Weeds are one of the major factors affecting agriculture. Currently, the damage caused by weeds to agriculture reaches as high as 34% (Oerke, 2006). For decades, herbicides have been widely adopted as the preferred method for weed management in global agriculture; however, herbicides often have adverse environmental side effects (Kudsk and Streibig, 2003). With the development of precision agriculture and the widespread use of intelligent agricultural machinery, robots for weed control hold great potential in building environmentally friendly agriculture and reducing labor demands (You et al., 2020). Intelligent robots rely on real-time weed detection systems with high accuracy to locate weeds, but weed detection in actual farmland environments still faces the following challenges.

Firstly, the coexistence of inter-class phenotypic similarity and intra-class morphological variation poses significant challenges. Weeds and crops exhibit substantial overlap in color features, particularly during the seedling stage, with notable similarities in leaf shape and texture characteristics during growth (Hasan et al., 2021). For instance, Italian ryegrass (Lolium multiflorum) and wheat are visually indistinguishable without expert knowledge (Bansal et al., 2024), rendering traditional detection methods ineffective. Compounding this complexity, dynamic variations in leaf color, morphology, and texture occur across different growth stages of the same weed species (Veeragandham and Santhi, 2021), hindering the establishment of stable feature representations in detection models. Studies have shown that YOLOX and YOLOv8 models experienced accuracy declines of 14.5% and 14.2%, respectively, in identifying eight cross-season weed categories common in cotton fields. This degradation primarily stems from seasonal variations in lighting, background conditions, and weed growth states (Deng et al., 2024).

Secondly, the sheer diversity of weed species (Chen et al., 2024), combined with the simultaneous presence of weeds at varying developmental stages within agroecosystems, creates scale differences spanning three orders of magnitude. Empirical analysis using the Weed25 dataset revealed significant disparities in recognition performance: Asiatic smartweed (Polygonum aviculare) achieved a mean average precision (mAP) of 62.92%, while velvetleaf (Abutilon theophrasti) reached 99.70% (Wang et al., 2022). Such scale heterogeneity complicates the development of algorithms capable of effectively detecting multi-category weeds. Furthermore, leaf overlap and weed occlusion in dense scenarios exacerbate differentiation and detection challenges (Wang et al., 2019). In cabbage fields, weed detection not only struggles with color similarity but also contends with illumination variations and leaf occlusion, leading to suboptimal performance in direct detection methods addressing these issues (Sun et al., 2024).

Finally, the high-density distribution and small size of weed targets frequently result in missed detections and false positives. Wu et al. (2023). proposed an enhanced YOLO-V4 model tailored for small weed detection in farmland, improving the mAP by 4.2%. However, despite advancements in lightweight performance, the parameter count of model remains high at 42.54 million, posing deployment challenges.

To address the aforementioned challenges, this paper proposes an PD-YOLO method based on a multi-scale feature fusion network, building on YOLOv8n. The method introduces an innovative Parallel Focusing Feature Pyramid (PF-FPN), which effectively improves the accurate classification and localization of different types of weeds in complex environments. The PF-FPN includes the Feature Filtering and Aggregation Module (FFAM) and the Hierarchical Adaptive Recalibration Fusion Module (HARFM). The FFAM module utilizes deep convolution and attention mechanisms to preliminarily filter and extract features, adaptively adjusting and fusing multi-scale features to capture rich semantic information. This enhances small object detection and multi-scale feature fusion, thereby improving the model’s accuracy and robustness. The HARFM module leverages attention mechanisms to fuse features at different levels, achieving adaptive optimization and enhancing the model’s expressive power. To further improve detection performance, a dynamic detection head (Dyhead) (Dai et al., 2021) is introduced, enhancing the model’s stability and accuracy in complex backgrounds. Ultimately, the proposed PD-YOLO model integrates the PF-FPN network and Dyhead architecture, using YOLOv8n as the base framework. This optimization of feature fusion and model representation capabilities significantly enhances the overall accuracy of weed detection.

The remainder of this paper is organized as follows: Section 2 introduces related research work; Section 3 presents the PD-YOLO model; Section 4 describes the experiments conducted on the model; Section 5 provides relevant discussions; and Section 6 concludes the paper.

Early research in weed detection primarily relied on traditional image processing techniques (Wu et al., 2021). These methods involved extracting features such as color, texture, and shape from images, which were then used in conjunction with machine learning algorithms like Random Forests or Support Vector Machines for weed identification (Sabzi et al., 2020). For example, Islam et al (Islam et al., 2021). achieved efficient weed detection in Unmanned Aerial Vehicle (UAV) imagery through image orthorectification combined with machine learning algorithms, reaching accuracy rates of 98.40% on the original dataset and 94.72% on an extended dataset. The success of these techniques heavily depended on the quality of image acquisition, preprocessing, and feature extraction, as these factors directly influenced the performance and generalization ability of the algorithms.

