1.1 The global burden of mosquito-borne diseases

Mosquito-borne diseases represent a persistent and severe health challenge. Over 700,000 people succumb to mosquito-borne diseases annually [5], with malaria alone claiming more than 600,000 lives and affecting 250 million people worldwide in 2022 [6]. The World Malaria Report 2023 reveals that climate change is exacerbating this challenge, as variations in temperature, humidity, and precipitation directly affect the behavior and viability of malaria-transmitting Anopheles mosquitoes [7].

Dengue fever presents an equally formidable threat, ranking as the second most lethal mosquito-borne disease with approximately 20,000–21,000 deaths annually worldwide, particularly devastating regions across Asia, the Americas, and Africa [6]. The scale of this epidemic became starkly evident on April 30, 2024, when the WHO reported more than 7.6 million dengue cases globally, including 3.4 million confirmed cases, over 16,000 severe cases, and more than 3,000 deaths [8]. The Aedes mosquito genus, particularly Aedes aegypti, serves as the primary vector for dengue, Zika virus, chikungunya, and urban yellow fever, with these mosquitoes typically feeding during daylight hours [9]. Despite available vaccines for yellow fever, the disease continues to cause 30,000 deaths annually worldwide, while lymphatic filariasis burdens approximately 657 million people across 39 countries [10].

1.2 Current control strategies and their limitations

Traditional approaches to controlling vector-borne diseases rely on an integrated strategy combining insecticide-treated nets, indoor residual spraying, environmental management, surveillance, and health education. However, these conventional methods face mounting challenges that limit their effectiveness. Vector populations are rapidly adapting to insecticides, creating resistance that undermines control efforts. Additionally, inadequate health infrastructure in endemic regions and the lack of efficient diagnostic tools further complicate disease management [10]. These persistent challenges necessitate innovative, technology-driven solutions that can adapt to changing vector behaviors and environmental conditions.

Recent scientific advances offer promising new avenues for addressing these limitations. Breakthroughs in remote sensing, geographic information systems (GIS), and environmental modeling have significantly improved our capacity to predict disease transmission patterns and implement preemptive measures. However, a substantial gap remains in developing accurate, efficient, and accessible tools for real-time mosquito identification and classification—a prerequisite for effective surveillance and targeted intervention strategies.

1.3 AI-driven solutions for mosquito classification

Artificial intelligence methods hold significant potential for revolutionizing mosquito surveillance and vector-borne disease control [11]. The integration of AI and machine learning has transformed vector-borne disease management by enabling more precise identification of disease vectors, mapping potential breeding sites, and optimizing resource allocation. When applied to diagnostic tools, environmental data analysis, and predictive modeling, these technologies offer new opportunities for early disease detection, real-time monitoring, and better-informed public health interventions.

Convolutional neural networks (CNNs) and attention mechanisms have emerged as particularly promising approaches for enhancing disease detection, surveillance, and risk mapping. CNNs excel at image classification tasks, and when combined with attention mechanisms, they can focus on the most relevant features of the input data, thereby boosting both accuracy and interpretability in disease classification. This capability becomes essential when dealing with the subtle morphological differences between mosquito species that are often indistinguishable to the human eye but carry vastly different disease transmission risks.

To address the pressing challenge of vector-borne disease classification, this study introduces a novel deep-learning model that integrates CNNs with attention mechanisms to classify mosquito species with high accuracy and efficiency. The proposed model incorporates explainable AI techniques and the RESTful API for easy integration into applications for real-time mosquito classification and dynamic disease risk mapping. This research presents the following key contributions:

• Development of MosQNet-SA: A novel deep learning architecture designed to classify mosquito species while reducing computational complexity and inference time, making it suitable for deployment on edge devices.
• Explainable AI Techniques: Integrating explainable AI methods such as Grad-CAM and SHAP ensures the model’s decision-making process is interpretable and transparent, enhancing its reliability in real-world applications.
• Real-Time Applications: The model is designed for scalability and real-time usage, with significantly fewer parameters than existing models, enabling dynamic risk mapping and mosquito classification through a RESTful API.

The subsequent sections of this work are organized as follows: section 2 reviews relevant literature and provides necessary background on the topic. section 3 details the methodology used in developing the MosQNet-SA model. section 4 presents experimental results and model evaluations, including performance comparisons with state-of-the-art models. Explainable AI techniques and their role in enhancing the model’s interpretability are demonstrated in section 5. section 6 presents RESTful API integration of the proposed model. section 7 discusses the findings and explores the implications within the broader research context. Finally, the paper concludes in section 8, summarizing key findings and potential future research directions.

2.1 CNN-based classification approaches

Convolutional neural networks have emerged as the dominant approach for image-based mosquito classification, demonstrating high accuracy in distinguishing different species and genera. Park et al. [12] achieved over 97% accuracy in classifying eight mosquito species using fine-tuned Deep CNNs, leveraging a large dataset of 3,600 images capturing various postures and deformations of mosquitoes. This study underscored the potential of transfer learning in this domain by applying data augmentation techniques and fine-tuning pre-trained models. Building on this foundation, researchers [13] proposed an innovative pipeline for low-cost IoT sensors, using models like VGG16, ResNet50, and custom CNNs to achieve 98% accuracy. This work is particularly notable for addressing the vital need for cost-effective, deployable solutions in resource-constrained settings where traditional laboratory-based identification methods are impractical.

