Fig 1.
Workflow for mosquito classification and disease risk assessment.
Fig 2.
The workflow depicts the creation of a balanced mosquito species dataset.
Images were sourced from MosquitoAlert.com, Mendeley Data, IEEE DataPort, and the Dryad Digital Repository, and then selected and quality-controlled to ensure high-quality images. The dataset includes 1000 images each for Anopheles, Aedes, and Culex. Image augmentation techniques enhanced diversity, leading to the final dataset used for model training.
Table 1.
Distribution of images across data sources in train/validation/test splits.
Fig 3.
Single instance of Aedes, Anopheles, and Culex mosquito.
Table 2.
The augmentation techniques that were applied to the merged dataset.
Fig 4.
Detailed workflow of MosQNet-SA architecture for classifying mosquito species images using convolutional layers, attention mechanisms, and a novel MosQNet block.
Fig 5.
Structural layout of MosQNet block components.
Fig 6.
This block contains two 3x3 convolutional layers with batch normalization and ReLU activation.
A skip connection adds the input to the output, facilitating gradient flow in deep networks.
Fig 7.
The Inception block applies parallel convolutions with different kernel sizes (1x1, 3x3, 5x5) to the input.
The outputs are concatenated and passed through batch normalization and ReLU activation, capturing multi-scale features.
Fig 8.
The MBConvBlock consists of an expansion phase, depthwise convolution, optional squeeze-and-excitation, and a projection phase.
It employs residual connections when input and output dimensions match.
Table 3.
Hyperparameters for the proposed classification model.
Table 4.
Detailed performances of baseline models (CNN architectures) and the proposed model with statistical validation.
Table 5.
Per-class performance metrics for MosQNet-SA with 95% confidence intervals.
Fig 9.
Training and validation metrics over 80 epochs, showing accuracy improvement and loss reduction.
Accuracy converges near 100% as loss decreases, and validation metrics exhibit higher volatility than training metrics.
Fig 10.
(a) Confusion matrix showing mosquito species classification results.
Aedes: 941 correct predictions, with 19 misclassifications as Culex. Anopheles: 847 correct predictions, with only one misclassification as Culex. Culex: 954 correct predictions, with six misclassifications as Aedes and eight as Anopheles. (b) ROC curve for each class.
Table 6.
Statistical comparison of MosQNet-SA with top baseline models.
Fig 11.
The bar chart showing the performance metrics comparison across CNN architectures, where MosQNet-SA achieves the highest scores.
Fig 12.
MosQNet-SA comes out as the best model considering the size, parameters, and F1 scores.
Fig 13.
AI-driven analysis of mosquito species reveals distinct anatomical focus areas: For Aedes, algorithms emphasize the thorax, legs, and wing patterns.
Anopheles classification relies heavily on the elongated body posture and proboscis. Culex identification centers on the abdomen shape and wing position. All species recognize the head-thorax junction and leg attachments as key distinguishing features.
Fig 14.
A four-step process illustrating the API workflow, including data submission, image processing, risk assessment, and verdict delivery for mosquito-borne disease risk.
Table 7.
API performance metrics across different hardware platforms.
Table 8.
Comparison of this study with existing works on mosquito classification using image data, including efficiency-oriented baselines.