2.1 Dataset

This study adopts a dataset derived from the IP102 benchmark, a publicly accessible and large-scale collection of insect pest images [26–28]. IP102 has been widely recognized as a standard reference dataset within the domain of agricultural pest classification. It is extensive in scale, containing over 75,000 images and encompassing 102 pest categories that affect various crops such as rice, wheat, maize, and fruits [26]. Images contained in the IP102 dataset mainly originate from real-world field environments, thus presenting rich real-world challenges such as complex backgrounds and variations in pest morphology across different growth stages and viewing angles, granting it significant research value [29,30].

However, as a general pest dataset, IP102 suffers from highly imbalanced class distribution and covers multiple crop types. For the specific task of maize pest detection, the large number of non-maize pest categories within the dataset can introduce irrelevant feature interference, which is detrimental to achieving optimal model performance.

To address these issues and focus specifically on maize pest detection, a dedicated subset was constructed from the IP102 dataset by retaining only the pest species associated with maize. This subset comprises 13 classes with a total of 4,533 images, which we refer to as the IPMaize dataset in this paper. The selected categories correspond to maize-related pest species included in the IP102 dataset and were determined according to the official category definitions of IP102 and previous studies on maize pests [3–6]. All images and their corresponding bounding-box annotations were directly adopted from the officially released IP102 object detection dataset. After category extraction, all retained images and annotations were manually reviewed to verify the consistency among image content, category labels, and bounding-box annotations. Details of the selected categories are listed in Table 1, and Fig 1 presents representative samples.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Pest species and sample numbers in IPMaize.

https://doi.org/10.1371/journal.pone.0357034.t001

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. A selection of raw images from the IPMaize.

(a) Sample(Grub). (b) Sample(Mole cricket). (c) Sample(White margined moth). (d) Sample(Black cutworm). (e) Larval sample(Peach borer). (f) Adult sample(Peach borer).

https://doi.org/10.1371/journal.pone.0357034.g001

The IPMaize dataset was divided into training, validation, and testing subsets at a ratio of 8:1:1 using stratified sampling according to pest categories, ensuring that each subset preserved the original class distribution. The same partition was adopted for all experiments. No additional offline class balancing strategy, such as oversampling or synthetic sample generation, was applied, and the original class distribution was retained to evaluate the proposed model under realistic agricultural conditions. During model training, online data augmentation provided by the Ultralytics YOLO11 framework was employed, including Mosaic augmentation, HSV color transformation, random horizontal flipping, scaling, and translation. Since these augmentation operations were dynamically performed during training, they did not generate additional permanent images or change the original class distribution of the IPMaize dataset. Therefore, no fixed augmented dataset or class-wise sample statistics after augmentation exist.

2.2 Baseline model YOLOv11

As shown in Fig 2, YOLOv11 adopts a modular architecture consisting of a backbone, a neck, and a head. In this design, feature extraction is mainly handled by the backbone, which incorporates four functional modules, namely Conv, C3k2, fast Spatial Pyramid Pooling (SPPF), and C2PSA. The C3k2 module serves as a substitute for the C2f module employed in YOLOv8, featuring an architecture built around a C3k module. The incoming feature map is split into two parallel processing streams, with one branch following a lightweight forward propagation process, while the other is processed through stacked 33 bottleneck residuals before being concatenated and fused. This design not only reduces repetitive gradient computation but also enhances semantic abstraction by providing an equivalent larger receptive field [31,32]. The SPPF module functions as a multi-scale pooling component, effectively enlarging the multi-scale receptive field without increasing computational load. Following the SPPF module, the C2PSA module is incorporated. The C2PSA module utilizes PSA (Partial Spatial Attention) to operate on different branches of the feature map, which are then concatenated. This mechanism ensures the model focuses on spatial information, refining its ability to selectively attend to regions of interest while maintaining a balance between computational cost and detection accuracy [33,34].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. The YOLO11 architecture is composed of several key components, including the BottleNeck, C3k, C3k2, PAS Block, SPPF, and C2PSA modules.

https://doi.org/10.1371/journal.pone.0357034.g002

The Neck module combines a Feature Pyramid Network (FPN) with a Path Aggregation Network (PAN) to enable the integration of feature representations across multiple hierarchical stages. The FPN establishes a top–down pathway that consolidates abstract semantic cues typically diminished in deeper layers of the architecture [35]. Complementing this, the PAN introduces a bottom–up information stream that strengthens the transmission of fine-grained spatial details toward higher-level layers, thereby enhancing the network’s capacity for precise localization and small-object recognition by sustaining high spatial resolution [36]. Through the collaborative interaction of the FPN’s descending path and the PAN’s ascending path, the Neck component achieves efficient fusion of features originating from diverse representational levels.

