Planthoppers suck the sap from the phloem of rice plants, causing malnutrition and slow growth of the plants, resulting in large-scale yield reduction. Therefore, timely and effective monitoring of planthopper pests and analysis of their occurrence degree are of vital importance for the prevention of rice diseases. The traditional detection of planthopper pests mainly relies on manual methods for diagnosis and identification. However, due to the tiny size of planthopper pests, on-site manual investigation is not only time-consuming and labor-intensive but also greatly influenced by human subjectivity, making it easy to misjudge. In response to the above issues, the intelligent light traps can be used to assist in the work. When using intelligent light traps to detect dense and occluded low-resolution and small-sized planthopper pests, problems such as low accuracy, false detection, and missed detection are prone to occur. For this purpose, based on YOLOv11x, a light-trapping rice planthopper detection method by combining spatial depth transform convolution and multi-scale attention mechanism was proposed in this research.

The image data in this research were collected by multiple light-induced pest monitoring devices installed in the experimental rice fields. The images included two types of planthopper pests, the brown planthopper and the white-backed planthopper. The image sizes were both 5472 pixels ×3648 pixels, totaling 998 images. The original dataset was divided into a training set and a validation set in a 4:1 ratio. To enhance the learning efficiency of the model during training, two data augmentation operations, horizontal flipping and vertical flipping, were performed on the images in the training set. A total of 2388 images in the training set were obtained for model training, and 200 images in the validation set were used for model inference validation. To improve the model performance, first of all, the C3k2 module in the original YOLOv11x network was improved by using the efficient multi-scale attention (EMA) mechanism to enhance the perception of the model and the fusion ability of small-volume pest features in dense and occlusions. Secondly, the space-to-depth-convolution (SPD-Conv) was used to replace the Conv common convolution module in the original model, further improving the extraction accuracy of the model for low-resolution and small-volume pest features and reducing the number of parameters. In addition, a P2 detection layer was added to the original network and the P5 detection layer was removed, thereby enhancing the model's detection performance for small targets in a targeted manner. Finally, by introducing the dynamic non-monotonic focusing mechanism loss function wise-intersection over union (WIoU)v3, the positioning ability of the model was enhanced, thereby reducing the false detection rate and missed detection rate.

The test results showed that the precision (P), recall (R), mean average precision at IoU equals 0.50 (mAP₅₀) and the mean average precision at IoU thresholded from 0.50 to 0.95 with a step size of 0.05 (mAP_50-95) of the improved model on the self-built rice planthopper dataset (dataset_Planthopper) reached 77.5%, 73.5%, 80.8%, and 44.9% respectively. Compared with the baseline model YOLOv11x, it has increased by 4.8, 3.5, 5.5 and 4.7 percent points, respectively. The number of parameters has been reduced from 56 M to 40 M, a reduction of 29%. Compared with the current mainstream object detection models YOLOv5x, YOLOv8x, YOLOv10x, YOLOv11x, YOLOv12x, Salience DETR-R50, Relation DETR-R50, RT-DETR-x, the mAP₅₀ of the improved model was 6.8, 7.8, 8.6, 5.5, 5.6, 8.7, 6.9 and 6.9 percentage points higher, respectively, and it had the best comprehensive performance.

The improved YOLOv11x model effectively enhances the performance of detecting low-resolution and small-sized planthopper pests under dense and occluded insect conditions, and reduces the probability of missed detection and false detection. In practical applications, it could assist in achieving precise monitoring of farmland pests and scientific prevention and control decisions, thereby reducing the use of chemical pesticides and promoting the intelligent development of agriculture. Although this method has achieved significant improvements in multiple indicators, it still had certain limitations. Firstly, the species of planthoppers were numerous and their forms were diverse. The current models mainly targeted some typical species, and their generalization ability needed to be further verified. Secondly, due to the limitations of the data collection environment, there was still room for improvement in the performance of the model under extreme lighting changes and extremely occluded scenarios. Finally, although the number of parameters had decreased, the real-time detection speed still needed to be optimized to meet the requirements of some low-power edge devices. Future research can focus on expanding the generalization, robustness and lightweighting of more types of rice planthopper models in more complex situations.

