Introduction
Vegetables play a critical role in human health by enhancing immunity, preventing chronic diseases, and addressing global “hidden hunger” through their rich nutritional content and phytochemicals (Asaduzzaman et al. Reference Asaduzzaman, Asao, Amao, Asaduzzaman and Asao2018). Alongside staple crops, vegetables rank among the most widely cultivated and economically significant crops worldwide (Dias and Ryder Reference Dias, Ryder and Janick2011). However, weeds pose a major challenge, not only by diminishing vegetable quality but also by causing yield losses of up to 45% to 90% (Mennan et al. Reference Mennan, Jabran, Zandstra and Pala2020). Herbicides remain effective tools for weed suppression; however, excessive application can leave chemical residues in vegetables, increase environmental risks, and accelerate the evolution of resistant weed populations, which further complicates management (Mennan et al. Reference Mennan, Jabran, Zandstra and Pala2020). Manual weeding, while reducing dependence on herbicides, is becoming increasingly unsustainable for large-scale agricultural production due to rising labor costs (Jin et al. Reference Jin, Che and Chen2021). Thus, developing automated, vision-based methods for distinguishing between crops and weeds has become increasingly essential for modern weed management (Jin et al. Reference Jin, Che and Chen2021).
Computer vision has shown significant potential for precision herbicide application in modern agriculture (Yu et al. Reference Yu, Schumann, Cao, Sharpe and Boyd2019a, Reference Yu, Schumann, Sharpe, Li and Boyd2020). Current approaches for distinguishing between crops and weeds primarily rely on either traditional image processing techniques or deep learning methods (Wu et al. Reference Wu, Chen, Zhao, Kang and Ding2021). Traditional methods rely on features such as texture (Bakhshipour et al. Reference Bakhshipour, Jafari, Nassiri and Zare2017; Ishak et al. Reference Ishak, Hussain and Mustafa2009), shape (Bakhshipour and Jafari Reference Bakhshipour and Jafari2018; Pereira et al. Reference Pereira, Nakamura, De Souza, Martins and Papa2012), spectral properties (Elstone et al. Reference Elstone, How, Brodie, Ghazali, Heath and Grieve2020; Pignatti et al. Reference Pignatti, Casa, Harfouche, Huang, Palombo and Pascucci2017), and color (Hamuda et al. Reference Hamuda, Glavin and Jones2016; Rasmussen et al. Reference Rasmussen, Nielsen, Streibig, Jensen, Pedersen and Olsen2019). However, relying on a single handcrafted feature (such as color or texture) is often insufficient for distinguishing crops from weeds, underscoring the need for multi-feature integration or deep learning–based approaches (Wu et al. Reference Wu, Chen, Zhao, Kang and Ding2021). To overcome this limitation, many studies have focused on integrating multiple features to improve detection accuracy (Sabzi et al. Reference Sabzi, Abbaspour-Gilandeh and Arribas2020). Machine learning techniques, such as support vector machines and artificial neural networks, have been widely employed for crop and weed classification (Behmann et al. Reference Behmann, Mahlein, Rumpf, Römer and Plümer2015; Tellaeche et al. Reference Tellaeche, Pajares, Burgos-Artizzu and Ribeiro2011). While these methods can accurately identify weeds under certain conditions, their reliance on single or manually designed features often limits their robustness and generalization ability, especially in complex and diverse agricultural environments (Kong et al. Reference Kong, Liu, Chen, Jin, Li and Yu2024; Wu et al. Reference Wu, Chen, Zhao, Kang and Ding2021).
