Front. Sustain. Food Syst.Frontiers in Sustainable Food SystemsFront. Sustain. Food Syst.2571-581XFrontiers Media S.A.10.3389/fsufs.2025.1734639Original ResearchGET-YOLO: a lightweight deep learning framework for real-time pest detection in marginal tea plantations to support green and sustainable tea productionYuChao12†Formal analysisMethodologyData curationWriting – original draftSoftwareVisualizationFunding acquisitionBaoQiang12†Writing – review & editingData curationWangYuanjiang12ConceptualizationWriting – original draftFunding acquisitionNiXiao12Writing – original draftInvestigationProject administrationXiaoLei12Writing – original draftInvestigationData curationZhouChao12*Writing – review & editingWriting – original draftMethodologyHunan Academy of Agricultural Sciences, Changsha, ChinaYuelushan Laboratory, Changsha, ChinaCorrespondence: Chao Zhou, chaozhou@hunaas.cn
The marginal land area in Hunan Province was about 638,800 hectares. Due to the limitation of acid sticky red soil and steep slope, the low grain production capacity was suitable for tea planting. However, outbreak pests such as Basilepta melanopus (Lefèvre) and other outbreak pests seriously threatened the yield and quality of tea, and the traditional manual inspection and chemical control were difficult to meet the needs of green and high-quality tea industry. Traditional manual inspection and chemical control methods were increasingly inadequate, which would lead to ecological risks (e.g., soil and water pollution) and fail to meet the demands of sustainable agroecosystem management for the green and high-quality tea industry. In this study, a GET-YOLO recognition method for marginal tea garden was proposed, which was based on YOLO11m depth model, embedding ECA attention mechanism to enhance key feature extraction, introducing GhostConv to achieve model lightweight, and integrating transfer learning to improve small sample adaptability. The precision, recall and mAP50 of the improved model were improved by 0.42%,4.36% and 1.83%, respectively. The mAP50 was up to 87.94%, the number of parameters was reduced by 28%, and the inference time of a single image was shortened to 23 ms, which achieved the balance of “high precision, lightweight and high real-time.” The model could automatically obtain the field growth and decline dynamics of Basilepta theobromae, provide real-time data support and key basis for the green prevention and control of diseases and insect pests in marginal tea gardens and the identification of other pests in tea plants, and has important theoretical and practical value for promoting the green and efficient upgrading of tea industry in marginal land.
green prevention and controlidentification methodinsect pestintelligent identificationmarginal landteaThe author(s) declared that financial support was received for this work and/or its publication. This research was supported by Key Research and Development Program of Hunan Province of China (2023NK2013), Agricultural Science and Technology Innovation Fund Project of Hunan Province (2024CX57) and National Key Research and Development Program of China (2022YFD1600803).section-at-acceptanceAgroecology and Ecosystem ServicesIntroduction
Marginal land is a special category of agricultural land characterized by low productivity and ecological fragility due to soil constraints, water and heat limitations, and topographical restrictions (Cao et al., 2021). Compared to conventional farmland, its core differences are showed in inherently poor soil quality (as exemplified by the southern lowland red soils dominated by marginal land in Hunan, which exhibit strong acidity, low organic matter, nutrient deficiency, and high clay content), complex topography (9.35% of the province’s arable land lies on slopes steeper than 15°, concentrated in remote mountainous areas of central-southern, western, and southern Hunan with scattered plots), and fragile ecosystems (strong geographical isolation and weak natural regulatory capacity in remote regions). Although unsuitable for grain cultivation, 638,800 hectares of marginal land has become the optimal path for efficient utilization in Hunan. The tea industry in Hunan has developed mature cultivation models and established well-known brands, laying the foundation for large-scale development of marginal tea plantations.
However, the fundamental differences are result in significantly higher pest risks in marginal tea plantations between marginal land and conventional soil directly, which become a core bottleneck for sustainable green tea production. Firstly, tea plants exhibits weak stress tolerance. The poor fertility of red soil leads to stunted growth, making them far less resistant to pests and diseases compared to high-quality farmland plantations. Secondly, monitoring is extremely challenging. The rugged mountainous terrain and poor transport infrastructure hinder traditional manual inspections, which is failing to cover scattered plots and mask early symptoms of pests, resulting in delayed control measures. Thirdly, ecological pest management capabilities are insufficient. The surrounding ecosystems of marginal tea plantations remains unstable, with limited natural pest control due to scarce natural predators. Finally, the risk of a vicious cycle is emerging. During pest outbreak events, farmers’ concerns regarding yield reduction drived a substantial increase in the application of chemical pesticides. This practice not only exacerbated the risks associated with pesticide residues but also further impaired the fragile soil conditions and ecological systems of marginal lands. A self-reinforcing cycle of “pest outbreaks—excessive pesticide use—soil degradation—weaker tea plant stress tolerance” would be created. Currently, the tea plantations are facing 58 types of annual pests in Hunan, with the Basilepta melanopus (Lefèvre) being the primary threat (Ou et al., 2018). The tea plantations is average 5.5–6.0 pesticide applications/year, using about 6,000 grams/ha of chemical pesticides. The unique conditions of marginal plantations exacerbate this issue, making precise pest monitoring and green control measures an urgent necessity for sustainable development.
