Insect Classification Using Mel-CSTFT: A Fusion of Mel Spectrogram and Chroma STFT Features

Insect audio classification has witnessed substantial advancements in recent years, facilitated by the development of CNNs and deep learning architectures, which have demonstrated improved performance. Typically used hand-crafted features, specifically MFCCs, have consistently achieved state-of-the-art results in various signal processing research works. However, there is still a lot of scope, and since CNNs and deep neural networks have great capabilities in learning essential information from image-based features, sound signals extracted as visual representations, such as in the form of mel spectrograms, are now being frequently used for classification and detection purposes. To enhance the robustness and reliability of these features, researchers have adopted various techniques, including feature fusion, denoising, data augmentation, transfer learning, and attention mechanisms, etc. Therefore, this study presents a fusion called Mel-CSTFT, comprising of two sound features - mel spectrograms and chroma STFT, and conducts a performance analysis of this combination against the individual features. Ultimately, the aim is to identify whether this combination can outperform the performance of mel spectrograms, which have demonstrated impressive performance in multiple works, and to assess its robustness and reliability across a diverse and relatively bigger dataset.