In recent years, deep learning has been increasingly applied to the agricultural domain (Fu et al. Reference Fu, Gao, Wu, Li, Karkee and Zhang2020; Too et al. Reference Too, Yujian, Njuki and Yingchun2019). Deep learning–based approaches for weed detection and classification have demonstrated promising results (Tiwari et al. Reference Tiwari, Goyal, Kumar and Vij2018; Yu et al. Reference Yu, Sharpe, Schumann and Boyd2019b). Commonly used deep learning methods for weed detection include convolutional neural networks (CNNs) (Dyrmann et al. Reference Dyrmann, Karstoft and Midtiby2016; Olsen et al. Reference Olsen, Konovalov, Philippa, Ridd, Wood, Johns, Banks, Girgenti, Kenny and Whinney2019; Yu et al. Reference Yu, Sharpe, Schumann and Boyd2019c, Reference Yu, Schumann, Sharpe, Li and Boyd2020) and fully convolutional networks (Huang et al. Reference Huang, Deng, Lan, Yang, Deng, Wen, Zhang and Zhang2018). Yu et al. (Reference Yu, Schumann, Cao, Sharpe and Boyd2019a) evaluated several deep convolutional neural network models, including AlexNet, Visual Geometry Group Network (VGGNet), GoogleNet, and DetectNet, for detecting dandelion (Taraxacum officinale F.H. Wigg.), ground ivy (Glechoma hederacea L.), and spotted spurge [Euphorbia maculata L.; syn.: Chamaesyce maculata (L.) Small] in annual bluegrass (Poa annua L.). Among these, VGGNet achieved a high F1 score of 92.78% and a recall of 99.52% in multi-class classification tasks, while DetectNet excelled in detecting T. officinale with an F1 score of 98.43% and a recall of 99.11%. Similarly, Jin et al. (Reference Jin, Bagavathiannan, McCullough, Chen and Yu2022b) evaluated DenseNet, EfficientNetV2, ResNet, RegNet, and VGGNet models for both multi-class and binary classification of weed species. In multi-class classification, VGGNet achieved an F1 score of 95.0% for detecting T. officinale and performed exceptionally well in identifying goosegrass [Eleusine indica (L.) Gaertn.], purple nutsedge (Cyperus rotundus L.), and white clover (Trifolium repens L.) in bermudagrass [Cynodon dactylon (L.) Pers.], with F1 scores ≥98.3%. In binary classification, where the goal was to distinguish weed-infested regions from turfgrass, the EfficientNetV2 model performed best, achieving F1 scores ≥98.1%. These results underscore the efficacy of deep learning–based models in addressing the challenges of weed detection and classification across different contexts and tasks.
Direct weed identification remains highly challenging due to the visual similarity between crops and weeds and the variability across field environments (Coleman et al. Reference Coleman, Bender, Hu, Sharpe, Schumann, Wang, Bagavathiannan, Boyd and Walsh2022; Jin et al. Reference Jin, Bagavathiannan, Maity, Chen and Yu2022a). Such morphological overlap and ecological variability introduce instability in feature extraction and reduce detection reliability (Coleman et al. Reference Coleman, Bender, Hu, Sharpe, Schumann, Wang, Bagavathiannan, Boyd and Walsh2022; Jin et al. Reference Jin, Bagavathiannan, Maity, Chen and Yu2022a; Zhuang et al. Reference Zhuang, Li, Bagavathiannan, Jin, Yang, Meng, Li, Li, Wang and Chen2022). Another major bottleneck lies in dataset construction. Collecting and annotating sufficient images to represent diverse weed species across regions, growth stages, and densities requires immense effort, and the lack of such comprehensive datasets limits the robustness and generalization of deep learning models (Kong et al. Reference Kong, Liu, Chen, Jin, Li and Yu2024; Wu et al. Reference Wu, Chen, Zhao, Kang and Ding2021). This highlights the need for innovative approaches to overcome the limitations posed by data diversity and availability in weed detection tasks. To tackle these challenges, this study proposes a novel deep learning–based approach for training a vegetable detection model. First identifying vegetables within an image allows all remaining green regions outside the detected area to be classified as weeds, providing a streamlined and efficient method for weed detection. In this study, we developed an effective two-step approach for weed detection:
-
1. Building upon the YOLOv10 (A Wang et al. Reference Wang, Chen, Liu, Chen, Lin and Han2024), we propose a novel detection model, CD-YOLOv10n (C2f-DBB_Dysample_YOLOv10n), specifically designed for efficient vegetable identification. This model not only demonstrates outstanding performance in terms of detection accuracy but also achieves significant lightweighting, making it suitable for deployment in resource-constrained environments.
