We design a novel Feature Aggregation Module (FAM) that enhances the representation of multiple pest objects by aggregating multi-scale features from support and query images using an attention mechanism.

We integrate SCL into the object detection framework to improve feature discrimination, promoting intra-class similarity and inter-class dissimilarity among pest species.

We introduce a balancing mechanism in the loss function using focal loss to mitigate the impact of class imbalance in pest detection datasets.

We conduct extensive experiments on benchmark pest detection datasets to validate the effectiveness of our proposed method, demonstrating significant improvements over baseline models.

The proposed method provides a practical solution for agricultural pest management by enabling accurate detection of critical pests with minimal annotated data (Ragu and Teo, 2023). Its ability to handle few-shot scenarios ensures timely responses to pest outbreaks, reducing reliance on pesticides and promoting sustainable practices.

The remainder of this paper is organized as follows: Section 2 reviews related work on pest detection, few-shot learning, and contrastive learning. Section 3 details our proposed methodology, including the Feature Aggregation Module (FAM), the SCL approach, and the multi-task loss function. Section 4 presents experimental setups and results, and Section 5 concludes with future directions for research.

We integrate a Feature Pyramid Network (FPN) into the backbone network to extract features from different scales. Specifically, we obtain feature maps from multiple levels (C2, C3, C4, C5) of the backbone network (e.g., ResNet) and generate multi-scale feature maps { P 2 , P 3 , P 4 , P 5 , P 6 } through 1×1 and 3×3 convolution operations. This multi-scale feature extraction ensures the model’s sensitivity to targets of various sizes.

For each Region of Interest (RoI) in the support set and query set, we perform RoI Align operations on these multi-scale feature maps to obtain fixed-size feature representations (e.g., 7×7). These feature representations preserve spatial information and contextual relationships, providing rich features for subsequent feature aggregation.

The Feature Aggregation Module (FAM) consists of the following components: Feature Mapping, maps the features of the support set and query set into query (Q), key (K), and value (V) spaces, as shown in Equation 1. Attention Mechanism, calculates the similarity between queries and keys to obtain the attention weight matrix. Feature Fusion, uses attention weights to perform weighted summation of values, achieving feature aggregation.

Implementation Details:

Through the feature aggregation process described above, the feature representation of the query set is enhanced in several ways. Multi-scale information fusion leverages features from the support set’s multi-scale feature maps, providing rich scale information that aids in detecting targets of various sizes. Multi object feature enhancement is achieved by supplementing high-level feature maps with low-level features from the support set, which preserves details that are often lost for multi objects. Additionally, the support set’s features provide valuable contextual information, helping the model understand the relationship between the target and its background. The module offers several advantages: it improves feature representation by effectively utilizing the rich information from the support set, enhances flexibility and scalability through an attention-based relational modeling approach that adaptively adjusts the influence of the support set on the query set, and allows for easy integration into existing object detection frameworks with minimal computational overhead.

We apply a projection head Head ( · ) to the enhanced features F e n h a n c e d output from the Feature Aggregation Module (FAM) to map them into the contrastive learning feature space, as expressed by Equation 5.

Where Normalize (·) denotes L 2 normalization to ensure the feature vectors lie on a unit hypersphere.

In SCL, label information is used to construct positive and negative sample pairs. Positive samples consist of a query sample i and support samples p ∈ P ( i ) that belong to the same class. Negative samples consist of the query sample i and support samples a ∈ A ( i ) from different classes. Here, P ( i ) represents the set of samples in the same class as sample i, while A ( i ) represents the set of all samples except sample i.

We adopt the supervised contrastive loss function to optimize the feature representation, which is formulated as Equation 6.

Where τ is the temperature parameter controlling the smoothness of the distribution.

By minimizing the supervised contrastive loss, the model is guided to achieve intra-class compactness, where features of the same class are closer together, enhancing similarity within each class. It also promotes inter-class separation, pushing features of different classes farther apart and increasing dissimilarity between classes. This optimization helps the model classify more accurately and reduces misclassification.

SCL may be affected by sample imbalance, where minority classes have insufficient samples, causing the model to bias toward majority classes. To address this, we introduce an imbalance correction mechanism.

Specifically, we incorporate class weights w y i into the supervised contrastive loss, adjusting according to the number of samples in each class, as defined in Equation 7.

where N y i is the number of samples in class y i . The loss function Equation 6 becomes Equation 8.