The existing technical framework struggle to adapt to the unique challenges of marginal tea plantations. Inadequate supply of non-chemical control technologies and incomplete prevention systems has led to excessive reliance on chemical pesticides among tea farmers. Traditional manual inspections are characterized by low efficiency and limited coverage, which fail to meet the real-time monitoring demands of these complex terrains. With advancements in deep learning and computer vision technologies, significant progress has been made in crop pest detection. Zhu et al. (2025) systematically reviewed cutting-edge developments in crop leaf disease detection using CNN (convolutional neural network) architectures, with particular emphasis on YOLO-based detection methods. Cao et al. (2025) conducted comprehensive analyses of various YOLO algorithm versions, evaluating their practical applications in crop recognition and pest identification. Su et al. (2025) developed the Pest-YOLOv8 detection model, enhancing precise pest localization through Triple-CA modules and Wise-IoU loss functions. Zhang et al. (2023) proposed an improved DCF-YOLOv8 model by optimizing the C2F module and adopting Mish functions to boost accuracy. Li et al. (2024) introduced the YOLOv8-SDPS model, which replaces standard convolutions with SD_Conv modules, incorporates GSConv and VoV-GSCSP architectures, and adds PSA self-attention mechanisms. This enhanced model achieves dual capabilities: not only identifying six major rice pests but also detecting and locating infested rice plants. Xu et al. (2022) developed a rapid identification method for small green leafhoppers using a color plate induction counting pattern combined with an improved deep learning model and sliding window algorithm, achieving an accuracy rate exceeding 89.29%. Zhu et al. (2022) enhanced the recognition accuracy of micro-insect pests and small sample pests by implementing a fine-tuned two-stage small sample learning strategy with the TPH-YOLO v5 recognition algorithm. Li (2024) successfully identified pests on Yunnan large-leaf tea plants using the “low-level freezing, high-level training” transfer learning approach with the Xception network model, reaching an accuracy of 99.17%. Liu et al. (2025) proposed an improved YOLO v8s-CFW model for detecting five types of tea pests, incorporating a Convolutional Attention Module (CBAM) to enhance feature representation. The model utilized FasterNet as the backbone network to maintain high precision while improving computational efficiency, and employed the WIoU loss function to boost detection accuracy. Li J. et al. (2025) proposed an improved YOLOv11n model called YOLO-MSLP to enhance multi-scale rice pest recognition accuracy. The model integrates the AP_BiFPN module and triplet attention mechanism MS-TAM, while employing RepViT and knowledge distillation techniques for model lightweighting. It achieves an average precision of 94.5%. Through optimizations of the YOLO series algorithm, research has improved recognition accuracy in crop pest detection and fruit surface defect identification across various crops including strawberries (Wang et al., 2025), grapes (Wu et al., 2025), tomatoes (Fan et al., 2025; Wu and Xu, 2025), citrus (Luo et al., 2023), corn (Yang et al., 2023; Guo et al., 2024), soybeans (Linqi et al., 2024; Fang et al., 2023), pears (Li C. et al., 2025), sugarcane (Xu et al., 2023), and flowers (Yang, 2023). However, existing studies still exhibited notable limitations. Most models were designed for conventional farmland crops without adequately addressing special scenarios in marginal tea gardens, such as complex terrain, pest species variations, and challenging sample collection. Some high-precision models suffered from excessive parameter complexity and computational costs, making them impractical for real-time deployment in edge devices at marginal tea gardens. There was a lack of specialized optimization models targeting dominant pests in marginal tea gardens, resulting in insufficient adaptability. This study addresses the core challenges of the tea horn beetle outbreak in Hunan’s marginal tea plantations, including inefficient manual inspections and excessive chemical control. By leveraging the fundamental differences between marginal land and conventional soil to tackle pest management complexities, we develop a high-precision, lightweight, and edge-deployable intelligent pest recognition framework. Through implementing the ECA attention mechanism, GhostConv lightweight convolution, and transfer learning improvements to the YOLO11m model, this framework enabled real-time dynamic monitoring of pests in marginal tea plantations. It would replace traditional empirical observation methods, reduce chemical pesticide usage, and resolve the conflict between ecological protection and industrial efficiency in marginal tea cultivation. The green and efficient transformation of marginal tea industries are drived, which provide a universal technical paradigm for ecological conservation and pest identification in hilly fragile areas and other small-scale crops, demonstrating significant theoretical and practical value.
Materials and methodsImage acquisition
The image acquisition site was located in Gaoqiao Base Tea Garden, located in Gaoqiao Town, Changsha County, Changsha City, Hunan Province, the geographical coordinates of this base are “113°4′33.632″ East longitude and 28°15′40.903″ North latitude.” The soil type here was typical hilly red soil. Excellent varieties such as “Zhuyeqi” and “Bixiangzao” were planted, which would provide raw materials for making famous and high-quality green teas like “Gaoqiao Yinfeng.” The main pests included Empoasca pirisuga Matumura which sucked the juice from tender shoots leading to curled leaves and withered buds, Ectropis grisescens that voraciously eats leaves, and Basilepta melanopus (Lefèvre) that feed on leaves creating holes, and the image acquisition methods were divided into two types: (1) Cannon EOS 700D camera with an image resolution of 3,456 pixels × 5,184 pixels was used to capture the yellow trap board image hanging on the tea tree (as shown in Figure 1); and (2) Carry out real-time image transmission under that tea tree for experimental identification through a special online monitor device of the insect situation (as shown in Figure 2). Through two collection methods, 806 pictures were collected. The data set was divided into the training set and the validation set in a ratio of 9:1, and the data set sample was shown in Figure 3.
Image of yellow trap plate of Basilepta melanopus (Lefèvre).
Based on the phototaxis of insects, yellow sticky traps exhibit strong attractiveness to insect pests. Hanging yellow sticky traps in pest-infested areas can achieve green prevention and control of diseases and insect pests to a certain extent.
On-line monitoring device for pest situation. (a) Presents the design drawing of the online pest monitoring device, while (b) shows the physical prototype of the device manufactured in accordance with the design drawing.
The diagram labeled “a” presents a 3D design model of an online pest monitoring device. The photograph labeled “b” shows the physical prototype of the device manufactured in accordance with the design model in “a”, which consists of core components including a yellow sticky trap, an image acquisition module, a data transmission unit, and a distribution box.
