Seven types of pakchoi pests and diseases that frequently occur in production were selected as research targets: Diamondback moth, Leaf Miner, Downy Mildew, Black Spot, Black Rot, White Rust, and White Spot. Detailed visual characteristics of these pests and diseases are summarized in Table 1, and representative image samples are shown in Figure 1. To minimize selection bias and ensure a representative coverage of different infection stages, a systematic scanning protocol was adopted during image acquisition. Images of pakchoi plants were captured sequentially along cultivation rows to include early-stage, mild, atypical, and late-stage symptomatic samples. Shooting distances were standardized between 20 and 50 cm to balance feature resolution and field of view. Images were primarily captured from a vertical top-down perspective, with additional 45° oblique views to account for leaf overlap and variations in plant morphology. Image acquisition was conducted from March 21 to May 8, 2025, in Cao County (Shandong Province), Wujiang District (Jiangsu Province), and Jiading District (Shanghai). Images were collected using multiple commonly used smartphone models, including the Xiaomi 12X, OPPO Reno8 Pro, and Samsung Galaxy A53. The corresponding shooting resolutions were 4000×3000, 3024×4032, and 5632×4224 pixels, respectively.

In this study, we collected images of seven common pests and diseases of pakchoi from three different field areas, covering various growth stages and disease manifestations. Given the significant variability in illumination and meteorological conditions in open-field environments, the data acquisition process was designed to encompass multiple diurnal phases (morning, noon, afternoon) and diverse weather scenarios (sunny, overcast, post-precipitation periods). The original dataset consisted of 1,782 images. To enhance the model’s generalization ability, we performed data augmentation on these images (as shown in Figure 2), resulting in a final dataset of 6,110 images. The sample distribution across the seven pest and disease categories is as follows: 1,085 images of Diamondback Moth damage; 842 images of Leaf Miner disease; 992 images of Downy Mildew disease; 753 images of Alternaria Leaf Spot disease; 745 images of Black Rot disease; 855 images of White Rust disease; and 838 images of White Spot disease. All images in the final dataset were standardized to a resolution of 640×640 pixels in JPG format. Furthermore, the dataset included images captured during “post-precipitation” periods, which naturally contained samples with water droplet reflections and soil splashes. Meanwhile, variations in handheld movement during image capture introduced natural motion blur effects, ensuring the model’s robustness against complex field challenges.

The dataset images were manually annotated using LabelImg software (https://github.com/HumanSignal/labelImg). The following categorical labels were assigned: “Backmoth” for Diamondback Moth, “Leafminer” for Leaf Miner, “Mildew” for Downy Mildew, “ALTERNARIA” for Alternaria, “BLACK-ROT” for Black Rot, “WHITE-RUST” for White Rust, “WHITE-SPOT” for White Spot. All annotations were saved in TXT files, each containing the corresponding object class and bounding box coordinates. Multi-instance annotations were preserved where applicable, with individual images containing simultaneous occurrences of multiple pathologies. The dataset was subsequently partitioned into training, validation, and test sets with an 8:1:1 ratio, resulting in 4,888 images for training, 611 for validation, and 611 for testing.

As a next-generation end-to-end object detection algorithm, YOLOv8 significantly enhances detection performance in complex scenarios through architectural refinements and technical innovations (Ma et al., 2024), while inheriting the computational efficiency characteristic of the YOLO series. The model employs a four-module architecture: Input → Backbone → Neck → Head (Figure 3), and its core design demonstrates in-depth optimizations for real-time performance, adaptability to multi-scale targets, and model lightweighting. The model has five scaled versions (n, s, m, l, x), which satisfy the adaptation requirements of diversified application scenarios (Liu et al., 2023; Ma and Pang, 2023; Wang et al., 2023; Chen et al., 2025; Long and Lin, 2025).