This adjustment prioritizes minority class samples in the loss function, prompting the model to focus more on learning these classes.

The introduction of SCL will enhance the discriminability of the feature space, reduce misclassification, and thus improve classification accuracy; at the same time, through the imbalance correction mechanism, the model can learn the minority classes more fully and adapt to imbalanced data; finally, better feature representation helps the model perform better on unknown data and enhances generalization capabilities.

To jointly optimize the model’s components, we design a comprehensive multi-task loss function that includes localization loss, classification loss, feature aggregation loss, and supervised contrastive loss, which is defined in Equation 9.

Among them, L cls is the classification loss, which measures the model’s prediction accuracy of the target class, L reg is the regression loss, which measures the model’s positioning accuracy of the target bounding box,   L agg is the loss of the Feature Aggregation Module (FAM), which may include the regularization term of the attention mechanism, L SCL is the supervised contrast loss, which enhances the discriminability of feature representation, λ 1  and λ 2 are trade-off coefficients that adjust the impact of each loss term.

We employ Focal Loss for the classification loss L cls   to address sample imbalance, especially in scenarios with imbalanced positive and negative samples. The Focal Loss is defined as Equation 10.

Among them, p t is the model’s predicted probability of the true class, α t is the class weight, which balances the impact of the number of samples in different classes, and γ  is the adjustment factor, which reduces the loss contribution of easy samples and focuses on hard samples. Through Focal Loss, we can reduce the impact of a large number of easy negative samples on the loss, so that the model can pay more attention to hard positive samples.

For the regression loss L reg , measuring bounding box localization accuracy, we use the Smooth L 1 Loss as expressed in Equation 11.

Where t i is the predicted bounding box parameter, v i is the truth bounding box parameter, { x , y , w , h } are location information of the box and smooth_L1 (·) is the Smooth loss function.

To ensure effective utilization of support set information and prevent overfitting or redundancy due to the attention mechanism, we introduce the feature aggregation loss L agg , consisting of attention regularization and sparsity constraint.

We employed a multi-task learning approach to facilitate collaborative optimization among various model components. In each training iteration, localization loss, classification loss, feature aggregation loss, and supervised contrastive loss were computed. These losses were then combined into a single cumulative loss L total . Backpropagation and parameter updates were performed based on this total loss, ensuring joint optimization of all components. This approach promotes better feature representation, faster convergence, and improved generalization by encouraging mutual information sharing among tasks.

To stabilize training and prevent initial oscillations, we adopted a piecewise or cosine annealing learning rate decay schedule. This strategy lowers the learning rate in a controlled manner, allowing the model to converge steadily. For the optimizer, we used Stochastic Gradient Descent (SGD) with momentum to accelerate convergence and smooth out gradients. The momentum factor was tuned on the validation set to achieve the best balance between convergence speed and stability.

To expedite convergence and leverage prior knowledge, we initialized model parameters using ImageNet-pretrained backbone weights. Newly added modules, such as the Feature Aggregation Module (FAM) and the projection head for contrastive learning, were initialized using Kaiming initialization. This approach ensures that important structural components inherit robust feature representations while newly introduced parameters adapt rapidly.

To mitigate overfitting, we applied L2 regularization (weight decay) in the optimizer. Additionally, dropout was introduced in fully connected layers and within the FAM to stochastically deactivate a fraction of neurons during training. This not only prevents the model from over-relying on specific neurons but also improves its capacity to generalize to unseen data.

We employed a variety of image augmentation techniques, including random cropping, rotation, flipping, and color jittering, to increase data diversity and reduce overfitting. For class imbalance issues—especially in few-shot scenarios—we performed sample balancing by oversampling minority classes or undersampling majority classes, aiming to achieve a more balanced and representative training set. This augmentation and balancing strategy is particularly critical in 3-shot, 5-shot, and 10-shot experiments, where the training samples are limited.

We continuously tracked training progress by observing the loss curves (localization, classification, feature aggregation, and supervised contrastive) to ensure stable convergence. Model performance was periodically evaluated on a validation set using metrics such as mean average precision (mAP) or recall. If the performance began to plateau or degrade, we adjusted hyperparameters—including learning rate, momentum, and regularization factors—accordingly.