Within the detection head, depthwise separable convolutions are adopted in YOLOv11 to enable the separation of localization and classification branches. To address variations in object scale, the network outputs feature representations at three distinct resolution levels, allowing effective recognition of targets with diverse sizes. This architectural arrangement decreases parameter count and overall model complexity, while simultaneously improving the network’s generalization capability and robustness [37]. Moreover, redundant bounding boxes are removed through the application of Non-Maximum Suppression (NMS), which helps to refine the final detection results [38].

Under a strict budget of 2.58 M parameters and 6.3 GFLOPs (input 640640), YOLOv11 achieves an effective balance between detection accuracy and computational cost, which supports its applicability in edge-device deployment scenarios. However, the performance of the YOLOv11 model still exhibits notable limitations. In particular, detection accuracy significantly degrades when the image background is complex, the distinction between the object and its surrounding background is relatively weak, or when target pests exhibit substantial morphological variations due to different developmental stages. These observations suggest that the current model’s robustness under complex scenarios, as well as its capability to capture variations in target shapes, still requires further improvement.

3.1 WindmillConv (WMConv)

Maize pests in field images often exhibit characteristics such as arbitrary posture, slender edges, and randomly oriented textures. The isotropic sampling grid of the standard 33 square convolutional kernel is unable to provide differentiated responses to textures of arbitrary directions. Through its four-directional asymmetric kernels and variable-phase padding, WMConv expands the effective receptive field by 177% without increasing the parameter budget. Simultaneously, its directional convolutional kernels and channel compression mechanism can effectively capture the fine-grained textural features of maize pests (e.g., wing venation, antennae, and body segment edges). Fig 4 depicts the structural design of the WMConv network. It demonstrates superior texture discriminative power on the IPMaize dataset, which features complex backgrounds and small inter-class variations. The specific implementation is as follows:

1. (1) Four branches of 13 and 31 convolutions are respectively paired with asymmetric padding schemes such as (1,0,0,3) and (0,3,0,1) to form a “windmill”-shaped sampling grid.
2. (2) The output channels are split into four equal parts, concatenated, and then compressed via a 22 group convolution. This maintains directional diversity while controlling computational overhead.
3. (3) The WMConv module is replacing all original 33 Conv modules while keeping the overall network architecture unchanged.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. The network architecture of WMConv.

https://doi.org/10.1371/journal.pone.0357034.g004

The padding strategy is denoted as padding (left, right, top, bottom) and the symbols H₁, W₁, C₁ correspond to the spatial dimensions and channel count of the input feature map X₁. First, asymmetric padding is employed to construct horizontal and vertical convolution kernels across different image regions. To promote stable and efficient training, convolutional layers are complemented with Batch Normalization and SiLU-based nonlinear activation. WMConv performs parallel convolutions as described in the following formulas:

(1)

(2)

(3)

(4)

(5)

where X_hor, X_ver, , and represent the features captured by the horizontal, vertical, right-up-to-left-down diagonal, and left-up-to-right-down diagonal kernels, respectively. denotes the intermediate feature map after concatenating the outputs of the four-path windmill kernels, where , denote the spatial dimensions of the intermediate feature map and its channel depth is (since is the output channels of each branch):

(6)

(7)

(8)

where s is the convolution stride to enable the plug-and-play deployment of the WMConv module and ensure its compatibility with the engineering design of convolutional neural networks (CNNs), the number of output channels in the 22 convolution layer is set to C₁, which is identical to the number of input channels. The final output feature map is X, here, H, W, and C denote the feature map’s height, width, and channel count, respectively:

(9)

(10)

(11)

the formula for computing the number of parameters of the convolution layer is:

(12)

if C₁ = C, the parameters of the 33 convolution and WMConv are calculated as follows:

(13)

(14)

by employing four groups of parallel asymmetric kernels for feature extraction, WMConv reduces the parameter count of an individual convolution operator by approximately 22% compared with a standard convolution, while expanding the theoretical receptive field from to . As shown in the upper right corner of Fig 4, WMConv expands the effective receptive field by approximately 177% without increasing the parameter cost of the convolution operator.