-
2. Once vegetables are identified, all green pixels outside the bounding box are classified as weeds. Weed detection and segmentation are then performed using image processing techniques, ensuring a streamlined and precise approach to weed detection.
Materials and Methods
Dataset
The images of vegetable seedlings used for training, validation, and testing in this study were collected in two batches from a vegetable farm on Bagua Island, Qixia District, Nanjing, China (32.2°N, 118.8°E), during July and September 2020. Each image had an original resolution of 4,032 × 3,024 pixels, and a total of 1,500 images were obtained. To enhance the diversity and generalization ability of the neural network, the dataset included images of bok choy [Brassica rapa subsp. chinensis (L.) Hanelt.] from vegetable fields with different sowing times, captured under various lighting conditions, including sunny and overcast weather. To reduce processing time and enhance real-time performance in field applications, all sample images were standardized to 1,400 × 1,050 pixels using a custom Python script. The collected images were annotated using LabelImg software, focusing on rectangular bounding boxes around vegetable seedlings. After annotation, corresponding extensible markup language (XML) label files were generated for each image, serving as training samples for the neural network model. The dataset was split into training (80%), validation (10%), and testing (10%) sets, as detailed in Table 1.
Table 1. Number of images used for training, validation, and testing.
Improvement of YOLOv10n
YOLO is a widely recognized single-stage object detection algorithm, demonstrating excellent performance across various detection tasks. Compared with two-stage object detection algorithms, single-stage algorithms like YOLO integrate object localization and candidate region generation (first stage) with the classification phase (second stage), significantly improving detection efficiency.
To meet the need for cost-effective implementation, this study selected YOLOv10n, the simplest and most lightweight model in the YOLOv10 series, for detecting vegetable seedlings. However, the YOLOv10n algorithm has certain limitations in specific application scenarios, including limited feature extraction capacity that may hinder detection accuracy under complex and variable environmental conditions. Additionally, the model’s ability to differentiate between morphologically similar objects may be insufficient, particularly in high-precision applications such as identifying specific plant species in diverse agricultural environments. To overcome these challenges and enhance the algorithm’s performance in detecting vegetable seedlings, this study introduces several optimizations to the YOLOv10n framework. These improvements were designed to address its limitations while maintaining its efficiency and lightweight structure, ensuring better suitability for complex agricultural tasks. The algorithm was improved in two key aspects, as illustrated in Figure 1:
-
1. The Cross Stage Partial Network Fusion (C2f) module in YOLOv10n’s backbone network was replaced with the C2f with Dynamic Block Branching (C2f-DBB) module (Zhang et al. Reference Zhang, Wang, Wang and Xia2024), resulting in a reduction in module size and improved efficiency.
-
2. The DySample module (Liu et al. Reference Liu, Lu, Fu and Cao2023) was integrated into the neck network to replace the original upsampling mechanism in YOLOv10n, thereby enhancing the network’s feature extraction capabilities.
Figure 1. CD-YOLOv10n architecture.
C2f-DBB Module
The C2f-DBB module is an optimized version of the original C2f design, in which Dynamic Block Branching (DBB) is integrated into the bottleneck structure (Zhang et al. Reference Zhang, Wang, Wang and Xia2024). DBB incorporates six transformation modes, all of which can be converted into convolutions during inference, thereby improving the model’s representation capability. In addition, the C2f-DBB module incorporates the Global Attention Mechanism, which strengthens feature interactions across both channel and spatial dimensions. By introducing DBB and attention mechanisms, the module significantly enhances feature extraction capabilities, leading to improved detection accuracy and model stability.