Similarly, Asgari et al. [14] modified the VGG16 architecture, achieving accuracy between 94.66% and 98.92% across multiple datasets. The study included extensive ablation experiments to identify optimal architectural modifications, further advancing CNN-based classification. Siddiqui and Jain [15] tackled the challenging task of distinguishing between mosquito genera, developing a CNN model that achieved 84.51% accuracy on a dataset of 1,800 images from three genera (Aedes, Anopheles, and Culex). Though the accuracy was lower than some other studies, this work highlighted the difficulties in classifying visually similar genera. A comparative study by Okayasu et al. [16] found that deep learning methods, especially ResNet, outperformed traditional feature-based methods, but that data augmentation was necessary to achieve this improvement.

2.2 Data modalities and feature optimization

Mosquito classification relies on various data types, including morphological, genetic, and image-based data. Mosquito classification using image data leverages advanced computer vision techniques, specifically CNNs, to improve species identification and lifecycle-stage classification. Studies have shown that images of wings provide better classification performance compared to body images, with CNNs achieving up to 99% accuracy in distinguishing Aedes species [17]. This finding is noteworthy as wing morphology contains species-specific characteristics that are less variable than body features across different environmental conditions and specimen preparations.

Researchers have demonstrated that focusing on prominent anatomical features can significantly improve classification accuracy. Studies using deep CNNs to classify mosquito photos have achieved 94% accuracy by focusing on distinctive features, such as white band stripes on the legs and thorax [18]. Kumar et al. [19] further validated the effectiveness of deep learning models, specifically DCNNs and pre-trained models, in accurately distinguishing mosquito species from image data through hyperparameter optimization and data augmentation, achieving enhanced precision and F1 scores.

2.3 Hybrid and multimodal approaches

As the field matures, hybrid and novel approaches have emerged, integrating multiple data types and new methodologies to overcome the limitations of single-modality approaches. De Lima et al. [20] integrated AI with wing geometric morphometry, achieving accuracy rates of 84%−95%. This work represents a significant advancement in combining traditional morphometric analysis with machine learning for mosquito classification, demonstrating how domain expertise can enhance automated approaches.

Beyond visual data, researchers have explored alternative sensing modalities. Genoud et al. [21] investigated optical signals from mosquitoes and found that Support Vector Machines were the most effective for complex classification tasks. Wei et al. [22] proposed WbNet, a novel ResNet-based model that integrates self-attention and data augmentation methods for classifying mosquito species based on wing-beating sounds. This approach achieved 89.9% accuracy on the WINGBEATS dataset, with 100% precision, recall, and F1 Scores for Aedes and Culex species. These hybrid approaches demonstrate the potential for combining various data modalities—visual, audio, and contextual—to improve classification accuracy and robustness further.

2.4 Explainable AI for mosquito classification

Explainable AI (XAI) has emerged as a crucial component for enhancing the interpretability and trustworthiness of machine learning models in mosquito classification. This capability becomes particularly important in public health applications where understanding model decisions is essential for practitioner acceptance and regulatory approval. Numerous studies have used XAI methods to improve classification accuracy while clarifying AI models’ decision-making processes.

Adhane et al. [18] employed Grad-CAM visualizations to develop an AI system capable of elucidating mosquito classification decisions. These visualizations emphasize the image components that influence classification decisions, facilitate error analysis, and enhance comprehension of the model’s predictions. Building on this work, the EfficientNet-B0 model used Grad-CAMs to highlight critical image pixels, focusing on the mosquito’s abdomen to precisely distinguish gonotrophic stages [23]. Goni et al. [24] focused specifically on malaria diagnosis through explainable AI, employing SHAP (SHapley Additive exPlanations) to clarify model decisions and provide insights into feature importance. In [25], researchers developed a formal concept lattice method for explainable AI, enabling mosquito classification by providing both accurate local and global explanations for classification decisions.

2.5 Research gaps and future directions

Despite significant progress, several critical challenges persist across the field. Limited dataset sizes and diversity, difficulties distinguishing visually similar species, issues with image quality, generalization across different environments, and the need to balance accuracy with computational efficiency represent prominent challenges identified in the literature. Joshi et al. [26] synthesized findings from 120 papers, offering valuable insights into the field’s current state and outlining potential future directions for machine learning in mosquito control. Rodriguez et al. [27] highlighted the potential of combining machine learning with citizen science initiatives, particularly for classifying disease-carrying mosquitoes using crowdsourced data.

These challenges highlight the need for more robust, efficient, and interpretable models that can operate effectively in real-world conditions while maintaining high accuracy and providing meaningful insights into their decision-making processes.