In our loss function, the coefficients λ 1 and λ 2 determine the relative importance of each sub-loss. We fine-tuned these values based on validation performance, ensuring that no single loss term dominated the training.

Of particular importance is the temperature parameter τ in the supervised contrastive loss, which controls the smoothness of the probability distribution when computing similarities among samples. Proper tuning of \(\tau\) helps stabilize contrastive learning by balancing the separation between positive and negative pairs. After addressing reviewer concerns, we corrected the temperature parameter usage by referencing the optimal settings reported in the official FSCE(Sun, B et al., 2021) experiment. Few-Shot Settings: For 3-shot, 5-shot, and 10-shot training, the positive sample IoU thresholds were set to 0.6, 0.7, and 0.8, respectively. Temperature Coefficients τ : Consistently set to 0.2 for 3-shot, 5-shot, and 10-shot. Aggregate Loss Weights λ 1   and Comparison Loss Weights λ 2 : Set to 0.2, 0.5, and 0.5, respectively, in the 3-shot, 5-shot, and 10-shot scenarios.

To safeguard against unexpected interruptions, we regularly saved model checkpoints during training. Each checkpoint contained the model weights, optimizer state, and current learning rate. After completing training, we selected the best-performing checkpoint based on validation metrics for final testing and deployment. This ensures that the model used in downstream tasks represents the most robust and accurate version learned during training.

By implementing the above multi-task joint training strategy with detailed hyperparameter tuning, our model demonstrated stable and efficient training, fully harnessing the benefits of collaborative optimization. The experimental results (presented in Section 4) indicate marked improvements in both convergence speed and overall performance, corroborating the effectiveness of these methodologies. Additionally, fine-tuning in the two-stage Faster R-CNN architecture proved essential for adapting the model to specific datasets and tasks, yielding enhanced robustness and accuracy. This tailored approach ensures alignment with the unique characteristics of real-world applications, thereby solidifying the model’s practical relevance.

Compared to few-shot classification and regular object detection tasks, FSOD faces more challenges. Its training dataset is mainly divided into two classes: base classes, with abundant annotated data, and novel classes, with limited annotated data. The main goal of FSOD is to significantly improve detection performance for novel classes while maintaining high detection accuracy for base class. FSOD effectively reduces the dependence of object detection models on large amounts of training data, solves the problem of imbalanced annotations in training data, and has significant practical value and a wide range of applications.

This study compared three classic FSOD algorithms: YOLO (Khanam and Hussain, 2024), TFA (Wang et al., 2020) VFAr43 (Han et al., 2023) and FSCE (Sun et al., 2021). Experiments were conducted on the Ubuntu operating system, using Python as the main development language, based on the PyTorch deep learning framework, with mmfewshot used for FSOD model training and testing. The hardware environment included two NVIDIA GeForce RTX 4090 GPUs with 24G VRAM each, an Intel(R) Xeon(R) CPU E5–2680 v3, and 64G of memory.

In experimental hyperparameter settings, SGD was selected as the optimizer, with an initial learning rate of 0.02, a batch size of 4, and 18,000 training iterations, with model evaluation intervals of 3,000 iterations. During the fine-tuning stage, the learning rate was adjusted to 0.001, and iteration numbers and evaluation intervals were adjusted according to different novel classes. During 3-shot, 5-shot, and 10-shot training, the IoU threshold for positive samples was set to 0.6, 0.7, and 0.8, respectively, the temperature coefficient was set to 0.2, and contrastive loss weights were set to 0.2, 0.5, and 0.5 respectively.

T P , F P , and F N represent true positive, false positive, and false negative, respectively. Precision and recall are defined as Equations 16, 17, respectively.

When the sum of IoU between the predicted box and the target box exceeds 0.5, the predicted box is positive, otherwise it is negative.

AP represents the area below the precision-recall curve, calculated as shown in Equation 18, with accuracy as the ordinate and recall as the abscissa.

In FSOD, base class performance is typically measured using bAP, while nAP is used to assess the performance of novel classes. Suppose class i ( i = 1 , 2 , … , N B ) belongs to base classes, and class j ( j = N B + 1 , N B + 2 , … , N ) belongs to novel classes (N denotes the number of the training classes), b A P and n A P can be expressed by Equations 19, 20.