DySample Block
DySample is an ultra-lightweight and efficient dynamic upsampling technique that redefines the perspective of point sampling in the upsampling process. It avoids the high computational complexity and resource consumption inherent in traditional convolution-based methods. Notably, DySample operates without relying on high-resolution guiding features and is not constrained by additional CUDA package requirements. This results in significantly lower inference latency, memory usage, floating-point operations (FLOPs), and parameter count, optimizing both performance and resource efficiency. Liu et al. (Reference Liu, Lu, Fu and Cao2023) demonstrated that DySample outperforms other upsampling techniques across five major dense prediction tasks: semantic segmentation, object detection, instance segmentation, panoptic segmentation, and monocular depth estimation. In addition to its exceptional performance, DySample achieves efficiency comparable to bilinear interpolation. This makes it a reliable alternative to traditional methods like nearest neighbor or bilinear interpolation, offering a practical solution to enhance the performance and efficiency of existing dense prediction models.
Experimental Configuration
In this study, all deep neural network models were trained and tested under a consistent hardware and software environment to ensure reproducibility. The hardware setup included Ubuntu 20.04.6 LTS as the operating system, an Intel® Xeon® W-2265 CPU, and an Nvidia GeForce RTX 3080 Ti GPU. For the software environment, a configured Conda environment was utilized, which included Python 3.8, PyTorch 2.3.1, and CUDA 11.3. Details of the hyperparameter settings for the deep learning models are provided in Table 2. Additionally, to achieve better convergence, higher precision, and enhanced adaptability to real-world agricultural scenarios, Mosaic data augmentation was disabled during the final 10 training epochs.
Table 2. The hyperparameters for deep learning training.a
a Hyperparameter settings follow common practice for YOLO-family detectors (A Wang et al. Reference Wang, Chen, Liu, Chen, Lin and Han2024) and general principles of hyperparameter optimization (Yang and Shami Reference Yang and Shami2020).
b A long dash (—) indicates a dimensionless parameter (no units).
c SGD, stochastic gradient descent.
Evaluation Metrics
This study employs multiple performance metrics, including precision, recall, mAP50, mAP50-95, and inference time, to comprehensively evaluate the performance of the deep learning models. The formulas for these metrics are as follows:
$$\rm{Precision = {{{\rm{TP}}} \over {{\rm{TP}} + {\rm{FP}}}}}$$
$$\rm{Recall = {{{\rm{TP}}} \over {{\rm{TP}} + {\rm{FN}}}}}$$
$${\rm mAP} = {\rm AP} = \int_{0}^{1} {\rm Precision}({\rm Recall})\, d({\rm Recall})$$
In this study, mAP is used to represent the average precision. Because our experimental case involves the identification of a single category, mAP specifically refers to the average precision (AP) for the vegetable category. True positive (TP) denotes the number of samples correctly identified as vegetables, false positive (FP) refers to the number of samples incorrectly identified as vegetables, and false negative (FN) represents the number of samples that were not correctly identified as vegetables.
Weed Segmentation
After crop bounding boxes were detected using the vegetable detector, green pixels outside the bounding boxes were segmented through a color-based image processing technique and marked as weeds. In this study, crops were detected but not segmented at the pixel level, which reduces annotation cost and computational complexity. However, this design may limit accurate identification of weeds located very close to seedlings, as bounding boxes cannot perfectly match crop boundaries. To enhance this process, the weed segmentation index proposed by Jin et al. (Reference Jin, Che and Chen2021) was adopted and further optimized. Specifically, pixels were evaluated using the color index technique only if their green (G) component exceeded the red (R) or blue (B) components; otherwise, they were directly classified as non-weed pixels. This optimization improved segmentation accuracy while reducing computational complexity, resulting in a more efficient and precise weed detection workflow.