Network structure of ECA attention mechanism.
This is an ECA attention mechanism structure: First, global average pooling is applied to the input feature map, compressing each channel’s spatial info into a value to get a channel descriptor. Then, 1D convolution models cross-channel dependencies—avoiding info loss from sudden dimensionality reduction and cutting computation. Finally, Sigmoid generates channel attention weights, which multiply element-wise with the original map to enhance key features and suppress minor ones.Image preprocessing
Labelimage was used for label annotation, and the position and category of Basilepta melanopus (Lefèvre) in each image were manually annotated. The labeled data are cleaned and verified to ensure the quality and accuracy of data. This study used three data enhancement techniques to enhance the dataset: vertical flipping, random resizing, and random rotation. Each image in the data set was enhanced twice, and there were a total of 2,410 images after data enhancement (including 2,169 images in the training set and 241 images in the validation set). The number of tags Basilepta melanopus (Lefèvre) in the image data set were 20,569 in the training set and 2063 in the validation set, respectively. By applying these data augmentation techniques, this study was able to train a more stable and accurate model to detect pests.
YOLO11 network model
YOLO algorithm had become one of the most important algorithms in the field of target detection because of its real-time, one-step, multi-scale feature fusion, prediction box design and multi-task learning. In this experiment, YOLO11m was used as the identification algorithm of Basilepta melanopus (Lefèvre) in the image, and it was improved. YOLO11 was a model launched by Ultralytics team on September 30, 2024. As the latest iteration of YOLO series, it had made many innovations and optimizations on the basis of its predecessor. These improvements significantly improved its performance and adaptability, making it the preferred tool in many computer vision applications, such as object detection and tracking, instance segmentation, image classification and pose estimation. YOLO11 offered the most exceptional detection accuracy, speed and efficiency, replacing the original C2f module with a C3k2 module in the backbone network and neck structure compared to YOLOv8. At the same time, a C2PSA module similar to the attention mechanism was added after the SPPF module to further enhance the ability of image feature extraction. In addition, by introducing the header design concept of YOLOv10, YOLO11 used a decoupling header based on anchor-free, in which the regression branch used ordinary convolution blocks and the classification branch uses depth-separable convolution (DWConv), which effectively reduced redundant computation and improves efficiency.
EGT-YOLO11 model designImprovement 1: add ECA attention mechanism
In order to enhance the model’s ability to focus on key features and improve the detection accuracy, this chapter introduces the ECA (Efficient Channel Attention) attention mechanism into YOLO11m and embeds it into the feature extraction stage of the backbone network. The core idea of ECA attention mechanism is to adaptively adjust the weights of different feature channels through efficient channel attention learning, so that the model pays more attention to the features that are more important to the detection task. The structure of the ECA attention mechanism is shown in Figure 3.
The implementation process of the ECA attention mechanism is simple and efficient. Firstly, the input feature map is globally averaged and pooled, and the spatial information of each channel is compressed into a numerical value to obtain the channel description vector; then, the dependency relationship between different channels is calculated by one-dimensional convolution, which avoids the information loss caused by the sudden drop of dimensions and significantly reduces the amount of calculation; Finally, the attention weight of each channel is generated by Sigmoid activation function, and multiplied with the original feature map channel by channel to enhance the important features and suppress the secondary features. For example, assuming that the shape of the input feature map is 64 × 64 × 32, a 32 × 1 channel description vector is obtained after global average pooling, and then a 32 × 1 attention weight is generated by capturing the local correlation between channels through a one-dimensional convolution with kernel _ size of 3. Finally, the weights are multiplied by the original 64 × 64 × 32 feature map to obtain the weighted feature map.
This method could effectively learn channel attention without dimensionality reduction, and significantly improve the sensitivity of the model to key features and enhance the ability of feature expression with little increase in parameters and computation. The ECA attention mechanism provided a more accurate feature screening capability for the model, especially in complex background or small target detection scenarios, which helped to further improve the detection performance of YOLO11m.
Improvement 2: add GhostConv
In order to reduce the amount of computation and the number of parameters while maintaining the performance of the model, GhostConv was added to YOLO11m in this chapter, replacing Conv in the backbone and neck networks with GhostConv. The process of GhostConv was shown in Figure 4. The core idea of GhostConv was to use a two-stage strategy for feature extraction. First, it used a small number of 1 × 1standard convolution kernels to generate basic feature maps; then, it used 5 × 5 deep convolution to further enrich these feature maps; the final output was the feature map spliced by these two stages. For example, assuming that the shape of the input feature map was 64 × 64 × 32, first use 1 × 1 convolution to reduce the number of channels of the input feature map, and the shape becomes 64 × 64 × 16; then used 5 × 5 depth convolution to extract features from each channel feature map, and the shape remains unchanged, which was still 64 × 64 × 16, which could be seen as a series of linear transformations of the previous layer; Finally, the output feature maps of the two convolutions were stacked in the channel dimension, and the shape was changed to 64 × 64 × 32, so as to achieve efficient feature representation with low computational cost. This approach not only significantly reduced the number of convolution kernels required, but also maintained the expressiveness of the model. GhostConv module provided an effective solution, which could significantly reduce the consumption of computing resources while ensuring or even improving the performance of the model, and opened up a new path for building a more efficient and flexible deep learning model. The method was particularly suitable for application scenes with higher requirements on real-time performance and efficiency.
GhostConv.