In the subsequent analysis, we also utilize mAP, expressed by Equation 21.

where c represents the class, n represents the number of classes, and m A P represents the average A P of multiple classes. The overall effect of multi-class target detection can be represented by m A P .

Figure 3 shows the overall loss curves of the three class split sets (split 1, 2, 3) during basic stage training. As can be seen from the figure, the loss curves of TFA, FSCE and the proposed method (OURS) are almost completely overlapped, indicating that the learning process of the three methods in the basic training stage is very similar. Since the variational autoencoder introduces additional loss terms during training, the loss of VFA is higher. Overall, the loss of the four methods is gradually decreasing with the increase in the number of training iterations, indicating that the model is constantly learning and improving.

Figure 4 shows the changes in mAP50 (average accuracy under IoU 0.5) of four different class splits during basic stage, and the test set is evaluated every 3000 iterations. In split1, VFA, FSCE, and the proposed method reach maximum mAP50 values of 86.6, 88.9, and 89.7 at 18,000 iterations, respectively; while TFA reaches a maximum mAP50 value of 89.1 at 15,000 iterations, but drops at 18,000 iterations, indicating possible overfitting. In split2, TFA and VFA reach maximum mAP50 values ​​of 88.6 and 85.6 at 18,000 iterations, respectively, while FSCE and the proposed method reach 89.0 and 89.8 at 15,000 iterations, and also show overfitting at 18,000 iterations. In split3, TFA and FSCE reached 85.8 and 85.3 respectively at 18,000 iterations, while VFA and the proposed method reached the maximum value of 84.5 and 85.9 at 15,000 iterations, but overfitting also occurred at 18,000 iterations. These results reflect the differences in the sensitivity of different methods to the number of training iterations and the stability in the later stages of training.

Based on these results, we will adopt the following strategy for subsequent fine-tuning: selecting the models saved at the point where mAP50 achieves the highest value in splits 1 to 3 as the starting point for fine-tuning. the proposed method is to leverage the model state that achieves optimal performance during the base stage to further enhance its performance in the few-shot object detection task.

Figures 5 – 7 show the visualization results of the relevant data after two rounds of random sampling and fine-tuning, and the results on the test set with different sample numbers (3, 5, and 10), covering splits 1, 2, and 3. The performance of each method (TFA, VFA, FSCE, and the proposed method) is measured by the average precision of the base class (bAP50), the average precision of the new class (nAP50), and the overall average precision (mAP50).As can be seen from Figures 5 – 7 , TFA and VFA show an inverse relationship in performance: TFA performs well on the base class (bAP50), but is relatively weak on the new class (nAP50), which indicates that TFA may not be able to effectively transfer knowledge to the new class. In contrast, VFA performs well in the new class but poorly in the base class, which indicates that the model may sacrifice the performance of the base class to adapt to the new class. In contrast, FSCE performs evenly in the two classes and shows better robustness. the proposed method performs better on the basis of FSCE. Under certain split and shot configurations, the proposed method even slightly outperforms VFA in terms of new classes and overall accuracy, indicating its excellent ability in balancing the performance difference between base and new classes.

In the nAP50 graph of new classes for split3, VFA outperforms other methods under 3-shot conditions; but its performance improves only slightly with the increase in sample size, increasing by only 11.04% from 3-shot to 10-shot. In contrast, the performance of the proposed method improves significantly, increasing by 23.95% from 3-shot to 10-shot. To further study this phenomenon, a third random sampling fine-tuning training experiment was conducted based on the split3 dataset.

Figure 7 shows the changing trends of bAP50 and nAP50 during split3 fine-tuning training under different shot conditions. The performance of FSCE and the proposed method in bAP50 is always between TFA and VFA, but its maximum nAP50 exceeds that of other methods, which highlights the advantage of the proposed method in balancing the performance of base and new classes.

To comprehensively compare the performance of detection methods across 20 tropical pest classes, we included the YOLO model (specifically the YOLO11x version) as a benchmark for FSOD tasks. Table 5 and relevant results incorporate the YOLO method alongside TFA, VFA, FSCE, and the proposed method (OURS). While YOLO is known for its efficiency in real-time detection tasks due to lower computational complexity, the results indicate that this advantage does not translate into better performance in FSOD scenarios. The results show that YOLO, TFA, VFA, FSCE and the proposed method all show high AP values ​​under 3-shot, 5-shot and 10-shot conditions, proving the stability of the methods. With the increase of sample size, the performance of the three methods in new classes gradually improves, especially when the sample size is small, the detection performance is significantly improved with a slight increase in sample size.