With the advent of deep learning, object detection methods have revolutionized weed detection due to their superior efficiency and accuracy. Tang et al. (2017). pioneered the use of K-means unsupervised feature learning in conjunction with Convolutional Neural Networks (CNNs), improving the identification accuracy of soybean seedlings and associated weeds to 92.89% through fine-tuning optimization. This approach effectively addressed the issues of instability and limited generalization found in manually designed feature-based methods. Similarly, dos Santos Ferreira et al (Tang et al., 2017) applied a ConvNets network to detect weeds in soybean crop images, classifying them into grass and broadleaf categories. This categorization enabled the targeted application of specific herbicides, achieving over 98% identification accuracy.

Currently, classical deep learning-based object detection methods are mainly categorized into single-stage and two-stage detection algorithms. Single-stage algorithms directly use neural networks to extract features from images and perform detection. Representative algorithms include SSD (Liu et al., 2016), the YOLO series (Redmon et al., 2016; Chen et al., 2021), and RT-DETR (Chen et al., 2021). Single-stage algorithms generally have the advantage of faster speed and better real-time detection performance, but they often suffer from lower localization accuracy. In contrast, two-stage algorithms first generate candidate regions and then classify and localize these regions. They usually achieve higher detection accuracy but are relatively slower and require more computational resources, with Faster R-CNN (Ren et al., 2015) being a representative example.

In weed detection, two-stage algorithms first generate potential weed-containing regions using object detection algorithms, followed by deep learning model classification to distinguish weeds from crops. Veeranampalayam Sivakumar et al (Veeranampalayam Sivakumar et al., 2020). constructed Faster R-CNN and SSD models and evaluated their performance for weed detection in soybean fields using UAV images. The results showed that both models performed well in weed detection, but Faster R-CNN outperformed SSD in terms of performance. This approach typically improves detection accuracy and reduces false positive rates, but its high computational complexity makes it difficult to meet real-time requirements.

To address the challenges of morphological diversity, scale variations, and complex backgrounds in farmland weed detection, researchers have proposed various improvements based on the YOLO series models. These methods enhance model performance in specific agricultural scenarios through strategies such as integrating attention mechanisms, optimizing multi-scale feature fusion, and improving small-target detection capabilities. Table 1 systematically compares representative models in terms of improvement methods, parameter counts, computational costs, and performance metrics.

Current YOLO-based farmland weed detection models generally suffer from scenario limitations and methodological homogenization. Most models, such as GTCBS-YOLOv5s (Shao et al., 2023) and YOLOv8-DMAS (Zheng et al., 2024), are optimized solely for single environments like rice or cotton fields, relying on repetitive technical approaches including attention mechanisms, multi-scale feature fusion, and loss function improvements, while lacking differentiated designs for weed morphological diversity and crop coexistence scenarios. Some models like RMS-DETR (Guo et al., 2024) with 40.8M parameters and 187 GFLOPs computational cost achieve only 85.1% accuracy, showing significant efficiency-accuracy imbalance. Lightweight models such as YOLOV7-G (Yu et al., 2024), though compressed to 0.57M parameters, suffer from high missed detection rates resulting in mAP as low as 56.6%, limiting practical applicability. Although a few models like YOLO-Riny (Xu et al., 2024) achieve edge device compatibility through structural lightweighting, their improvements remain confined to specific weed types Corydalis edulis and Setaria viridis in cornfields, failing to address complex multi-target interaction detection needs. While YOLO-CWD (Ma et al., 2025) achieves lightweight design with 75.1% mAP@50 and 9.6 GFLOPS in cornfield weed detection through hybrid attention mechanisms and PIoU loss function, its detection accuracy and model compactness still require further optimization. Existing studies generally lack cross-scenario generalization validation, showing weak support for multi-category weed interaction detection, environmental robustness, and crop-weed coexistence mechanisms, which constrains practical agricultural applications.

Existing weed detection methods based on general deep learning architectures face challenges due to the diversity in weed morphology and environmental conditions, making algorithm development for different plant species difficult (Hu et al., 2024). Additionally, Convolutional Neural Networks (CNNs), while extracting image features, are limited by the local receptive field of convolutional operators, making it hard to capture global information, which affects accurate image localization and classification (Luo et al., 2016). To overcome this limitation, researchers often employ multi-scale feature fusion techniques, with parallel multi-branch networks and serial skip-connection structures being two commonly used approaches.