Workflow Diagram of GhostConv: Its core idea is a two-stage feature extraction strategy. First, a small number of 1×1 standard convolution kernels generate base feature maps. Then, 5×5 depthwise convolutions further enrich these maps. The final output is the concatenation of feature maps from the two stages.Improvement three: transfer learning
Deep learning needs a lot of data to drive, so it was a very time-consuming and costly method to train a model from scratch. A common training technique in deep learning was transfer learning, which transplanted the model weight parameters trained on the source domain to the model that needed new target learning, which could not only improve the convergence speed of the model, but also reduce the occurrence of model overfitting to a certain extent. This study uses transfer learning to further improve the accuracy of the model. The process is shown in Figure 5. The pest part of the Pest24 data set had many similar characteristics with the self-made pest data set, so the Pest24 data set was used as the source domain to train the model, so that the model could fully learn the general characteristics of pests, and then the self-made Basilepta melanopus (Lefèvre) data set was used as the target domain. Then the GET-YOLO11 model was trained to make the model have the ability to identify pests in the field, and finally the GET-YOLO11 model with higher accuracy was obtained. The GET-YOLO11 model structure diagram was shown in Figure 6.
Transfer learning.
Transfer Learning Process: The Pest24 dataset is used as the source domain to train the model for learning general pest features. Then, the self-built Basilepta melanopus (Lefèvre) dataset serves as the target domain to fine-tune the GET-YOLO11 model, enabling it to identify field pests and yielding a more accurate GET-YOLO11 model.
Network structure of EGT-YOLO11.
Network structure of EGT-YOLO11,This model consists of three parts: Backbone, Neck, and Detection Head. The Backbone uses GhostConv, CSk2, SPPF, ECA, and C2PSA for feature extraction. The Neck employs Contact and Upsample to fuse features. The Detection Head (Head1-3) outputs detection results via CSk2 and GhostConv.Model evaluation index
The test was conducted under the 64-bit operating system Ubuntu 22.04.4 LTS, and the server configuration was as follows: GPU was RTX4090, memory was 64GB, and CPU is AMD EPYC 9354. It was completed in the deep learning environment built by Python-3.10, PyTorch-2.2.2 and CUDA-12.8. In the training process, the number of training batches was 32, and the total number of training rounds was 100. SGD was used to optimize the parameters, the learning rate was 0.01, the weight attenuation coefficient was 0.0005, and the learning rate momentum was 0.937.
In this study, precision (P), recall (R), average precision meant (mean average precisoin, mAP), required floating-point operations [floating-point operations per second, FLOPS, Params and FPS (frames per seconds)] to evaluate the performance of the model. The calculation formula of precision is shown in Equation 1, that of recall in Equation 2, and that of mean average precision in Equations 3,4.
P=TPTP+FPR=TPTP+FNAP=∫01P(R)dRmAP=∑i=1NAPiN
TP was a true positive sample, indicating the number of correctly identified Basilepta melanopus (Lefèvre), the prediction box was the same as the label box and the intersection ratio was greater than 0.5; FP was a false positive sample, indicating the number of false identification of Basilepta melanopus(Lefèvre); FN was a false-negative samples, indicating the number of unrecognized Basilepta melanopus (Lefèvre); N was the number of detection categories, which was 1.
Results and discussionComparative test and analysis of ECA and other attention mechanisms
In order to verify the effectiveness of ECA attention mechanism, this study analyzed the relationship between ECA attention mechanism and global attention (global attention mechanism, GAM), convolutional attention module [convolutional block attention module, CBAM and SE (squeeze and excitation networks)]. The comparison results are shown in Table 1.
Comparison of results of different attention mechanisms.
Model
P/%
R/%
mAP50/%
Floating-point operations per second/G
Params/M
YOLO11m
82.10
82.84
86.11
67.60
20.03
YOLO11m + GAM
81.50
86.14
84.14
72.90
26.57
YOLO11m + CBAM
83.39
83.69
86.19
67.90
20.29
YOLO11m + SE
80.29
86.86
86.53
67.70
20.06
YOLO11m + ECA
83.95
84.96
87.05
67.70
20.03
Values in bold represent the optimal performance data among the comparative datasets.
The mAP50 of YOLO11m + ECA reached 87.05%, which was 0.94% higher than that of the baseline model YOLO11m, ranking first among all the comparison models. Although the recall rate of YOLO11m + SE is high, ECA further improves the precision rate to 83. 95% while maintaining a high recall rate through efficient channel attention learning, which achieves the collaborative optimization of precision and recall rate. The data set of this study is based on the yellow plate collection of tea garden in Hunan, and the image of the thorax beetle is easily disturbed by the shadow of leaves and the background of yellow plate, and some samples have the characteristics of small target size. In this case, YOLO11m + GAM greatly increases the amount of calculation and the number of parameters, resulting in the imbalance of model performance, and the mAP50 is reduced to 84.14%. Although the computational cost of YOLO11m + CBAM was close to the baseline, the mAP50 was only slightly improved to 86.19% due to the simple stacking channels and spatial attention module, which did not adapt to the complex characteristics of field targets, while the accuracy of YOLO11m + SE was reduced to 80.29% due to the loss of information in the process of dimensional processing. In sharp contrast, ECA can adaptively adjust the weight of feature channels without dimensionality reduction by means of the design of one-dimensional convolution channel dependence, which not only avoids the loss of information, but also does not significantly increase the computational cost, and can accurately capture the key features of Basilepta melanopus (Lefèvre), effectively avoid the problem of feature redundancy, and avoid the problem of feature redundancy. This method thereby shows better detection performance on a field-collected dataset of Basilepta melanopus (Lefèvre).
The experimental results of the ECA attention mechanism aligned with those reported by Tang et al. (2023). Their study implemented the ECA attention mechanism to enhance key feature selection on agricultural pest detection, achieving a 73.4% mAP0.5 score on the Pest24 dataset. By applying YOLO11m + ECA, we improved the mAP0.5 to 87.05%, further demonstrating the feasibility of the ECA attention mechanism in identifying small agricultural pests. In their tea disease detection study Song et al. (2025), achieved 98.0% accuracy by integrating the EMA attention mechanism with deformable attention. However, since the dataset primarily contains single-leaf images and disease leaves exhibit large-scale contiguous features, which differ from the small target pests in this research, the EMA attention mechanism struggles to improve detection performance for small target pests.
