A Parameter-Efficient Cascaded Framework for Mosquito Species Identification using MFCC Features
Abstract
Effective mosquito-borne disease surveillance requires complex and scalable solutions, but current acoustic deep learning models are computationally demanding for resource-constrained IoT devices. This study proposes a lightweight, two-stage cascaded Two-Dimensional Convolutional Neural Network (2D-CNN) for mosquito species identification using Mel-Frequency Cepstral Coefficients (MFCCs) as input. The framework first employs a lightweight binary classifier to distinguish mosquito sounds from background noise, followed by a specialized classifier to identify six distinct mosquito species. Evaluated against a diverse acoustic dataset, the proposed system achieves an overall accuracy of 89%. Furthermore, the combined model requires only 51,351 trainable parameters, representing a significant reduction compared to traditional models. These results demonstrate a scalable approach for real-time, high-precision vector monitoring in resource-limited environments.
Authors: Natchapol Kiriwanna, Chakarida Nukoolkit, Tuul Triyason