To quantitatively illustrate the receptive field enlargement introduced by WMConv, the theoretical receptive field area is calculated as:

(15)

where A_Conv denotes the theoretical receptive field area of standard convolution, and A_WMConv represents the theoretical receptive field area of the proposed WMConv module; refers to the growth rate of receptive field area.

The effective receptive field expands from approximately to , the receptive field area increases from 9 to 25, resulting in:

(16)

To further verify that the enlarged receptive field is not merely a theoretical derivation, the Effective Receptive Field (ERF) was visualized using the gradient-based ERF analysis method. Fig 5 compares the ERF distributions of the baseline YOLOv11n and the proposed WSD-YOLO at both the shallow feature extraction stage (Layer 7) and the deep feature extraction stage (Layer 22).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Effective receptive field (ERF) visualization of the baseline YOLOv11n and the proposed WSD-YOLO.

(a) YOLOv11n (Layer 7). (b) WMConv (Layer 7). (c) YOLOv11n (Layer 22). (d) WMConv (Layer 22).

https://doi.org/10.1371/journal.pone.0357034.g005

As shown in Fig 5(a) and Fig 5(b), compared with the baseline YOLOv11n, WSD-YOLO exhibits a noticeably broader activation region while maintaining strong responses around the target center in the shallow feature layer, indicating enhanced local structural perception. Furthermore, Fig 5(c) and Fig 5(d) show that the effective receptive field of the deep feature layer is also significantly enlarged, enabling richer contextual information to participate in feature extraction. These observations are consistent with the theoretical analysis presented above and qualitatively demonstrate that the proposed WMConv effectively enlarges the effective receptive field in the trained network.

WMConv is designed to strengthen multi-directional structural perception for arbitrarily oriented maize pests while maintaining lightweight computation. Representative convolution enhancement methods mainly improve feature extraction through different receptive-field modeling strategies. The Asymmetric Convolution Block (ACB) proposed in the Asymmetric Convolutional Network (ACNet) enhances kernel representation by combining horizontal and vertical asymmetric branches [39], whereas Wavelet Transform Convolution (WTConv) enlarges the effective receptive field through wavelet-domain frequency decomposition [40]. Different from these methods, WMConv constructs a windmill-shaped sampling pattern using four complementary asymmetric kernels with phase-shifted padding, enabling simultaneous modeling of horizontal, vertical, and diagonal texture responses within a unified lightweight operator. This design is particularly suitable for capturing arbitrary-oriented pest structures and fine-grained edge details in complex field environments.

3.2 Single-Scale High-Level Transformer (SHLT)

In field environments, maize pests often exhibit characteristics such as color similarity between the target and leaves, and abundant repetitive textures. The SPPF module obtains different receptive fields through cascaded max pooling while being computationally efficient, it merely concatenates local contexts and lacks the ability to model global dependencies. The feature processing of SPPF is illustrated by the lower branch in Fig 6.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Comparative diagram of the network architectures of SPPF and SHLT.

The upper branch depicts the SHLT architecture; the lower branch illustrates the network architecture of SPPF.

https://doi.org/10.1371/journal.pone.0357034.g006

This paper introduces the SHLT module to replace the original SPPF architecture in YOLOv11n. Its core concept is to apply self-attention solely to the single feature layer with the richest semantics and lowest resolution, because the semantically richest hierarchy captures sufficient global context, rendering multi-scale modeling unnecessary. This enables the capture of long-range dependencies at a minimal computational cost, enhancing the salience of pest regions while suppressing complex backgrounds, as demonstrated by the upper branch in Fig 6.

The input features are selected exclusively from the deepest layer of the backbone, the P5 feature map (where C = 1024, H = W = 20). This layer possesses the richest semantic information but the lowest spatial resolution, adhering to the single-scale design principle.

First, the two-dimensional feature map is transformed into a one-dimensional token sequence to prepare for subsequent Transformer operations.