Ablation experiment
In order to verify the effectiveness of each improvement proposed in this study, the original model YOLO11m is used as the baseline model, and a series of ablation tests are carried out for each improvement proposed. The ablation tests are carried out through different combinations of each improved module, and the results are shown in Table 2.
Ablation experiment.
ECA
GhostConv
Transfer learning
P/%
R/%
mAP50/%
Floating-point operations per second/G
Params/M
Frames per seconds /frames·s−1
×
×
×
82.10
82.84
86.11
67.60
20.03
161.29
✓
×
×
83.95
84.96
87.05
67.70
20.03
159.25
✓
✓
×
83.20
85.70
87.71
54.50
14.65
185.16
✓
✓
✓
82.52
87.20
87.94
54.50
14.65
182.34
Values in bold represent the optimal performance data among the comparative datasets.
When only ECA attention mechanism was introduced, the precision of the model increases by 83.95%, the recall increased to 84.96%, mAP50 increases to 87.05%, and the number of floating-point operations increased only slightly to 67.70 G. This showed that the ECA module captures the channel dependence through one-dimensional convolution, and could adaptively enhance the key features of the thorax beetle without dimensionality reduction, such as the edge and texture details of the beetle, which effectively reduced the false detection and missed detection caused by the background interference of the yellow board, and would not significantly increase the complexity of the model. When the combination of ECA and GhostConv was introduced, the number of model parameters was reduced from 20.03 M to 14.65 M, the number of floating-point operations was reduced from 67.70 G to 54.50 G, and the frame rate was increased to 185.16 frames. At the same time, mAP50 further increased to 87.71%, and the recall rate increased to 85.70%. Although the accuracy was slightly reduced to 83.20%, the significant improvement of recall and computational efficiency showed that GhostConv achieved the initial balance of “accuracy-efficiency” by reducing redundant computation while retaining small target features through the two-stage strategy of “1 × 1 convolution dimension reduction + 5 × 5 deep convolution complementary features.” When ECA, GhostConv and transfer learning steps were integrated at the same time, the mAP50 of the model reaches 87.94%, the recall rate was significantly improved to 87.20%, and a further breakthrough in detection accuracy was achieved, which proved that transfer learning could learn the general morphological characteristics of pests through the pre-training of Pest24 source domain data set. The overfitting caused by sparse samples in the field was effectively alleviated and the slight accuracy loss of ECA + GhostConv combination was compensated by adapting the target domain data of B. theobromae, while GhostConv continued to play a lightweight role to control the computational cost and ensure the real-time of the model. To sum up, ECA is responsible for feature selection, GhostConv optimizes computational efficiency, and transfer learning improves sample adaptability. The three modules work together to promote the improved model to achieve a comprehensive optimization of “high precision-lightweight-high real-time” in the identification task of the Basilepta melanopus (Lefèvre), which verifies the effectiveness and rationality of the improvement strategy in this study.
In their tomato detection study, Tian et al. addressed challenges including small targets, severe occlusion, and complex background interference by integrating GhostConv and C2fGhost modules into a neck network, achieving a 96.31% mAP0.5 and 2.7 M ultra-lightweight parameter count (Tian et al., 2024). This validated GhostConv’s core value in agricultural small target detection. For the micro-sized tea leaf beetle, the study further fused ECA attention with lightweight GhostConv convolution at the module level, while employing transfer learning to mitigate overfitting. These enhancements significantly improved feature extraction precision and anti-interference capabilities, demonstrating the advantages of fine-tuning models for specific tea pest morphological characteristics.
Comparative test
RT-DETR, YOLOv5, YOLOv8, YOLOv9, YOLOv10, YOLO11, YOLO12 and GET-YOLO were successively used in the detection of Thorax cornutus, and the experimental results were shown in Table 3.
Comparative experiment.
Model
P/%
R/%
mAP50/%
Floating-point operations per second /G
Params/M
Frames per seconds/frames·s−1
RT-DETR
70.79
76.44
74.48
103.40
31.99
92.91
YOLOv5m
80.84
81.77
84.59
64.00
20.06
155.87
YOLOv8m
81.80
85.41
85.17
78.70
25.84
151.29
YOLOv9c
82.68
85.50
85.42
102.30
25.32
117.34
YOLOv10m
80.68
81.92
83.83
58.90
15.31
185.19
YOLO11m
82.10
82.84
86.11
67.60
20.03
161.29
YOLO12m
82.34
84.05
85.95
67.10
20.11
159.43
GET-YOLO
82.52
87.20
87.94
54.50
14.65
182.34
Values in bold represent the optimal performance data among the comparative datasets.
