Authors - P.N. Deorukhakar, V.B. Waghmare, I.K. Mujawar, R.Y. Patil Abstract - Convolutional Neural Networks (CNNs) have been widely and successfully applied to bioacoustic and passive acoustic monitoring tasks, including soundscape classification. However, the high dimension ality of CNN-derived embeddings often results in increased computa tional cost and reduced efficiency, particularly in iterative learning frame works such as Active Learning (AL) and in scenarios with limited labeled data. This work addresses these limitations by proposing a method for adapting CNN architectures to generate compact and discriminative em beddings tailored to soundscape data classification. The proposed ap proach leverages transfer learning and incorporates three progressively reduced dense layers (512, 256, and 128 neurons), enabling dimensional ity reduction to be learned intrinsically during network training rather than applied as a post-processing step. Experimental evaluations con ducted across multiple soundscapes datasets under the Active Learning paradigm demonstrate that the proposed embeddings consistently out perform conventional CNN embeddings (CNNE) in terms of classification performance and the efficient use of labeled data. These results indicate that integrating dimensionality reduction directly into CNN training en hances representation quality and robustness, offering an effective solu tion for soundscape data classification in labeling-constrained environ ments.
Open Zoom