IMPLEMENTATION OF TRANSFER LEARNING ON DEEP NEURAL NETWORK FOR IDENTIFICATION BIRDS SOUND IN SUMATERA ISLAND

• Main Author: Satria, Ardika
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/68970
• Institution: Institut Teknologi Bandung

The impact of deforestation has made it several animals difficult to communicate with each other. Communication is usually built through the ability to transmit acoustic signals that can be understood by fellow species. The activity of nature sounds produced by animal sounds (biophony) can be used as an indicator of changes in their habitat environment. Therefore, it is important for monitoring animal sounds to determine whether these species can adapt to new niches or migrate to other niches. In the past few years, the remaining conservation forest that protects bird species on Sumatera Island is only 5 million ha. In this study, modeling of bird sound classification on Sumatra Island was carried out using the Convolution Neural Network (CNN) method through the implementation of transfer learning on a deep neural network. The transfer learning models used were ResNet50, DenseNet169, and VGG19 which were tested on three optimizers such as Adam, RMSProp, and SGD. The results obtained are that VGG19 is the best model with an accuracy reached 91 percent for the mels spectrogram feature extraction, at a learning rate of 10-4, dropout of 50 percent, batch size 32 with Adam optimizer. In the Mels Frequency Cepstral Coefficient (MFCC) feature extraction, the accuracy obtained is 84 percent at a learning rate of 10-3, dropout of 20 percent, batch size 16 with Adam optimizer. The F1 score classification values obtained for both feature extractions reached 0.91 and 0.84.