Results and Discussion
Ablation Experiments of Each Module
Table 3 summarizes the performance metrics of the model after replacing individual components. The experimental results demonstrate that both the C2f-DBB and DySample strategies positively impacted model performance, with the C2f-DBB module achieving a mAP of 97.3% and the DySample module reaching 97.6% compared with 96.9% for the baseline YOLOv10n before adding these modules. However, it is noteworthy that while the DySample module improved overall detection accuracy, it led to a 1.2% reduction in recall. Both optimization components effectively reduced the model complexity, significantly decreasing the number of parameters and giga floating-point operations per second (GFLOPs). DySample exhibited the most pronounced impact on performance when replacing the neck network, whereas C2f-DBB achieved the greatest reduction in parameter count when replacing the backbone network. Both components contributed to a decrease in GFLOPs, indicating a reduction in the computational load of the neural network.
Table 3. Performance improvements achieved through the replacement of each component.
a C2f-DBB, C2f with Dynamic Block Branching.
b GFLOPs, giga floating-point operations per second.
To validate the effectiveness of the proposed optimization strategies, ablation experiments were conducted for each module, with detailed results presented in Table 4. The findings reveal that the improved CD-YOLOv10n model demonstrates exceptional and efficient feature extraction capabilities. Built on a lightweight core architecture, the model achieves significant improvements in operational efficiency without compromising performance. Additionally, both computational costs and parameter counts are effectively reduced.
Table 4. Results of ablation experiments.
a PM, proposed method.
b C2f-DBB, C2f with Dynamic Block Branching.
c GFLOPs, giga floating-point operations per second.
While the DySample module enhances overall performance, it comes at the expense of a slight reduction in recall value. In contrast, the C2f-DBB module improved performance without sacrificing recall. By leveraging the strengths of both components, the proposed model achieved a balanced trade-off, reducing overall parameters while enhancing performance. The proposed model outperformed YOLOv10n across key metrics, achieving improvements of +1.2% in mAP50, +2.1% in mAP50-95, +1.4% in precision, and +1.6% in recall. Meanwhile, the parameter counts and GFLOPs were reduced by 15.87% and 22.62%, respectively. These results highlight significant advances in both model performance and lightweight design, indicating the effectiveness of the proposed optimization strategies.
Training Performance of the Proposed Method Compared with YOLOv10n
As shown in Figure 2, CD-YOLOv10n achieved higher accuracy than YOLOv10n throughout training, with a consistently superior mAP50 curve. First, the model exhibited a faster convergence rate, with its mAP50 value surpassing that of YOLOv10n during the early training stages (around the 7th to 10th epoch). This indicates that CD-YOLOv10n is capable of extracting effective features at an earlier stage. Second, the training process of CD-YOLOv10n was notably more stable: as the mAP50 value approached 1, the curve showed smaller fluctuations and greater smoothness, suggesting improved reliability in later-stage predictions. Finally, although the final mAP50 values of the two models were very close, CD-YOLOv10n consistently maintained a slight advantage throughout most of the training process, consistently outperforming YOLOv10n. Overall, CD-YOLOv10n demonstrates superior performance in terms of convergence speed, stability, and training efficiency, highlighting the effectiveness of the proposed optimizations.
Figure 2. Training accuracy (mAP50) versus epoch (0–100) for YOLOv10n and CD-YOLOv10n. The x axis shows training epochs, and the y axis shows mAP50. Curves are averaged across epochs and smoothed with a three-epoch moving average for clarity.
CD-YOLOv10n demonstrated significant advantages in training loss performance. The training loss curve reflects the optimization process of bounding box regression, where lower values indicate more accurate localization of objects. As shown in the training loss curve in Figure 3, CD-YOLOv10n exhibited a faster decline in loss during the initial training stages. Although its initial loss was slightly higher than that of YOLOv10n, it quickly surpassed YOLOv10n, indicating higher efficiency in the early learning phase. Further analysis revealed that the loss curve of CD-YOLOv10n remained smoother throughout the training process. In particular, during the mid- to late stages of training, the fluctuation amplitude was noticeably reduced, reflecting greater stability in the training process. Training stability refers to the smoothness and consistency of the optimization process, wherein fewer oscillations in the loss curve indicate more reliable convergence and reduced risk of overfitting. By the end of training, the loss value of CD-YOLOv10n was slightly lower than that of YOLOv10n, suggesting superior performance in bounding box precision and optimization. Overall, these findings highlight that CD-YOLOv10n outperforms YOLOv10n in both localization accuracy and training stability.