In the aspect of target detection core index mAP50, GET-YOLO was superior to all comparison models with 87.94% detection accuracy, which was 1.83% higher than baseline model YOLO11m, 2.77% higher than YOLOv8m and 4.11% higher than YOLOv10m. Compared with the traditional two-stage model RT-DETR, it was improved by 13.46%, which fully verified the improvement effect of the improved strategy in this study on the identification accuracy of the thorax tea beetle. On the premise of maintaining the highest mAP50, GET-YOLO showed excellent precision-recall balance: its recall rate reached 87.20%, which was the highest among all models, and effectively reduced the missed detection rate of the sparse tea pod beetle in the field. The accuracy is 82.52%, which is slightly lower than that of YOLOv9c, but higher than that of most models such as YOLOv5m and YOLOv10m. Compared with YOLO12m, GET-YOLO improves the recall by 3.15% and the precision by 0.18%, which indicates that the improved structure can effectively reduce the false detection rate caused by the background interference of yellow boards and enhance the adaptability of complex field scenes. In terms of model lightweight, GET-YOLO is the lightest scheme among all the comparison models with a parameter quantity of 14.65 M, which is 26.9% lower than benchmark YOLO11m and 54.2% lower than RT-DETR. The lightweight advantage is significant, and it has high practicability in the field deployment of tea garden. In terms of frame rate, GET-YOLO reaches 182.34 frames·s−1, which is 89.43 and 65.00 frames·s−1 higher than RT-DETR and YOLOv9c, respectively. Although it is 2.85 frames·s−1 lower than fastest YOLOv10m, it still fully meets the speed requirements of real-time pest monitoring in tea gardens on the premise of maintaining higher detection accuracy, reflecting a good balance between “speed and accuracy.” To sum up, through the collaborative optimization of ECA attention mechanism, GhostConv lightweight convolution and transfer learning, GET-YOLO is superior to the comparison model in terms of detection accuracy, model efficiency and adaptability to complex field scenes, which verifies the effectiveness and advancement of the improved strategy proposed in this study. And a bet technical scheme is provided for field real-time monitor of that Sterculia theae.
In order to visually verify the key role of the improved module in the identification of Basilepta melanopus (Lefèvre), this study shows the real bounding box through visual comparison experiments, and chooses YOLOv8m, YOLOv9c, YOLO11m and YOLO12m with better effect for visual comparison with GET-YOLO. The comparison results are shown in Figure 7. It can be seen from Figure 7 that GET-YOLO shows a lower missed detection rate and a more stable detection confidence in the detection of Basilepta melanopus (Lefèvre), especially in small target recognition and complex background adaptation scenarios, and can accurately detect Basilepta melanopus (Lefèvre) with sparse distribution, dense arrangement and multi-scale. In Figure 7a, the detection confidence of YOLOv8m and YOLOv9c was low, which reflected the lack of feature characterization ability. Although the confidence of YOLO11m and YOLO12m was improved, the confidence of most of the beetles was lower than that of GET-YOLO. GET-YOLO strengthens the key features such as texture and edge of the worm body through the ECA attention module, which effectively improves the confidence of prediction and makes the detection more accurate. In the type of images shown in Figure 7b, some individuals of the Basilepta melanopus (Lefèvre) are distributed in the top area of the sticky traps. After being stuck, the pests tend to struggle, which easily leads to the damage or incompleteness of their body morphology. This makes the identification of incomplete pest individuals located at the image boundaries a technical difficulty in practical monitoring: YOLOv8m and YOLOv9c miss the target; By virtue of GhostConv’s ability to retain small target features, combined with the generalization optimization of transfer learning, GET-YOLO not only successfully detects the target, but also has a higher confidence than YOLO11m and YOLO12m. In Figure 7c, YOLOv8m, YOLOv9c, YOLO11m and YOLO12m were not detected in the lower left, which was close to the background noise. GET-YOLO accurately captures this goal through the synergy of feature enhancement and lightweight of ECA + GhostConv. To sum up, GET-YOLO is superior to the comparison model in visualization, which intuitively verifies the synergistic value of ECA attention, GhostConv and transfer learning.
Comparison of detection effects of different models. (a-c) Present the experimental results of visual comparison performed by five models—YOLOv8m, YOLOv9c, YOLO11m, YOLO12m and GET-YOLO—on three images, where each row corresponds to one group of comparison results.
Comparison of object detection outputs from five models: YOLOv8m, YOLOv9c, YOLO11m, YOLO12m, and GET-YOLO. Three rows labeled a, b, and c display images with marked Basilepta melanopus detections and their confidence scores. Each column shows results from different models, highlighting variations in detection accuracy and precision across the images.
In this study, 806 images of marginal tea plantations in the main producing areas of Hunan were collected, and the YOLO11m model was optimized by combining ECA attention, GhostConv lightweight convolution and transfer learning, and then the SPXY method was used to divide the training set and the validation set, and the identification model of the population density of the tea pod beetle was established by deep learning method. The Results showed that the model optimized by GET-TOLO had the best performance in pest identification, the mAP50, precision and recall of the validation set were 87.94%, 89.13% and 88.47%, respectively, the single inference time was 23 ms, and the number of parameters was reduced by 28%. It showed that the lightweight deep learning technology could be used to quickly and accurately identify pests in marginal tea plantations.
However, due to the great differences in terrain, illumination and shelter degree of marginal tea plantations, the existing general pest detection model cannot be directly used, so a lightweight identification model suitable for regional characteristics must be established. There were still some deviations in the identification results of this study, which might be due to the fact that the samples were not cleaned when they were labeled, and the extreme light outliers were not eliminated in the data analysis, so the role of outlier elimination method in improving the performance of the model needs further study. Currently, integrated pest management (IPM) strategies for the Basilepta melanopus (Lefèvre) remain predominantly chemical-based, with limited research on green control technologies. Zhang et al. (2025) systematically summarized existing control methods, categorizing them into six approaches: agricultural management and cultivation practices, pathogen control, plant-derived insecticides, natural enemy insect utilization, chemical control, and color trap trapping. Qiu et al. (2025) conducted leaf immersion tests to evaluate the efficacy of 8 chemical agents and 6 biological agents against adult Basilepta melanopus (Lefèvre) in laboratory settings, followed by field trials to identify the optimal formulation for practical application.
