental results show that the best-averaged macro F1-score and accuracy obtained was 98.41% and 97.99%, obtained using Mel-spectrogram and VGG-style CNN. In addition, when the training and test data were from different recording days, the averaged macro F1-score reduced to 82.12% and 71.38%, respectively.