Identification of Aedes Mosquitoes with Audio Data Using Gammatone Filter

Studies show that mosquitos could be identified using their recording during their flights with machine learning. However, differences in environmental conditions and the use of various recording devices can cause significant variations in audio channels making their performances drop. Models created with Machine Learning are usually less robust to audio channel variations. This research aims to increase the robustness of the model in identifying Aedes Aegypti and Aedes Albopictus through audio spectral analysis. The approach used involves audio manipulation techniques such as normalization, channel equalization, and histogram equalization. Modeling is performed using various machine learning architectures, including Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN), Gated CNN, Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). Feature extraction involves ShortTime Fourier Transform (STFT), Filter Bank (FBank), MelFrequency Cepstral Coefficients (MFCC), Gammatone Frequency Cepstral Coefficients (GFCC), and Power Normalized Cepstral Coefficients (PNCC). Test results show that the combination of histogram equalization, MLP, and GFCC produces optimal performance. On the Wingbeats dataset, the highest accuracy rate reached $\mathbf{9 3. 8 2 \%}$. Apart from that, this model also shows good performance on the Humbugdb dataset with an accuracy of $\mathbf{8 3. 0 8 \%}$ and on the Dryad dataset with an accuracy of $\mathbf{7 5. 8 6 \%}$. These findings provide a valuable contribution in developing a robust model and can be a solid basis for supporting efforts to monitor and control diseases caused by Aedes Aegypti and Aedes Albopictus mosquitoes.