The research would promote: (1) Iteratively optimize the existing identification algorithm, improve the identification accuracy and generalization ability of the model in the complex environment of marginal land, and expand to other categories of tea diseases and insect pests, such as tea aphids and tea leafhoppers, to build an intelligent identification system of tea diseases and insect pests covering marginal land tea gardens. (2) Taking the tea garden of typical marginal land such as steep slope and mountain in hilly red soil area of Xiangxi Prefecture, Hunan Province as the study area, the field application and verification of the model were carried out to clarify the adaptability and improvement direction of the model in the special habitat of marginal land in Hunan Province (acid clay red soil, steep mountain slope); (3) Integrating the multi-source data of environmental factors (such as red soil fertility, steep slope light and water), tea physiological and ecological indicators (such as stress resistance related indicators, new shoot growth rate) and other data of marginal land tea garden, exploring and establishing an information management model suitable for the whole production process of marginal land tea garden in Hunan Province. It provides technical support and key foundation for the efficient utilization, precise management and sustainable development of marginal land tea plantations (such as steep slopes and mountain tea plantations in hilly red soil areas of Xiangxi Prefecture, Hunan Province).
Current research in intelligent identification systems and real-time population monitoring of the Basilepta melanopus (Lefèvre) remains underdeveloped, posing significant challenges to implementing precise pest management strategies. The proposed intelligent identification method not only enhances technical accuracy and efficiency in pest detection but also holds practical significance for ecological conservation and sustainable management of tea plantations. Traditional tea garden pest control heavily relies on broad-spectrum chemical pesticides, whose excessive use leads to soil contamination, excessive pesticide residues in tea leaves, and disruption of natural ecosystem regulation. The GET-TOLO model developed in this study enables accurate monitoring of Basilepta melanopus (Lefèvre) population dynamics and identifies critical density thresholds for pest outbreaks. Tea farmers can implement targeted control measures based on this information, significantly reducing chemical pesticide usage frequency and dosage. Notably, this approach aligns with the perspective proposed by Zhang et al. (2025), who emphasized the urgent need to develop intelligent identification technologies and related products for automated population monitoring to support timely and precise pest control decisions. Therefore, this research holds substantial application value in advancing green pest management practices and promoting ecological sustainability in tea cultivation systems.
Conclusion
This study showed the pest threats faced by tea cultivation on approximately 638,800 hectares of marginal land in Hunan Province. It proposed an innovative GET-YOLO lightweight model to overcome the limitations of traditional manual detection and chemical control methods, thereby meeting the demands of a green and high-quality tea industry. The GET-YOLO model significantly enhanced key feature extraction for tea tree pests by embedding an ECA attention mechanism which built upon the YOLO11m deep learning framework, thereby improving both detection accuracy and recall rate. The model achieved an mAP50 score of 87.94%.
To achieve model lightweight, the research introduced GhostConv modle, which reduced the model parameters by 28%. Meanwhile, the transfer learning strategy improved the model’s adaptability to small sample datasets and shortened the inference time of a single image to 23 ms, achieving a balance of “high precision, lightweight, and high real-time performance.”
The results demonstrated that the GET-YOLO model outperforms traditional methods by achieving 0.42% improvement in precision, 4.36% increase in recall rate, and 1.83% enhancement in mAP50, which showed superior performance in pest detection. The GET-YOLO model automatically tracks pest population dynamics, including the Empoasca onukii Matsuda (tea green leafhopper), providing real-time data for green pest control in tea plantations. This system holds significant theoretical and practical value for advancing the green transformation of the tea industry and optimizing marginal agricultural land use, effectively driving the sustainable and efficient upgrading of marginal tea cultivation.
Data availability statement
The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author.
Author contributions
CY: Formal analysis, Methodology, Data curation, Writing – original draft, Software, Visualization, Funding acquisition. QB: Writing – review & editing, Data curation. YW: Conceptualization, Writing – original draft, Funding acquisition. XN: Writing – original draft, Investigation, Project administration. LX: Writing – original draft, Investigation, Data curation. CZ: Writing – review & editing, Writing – original draft, Methodology.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that Generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