In the Inception module of GoogLeNet, the parallel multi-branch network extracts multi-scale and hierarchical feature information from the same feature map using convolutional kernels of different sizes (Szegedy et al., 2015). Although this method takes advantage of convolution kernels with different receptive fields and carefully designed modules to learn rich multi-scale features, it overlooks semantic differences between features of different scales, which can lead to the loss of semantic information. High-level feature maps generated by the backbone network contain rich semantic information but lack the detailed information of objects, while low-level features, although containing precise object locations, lack sufficient semantic information.

To address this issue, high-dimensional features are typically upsampled and aligned with downsampled low-dimensional features, followed by pixel-wise summation to enhance semantic information. However, this method does not perform feature selection, merely summing pixel values across multiple feature layers, which may lead to redundant and repetitive information. Consequently, this approach fails to fully integrate and utilize diverse features (Chen et al., 2024).

In addition, feature fusion networks represent an efficient method for multi-scale fusion. The classic Feature Pyramid Network (FPN) (Lin et al., 2017) employs a top-down pyramid structure to achieve multi-scale feature fusion. However, due to its structural characteristics, FPN lacks sufficient high-level semantic information. To enhance local localization information, PANet (Liu et al., . 2018) added a bidirectional feature fusion module on top of FPN. Building on these approaches, BiFPN (Tan et al., 2020) introduced bidirectional connections in the process of information propagation, allowing information to flow both top-down and bottom-up within the network. This bidirectional flow effectively addresses issues of information loss and blurring in feature pyramid networks.

The Gold-YOLO model introduced an advanced gather and distribute mechanism (GD mechanism), which uses a unified module to collect and fuse information from all levels and distribute it to different levels, addressing the information fusion issues in traditional object detection models (Wang et al., 2024). Although multi-scale feature fusion methods have significant reference value for image processing tasks, current networks struggle to meet the practical requirements of weed detection tasks due to challenges such as limited weed features, varying lighting conditions, and occlusion problems. Therefore, developing more efficient feature fusion networks is crucial for improving the accuracy of weed detection.

This study improves and optimizes the YOLOv8 model, enhancing its performance in weed detection tasks and developing a novel weed detection model, PD-YOLO. The FAFM and HARFM modules were introduced in this study, and based on these modules, a Parallel Focusing Feature Pyramid was proposed to replace the original PA-FPN, improving the model’s feature fusion capability. Additionally, a dynamic detection head was incorporated to enhance the model’s detection stability. These methods improve the classification accuracy of the traditional YOLOv8n model in weed detection and reduce both missed detections and false positives.

Section 4.1 focuses on the performance comparison of different object detection models. The results in Tables 1 and 2 show that FD-YOLO achieves an mAP@0.5 of 95.0% on the CottonWeedDet12 dataset and 76.8% on the Lincoln Beet dataset. This difference may be attributed to several factors: first, the Lincoln Beet dataset only contains two labeled categories—”weeds” and “beetroot”—while CottonWeedDet12 includes 12 subcategories of weeds, which tests the model’s fine-grained classification ability; second, the weed distribution in the Lincoln Beet dataset is denser, and the target sizes are smaller, leading to increased difficulty in localization in dense occlusion scenarios; furthermore, the images in the Lincoln Beet dataset were captured under the variable field lighting conditions in the UK, with diverse background soil types, further increasing the detection complexity. FD-YOLO’s PF-FPN enhances multi-category feature discrimination through the FFAM and HARFM modules, demonstrating clear advantages on CottonWeedDet12, but the high-density targets in Lincoln Beet require stronger spatial context modeling capabilities. The current model still has room for improvement in the spatial awareness attention mechanism of the Dynamic Head (DyHead). Future work could focus on introducing adaptive resolution adjustment strategies or enhancing spatial attention weight distribution to further improve the model’s performance in high-density scenarios. These differences indicate that FD-YOLO requires fine-tuning or data augmentation for specific environments in cross-regional and multi-crop scenarios.

The results in Table 4 show that PD-YOLO has fewer parameters and lower computational complexity than lightweight models like YOLOv7-tiny, positioning it as an efficient lightweight model. However, its FPS performance still lags significantly, and it may face challenges such as insufficient frame rates and image blur in high-speed agricultural robots. Further optimization of computational efficiency or the adoption of hardware acceleration solutions is needed. Additionally, future research could explore techniques such as model pruning, quantization, and hardware acceleration to better adapt to low-power embedded devices, ensuring its wide applicability in real-time agricultural applications.