In crop cultivation and production, pests have gradually become one of the main issues affecting agricultural yield. Traditional models often focus on achieving high accuracy, however, to facilitate model application, lightweighting is necessary. The targets in yellow sticky trap images are often very small with low pixel resolution, so modifications in network structure, loss functions, and lightweight convolutions need to adapt to the detection of small-object pests. Ensuring a balance between model lightweighting and small-object pest detection is particularly important. To improve the detection accuracy of small target pests on sticky trap images from multi-source scenarios, a lightweight detection model named MobileNetV4+VN-YOLOv5s was proposed in this research to detect two main small target pests in agricultural production, whiteflies and thrips.

In the backbone layer of MobileNetV4+VN-YOLOv5s, an EM block constructed with the MobileNetV4 backbone network was introduced for detecting small, high-density, and overlapping targets, making it suitable for deployment on mobile devices. Additionally, the Neck layer of MobileNetV4+VN-YOLOv5s incorporates the GSConv and VoV-GSCSP modules to replace regular convolutional modules with lightweight design, effectively reducing the parameter size of the model while improving detection accuracy. Lastly, a normalized wasserstein distance (NWD)loss function was introduced into the framework to enhance the sensitivity for lowresolution small target pests. Extensive experiments including state-of-the-art comparison, ablation evaluation, performance analysis on image splitting, pest density and multi-source data were conducted.

Through ablation tests, it was concluded that the EM module and the VoV-GSCSP convolution module had significant effects in reducing the model parameter size and frame rate, the NWD loss function significantly improved the mean average precision (mAP) of the model. By comparing tests with different loss functions, the NWD loss function improves the mAP by 6.1, 10.8 and 8.2 percentage compared to the DIoU, GIoU and EIoU loss functions, respectively, so the addition of the NWD loss function achieved good results. Comparative performance tests were detected wiht different light weighting models, the experimental results showed that the mAP of the proposed MobileNetV4+VN-YOLOv5s model in three scenarios (Indoor, Outdoor, Indoor&Outdoor) was 82.5%, 70.8%, and 74.7%, respectively. Particularly, the MobileNetV4+VN-YOLOv5s model had a parameter size of only 4.2 M, 58% of the YOLOv5s model, the frame rate was 153.2 fps, an increase of 6.0 fps compared to the YOLOv5s model. Moreover, the precision and mean average precision reach 79.7% and 82.5%, which were 5.6 and 8.4 percentage points higher than the YOLOv5s model, respectively. Comparative tests were conducted in the upper scenarios based on four splitting ratios: 1×1, 2×2, 5×5, and 10×10. The most superior was the result by using 5×5 ratio in indoor scenario, and the mAP of this case reached 82.5%. The mAP of the indoor scenario was the highest in the low-density case, reaching 83.8%, and the model trained based on the dataset from indoor condition achieves the best performance. Comparative tests under different densities of pest data resulted in a decreasing trend in mAP from low to high densities for the MobileNetV4+VN-YOLOv5s model in the three scenarios. Based on the comparison of the experimental results of different test sets in different scenarios, all three models achieved the best detection accuracy on the IN dataset. Specifically, the IN-model had the highest mAP at 82.5%, followed by the IO-model. At the same time, the detection performance showed the same trend across all three test datasets: The IN model performed the best, followed by the IO-model, and the OUT-model performed the lowest. By comparing the tests with different YOLO improvement models, it was concluded that MobileNetV4+VN-YOLOv5s had the highest mAP, EVN-YOLOv8s was the second highest, and EVN-YOLOv11s was the lowest. Besides, after deploying the model to the Raspberry Pi 4B motherboard, it was concluded that the detection results of the YOLOv5s model had more misdetections and omissions than those of the MobileNetV4+VN-YOLOv5s model, and the time of the model was shortened by about 33% compared to that of the YOLOv5s model, which demonstrated that the model had a good prospect of being deployed in the application.

The MobileNetV4+VN-YOLOv5s model proposed in this study achieved a balance between lightweight design and accuracy. It can be deployed on embedded devices, facilitating practical applications. The model can provide a reference for detecting small target pests in sticky trap images under various multi-source scenarios.