This study adopts YOLOv8 as the baseline network model, following the canonical “Backbone-Neck-Head” hierarchical design paradigm, and is collaboratively composed of three core functional modules to form an efficient object detection framework (Huang et al., 2024). Among these, the backbone network, serving as the primary structure for feature extraction, incorporates the Basic Convolution (Conv) module, the Cross-Stage Partial Feature Fusion (C2F) unit, and the Spatial Pyramid Pooling Fast (SPPF) module (Xiao et al., 2024). The neck network adopts a bidirectional architecture that integrates the Feature Pyramid Network (FPN) (Xie et al., 2023) and the Path Aggregation Network (PAN) (Roy et al., 2022). Via a bidirectional connection mechanism involving top-down semantic feature transmission and bottom-up detailed feature feedback, it realizes cross-scale fusion of feature maps across different levels. The detection head employs an Anchor-Free detection approach, doing away with the reliance of traditional anchor-based mechanisms on prior target sizes (Wang et al., 2023).

However, during training on the pakchoi pest/disease dataset, the original YOLOv8 model exhibited insufficient detection accuracy and a low target recognition rate (Yue et al., 2024). To evaluate the performance differences among various model versions, comparative experiments were conducted on the YOLOv8 series (n/s/m/l/x). Mean Average Precision (mAP) served as the core evaluation metric to assess the detection performance variations across models under conditions of multi-scale target distribution and leaf occlusion scenarios. The results are detailed in Table 2.

As shown in Table 2, under the unified training configuration (200 epochs, RTX 4090 GPU, and identical hyperparameters), despite their larger parameter counts and more complex architectures, the YOLOv8m and YOLOv8l models exhibited lower detection accuracy on the pakchoi pest/disease dataset compared to the lightweight YOLOv8n model. This indicates that merely increasing model complexity failed to yield accuracy gains on this specific dataset, while significantly increasing the computational burden and inference time. Although YOLOv8x achieved the highest accuracy, its substantial parameter count resulted in excessively slow inference speeds, rendering it impractical for low-cost, high-efficiency real-world applications. In addition, the YOLOv8s model also demonstrated slightly lower accuracy than YOLOv8n. This phenomenon may be attributed to the fact that larger models with higher parameter counts typically require larger datasets or different convergence schedules to avoid redundancy and potential overfitting in specific agricultural scenarios. Considering the balance between detection accuracy, computational cost, and inference speed, YOLOv8n was selected as the baseline model for subsequent improvements. This model maintains relatively high detection accuracy while possessing the lowest parameter count and highest inference efficiency, serving as a solid foundation for algorithmic optimization.

In natural settings, pakchoi exhibits high-density planting, leading to challenges such as mutual leaf occlusion, weed interference, and overlapping leaves across different growth stages. Concurrently, pest and disease regions display high diversity in characteristics: pathogen infection manifests as lesions with distinct textures, morphologies, and colors, while insect damage presents as mechanical injuries like mines and holes. This complex background interference coupled with significant morphological variations in pests and diseases complicates the precise identification and localization of target regions by detection models. To address these challenges, this study proposes the YOLOv8-DBW model, with the improvement strategy comprising the following three key aspects:

Visualization of loss curves can intuitively reflect the convergence process of the model, thereby facilitating better adjustment of training strategies. The loss values include bounding box loss and distribution focal loss (used to evaluate the regression performance of object detection bounding boxes) as well as classification loss (used to evaluate classification performance) (Yang et al., 2023).

To systematically evaluate the stability of model performance and mitigate interference from random factors during training, this study conducted independent and repeated training and evaluation experiments for each model. The specific procedures were as follows: All experiments were performed under identical hardware (e.g., GPU model, memory configuration) and software (e.g., deep learning framework version, operating system) environments. Each model was trained independently for five repeated runs. Before each run, the model parameters and weights were reinitialized, and the input order of the training dataset was randomly shuffled to eliminate the influence of initial weights and data sequence on the results. The final reported performance metrics (e.g., accuracy, mAP) were calculated as the arithmetic mean of the results from the five runs, serving to quantify the stability of model performance. All model comparisons were based on these averaged metrics to ensure fairness and reliability in the evaluation.

