Abstract:
Mosquitoes are the cause of the most illnesses and deaths worldwide each year. Many Ethiopians
lose their lives as a result of diseases spread by mosquitoes. Dangerous mosquito bites are the
source of diseases like malaria, yellow-fever and chikungunya. The three main mosquito species
found in Ethiopia are Ades, Anopheles, and Culex. To take the appropriate action to monitor them
in a location, it is essential to identify them. Prior approaches to mosquito species classification
have not achieved accuracy levels comparable to human specialists due to mosquitoes' complicated
morphological structure. Several studies are exploring strategies such as error masking,
particularly in traditional CNN-based approaches, to improve classification accuracy. However,
transfer learning is a more effective technique for mosquito image classification, leveraging pre trained models and reducing overfitting risks. This approach results in lower training error,
improved generalization, and lower error rates in mosquito species image classification tasks.
In this study, we examined the capacity of state-of-the-art deep learning models to classify
mosquito species with significant levels of intra and inter-species heterogeneity. To accomplish
our study, we employed an experimental research design, which includes dataset preparation,
classification model design, and evaluation. We have collected 8250 images, with 2750 images in
each class. The mosquito species datasets were collected from Oromia Health Bureau Adama
Public Health Research and Referral Laboratory Center, Tropical and Infectious Diseases
Research Center of Jimma University and online data sources. We have applied different image
preprocessing techniques to enhance the quality of images. Additionally, augmentation is used to
increase the quantity of images. Running the proposed model on the collected dataset resulted in
an accuracy of 71.17%. However, after augmentation of the dataset to 21450 images, the accuracy
increases to 94.8%. The study conducted a comparison of four pre-trained models, namely
EfficientNetB0, MobileNetV2, ResNet50, and VGG19, while also implementing a customized
CNN architecture, and found that MobileNetV2, which has a filtering mechanism, performed the
best in classification and was highly accurate. The study could benefit the vector-borne diseases
investigator by offering a humanitarian technology in which data science can be used to support
mosquito classification, allowing for the treatment of several mosquito-borne diseases.