Figure 3. Training loss versus epoch (0–100) for YOLOv10n and CD-YOLOv10n. Loss represents the weighted sum of box regression, objectness, and classification components. Lower values indicate more accurate bounding box regression.
Comparison of CD-YOLOv10n and YOLOv10n in Vegetable Detection
As illustrated in Figure 4, the experimental results clearly demonstrate that CD-YOLOv10n achieved higher accuracy compared with YOLOv10n, particularly in terms of bounding box localization. Moreover, CD-YOLOv10n exhibited significantly greater robustness in handling complex scenarios, such as occlusion and overlapping targets. These improvements highlight the model’s enhanced capability for precise detection under challenging conditions.
Figure 4. Vegetable detection results of YOLOv10n and CD-YOLOv10n on challenging field scenes. Columns show the original image, YOLOv10n output, and CD-YOLOv10n output. Smaller inset boxes highlight regions where the two models differ, with bounding boxes indicating predictions in difficult areas.
Weed Detection
As shown in Figure 5, the proposed method accurately detected vegetable seedlings and effectively segmented weeds, even in visually complex agricultural scenes. The pipeline maintained robustness under challenging conditions such as occlusion, illumination variation, and overlapping plants, highlighting its potential for practical weed management in real field environments.
Figure 5. Vegetable detection and weed segmentation results. Columns show the original image, CD-YOLOv10n detection (vegetable bounding boxes), and the segmentation output where green pixels outside the boxes are classified as weeds.
Model Comparison Experiments
To comprehensively evaluate the efficiency, accuracy, and superiority of the CD-YOLOv10n model, this study compared its performance with various object detection models for vegetable detection. The results are summarized in Table 5. To gain deeper insights into the performance of each model, a comparative analysis was conducted using key metrics, including mAP, recall, parameter count, and GFLOPs. The reported values correspond to the worst-case results among 10 independent runs for each model, thereby providing a conservative evaluation of performance.
Table 5. Performance comparison of detection models.a
a Values represent the worst results from 10 independent runs for each model, providing a conservative assessment of performance.
b ATSS, adaptive training sample selection; R50, ResNet-50; R-CNN, region-based convolutional neural network; FPN, Feature Pyramid Network; DINO, DETR with improved denoising anchor boxes for end-to-end object detection; CIOU, complete intersection over union; FCOS, fully convolutional one-stage; FSAF, feature selective anchor-free; GFL, generalized focal loss; TOOD, task-aligned one-stage object detection; YOLO, You Only Look Once; RT-DETR-R18, Real-Time Detection Transformer ResNet-18.
c GFLOPs, giga floating-point operations per second.
Although some models, such as YOLOX-Tiny (mAP50= 98.2%) and Real-Time Detection Transformer ResNet-18 (RT-DETR-R18) (mAP50 98.2%), achieved slightly higher mAP50 values compared with CD-YOLOv10n (mAP50 98.1%), they exhibited clear disadvantages in other critical areas. Specifically, YOLOX-Tiny had a 13.8% lower recall, along with higher parameter counts and GFLOPs, thereby limiting its suitability for resource-constrained environments. RT-DETR-R18, despite its impressive mAP50 and recall (94.7%), required significantly higher parameter counts and GFLOPs than CD-YOLOv10n, leading to increased computational burdens and reduced efficiency for real-time applications. For other models, the performance gap with CD-YOLOv10n was even more pronounced, particularly in terms of recall, parameter count, and GFLOPs, further underscoring the advantages of CD-YOLOv10n.
