dataset.rar

Abstract—Birds reflect environmental health as the speciesdiversity is affected by pollution and climate change. Thus, largescaleautomated biodiversity monitoring and machine learninganalysis benefits the experts and citizen scientist. However, manydevices intended for these models have limited computationresources and strict power consumption constraints. Therefore,this paper proposes an optimized CNN-based bird call classifiertargeting lightweight embedded platforms. The proposed classifieris using low complexity CNN model, MobileNetV2, wherethe input is the spectrogram image of bird sounds. The imagesare derived by Short Time Fourier Transform (STFT) and MelFrequency Cepstral Coefficient (MFCC) algorithms. To validatethe classifier accuracy, 1,000 spectrogram images from each often bird species are generated and run into the classifier, andthen compared with the accuracy of the same inputs run intohigh complexity CNN model classifier, ResNet-50. The resultsshow that the accuracy of low complexity classifier is about thesame compared to high complexity classifier, which is 79% and81% respectively