Given these challenges, our novel MosQNet-SA model emerges as a promising solution to advance mosquito classification. The model incorporates many data modalities to enhance robustness. Key domains for continued advancement include multimodal data integration, adaptive learning for continuous model improvement, intuitive interfaces for academics and citizen scientists, and a scalable architecture capable of accommodating new species and geographic regions. Utilizing various explainable AI (XAI) methodologies for mosquito species categorization provides supplementary insights into model decision-making. Multiple explainability methods, such as Saliency, GradCAM, LIME (Local Interpretable Model-agnostic Explanations), and Kernel SHAP (SHapley Additive Explanations) were utilized. Furthermore, including MosQNet-SA into a RESTful API provides a scalable, stateless framework for real-time mosquito categorization and disease risk evaluation. Fig 1 illustrates the comprehensive process used in this research.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Workflow for mosquito classification and disease risk assessment.

https://doi.org/10.1371/journal.pone.0344970.g001

3.1 Dataset acquisition and preparation

The data set was compiled from four distinct sources to ensure a diverse and comprehensive representation of mosquito species. Fig 2 shows the detailed workflow to create a balanced dataset of mosquito species for this research.

• Dataset A: This citizen science platform [28], dedicated to monitoring and controlling mosquito populations, contributed 1,234 images of various mosquito species, including Anopheles, Aedes, and Culex. The photos, submitted by participants worldwide, were meticulously screened and vetted for quality and accuracy.
• Dataset B: A publicly accessible dataset from the Mendeley Data repository [29] provided 876 well-annotated images of Aedes and Culex mosquitoes. The images selected by researchers are particularly well-suited for machine learning tasks.
• Dataset C: The IEEE DataPort repository [30], maintained by the Institute of Electrical and Electronics Engineers (IEEE), offered 748 images specifically focused on Aedes and Culex mosquito species. The research community’s creation of this dataset ensured high relevance and reliability.
• Dataset D: Sourced from the Dryad Digital Repository [31], which hosts research data associated with scholarly publications, 600 images of Anopheles mosquitoes were included. Extensive metadata and annotations were present in these images.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

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.

https://doi.org/10.1371/journal.pone.0344970.g002

Images were selected based on quality metrics including resolution (minimum 224x224 pixels), clarity (sharpness score >0.7), and proper species identification verified by entomologists.

3.1.1 Merged dataset.

The images from these diverse sources were carefully inspected and selected based on the quality, clarity, and suitability for the research objectives of this study. The research dataset [32] includes three significant mosquito species: Anopheles, Aedes, and Culex, each represented by 1,000 images. Fig 3 represents the single instances of the three species of mosquitoes. These species were selected for their significant role in transmitting various mosquito-borne diseases and for their prevalence across different regions worldwide. Ensuring an equal distribution of 1,000 images per species was a deliberate choice to maintain a balanced representation and prevent potential biases during model training and evaluation (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Distribution of images across data sources in train/validation/test splits.

https://doi.org/10.1371/journal.pone.0344970.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Single instance of Aedes, Anopheles, and Culex mosquito.

https://doi.org/10.1371/journal.pone.0344970.g003

3.2 Augmented merged dataset

Multiple image augmentation techniques were applied to enhance the dataset’s diversity and robustness. These augmentation techniques were used to introduce variability and improve the generalization capabilities of the machine learning models. Table 2 demonstrates the summary of applied augmentation techniques.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. The augmentation techniques that were applied to the merged dataset.

https://doi.org/10.1371/journal.pone.0344970.t002

3.3 Data splitting strategy

To prevent data leakage and ensure robust evaluation, we implemented a stratified splitting strategy that respects data source boundaries. The dataset was divided into training (80%), validation (10%), and test (10%) sets using the following methodology:

• Per-source stratification: Each data source (A, B, C, D) was split independently before merging, ensuring no cross-source contamination between training and test sets.
• Near-duplicate removal: We applied perceptual hashing (pHash) with a similarity threshold of 0.95 to detect and remove near-duplicate images across all splits. A total of 47 near-duplicates were identified and removed.
• Species and pose stratification: Within each source split, images were balanced across species and pose variations to maintain representative distributions.
• Background stratification: Images with different background types (natural, laboratory, plain) were proportionally distributed across splits.

3.4 Baseline model development

For experiments, a range of deep learning architectures, from classic models to advanced designs, have been explored and chosen for effectiveness in image classification tasks. VGG models, such as VGG16 and VGG19, use small receptive fields and a uniform architecture, though they require substantial computational power. ResNet, with its residual connections, solves the vanishing gradient problem, enabling deeper networks such as ResNet50 and ResNet101. Inception and its successor, Xception, improve feature extraction via multi-scale convolutions, with Xception additionally using depthwise separable convolutions for greater efficiency. Inception-ResNet combines these strengths by merging Inception modules with residual connections for improved training stability and accuracy.