In practical agricultural robotics applications, FD-YOLO can integrate with SLAM technology to enable “detection-navigation-operation” integration. Following Zhang W et al (Zhang et al., 2024), combining 2D LiDAR and visual sensors with YOLOv3 algorithm allows target detection and information mapping onto 2D grid maps for efficient path planning, eliminating computational latency-induced trajectory deviations in traditional models. Researchers applied the trained DIN-LW-YOLO model to autonomous laser weeding robots in strawberry fields, with robot speed set at 0.50 m/s and Intel Realsense D435i camera mounted 600 mm above ground at 30 fps. Field tests demonstrated 92.6% weed control rate and 1.2% seedling damage rate (Zhao et al., 2025). Moreover, FD-YOLO’s lightweight design enables edge device deployment for future weed management. Similar studies deployed customized YOLOv7 models on NVIDIA Jetson Xavier NX platforms, integrating robotic frame spraying systems that recognize Amaranthus palmeri in cornfields for real-time spot spraying (Balabantaray et al., 2024).

Furthermore, the ablation experiments validated the roles and necessity of each module and their impact on model size, as discussed in Section 4.2. The experimental results demonstrated the impact of each module on the model’s recognition accuracy, as well as a comparison of different feature pyramids. In PF-FPN, the excessive upsampling and downsampling during feature aggregation and distribution by the FFAM module led to a loss of detailed features. Therefore, the introduction of the HARFM module reduced this loss, showcasing efficient feature fusion capabilities. Despite the increase in computational cost and detection speed, the improvements in multi-scale feature fusion and small object detection enable the PD-YOLO model to perform well in scenarios involving high-density, partial occlusion, and multi-class weeds. However, in practical applications, environmental adaptability still needs to be considered. The robustness of the model under low-light conditions or extreme weather has not been validated. The current dataset is primarily based on natural lighting conditions, which may limit the model’s stability in complex lighting scenarios.

The systematic analysis of the TIDE metrics provides clear directions for model optimization. Experimental results show that FD-YOLO significantly outperforms the baseline model in terms of classification errors and localization errors, but there is still room for further optimization. To address classification errors, future work could introduce fine-grained feature alignment strategies to enhance the model’s ability to distinguish between morphologically similar weeds. The residual localization errors may stem from blurred object boundaries in complex occlusion scenarios, which could be improved by integrating deformable convolutions or refining the bounding box regression loss function to boost localization accuracy. Additionally, the background misdetection rate remains relatively high, indicating that the current model lacks sufficient suppression of background noise such as soil texture. To address this, more diverse background samples should be included during data augmentation, or a lightweight background-aware attention module could be designed.

This study proposes PD-YOLO, a novel computer vision method specifically designed for real-time weed detection. The architecture of PD-YOLO combines the Parallel Focusing Feature Pyramid (PF-FPN) and Dyhead framework, built upon the YOLOv8n framework. The FAFM module optimizes feature fusion, enhancing the model’s representational capabilities, while the HARFM module strengthens weed-specific features, improving weed identification. The PF-FPN network, developed with consideration of weed morphological characteristics, serves as an effective feature fusion network for weed detection. The Dyhead framework improves the design of the detection head, ensuring both accuracy and stability in detection results.

The research results show that, compared to the baseline model, PD-YOLO improves mAP by 1.7% and 1.8% (at thresholds of 0.5 and 0.5-0.95, respectively). While maintaining a lightweight structure, PD-YOLO outperforms current mainstream object detection algorithms, demonstrating superior performance. Moreover, although the model’s detection speed meets real-time detection requirements, there is potential for further optimization in real-world field environments.

Future research will focus on the following directions:

(1)Data Augmentation and Multimodal Fusion: Integrating multispectral imaging data to enhance the model’s detection capability under complex lighting and occlusion conditions, and expanding data diversity through synthetic data augmentation (e.g., simulating rain, fog, and shadows).

Lightweight and Efficiency Optimization: Developing an FD-YOLO-Tiny variant that combines the vMamba (Zhu et al., 2024) architecture to improve the model’s backbone, reducing computational overhead while maintaining accuracy, and adapting it for deployment on edge devices.

(2)Error-Driven Model Improvement: Based on the analysis results from the TIDE metrics, specifically optimizing the false negative and false positive modules. This can be achieved by strengthening the training of negative samples, improving the loss function, or adjusting the post-processing stage of the detection algorithm to reduce misdetections and missed detections.

(3)Cross-Scene Validation and Transfer Learning: Expanding experimental validation to include different crops, such as corn and wheat, and varying agricultural environments. Combining transfer learning techniques to enhance the model’s generalization ability, ensuring its practicality in diverse agricultural settings.