ReferencesCaoX. M.SunB.ChenH. B.ZhouJ. M.SongX.LiuX. W.. (2021). Approaches and research progress on expanding marginal land productivity and enhancing ecological benefits in China. Chin. Acad. Sci. J.36, 336–348. doi: 10.16418/j.issn.1000-3045.20201228002CaoL.XiaoW.LiX. (2025) Overview of the YOLO algorithm and its application in crop recognition and pest detection. J. Zhongkai Univ. Agric. Eng. doi: 10.3969/j.issn.1674-5663.2025.06.010FanX. P.ZhangY. Q.ZhouS.RenM. F.WangY. W.CaiX. L.. (2025). Recognition and position method of tomato picking robot based on improved YOLOv8s and RGB-D information fusion. Trans. Chin. Soc. Agric. Eng.41, 106–116. (in Chinese with English abstract). doi: 10.11975/j.issn.1002-6819.202503181FangW. B.GuoY. G.GuanF. C.ZhangW.LiuQ. Q.WangS. W.. (2023). Soybean leaf wormhole recognition based on improved YOLO v5s algorithm. J. Hunan Agric. Univ.49, 127–132. doi: 10.13331/j.cnki.jhau.2023.01.018GuoJ.WangR.NanJ.XiaohuL. I.ChangzheJ. (2024). YOLOv5 model integrated with GRN attention mechanism for insect pest recognition. Trans. Chin. Soc. Agric. Eng.40, 159–170. (in Chinese with English abstract). doi: 10.11975/j.issn.1002-6819.202310226LiZ. (2024). Research on tea disease and pest identification based on transfer learning and convolutional neural networks. China: Yunnan Agricultural University.LiL.LiM.LiZ. (2024). Method for rice pest identification based on improved YOLO v8. Jiangsu Agric. Sci.52, 209–219. doi: 10.15889/j.issn.1002-1302.2024.20.025LiJ. H.XuY. L.LuJ.LiS. H.CaiX.. (2025). Lightweight multi-scale rice pest recognition model based on improved YOLOv11n. Trans. Chin. Soc. Agric. Eng.41, 175–183(in Chinese with English abstract). doi: 10.11975/j.issn.1002-6819.202504222LiC. X.ZhaoC. J.ZhangC.HuangW. Q.LiJ. Q.HeX.. (2025). Multi-surface defect detection method for Dangshan pears based on AICYOLOv11n model. Trans. Chin. Soc. Agric. Eng.41, 320–328. (in Chinese with English abstract). doi: 10.11975/j.issn.1002-6819.202504201LinqiZ.XiaomingL.HongminS.YingpengH. (2024). Research on CBF-YOLO detection model for common soybean pests in complex environment. Comput. Electron. Agric.216:108515. doi: 10.1016/j.compag.2023.108515LiuR.WangL. J.WangQ. H.LinX. D.GuoQ. H.XuD. L.. (2025). Detection of tea pests and diseases based on improved YOLOv8s. Acta Agric. Zhejiangensis37, 1933–1942. doi: 10.3969/j.issn.1004-1524.20240836LuoD.XueY.DengX.YangB.ChenH.MoZ. (2023). Citrus diseases and pests detection model based on self-attention YOLOV8. IEEE Access11, 139872–139881. doi: 10.1109/ACCESS.2023.3340148OuG. C.WangY. J.YangF. C.LiG. H.BaoQ.XiaoL.. (2018). Integration and application of green prevention and control techniques for tea pests and diseases in Hunan Province. China Plant Protect. Guide38, 78–80.QiuH.TanR. R.ChenX.WangH.HuangD.WuP.. (2025). Study on safe chemical control of Basilepta melanopus. J. Tea Sci., 1–12. doi: 10.13305/j.cnki.jts.20251103.002SongJ.ZhangY.LinS.HanH.YuX. (2025). TLDDM: an enhanced tea leaf pest and disease detection model based on YOLOv8. Agronomy15:727. doi: 10.3390/agronomy15030727SuB.ZhuY. Y.LinY.ZhuY. (2025). Pest-YOLOv8: enhanced detection for small and complex agricultural pests using triple attention and wise-IoU. J. Plant Dis. Protect.132:135. doi: 10.1007/s41348-025-01134-wTangZ.LuJ.ChenZ.QiF.ZhangL. (2023). Improved Pest-YOLO: real-time pest detection based on efficient channel attention mechanism and transformer encoder. Ecol. Inform.78:102340. doi: 10.1016/j.ecoinf.2023.102340.TianZ.HaoH.DaiG.LiY. (2024). Optimizing tomato detection and counting in smart greenhouses: a lightweight YOLOv8 model incorporating high-and low-frequency feature transformer structures. Network, 1–37. doi: 10.1080/0954898X.2024.2428713WangS. C.WangH. Q.DingX. M.DuX. P.YinY. L.CuiJ. Y.. (2025). Pest monitoring in strawberry greenhouses using improved YOLOv8n. Trans. Chin. Soc. Agric. Eng.41, 184–193(in Chinese with English abstract). doi: 10.11975/j.issn.1002-6819.202411131WuS. C.MaoY. M.HuH. Z.ZhangW. K.XingS. Y.DuanH. Q.. (2025). Detecting grape leaf diseases using improved YOLO11n. Trans. Chin. Soc. Agric. Eng.41, 140–147. (in Chinese with English abstract). doi: 10.11975/j.issn.1002-6819.202503032WuL.XuX. (2025). Lightweight tomato leaf disease and pest detection method based on improved YOLOv10n. Smart Agric7, 1–10. doi: 10.12133/j.smartag.SA202410023(in Chinese with English abstract)XuR.JinZ. H.LuoL. W.FengH. Q.FangH. H.WangX. C.. (2022). Monitoring and precise prevention and control of small green leafhopper with sound and light. China Tea Process.3, 28–33. doi: 10.15905/j.cnki.33-1157/ts.2022.03.005XuW.XuT.ThomassonJ. A.ChenW.KarthikeyanR.TianG.. (2023). A lightweight SSV2-YOLO based model for detection of sugarcane aphids in unstructured natural environments. Comput. Electron. Agric.211:107961. doi: 10.1016/j.compag.2023.107961YangH. (2023). Research on the identification method of clivia shell worm pests based on deep learning. China: Guizhou University.YangS.XingZ.WangH.DongX.GaoX.LiuZ.. (2023). Maize-YOLO: a new high-precision and real-time method for maize pest detection. Insects14:278. doi: 10.3390/insects14030278, 36975962ZhangL.DingG.LiC.LiD. (2023). DCF-Yolov8: an improved algorithm for aggregating low-level features to detect agricultural pests and diseases. Agronomy13. doi: 10.3390/agronomy13082012ZhangY. P.PanZ. P.YuX. Q.HuangS. H.LiC. Y.LiaoZ. X.. (2025). Research progress on biology, ecology and control of Basilepta melanopus (Coleoptera: Chrysomelidae). J. Environ. Entomol., 1–10. Available online at: https://link.cnki.net/urlid/44.1640.q.20250314.1714.004ZhuX.NieH.ZhouX. (2022). Pest recognition method based on TPH-YOLOv5 and small sample learning. Computer Science49, 257–263. doi: 10.11896/jsjkx.221000203ZhuR.ZhangJ. Y.HuangJ. C.KangR.ChenK. J.. (2025). Research progress of crop leaf disease detection based on convolutional neural network. Trans. Chin. Soc. Agric. Eng.(in Chinese with English abstract)41, 15–28. doi: 10.11975/j.issn.1002-6819.202502117
Edited by: Tongcheng Fu, Hunan Agricultural University, China
Reviewed by: Chaochen Tang, Guangdong Academy of Agricultural Sciences (GDAAS), China