DenseNet introduces dense connectivity, in which each layer connects to all previous layers, thereby optimizing feature reuse. MobileNet and EfficientNet focus on reducing computational costs. MobileNet employs depthwise separable convolutions for real-time applications on resource-constrained devices, while EfficientNet uses a compound scaling method to balance network depth, width, and resolution. NASNetMobile, built through neural architecture search, optimizes model performance on mobile platforms by reducing computational demand while maintaining high accuracy. Together, these models provide a diverse toolkit for evaluating performance across varying levels of accuracy, computational efficiency, and suitability for real-time and resource-constrained environments.

3.5 Baseline model construction and training setup

Sixteen models with diverse architectures were selected to comprehensively evaluate model performance. This selection includes classical architectures like VGG16 and VGG19, along with more advanced models such as ResNet variants (ResNet50, ResNet101, ResNet152) and others like Xception, InceptionV3, and InceptionResNetV2. Efficient models for mobile environments, such as MobileNet, MobileNetV2, and NASNetMobile, were also included. The DenseNet (121, 169, 201) and EfficientNet (B0, B1, B2) families were selected because they employ dense connectivity patterns (where each layer receives inputs from all previous layers) and compound scaling (systematically increasing network depth, width, and resolution), making them particularly effective for complex classification tasks while maintaining computational efficiency.

All baseline models were trained using consistent procedures to ensure fair comparison: identical data augmentation, same train/validation/test splits, consistent learning rate schedules (cosine annealing), and early stopping with patience of 15 epochs based on validation loss. All models were trained with a learning rate of 0.001, with the number of epochs varying between 38 and 121 depending on the convergence.

3.6 MosQNet-SA: Proposed classification model

MosQNet-SA is a novel CNN designed specifically for mosquito classification. Its architecture integrates several vital components to enhance feature extraction and improve classification accuracy. In Fig 4, the overall architecture of the MosQNet-SA model is presented for the classification of mosquito species.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Detailed workflow of MosQNet-SA architecture for classifying mosquito species images using convolutional layers, attention mechanisms, and a novel MosQNet block.

https://doi.org/10.1371/journal.pone.0344970.g004

The model begins with a convolutional layer for extracting low-level features (edges, textures) and incorporates residual blocks (which allow gradients to flow directly through skip connections), enabling deeper training without vanishing gradient problems. Inception-like blocks perform multi-scale feature extraction (capturing both fine details and broader patterns), while MBConv blocks (mobile-optimized convolutions that separate spatial and channel-wise operations) enhance computational efficiency. The spatial attention block (which learns to focus on the most relevant image regions) allocates computational resources to key areas, improving accuracy in noisy or cluttered environments. The final classification layers aggregate these processed features to make accurate species predictions.

3.6.1 Spatial attention block.

The spatial attention block refines feature maps by focusing on the most critical spatial locations. Let the input tensor be:

(1)

Where H, W, and are the height, width, and number of input channels, respectively.

Max Across Channels:

The maximum value across channels for each spatial location:

(2)

Convolution and Sigmoid Activation:

Next, a convolution and batch normalization are applied:

(3)

Spatial Attention Mask Application:

Finally, the attention mask is applied to the input tensor:

(4)

3.6.2 MosQNet block.

The MosqNet block architecture improves the efficiency of neural network feature extraction and processing. It combines three other blocks, residual and inception modules, alongside pooling and MBConv blocks. The interconnected blocks work together seamlessly, providing a balanced structure that supports effective learning and performance across complex tasks. Fig 5 illustrates the MosqNet block architecture.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Structural layout of MosQNet block components.

https://doi.org/10.1371/journal.pone.0344970.g005

Residual block: The residual block introduces skip connections to preserve the input’s identity across layers, thus allowing better gradient flow during backpropagation. Fig 6 illustrates the detailed block architecture. Let the input tensor be denoted as Equation 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

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.

https://doi.org/10.1371/journal.pone.0344970.g006

First Convolution:

(5)

Where F is the number of filters.

Second Convolution:

(6)

If the input tensor and output tensor do not match in dimension, the shortcut connection is transformed as follows:

(7)

Finally, the output is computed as:

(8)

Inception block: The inception block applies parallel convolution operations with different kernel sizes to capture multiple levels of detail. Let Equation 1 represent the input tensor, and be the convolutional kernel sizes.

Parallel Convolutions:

(9)

Concatenation: The output is the concatenation of the results from each convolutional path-

(10)

The Inception block (Fig 7) incorporates multiple convolutional layers with different kernel sizes, capturing features at various scales. This parallel structure enhances feature diversity and representation, improving image classification accuracy.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

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.

https://doi.org/10.1371/journal.pone.0344970.g007

Mobile bottleneck convolution block: The MBConv (Mobile Bottleneck Convolution) block is designed to balance computational efficiency and accuracy. It expands the input tensor, applies depthwise separable convolutions, and potentially uses squeeze-and-excitation (SE) for feature recalibration. Fig 8 illustrates the block architecture of the MBConv Block. Let the input tensor be denoted by Equation 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

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.

https://doi.org/10.1371/journal.pone.0344970.g008

Expansion Phase: The expansion phase applies a convolution to increase the number of channels:

(11)

where , and r is the expansion ratio.