In summary, although models like YOLOX-Tiny and RT-DETR-R18 achieve slightly higher mAP50, their lower recall or significantly higher computational demands make them less suitable for lightweight and deployment solutions. Considering all performance indicators, and given that even the worst-case results of CD-YOLOv10n remain competitive, the model achieves an optimal balance between accuracy, recall, and resource efficiency, making it the most robust and practical detection model for precision agriculture applications.
The proposed method demonstrates strong robustness and adaptability by combining deep learning–based vegetable detection with a color-based segmentation approach for weed identification. By leveraging bounding box information to isolate non-crop areas, the method narrows the scope of weed segmentation, effectively avoiding the challenges associated with directly recognizing diverse weed species. This design reduces reliance on large-scale annotated weed datasets, simplifying the data-collection process and improving the method’s practicality in various agricultural scenarios.
Furthermore, the two-stage framework reduces potential errors in weed identification by focusing segmentation efforts on non-crop areas. This targeted approach enhances the method’s reliability and ensures its applicability across diverse farming environments. By streamlining the weed identification process and eliminating the need for extensive weed datasets, the proposed method provides an efficient and practical solution for precision agriculture, addressing critical challenges in weed management with high accuracy.
It should be noted that a limitation of the proposed color-based segmentation approach is its sensitivity to illumination variations and the presence of non-weed green objects such as crop residues or algae. While the current implementation incorporates a green-dominance guard condition to mitigate some false positives, further improvements are needed. Future work could explore adaptive thresholding or color normalization in hue-saturation-value and hue-saturation-lightness color spaces to enhance robustness under varying light conditions. In addition, integrating multispectral or near-infrared information may further help discriminate weeds from non-weed vegetation in complex field environments.
This study developed an efficient vegetable recognition model, CD-YOLOv10n, which demonstrated exceptional performance with a mAP50 of 98.1% and a recall of 93.4%. The model also significantly reduced computational costs, as demonstrated by a notable decrease in GFLOPs, improving its resource efficiency and suitability for practical deployment. Furthermore, an innovative indirect weed detection strategy was introduced, requiring only crop annotations during training. By combining crop identification with image processing techniques, this approach effectively detected weeds in non-crop regions, addressing limitations of traditional direct weed detection methods that rely heavily on labor-intensive annotations. This proposed strategy improved robustness and adaptability to varying field conditions, effectively handling challenges such as species diversity, weed density, and growth stages. Future research should focus on validating the method across diverse agricultural scenarios to further enhance its practical applicability, including other vegetable species and mixed cropping systems, as differences in morphology, planting patterns, and canopy structure may affect model performance.
This study proposed CD-YOLOv10n, a lightweight detection model for indirect weed identification. By integrating C2f-DBB and DySample, the model achieved superior accuracy (mAP50 98.1%, recall 93.4%) while reducing parameters and GFLOPs compared with YOLOv10n. The indirect weed detection strategy, based on crop detection followed by optimized color-index segmentation, reduced annotation costs and improved robustness under complex conditions.
While this study validated the approach on bok choy, the proposed pipeline has broader applicability. Because the method relies on detecting crops rather than classifying diverse weed species, it is less sensitive to the variability of weeds across environments. For adaptation to other vegetables or mixed cropping systems, the model would only require retraining on the limited set of crop classes relevant to the target field. Once the crop(s) are reliably identified, all non-crop vegetation can still be indirectly classified as weeds, regardless of species composition. This design reduces the need for extensive weed-specific annotations and highlights the scalability of the method. Nonetheless, additional validation across different vegetables and cropping patterns is necessary to confirm robustness under more diverse agronomic conditions.
Funding statement
This work was supported by the Weifang Science and Technology Development Plan Project (grant no. 2024ZJ1097), Shandong Provincial Natural Science Foundation (grant no. SYS202206), the National Natural Science Foundation of China (grant no. 32072498), Taishan Scholar Program of Shandong Province, and the Yuandu Scholar, Program of Weifang, Shandong China.
Competing interests
The authors declare no conflicts of interest.
Open access
