Figure - available from: PLOS One
This content is subject to copyright.
Confusion matrix of our MosquitoSong+ model from one cross-validation fold Most of the misclassifications were from between Ae. aegypti and Cx. quinquefasciatus species, which could be due to the overlap between the wingbeat frequency components among these species and sex, with the noise possibly further blurring the distinction. Note: Cx. quin refers to Cx. quinquefasciatus species.
Source publication
In order to assess risk of mosquito-vector borne disease and to effectively target and monitor vector control efforts, accurate information about mosquito vector population densities is needed. The traditional and still most common approach to this involves the use of traps along with manual counting and classification of mosquito species, but the...
Citations
... The initial models used pseudo-acoustic optical data with artificial neural networks (ANN) to identify mosquito species and sex [16]. Subsequent research has advanced to deep learning-based acoustic classification models for species identification [5,17,18]. A recent study explored machine learning, specifically convolutional neural networks (CNN), to identify Aedes aegypti mosquitoes from wingbeat recordings [7]. ...
The increasing spread of diseases transmitted by mosquitoes, including malaria and dengue, poses a major global health challenge. Traditional mosquito detection methods, which are based on manual trapping and counting, are time-consuming and inefficient for continuous monitoring. Recently, sensor-based systems have been developed that utilize acoustic signatures. Still, their effectiveness is limited by deep learning models that struggle with noisy environments and fail to adapt to new conditions. Additionally, the scarcity of labeled data for training these models remains a significant obstacle, further reducing their accuracy and generalizability. This paper proposes a novel approach to overcome these limitations by developing an adaptable pipeline to create environment-specific deep-learning models for mosquito detection at diverse locations. This study addresses the challenge of data scarcity and evaluates various feature extraction strategies, such as log-mel and per-channel energy normalization (PCEN), can enhance model robustness in different environmental settings. Our proposed solution successfully creates models that achieve accuracy greater than 90% for any given environment, improving adaptability and supporting public health efforts to control vector-borne diseases. Experimental results confirm this by testing CNN and TCN models in different environments. PCEN preprocessing outperformed log Mel, with the CNN model achieving the highest accuracy of 93.25% in the open environment. Cross-testing results further justify the approach of using environment-specific models.