In terms of detection performance in each class, the proposed method shows an upward or stable trend in mAP value with the increase of sample size, while YOLO, TFA, VFA and FSCE have certain fluctuations. Especially in the new class, the proposed method achieved the maximum mAP value of 82% in the 10-shot experiment, which is significantly better than YOLO, TFA, VFA and FSCE. In addition, the proposed method shows particularly excellent performance in specific classes such as 15 and 95, and significantly improves AP in the challenging 101 class (Cicadellidae). Compared with other methods, its mAP value is nearly 3 times higher, reflecting the powerful feature aggregation and migration capabilities of the proposed method.

Although the mAP values ​​of most classes are above 70, indicating that the proposed method can effectively detect these pests, the mAP values ​​of other methods are relatively low for classes such as 48, 101, 76, and 95. As can be seen from the relevant images in Figure 8 , the visual features of these pests are highly similar to the background, or have features that are difficult to distinguish from other classes, making it difficult for YOLO, TFA, VFA, and FSCE methods to accurately identify them. Overall, the mAP value of the proposed method in the new class is nearly 10 percentage points higher than that of YOLO, TFA, VFA, and FSCE on average, showing its significant advantage in the tropical pest detection task.

In terms of detection performance under 3-shot, 5-shot, and 10-shot conditions, YOLO demonstrates relatively lower AP values compared to the other methods. For example, in the 10-shot experiment, YOLO achieves an mAP of 77.5%, whereas the proposed method achieves a significantly higher mAP of 82.6%. Notably, in challenging classes such as 48 and 101, YOLO struggles to distinguish pests with features similar to the background, resulting in mAP values below 60%, significantly lower than the corresponding performance of the proposed method. Overall, while YOLO provides a computationally efficient solution, the trade-off between speed and accuracy limits its applicability in FSOD tasks that prioritize precise detection over real-time processing. The proposed method strikes a better balance by achieving state-of-the-art detection performance, justifying the slightly higher computational cost for critical applications like pest management in tropical agricultural settings.

As shown in Figure 9 , from the 10-shot confusion matrix analysis of TFA, VFA, FSCE and the proposed method in split3, the proposed method has obvious advantages in terms of accuracy, missed detection rate and recall rate. First, in terms of accuracy, the proposed method presents higher values ​​on the diagonal, indicating that the model has higher classification accuracy on multiple classes. In contrast, TFA and FSCE methods have lower diagonal accuracy in some classes, showing that the recognition of some classes is not accurate enough under few-sample conditions. In particular, the TFA method has serious misclassification in some classes, while the proposed method is relatively balanced in overall accuracy. In addition, VFA has some misclassification in the background class, while the proposed method is better at distinguishing between targets and backgrounds. Secondly, in terms of missed detection rate performance, the off-diagonal misclassification rate of the proposed method is lower, which means that it has fewer missed detections. In contrast, the FSCE and VFA methods have high missed detection rates in some classes, especially between difficult-to-distinguish classes, which are prone to prediction deviation. FSCE has more obvious misclassification in medium-complexity classes, while VFA shows a tendency to misdetect when the background interference is strong, resulting in an increase in missed detection rate. the proposed method significantly reduces the missed detection rate and improves overall reliability by improving feature extraction. Finally, in terms of recall rate, the proposed method has a higher recall rate in most classes. With fewer misclassifications, the proposed method can effectively identify more real samples, especially in complex backgrounds or with few samples, and the recall performance is more stable. In contrast, the recall rate of the TFA method is low, and it is easy to make recognition errors when the class boundaries are blurred. The recall rate of FSCE is also slightly insufficient when dealing with some subdivided classes.

In summary, the proposed method is superior to other methods in accuracy, missed detection rate and recall rate. Its advantages lie in better feature extraction ability, lower misclassification rate and higher recall rate, making it a more robust model in the case of few samples and complex backgrounds. These improvements enable the proposed method to perform better classification results in the split3 10-shot scenario.