Depthwise Convolution:

Next, a depthwise convolution is applied:

(12)

where k is the kernel size and s is the stride.

Squeeze-and-Excitation (Optional): If squeeze-and-excitation (SE) is used, it performs global average pooling followed by two fully connected layers:

(13)

The output is recalibrated:

(14)

where .

Projection Phase: The final phase projects the features back to the desired output channels:

(15)

where is the number of output channels.

If the input and output dimensions match, a residual connection is applied:

(16)

if and .

3.7 Hyperparameters for proposed classification model

Hyperparameters for the classification models, detailing key settings for optimizing performance. Table 3 outlines the chosen hyperparameters for the proposed classification model. The Adam optimizer is selected, offering adaptive learning rates and improved weight decay to help counteract overfitting by regularizing the model’s parameters. A relatively high learning rate of 0.001 suggests a balanced approach that converges efficiently without risking overshooting the optimal point during training. The weight decay parameter, set to 0.01, further enhances regularization by imposing a minor penalty on larger weights. ReLU is used as an activation function in hidden layers and is known for mitigating vanishing gradient issues. In contrast, Softmax is used at the output layer to ensure that class scores are probabilistic. A larger batch size of 256 is employed, potentially improving training stability and reducing noise in gradient estimates. The model trains for over 80 epochs, striking a balance between learning capacity and computational efficiency, and uses categorical cross-entropy as the loss function, which is adequate for multi-class classification. Collectively, these hyperparameters were chosen to ensure efficient learning, robust generalization, and optimal classification accuracy.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Hyperparameters for the proposed classification model.

https://doi.org/10.1371/journal.pone.0344970.t003

4.1 Experimental setup

The experiments were conducted on Google Colab, utilizing its GPU resources for efficient model training. The software environment was based on Python 3.8, with key libraries including TensorFlow for model development and training, Scikit-learn for evaluation metrics, and OpenCV for image preprocessing. Additionally, NumPy and Pandas were used for data manipulation, while Matplotlib and Seaborn were employed to visualize training progress and results. The TensorFlow Explain library was used to implement explainable AI (XAI) techniques to provide insights into model interpretability. GPU acceleration was enabled via CUDA 11.2, and all experiments were run in a Jupyter Notebook to ensure a streamlined workflow and reproducibility.

4.2 Statistical validation methodology

To ensure robust and statistically rigorous evaluation, we employed multiple validation strategies:

• K-fold Cross-Validation: 5-fold stratified cross-validation was performed on the training set to assess model stability and generalization. Each fold maintained consistency in species and source distributions.
• Multi-seed Experiments: All experiments were repeated with five different random seeds (42, 123, 456, 789, 1024) to quantify variability in model performance.
• Confidence Intervals: 95% confidence intervals were calculated using bootstrap resampling (n = 1000) for all performance metrics.
• Statistical Significance Testing: We employed McNemar’s test to compare paired predictions between MosQNet-SA and baseline models, applying a Bonferroni correction for multiple comparisons ().

4.3 Evaluation metrics

Evaluation metrics are quantitative indicators that provide significant insights into the model’s efficacy. The following parameters are examined for classification and comparison tasks.

Test loss: Test Loss quantifies the error between predicted values and actual labels during the testing phase. It calculates how well the model’s predictions match the true labels on unseen data. The goal is to minimize test loss, as lower values indicate that the model’s predictions are closer to the actual labels. Test loss is computed using a suitable loss function :

(17)

where is the number of samples in the test set, represents the true label of the i-th sample, and represents the predicted value by the model.

Test accuracy: Test Accuracy measures the proportion of correctly classified samples in the test set relative to the total test samples. It provides a clear indication of overall correctness. The accuracy score is calculated by dividing the number of correctly predicted samples by the total number of samples:

(18)

Macro precision: Macro Precision calculates the average precision across all classes without considering class imbalance. Precision for each class i is the ratio of true positives to the sum of true positives and false positives :

(19)

where is the number of true positives for class i and is the number of false positives for class i.

Macro recall (sensitivity): Macro Recall computes the average recall across all classes, ignoring class imbalance. Recall for each class i is the ratio of true positives to the sum of true positives and false negatives :

(20)

where is the number of false negatives for class i.

F1-Score: F1-Score is the harmonic mean of precision and recall, providing a metric that balances both measures. It is useful in applications where balancing precision and recall is essential. F1-Score is calculated as:

(21)

4.4 Baseline model evaluation

Synthesis of Baseline Model Performance: The comprehensive evaluation of 15 baseline models reveals several important patterns that inform our understanding of mosquito classification challenges. DenseNet architectures (DenseNet201, DenseNet169, DenseNet121) consistently achieved the highest accuracy among baseline models, with DenseNet201 reaching 99.1% test accuracy. This performance suggests that dense connectivity patterns are particularly effective for capturing the subtle morphological features that distinguish mosquito species. However, these high-performing models require substantial computational resources, with DenseNet201 containing 26.2M parameters—nearly 70 times as many as our proposed MosQNet-SA model.