As shown in Figure 10, throughout the training process, the model exhibited no signs of overfitting or underfitting, indicating that it possesses good generalization ability and the capability to capture data patterns. As the number of training epochs increased, all three types of loss values decreased continuously. After the 130th epoch, the loss curves tended to converge and stabilize, suggesting that the model had reached the optimal state and could proceed to the stage of model performance evaluation.

To verify the rationality of introducing the EMA attention mechanism, this study independently embedded it into the backbone network of the original YOLOv8n model. Meanwhile, the four attention mechanisms (SE, CA, ECA, and CBAM) were respectively introduced at the same position to conduct comparative experiments, and the experimental results are shown in Table 5. Data indicate that the model with the EMA attention mechanism introduced achieves an Accuracy of 83.2% and a mean average precision (mAP) of 80.1%, with both indicators being higher than those of the models incorporating the other four attention mechanisms. In addition, the Recall of this model is 71.6%, which is only 0.4 percentage points lower than that of the model with the SE attention mechanism introduced. Overall, the EMA mechanism shows obvious advantages in terms of core detection accuracy indicators.

To verify the effectiveness of each improvement in the modified YOLOv8n, this study set up 8 ablation experiment schemes to validate the effectiveness of the modified YOLOv8n modules, with the results shown in Table 6.

As shown in Table 6, after introducing the C2f-PE module, the feature extraction capability of the model was notably enhanced compared with the original baseline. Specifically, Precision (P), Recall (R), and mean Average Precision (mAP) increased by 1.9, 1.0, and 2.7 percentage points, respectively. This improvement can be attributed to the EMA attention mechanism, which effectively suppresses interference caused by occlusion and enhances the model’s focus on small-target features. Meanwhile, due to the lightweight convolutional design of the PConv module, the number of floating-point operations (FLOPs) was reduced by 0.8 G, while the model size and parameter count decreased by 0.3 MB and 0.5 M, respectively. After further integrating the BiFPN module, the model achieved additional performance gains and improved computational efficiency. Compared with the original model, P, R, and mAP increased by 4.3, 4.3, and 6.5 percentage points, respectively. At the same time, the number of parameters and model size were reduced by 33.3% and 31.8%, while FLOPs decreased by 13.6%, indicating a significant improvement in lightweight performance. These results demonstrate that BiFPN effectively enhances multi-scale feature fusion while reducing redundant computations. Finally, after replacing the original loss function with the Wise-IoU loss, the model’s detection performance was further improved, with P, R, and mAP increasing by 5.0, 5.5, and 7.5 percentage points, respectively. This result suggests that Wise-IoU improves the accuracy and stability of bounding box regression, thereby enhancing overall detection robustness. Based on the ablation experiment results, each individual module contributes positively to performance improvement. When all proposed modules are combined, the model achieves optimal performance across all evaluation metrics, confirming the effectiveness of the proposed improvements.

To verify the effectiveness of the proposed Wise-IoU loss function in the pakchoi pest and disease detection task, training was conducted using YOLOv8’s default CIoU as well as existing mainstream regression loss functions including DIoU, GIoU, SIoU, and EIoU. The evaluation metrics adopted mAP at IoU thresholds of 0.5 and 0.5–0.95. The experimental results are shown in Table 7. As indicated in Table 6, compared with the default CIoU of YOLOv8n, the proposed Wise-IoU increased mAP@0.5 and mAP@0.5:0.95 by 1.5 and 1.3 percentage points respectively. Among all comparative methods, Wise-IoU achieved the optimal precision, verifying its superiority in agricultural disease detection scenarios.

