[Paper Review] Acoustic Identification of Ae. aegypti Mosquitoes using Smartphone Apps and Residual Convolutional Neural Networks

Kayuã Oleques Paim, Ricardo Rohweder+3 more|arXiv (Cornell University)|Jun 16, 2023

Animal Vocal Communication and BehaviorBiochemistry, Genetics and Molecular Biology3 citations

TL;DR

This paper presents a lightweight residual convolutional neural network (ResNet-based) for smartphone-based acoustic identification of Aedes aegypti mosquitoes using wingbeat sound recordings. It achieves high accuracy and robustness to background noise through a novel noise-reduction training strategy and a new benchmark dataset, enabling real-time, efficient inference on low-end smartphones with 18.5% smaller model size and 2x faster inference time via quantization and multi-threaded inference.

ABSTRACT

In this paper, we advocate in favor of smartphone apps as low-cost, easy-to-deploy solution for raising awareness among the population on the proliferation of Aedes aegypti mosquitoes. Nevertheless, devising such a smartphone app is challenging, for many reasons, including the required maturity level of techniques for identifying mosquitoes based on features that can be captured using smartphone resources. In this paper, we identify a set of (non-exhaustive) requirements that smartphone apps must meet to become an effective tooling in the fight against Ae. aegypti, and advance the state-of-the-art with (i) a residual convolutional neural network for classifying Ae. aegypti mosquitoes from their wingbeat sound, (ii) a methodology for reducing the influence of background noise in the classification process, and (iii) a dataset for benchmarking solutions for detecting Ae. aegypti mosquitoes from wingbeat sound recordings. From the analysis of accuracy and recall, we provide evidence that convolutional neural networks have potential as a cornerstone for tracking mosquito apps for smartphones.

Motivation & Objective

• To develop a low-cost, scalable smartphone application for real-time detection of Aedes aegypti mosquitoes using acoustic signals.
• To address computational and environmental noise challenges in deploying machine learning models on consumer-grade smartphones.
• To create a publicly available, high-quality dataset for benchmarking mosquito sound classification models.
• To improve model efficiency and robustness to background noise without sacrificing classification accuracy.
• To enable widespread community-based monitoring of Aedes aegypti populations through accessible mobile technology.

Proposed method

• Proposes a residual convolutional neural network (ResNet-based) architecture optimized for mobile inference, with residual blocks to reduce computational overhead and dropout layers to prevent overfitting.
• Applies dynamic range quantization during training to reduce model size by 18.5% and halve inference time while preserving accuracy.
• Implements a multi-threaded inference pipeline that decouples feature extraction, classification, and result display for improved responsiveness.
• Introduces a noise-aware training strategy that reduces the influence of environmental background noise on classification performance.
• Uses spectrograms of short audio segments (1–2 seconds) as input to the CNN for end-to-end mosquito classification.
• Validates the model on both real-world recordings and a newly collected dataset, testing performance across diverse acoustic conditions.

Experimental results

Research questions

• RQ1 Can a lightweight residual CNN achieve high accuracy in classifying Aedes aegypti wingbeat sounds using only smartphone microphone input?
• RQ2 How effective is dynamic range quantization in reducing model size and inference time without degrading classification performance?
• RQ3 To what extent can the proposed training strategy mitigate the impact of background noise on mosquito detection accuracy?
• RQ4 Can the model maintain high performance on low-end smartphones commonly used in endemic regions?
• RQ5 How does the proposed model compare to state-of-the-art models in terms of accuracy, inference speed, and model size?

Key findings

• The proposed model achieves high classification accuracy and recall, demonstrating strong potential for real-world deployment in mosquito tracking apps.
• The model size was reduced by 18.5% through dynamic range quantization, significantly improving computational efficiency.
• Inference time was reduced by approximately 2x using multi-threaded processing and model optimization.
• The model demonstrated high stability in identifying Aedes aegypti in semi-controlled laboratory environments with varying background noise levels.
• The application ran successfully and without crashes on entry-level smartphones (e.g., Asus Zenfone Max M3 and Motorola Moto E6 Play), with average inference time of ~320 ms.
• The newly introduced dataset provides a standardized benchmark for future research in acoustic mosquito detection using smartphone-based systems.