EfficientNet variants (B0, B1, B2) demonstrated strong performance while maintaining relatively low parameter counts, achieving 97.4–97.7% accuracy with 9.3−13.5M parameters. This efficiency makes them particularly relevant for comparison with our lightweight approach. The consistent performance across different EfficientNet scales suggests that the compound scaling approach is practical for mosquito classification tasks. MobileNet variants showed competitive performance (97.1–97.4%) with the lowest parameter counts (7.4−7.5M), indicating that depthwise separable convolutions are well-suited for this domain.

These results establish a strong foundation for evaluating MosQNet-SA’s performance, demonstrating that achieving high accuracy in mosquito classification is possible but typically requires significant computational resources. The challenge lies in maintaining this high performance while dramatically reducing model complexity—a gap that MosQNet-SA addresses through its innovative architectural design.

MosQNet-SA Performance Analysis: The per-class performance metrics reveal that MosQNet-SA achieves high classification accuracy across all three mosquito species, with robust performance for Anopheles (99.8% precision and recall) and consistently high performance for Aedes and Culex (98.4% and 98.2% F1-scores, respectively). The narrow confidence intervals (all within ±0.5%) demonstrate the model’s stability and reliability across multiple experimental runs. This performance is noteworthy given that MosQNet-SA achieves these results with only 388K parameters—a 70-fold reduction compared to the best-performing baseline model (DenseNet201 with 26.2M parameters).

The balanced performance across all species (macro average F1-score of 98.8%) indicates that MosQNet-SA successfully captures the distinctive morphological features of each mosquito genus without bias toward any particular class. This balanced performance is essential for public health applications, where misclassification of any species could have serious consequences for disease surveillance and control efforts.

The evaluation process included a detailed performance analysis of various deep learning models, as presented in Table 4. Among the top performers, Xception, InceptionV3, and InceptionResNetV2 consistently excelled across key metrics, including training and validation losses and accuracy on both validation and test datasets. These models demonstrated strong generalization, reflecting effective feature extraction and robust architectures.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Detailed performances of baseline models (CNN architectures) and the proposed model with statistical validation.

https://doi.org/10.1371/journal.pone.0344970.t004

In contrast, models like VGG19 and ResNet101 showed higher losses and lower accuracies, suggesting challenges in capturing complex patterns or potential overfitting. Simpler architectures such as VGG16 and MobileNet achieved competitive results but were limited in capturing fine-grained details due to their lower parameter count. The advanced techniques in models like MobileNet (depthwise separable convolutions) and DenseNet (dense connectivity) contributed significantly to the observed performance improvements. EfficientNet models (B0, B1, and B2) demonstrated balanced performance across various metrics, validating the effectiveness of the compound scaling method in optimizing network depth, width, and resolution. It’s important to note that dataset biases, hyperparameter variations, and optimizer choices also contributed to overall performance.

This comprehensive evaluation highlights the strengths and limitations of different architectures, offering valuable insights for future model development tailored to specific tasks and datasets (Table 5).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Per-class performance metrics for MosQNet-SA with 95% confidence intervals.

https://doi.org/10.1371/journal.pone.0344970.t005

4.5 Performance analysis of MosQNet-SA

The loss curves in training and validation, shown in Fig 9, illustrate the learning progression of MosQNet-SA over 81 epochs. The model shows a steady decrease in both training and validation loss, indicating effective learning without significant overfitting. The convergence of these curves towards the later epochs suggests that the model has reached a stable performance state.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

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.

https://doi.org/10.1371/journal.pone.0344970.g009

To provide a more granular view of the performance of MosQNet-SA, the confusion matrix is analyzed: Fig 10, the confusion matrix demonstrates strong classification performance with 98.1% (95% CI: 97.5–98.7%) correct predictions for Aedes, 99.8% (95% CI: 99.4–100%) for Anopheles, and 98.5% (95% CI: 97.9–99.1%) for Culex. Most misclassifications occur between Aedes and Culex, suggesting a potential area for further investigation and improvement.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

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.

https://doi.org/10.1371/journal.pone.0344970.g010

The Receiver Operating Characteristic (ROC) curves for each class in Fig 10 provide evidence of the strong performance of MosQNet-SA. All three classes show high Area Under the Curve (AUC) values, with Aedes demonstrating excellent classification performance (AUC: 0.998, 95% CI: 0.996–1.000). The ROC curves for Anopheles and Culex show slightly lower but still excellent performance, consistent with the confusion matrix.

4.6 Statistical comparison with baseline models

McNemar’s test was conducted to assess the statistical significance of performance differences between MosQNet-SA and the top three baseline models (DenseNet201, MobileNetV2, EfficientNetB2). Results demonstrate that MosQNet-SA significantly outperforms all baselines (p < 0.001 for all comparisons after Bonferroni correction) (Table 6).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Statistical comparison of MosQNet-SA with top baseline models.

https://doi.org/10.1371/journal.pone.0344970.t006

The effect sizes indicate small to medium practical significance, with MosQNet-SA achieving meaningfully better classification accuracy while maintaining substantially fewer parameters.