Accurate identification of insect pests is crucial for the effective prevention and control of crop infestations. However, existing pest identification methods primarily rely on traditional machine learning or deep learning techniques that are trained on seen classes. These methods falter when they encounter unseen pest species not included in the training set, due to the absence of image samples. An innovative method was proposed to address the zero-shot recognition challenge for pests.

The novel zero-shot learning (ZSL) method proposed in this study was capable of identifying unseen pest species. First, a comprehensive pest image dataset was assembled, sourced from field photography conducted around Beijing over several years, and from web crawling. The final dataset consisted of 2000 images across 20 classes of adult Lepidoptera insects, with 100 images per class. During data preprocessing, a semantic dataset was manually curated by defining attributes related to color, pattern, size, and shape for six parts: antennae, back, tail, legs, wings, and overall appearance. Each image was annotated to form a 65-dimensional attribute vector for each class, resulting in a 20×65 semantic attribute matrix with rows representing each class and columns representing attribute values. Subsequently, 16 classes were designated as seen classes, and 4 as unseen classes. Next, a novel zero-shot pest recognition method was proposed, focusing on synthesizing high-quality pseudo-visual features aligned with semantic information using a generator. The wasserstein generative adversarial networks (WGAN) architecture was strategically employed as the fundamental network backbone. Conventional generative adversarial networks (GANs) have been known to suffer from training instabilities, mode collapse, and convergence issues, which can severely hinder their performance and applicability. The WGAN architecture addresses these inherent limitations through a principled reformulation of the objective function. In the proposed method, the contrastive module was designed to capture highly discriminative visual features that could effectively distinguish between different insect classes. It operated by creating positive and negative pairs of instances within a batch. Positive pairs consisted of different views of the same class, while negative pairs were formed from instances belonging to different classes. The contrastive loss function encouraged the learned representations of positive pairs to be similar while pushing the representations of negative pairs apart. Tightly integrated with the WGAN structure, this module substantially improved the generation quality of the generator. Furthermore, the visual-semantic alignment module enforced consistency constraints from both visual and semantic perspectives. This module constructed a cross-modal embedding space, mapping visual and semantic features via two projection layers: One for mapping visual features into the cross-modal space, and another for mapping semantic features. The visual projection layer took the synthesized pseudo-visual features from the generator as input, while the semantic projection layer ingested the class-level semantic vectors. Within this cross-modal embedding space, the module enforced two key constraints: Maximizing the similarity between same-class visual-semantic pairs and minimizing the similarity between different-class pairs. This was achieved through a carefully designed loss function that encourages the projected visual and semantic representations to be closely aligned for instances belonging to the same class, while pushing apart the representations of different classes. The visual-semantic alignment module acted as a regularizer, preventing the generator from producing features that deviated from the desired semantic information. This regularization effect complemented the discriminative power gained from the contrastive module, resulting in a generator that produces high-quality, diverse, and semantically aligned pseudovisual features.

The proposed method was evaluated on several popular ZSL benchmarks, including CUB, AWA, FLO, and SUN. The results demonstrated that the proposed method achieved state-of-the-art performance across these datasets, with a maximum improvement of 2.8% over the previous best method, CE-GZSL. This outcome fully demonstrated the method's broad effectiveness in different benchmarks and its outstanding generalization ability. On the self-constructed 20-class insect dataset, the method also exhibited exceptional recognition accuracy. Under the standard ZSL setting, it achieved a precise recognition rate of 77.4%, outperforming CE-GZSL by 2.1%. Under the generalized ZSL setting, it achieved a harmonic mean accuracy of 78.3%, making a notable 1.2% improvement. This metric provided a balanced assessment of the model's performance across seen and unseen classes, ensuring that high accuracy on unseen classes does not come at the cost of forgetting seen classes. These results on the pest dataset, coupled with the performance on public benchmarks, firmly validated the effectiveness of the proposed method.

The proposed zero-shot pest recognition method represents a step forward in addressing the challenges of pest management. It effectively generalized pest visual features to unseen classes, enabling zero-shot pest recognition. It can facilitate pests identification tasks that lack training samples, thereby assisting in the discovery and prevention of novel crop pests. Future research will focus on expanding the range of pest species to further enhance the model's practical applicability.