To evaluate the performance of the proposed YOLOv8-DBW model, comparative experiments were conducted against several mainstream object detection methods, including Faster R-CNN, SSD, YOLOv5s, YOLOv5n, YOLOv7-tiny, YOLOv10n, YOLOv11n, and YOLOv12n. To ensure fairness and scientific rigor, all benchmark models were retrained on the same pakchoi pest and disease dataset using identical hardware environments and hyperparameter configurations, as specified in Tables 3 and 4. To minimize the effects of experimental randomness, each model was independently trained five times, and the reported performance metrics represent the arithmetic mean of the five runs. The comparison results are summarized in Table 8. As shown in Table 8, the proposed YOLOv8-DBW model achieved superior detection performance compared with all benchmark models. Specifically, its mean average precision (mAP) exceeded that of Faster R-CNN, SSD, YOLOv5s, YOLOv5n, YOLOv7-tiny, YOLOv10n, YOLOv11n, and YOLOv12n by 23.4, 19.1, 10.0, 11.5, 14.0, 6.9, 7.8, and 12.1 percentage points, respectively. Meanwhile, the number of model parameters was reduced by 96.7%, 94.4%, 71.5%, 20.0%, 66.7%, 25.9%, 20.0%, and 20.0% compared with the corresponding models. Although the FLOPs of YOLOv8-DBW are slightly higher than those of YOLOv5n and approximately 0.6 G and 1.0 G higher than those of YOLOv11n and YOLOv12n, respectively, they remain substantially lower than those of the other compared models. In addition, the model size of YOLOv8-DBW is reduced by 96.7%, 91.3%, 73.7%, 5.3%, 69.5%, 33.3%, 30.8%, and 29.5%, respectively, meeting the requirements for lightweight deployment. Although YOLOv5n, YOLOv11n, and YOLOv12n achieve higher inference speeds in terms of frames per second (FPS), their parameter counts and model sizes are larger than those of YOLOv8-DBW, and their detection accuracy remains lower. Overall, the proposed YOLOv8-DBW model demonstrates a more favorable balance among detection accuracy, computational efficiency, and model compactness.

The radar chart results characterizing the comprehensive performance of the models (Figure 11) show that the improved YOLOv8-DBW model has the most full and complete area, indicating that its performance in all aspects is closer to the ideal state compared with other models. In summary, the YOLOv8-DBW algorithm proposed in this study has demonstrated its superiority in multiple metrics.

To further evaluate the robustness of the proposed model under class imbalance, detailed performance metrics for each of the seven pest and disease categories are summarized in Table 9. Despite the variation in sample counts among different categories, the YOLOv8-DBW model achieved consistently strong performance across all classes. Specifically, the minority class Black Rot attained an mAP of 83.9%, which is highly comparable to that of the majority class Diamondback Moth at 87.4%. This relatively balanced performance across categories suggests that the proposed model is less sensitive to sample imbalance. Such robustness can be attributed to the synergistic effect of the enhanced feature extraction capability provided by the C2f-PE module and the dynamic sample weighting mechanism introduced by the Wise-IoU loss function, which together help mitigate potential bias toward majority classes.

Based on the experimental results of mainstream models, by selecting and verifying better model algorithms for comparative analysis, the YOLOv5n, YOLOv10n, YOLOv11n, YOLOv12n, and the improved YOLOv8-DBW algorithm proposed in this study were chosen for visual comparison in the detection results of pakchoi pests and diseases, with the detection results shown in Figure 12.

As shown in Figure 12, the comparison of detection confidence on the test set reveals that some models exhibit noticeable cross-misclassification and background false detection issues in the identification of pakchoi pests and diseases. Specifically, YOLOv12 misidentifies Alternaria Leaf Spot as White Rust, and YOLOv11n makes a similar error in detecting White Spot. Moreover, in the detection of Black Rot, YOLOv5s, YOLOv10n, and YOLOv11n all misclassify background areas as Diamondback Moth. In contrast, the YOLOv8-DBW model shows no misclassification across all cases and achieves significantly higher detection accuracy than the other models. These results confirm that the improvements made to YOLOv8n effectively enhance the detection performance for pakchoi pests and diseases, addressing the insufficient accuracy of existing models.