4.7 Comparative analysis of model performance with the novel model

To thoroughly assess MosQNet-SA’s performance, a comprehensive comparison with 16 established neural network architectures was conducted, with results presented in Table 4. This table highlights critical metrics, including total parameters, test accuracy, and loss. Notably, MosQNet-SA stands out with a meager parameter count of 388,349 yet achieves the highest test accuracy of 99.42% among all models considered. For further insights into performance, Fig 11 provides additional metrics such as precision, recall, and F1-score. These metrics highlight MosQNet-SA’s exceptional performance with a precision of 0.987925, a recall of 0.988189, and an F1-score of 0.988033, underscoring its efficacy despite its compact architecture compared to larger models.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 11. The bar chart showing the performance metrics comparison across CNN architectures, where MosQNet-SA achieves the highest scores.

https://doi.org/10.1371/journal.pone.0344970.g011

MosQNet-SA stands out as a paradigm of efficiency in neural network design, boasting 388,349 parameters—a stark contrast to its larger counterparts, as shown in Figs 11 and 12. For instance, MobileNet, the next most miniature model, employs 7,458,243 parameters, making MosQNet-SA nearly 19 times more parameter-efficient. Even more substantial models like ResNet152 escalate parameter counts to 66,794,627, dwarfing MosQNet-SA by 172. This reduction in parameter complexity is not just a numerical feat but holds profound implications for practical deployment in resource-constrained environments. Even though MosQNet-SA is small, it performs very well. It achieves a remarkable 99.42% test accuracy and high precision, recall, and F1-score, beating even larger, more parameter-heavy models like DenseNet201 and InceptionResNetV2. These findings underscore MosQNet-SA’s efficacy in balancing model size with performance, setting a new benchmark for efficient neural network architectures.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 12. MosQNet-SA comes out as the best model considering the size, parameters, and F1 scores.

https://doi.org/10.1371/journal.pone.0344970.g012

Table 4 shows the performance of ResNet models (ResNet50, ResNet101, ResNet152) shows diminishing returns with increased depth. ResNet152, despite having the highest parameter count, achieves lower accuracy (90.49%) compared to ResNet50 (91.64%). This observation suggests that adding more layers does not guarantee improved performance and may even degrade performance. DenseNet201 achieves the second-highest accuracy (99.14%) and firm performance across other metrics, suggesting its dense connectivity pattern is highly effective. However, it requires significantly more parameters (26,221,379) than MosQNet-SA, highlighting the efficiency gap between the two architectures.

EfficientNet models (B0, B1, B2) show consistent performance improvements with increased model size, validating the compound scaling approach. Nevertheless, they are outperformed by MosQNet-SA in both efficiency and accuracy, suggesting that uniform scaling strategies may have limitations in effectiveness. MobileNet and MobileNetV2, designed for efficiency, demonstrate strong performance considering the focus. However, MosQNet-SA surpasses them in both parameter count and all performance metrics, suggesting that MosQNet-SA’s architecture may offer a superior approach to designing efficient models.

5.1 GradCAM

Gradient-weighted Class Activation Mapping (GradCAM) is a visualization method for CNNs that is instrumental in computer vision tasks. GradCAM computes the gradient of the target class score with respect to the feature maps of a convolutional layer:

(22)

Here, represents the importance of the -th feature map for the class , and is the number of pixels in the feature map. The weighted sum of the feature maps, followed by a ReLU activation, produces the final class activation map :

(23)

This heatmap highlights the regions of the input image most influential to the model’s decision-making process.

5.2 Saliency

Saliency methods aim to identify which pixels or input features significantly influence a model’s output. Given a model , saliency methods typically compute the gradient , where represents the input features. The magnitude of the gradient indicates the sensitivity of the model’s output to the input. This method identifies the most important pixels in an input image, highlighting regions that have the greatest influence on the model’s prediction. Mathematically, for an input feature , the saliency score is expressed as:

(24)

This captures the impact of small changes in on the model’s output.

5.3 LIME

Local Interpretable Model-agnostic Explanations (LIME) is a technique that approximates the complex model locally around the input of interest . LIME samples perturbed versions of the input and generates predictions using the original model. A simpler, interpretable surrogate model , such as a linear model, is then trained to mimic the behavior of around . The surrogate model minimizes the following weighted objective function:

(25)

Here, is a locality-aware kernel that weighs perturbed samples based on the proximity to . The resulting importance score for the surrogate model’s characteristics provides local interpretability of the model’s behavior.

5.4 KernelSHAP

KernelSHAP is a model-agnostic interpretation method based on Shapley values, which come from cooperative game theory. Shapley values provide a fair allocation of the model’s output to each feature based on the contributions in all possible coalitions of features. For a model with features, the Shapley value for a feature is defined as:

(26)

Here, is a subset of all features except , and is the prediction of the model when only the features in are present. KernelSHAP approximates the Shapley values by learning a weighted linear regression model that assigns feature importance scores, ensuring a fair representation of each feature’s contribution to the output.