We have evaluated the effectiveness of the modules used in the study, such as the feature aggregation module(FAM), the SCL module SCL, and the multi-task loss optimization MTLF, in detail through ablation experiments. In the ablation study show in Figure 10 , we systematically introduced three key modules based on the baseline method TFA. By incorporating these modules into the baseline separately, we conducted 3-shot, 5-shot, and 10-shot experiments in the few-shot scenario, and evaluated them in terms of bAP50, nAP50, and mAP50. Through this comprehensive evaluation, we can thoroughly investigate and verify the effectiveness of each component in the framework, which helps to further fully understand the proposed method. The ablation results are shown in Figure 10 , showing the effect of the key modules. The performance is significantly improved by about 1.1% by introducing FAM alone. Specifically, mAP increases from 0.741 to 0.751 in the case of 3 shots, from 0.772 to 0.779 in the case of 5 shots, and from 0.794 to 0.805 in the case of 10 shots. In addition, the inclusion of the SCL module alone can improve its performance by about 1.5%. In the case of 3 shots, mAP increases from 0.741 to 0.753, in the case of 5 shots, from 0.772 to 0.787, and in the case of 10 shots, from 0.794 to 0.809, highlighting the effectiveness of the SCL module in addressing the multi-scale challenges encountered in pest object detection. In addition, adopting the multi-task loss optimization module as a standalone ensemble on the baseline improves the results by about 3.5%. This improvement is evident in the case of 3 shots, where mAP increases from 0.741 to 0.776, in the case of 5 shots, from 0.772 to 0.807, and in the case of 10 shots, from 0.794 to 0.826.

Figure 8 shows an example of the comparison of the new class detection results of the proposed method with those of TFA, VFA, and FSCE methods in the dataset. As shown in Figure 8 , most of the new class objects are correctly detected, demonstrating the efficiency of our model. Other methods have difficulty in effectively detecting new class multi-target situations. In Figure 5 , we can see that although the insect is similar to the background, our model correctly identifies the background and does not misidentify the insect. Similarly, although there are multiple insect targets in the image, our model can still correctly identify all the targets. Our model can effectively handle size variations and multiple targets, and correctly identify single targets and multiple targets of different sizes. Edge cases, such as overlapping pests or those camouflaged within cluttered backgrounds, posed challenges for all tested models. While the proposed method outperformed others in these scenarios, future work could explore adaptive feature learning techniques or advanced data preprocessing to further improve performance in such cases”.

To evaluate the stability and differences of the proposed method compared to other methods, statistical analysis and significance tests were conducted. As shown in Table 6 , our method outperforms the comparison methods (TFA, VFA, and FSCE) in terms of statistical metrics such as mean (Mean), standard deviation (Std), and confidence interval (CI). The mean value of OURS is 79.13, which is higher than TFA (75.30), VFA (75.07), and FSCE (77.26), indicating its superior overall performance. Furthermore, the standard deviation of the proposed method is 1.920, lower than those of VFA (2.471) and FSCE (2.233), demonstrating greater stability. Within the 95% confidence interval, the proposed method exhibits a range of (78.178, 80.088), which is significantly higher than the intervals of other methods, such as TFA (74.541, 76.059). This indicates that the proposed method holds a clear statistical advantage.

As presented in Table 7 , statistical significance tests based on multiple independent experimental results further confirm the advantages of the proposed method compared to TFA, VFA, and FSCE. Using independent t-tests at a significance level of 0.05, the results show that the p-value for the proposed method versus TFA is 0.00116, versus FSCE is 0.03284, and versus VFA is 0.00288—all below 0.05. This demonstrates that the performance of the proposed method is statistically significantly different from the other models. Additionally, the mean value of OURS is 79.13, which is higher than TFA (77.05), FSCE (77.62), and VFA (76.87). These results indicate that the proposed method not only outperforms other models in overall performance but also achieves statistically significant differences across multiple experiments. In summary, the proposed method demonstrates superior stability and performance compared to other models, highlighting its statistical advantages.

The analysis of computational complexity and detection performance highlights the trade-offs made in this study. YOLO, known for its efficiency in real-time detection tasks, achieves the lowest computational complexity with 114.5 GFLOPs and relatively moderate mAP values (66.3, 75.5, 77.5 for 3-shot, 5-shot, and 10-shot tasks, respectively). In contrast, OURS, a model based on the Faster R-CNN framework with enhancements such as FAM and SCL, achieves the highest mAP values across all settings (75.9, 79.5, 82.6) at a slightly higher computational cost of 130.2 GFLOPs. These results, summarized in Table 8 , clearly demonstrate the performance and computational trade-offs between YOLO and OURS. This demonstrates that OURS leverages the computational resources to achieve significant performance gains, particularly in few-shot detection tasks, where accuracy and robustness are critical. While YOLO is more suitable for real-time applications, its lower performance in few-shot tasks highlights its limitations in capturing fine-grained and diverse pest characteristics. Models like TFA, FSCE, and VFA strike a balance between complexity and performance, but they fall short of the proposed method in overall accuracy.

By choosing Faster R-CNN as the base framework, this study prioritizes higher detection accuracy over real-time speed, a trade-off that is justified for applications requiring precise pest management. This approach demonstrates that slight increases in computational complexity are acceptable to achieve substantial performance improvements, aligning with the study’s goal of advancing few-shot object detection in complex agricultural environments.

The proposed algorithm has been integrated into a practical pest management system, whose architectural design (as depicted at the top of Figure 10 ) addresses three specific application scenarios: Under weak network conditions, front-end devices with edge computing capabilities perform local pest detection in real-time and autonomously activate laser-based capture mechanisms. Under stable network conditions, low-cost front-end visual sensors transmit images to a backend cloud platform for rapid pest identification, subsequently triggering front-end laser capture devices, thus optimizing deployment costs. Agricultural technicians or unmanned aerial vehicles (UAVs) upload images to the backend platform, enabling precise identification and geolocation-based positioning, supporting flexible mobile monitoring. The backend cloud platform employs parallel computing to achieve millisecond-level processing and feedback, effectively fulfilling diverse scenario requirements and establishing a comprehensive intelligent pest management system encompassing real-time monitoring, rapid identification, precise localization, and targeted pest control.

As shown in the lower-left section of Figure 11 , the pest induction and laser capture device comprises key modules including a core computing board, laser emitter, galvanometer controller, and visual sensing components. Specific attractants or optical methods accurately lure pests onto designated induction panel areas. Real-time visual data captured by onboard cameras is swiftly processed by a lightweight detection algorithm developed in this research, which can also be deployed in parallel on cloud platforms to handle large volumes of data from multiple devices simultaneously. The coordinate conversion module precisely calculates the physical positions of detected pests, guiding the laser galvanometer to accurately target and activate the laser for pest capture. Captured pests are subsequently collected in designated containers for further identification and analysis. This approach effectively minimizes environmental interference and protects beneficial insects, significantly enhancing the precision and effectiveness of pest monitoring and control.

The backend platform, based on our proposed algorithm, provides a comprehensive management interface, facilitating efficient, real-time collection of pest monitoring data from greenhouses and farms. Data can be flexibly submitted by agricultural technicians via smartphones or automatically uploaded by pest induction and laser capture devices. The backend management system automatically identifies pests, clearly visualizes real-time identification results, and assigns data to corresponding greenhouse or farmland regions according to geographic locations. The detailed system processing workflow is presented in the lower-right section of Figure 10 .

To further validate the practical efficacy of our proposed few-shot pest insect detection model, we conducted an extensive field evaluation over a three-month period in vegetable greenhouses located in Haikou, Hainan Province. Situated in a tropical region, Hainan faces significant pest challenges. The evaluation specifically targeted eight prevalent pest species in this region: flea beetles, aphids, whiteflies, thrips, diamondback moths, armyworms, fruit flies, and leaf miners. We deployed a detection platform utilizing our proposed algorithm, continuously monitoring pest instances captured through smartphone images provided by agricultural technicians and integrated intelligent trapping devices. Throughout the evaluation period, a total of 563 pest instances were captured across all monitored areas. Among these, the AI model successfully identified 534 instances, yielding an overall accuracy of 94.84%. Notably, aphids and whiteflies demonstrated the highest detection accuracy, each exceeding 96%. In contrast, flea beetles exhibited slightly lower accuracy at 89.7% due to their smaller size and higher mobility.

Our methodology comprehensively addresses the dynamic and complex nature of pest monitoring environments by employing targeted detection strategies that integrate crop types, regional characteristics, and seasonal factors, significantly reducing data collection and labeling costs through few-shot learning techniques. The lightweight model design ensures effective deployment even in agricultural scenarios with limited computational resources or poor network connectivity, exhibiting robust and stable performance in greenhouse monitoring environments.