5.5 Data loading and preprocessing of XAI

This section details the loading and preprocessing of images for model evaluation, focusing on the “test” folder that contains three mosquito classes: Aedes, Anopheles, and Culex. We implemented a preprocessing function, central_crop_and_resize, to extract and resize each image’s central square portion to 128x128 pixels, ensuring uniform input dimensions, which are essential for model performance.

After setting up the directory paths, the lists were initialized for the image data (X) and the labels (Y). A label mapping was created to convert class names into numeric indices. One image per class was randomly selected from the “test” category to ensure sample diversity. Each image was read, converted from BGR to RGB, and processed using central_crop_and_resize, with the results appended to X and the corresponding labels to Y.

Post-processing involved converting X and Y into Numpy arrays and normalizing the pixel values by dividing them by 255.0. Subsequently, the preprocessed images and labels were visualized using Matplotlib, and the figure was saved as “the mosquitoes.png” to verify the correctness of the preprocessing before proceeding with model evaluation.

The system processes mosquito images, classifies the species, and calculates disease risk using historical data. It also includes error handling, alerts health authorities when a high risk is detected, and stores data for future research. In addition, the system can generate real-time analytics dashboards on request, making the API highly suitable for global vector-borne disease surveillance and control efforts.

5.6 The analysis on explainability

Applying multiple explainable AI (XAI) techniques to mosquito species classification reveals complementary perspectives on model decision-making, as seen from Fig 13. GradCAM generates class-discriminative localization by using gradients from the target class flowing into the final convolutional layer, highlighting anatomical features such as wing positions and species-specific body structures. Saliency maps, which compute gradients of class scores with respect to the input image, emphasize fine-grained details such as leg positioning and proboscis angles characteristic of each genus. LIME and KernelShap, both model-agnostic approaches, provide interpretations through different methodologies: LIME approximates the model locally by perturbing the input and fitting an interpretable linear model, yielding superpixel-based explanations, while KernelShap uses coalitional game theory to distribute feature importance values across image regions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

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.

https://doi.org/10.1371/journal.pone.0344970.g013

Notably, across all methods, the visualizations consistently emphasize distinctive taxonomic features used by entomologists for species identification: the resting position and scaling patterns in Aedes, the characteristic angular posture of Anopheles, and the robust body structure of Culex. This multi-technique validation suggests that the deep learning model has learned biologically relevant features rather than spurious correlations, enhancing trust in the model’s classification decisions.

6.1 API limitations and future work

Several limitations must be addressed before operational deployment:

• Validation Requirements: Prospective validation with real-world disease surveillance data and epidemiological outcomes is essential to establish clinical utility.
• Infrastructure Needs: Production deployment requires robust infrastructure, including load balancing, auto-scaling, and geographic distribution for global accessibility.
• Data Privacy: Implementation must comply with health data regulations (HIPAA, GDPR) and establish secure data handling protocols.
• Scalability: Performance during peak mosquito seasons requires stress testing and capacity planning.
• Integration: Seamless integration with existing public health surveillance systems and workflows needs development.

Future work will focus on pilot studies with public health agencies, integration with geographic information systems, and real-time validation against disease case reports.

Integrating MosQNet-SA into a RESTful API offers a scalable, stateless solution for real-time mosquito classification and disease risk assessment. This API enables researchers, health organizations, and the public to access MosQNet-SA’s classification capabilities without requiring extensive local resources or machine learning expertise. The system aims to enhance disease surveillance by analyzing mosquito species distribution and offering contextualized risk assessments for dengue and malaria.

The RESTful API provides the following endpoints:

• POST /classify: Accepts mosquito images (JPEG/PNG, max 10MB) and returns species classification with confidence scores. Request format: multipart/form-data with ’image’ field. Response: JSON with species, confidence, and inference time.
• GET /health: Returns API health status, uptime, and model version information.
• POST /risk-assessment: Accepts classification results with geolocation and temporal data. Returns a conceptual disease risk assessment based on species distribution.

6.2 Flowchart of the process

Fig 14 illustrates the MosQNet-SA API’s overall workflow, from data submission to risk verdict generation. This streamlined approach underscores the practicality of AI-driven mosquito classification systems for real-time applications, supporting global efforts to combat vector-borne diseases.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

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.

https://doi.org/10.1371/journal.pone.0344970.g014

The API operates through several key stages: users upload mosquito images and relevant location and environmental data. The backend validates the data, processes the mosquito images using MosQNet-SA, and classifies the species. The system then calculates the disease risk based on species frequency, temporal factors, and historical disease data. The final verdict, indicating the risk level for dengue and malaria, is returned to the user.

6.3 API performance benchmarks

Performance benchmarks were conducted across multiple hardware configurations to assess deployment feasibility (Table 7):

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. API performance metrics across different hardware platforms.

https://doi.org/10.1371/journal.pone.0344970.t007

Testing methodology: Each configuration was tested with 1000 consecutive requests, measuring end-to-end latency including network overhead. Values represent mean±standard deviation across 5 runs.