(17)

where X_encoded represents the features after serialization and position encoding, where the complete formulation for Position Embedding (PosEmbed) is as follows:

(18)

For the position index of the sequence, denoted as , and the dimension index , the position-encoding matrix can be expressed as:

(19)

where represents the sequence length. The query, key, and value representations are obtained via linear projection as follows:

(20)

(21)

(22)

where are learnable projection matrices, with denoting the projection dimension. Subsequently, long-range contextual relationships across the feature space are captured using a multi-head self-attention (MHSA) module. For the i-th attention head:

(23)

where denote the feature projections associated with the i-th attention head. Subsequently, representations from multiple attention heads are integrated and further processed using a linear transformation:

(24)

where Z is the output of the multi-head self-attention mechanism, with being the output projection matrix and h the number of attention heads. Subsequently, a residual connection is applied followed by layer normalization to stabilize the training process:

(25)

where is the enhanced feature obtained after stable training, where the specific computation of layer normalization is as follows:

(26)

where and are the mean and variance, respectively, and and are learnable parameters. This is followed by the Feed-Forward Network (FFN) with GELU activation:

(27)

where represents the output of the Feed-Forward Neural Network (FFN), which is composed of two linear transformation layers with a GELU activation function applied between them. The hidden layer dimension dff is typically set to four times the input dimension. This is followed by a second residual connection and layer normalization:

(28)

where represents the final result after the second residual connection and layer normalization. Finally, the processed sequence is transformed back into a feature map format, yielding the reshaped feature map X:

(29)

where denotes the spatial dimension reshaping operation. The final features retain the original spatial resolution while integrating global contextual information. This integration can effectively suppress the noise from repetitive leaf textures and enhance the subsequent detection head’s response to weak-contrast regions, such as pest edges and spots.

SHLT is proposed to establish efficient long-range contextual modeling by applying global self-attention only to the highest semantic feature map. Representative Transformer-based feature enhancement methods mainly improve contextual representation through different attention mechanisms. Large Separable Kernel Attention (LSKA) captures long-range spatial dependencies by employing large separable kernels [41], whereas Multi-cognitive Visual Adapter (Mona) enhances feature representation through lightweight visual adaptation [42]. Different from these approaches, SHLT performs global self-attention only on the deepest semantic feature map. This single-scale design preserves sufficient global contextual information while avoiding redundant attention computation on high-resolution feature maps, resulting in a better balance between detection accuracy and computational efficiency.

3.3 Dual-Attention Weighted Aggregation (DAWA)

To address the imbalance between target and background weights within same-scale features, which leads to high miss rates for morphologically diverse maize pests, this paper proposes DAWA (Dual-Attention Weighted Aggregation). Fig 7 provides a depiction of the architectural design of the DAWA network.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. The network architecture of DAWA.

https://doi.org/10.1371/journal.pone.0357034.g007

It takes two feature maps from the same scale as input: X₁ (from the Backbone layer, rich in texture details) and X₂(from the Neck layer, strong in semantic information). LMG (Local-Max and Global-Avg Pooling) combines local maximum and global average pooling, enabling the capture of both local image details and holistic global context. This provides richer feature representations, which helps the model not only enhance target features but also effectively handle complex backgrounds, thereby improving model robustness and generalization. Additionally, it supplies more informative features for subsequent attention mechanisms. It then achieves detail enhancement in complex backgrounds through Dual-Path Channel-Spatial Attention and Adaptive Weighted Fusion.

The represents the shallow features from the Backbone (rich in texture details), and represents the deep features from the Neck (strong in semantics). They are input for subsequent weight generation. The feature enhancement process is defined as follows:

(30)

(31)

where and represent the enhanced features generated after feature reinforcement, here, ⊕ represents addition performed on corresponding elements, indicates element-wise multiplication, and Concat signifies the operation of merging features along the channel axis. and LMG refer to the Sigmoid activation function and Local-Max and Global-Avg Pooling, respectively. The LMG layer fuses the results of local max pooling and global average pooling. indicates two cascaded pointwise convolutional layers. The enhanced features are aggregated in the channel dimension before being processed by parallel attention mechanisms for fusion weight generation. The fusion weight generation process can be expressed as:

(32)

(33)

(34)

where and represent the weights corresponding to channel-wise and spatial attention, respectively, and W denotes the fusion weight. GAP stands for Global Average Pooling. Since features extracted at early stages and those derived from deeper layers are complementary, we generate the fusion weight W based on and apply its complementary weight to . Accordingly, the resulting feature integration procedure can be formulated as follows:

(35)

DAWA is designed to adaptively aggregate complementary backbone and neck features at the same spatial scale before prediction. Representative attention-based feature fusion methods mainly improve feature interaction through different attention strategies. The Convolution and Attention Fusion Module (CAFM) jointly models local and global information by integrating convolution and attention branches [43], whereas Content-Guided Attention Fusion (CGAFusion) generates adaptive attention weights according to feature content [44]. Different from these approaches, DAWA performs adaptive complementary aggregation of backbone and neck features through channel-spatial dual attention. This design generates more discriminative feature representations before prediction while maintaining a lightweight architecture.

3.4 Design rationale and comparison with existing methods

To further clarify the novelty of the proposed modules, Table 2 summarizes the fundamental differences between WMConv, SHLT, and DAWA and representative feature enhancement methods from the perspectives of design objective, information modeling strategy, and architectural characteristics. As can be seen, although these representative methods improve feature representation through different mechanisms, the proposed modules are designed specifically for maize pest detection under complex field environments and therefore differ fundamentally in both design motivation and implementation strategy.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Fundamental differences between the proposed modules and representative feature enhancement methods.

https://doi.org/10.1371/journal.pone.0357034.t002

4.1 Experimental environment and parameter settings

The experiments were conducted using the Windows 10 Professional operating system, version 22H2. The CPU was an Intel Core i9-12900K @ 3.2 GHz with 128 GB of RAM. The GPU configuration consisted of two NVIDIA GeForce RTX 3090 cards, each with 24 GB of memory. The software environment included CUDA version 12.1, PyTorch version 2.2.2, and Python version 3.10. The detailed experimental configuration is shown in Table 3.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Experimental setup and environment.

https://doi.org/10.1371/journal.pone.0357034.t003

Proper configuration of hyperparameters is essential for stable model training and fair performance comparison, and the detailed settings are presented in Table 4. Considering the pronounced domain differences between the IPMaize dataset and generic object-detection datasets, pretrained weights may introduce feature bias unrelated to maize pests. Therefore, all models were trained from scratch under identical settings to ensure that the observed performance differences were primarily attributable to their architectural designs.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Hyperparameters for model training.

https://doi.org/10.1371/journal.pone.0357034.t004

All input images were resized to pixels, which preserves sufficient spatial details for small pest targets while maintaining manageable computational cost. The batch size was set to 32 to ensure stable gradient estimation and efficient utilization of the available GPU memory. SGD was selected as the optimizer because it provided stable convergence during preliminary training, with a momentum of 0.937 to accelerate optimization and reduce gradient oscillation. The initial learning rate (lr0) was set to 0.01, while the final learning-rate factor (lrf) was set to 0.01, resulting in a final learning rate of approximately .

An IoU threshold of 0.7 was adopted to provide a relatively strict matching criterion between predicted and ground-truth bounding boxes. Since the models were trained from randomly initialized weights, the maximum number of training epochs was set to 300 to allow sufficient convergence. Meanwhile, early stopping with a patience of 100 epochs was employed to terminate training when the validation performance no longer improved, thereby avoiding unnecessary computation and potential overfitting. The same hyperparameter settings were applied to all compared models to ensure experimental fairness.

To improve the reproducibility of the experimental results, three random seeds (42, 2026, and 3407) were selected for repeatability analysis.To ensure experimental reproducibility, we performed repeated experiments using three independent random seeds (42, 2026, and 3407) on the proposed model, baseline model and ablation studies to further validate the reliability of performance improvements. The repeated results were reported as mean standard deviation where applicable. For all other main experiments, random seed 42 was uniformly adopted to guarantee consistent experimental conditions across different detection frameworks. Since YOLOv11n and WSD-YOLO were evaluated under identical random seeds, a two-tailed paired Student’s t-test was employed to evaluate whether the observed performance differences were consistent across repeated experiments. A p-value smaller than 0.05 was considered indicative of statistically significant differences.

4.2 Performance evaluation indicators

To achieve a thorough and unbiased assessment of the proposed object detection approach, several evaluation indicators were employed,including Precision (P), Recall (R), F1-score, mean Average Precision (mAP) and frames per second (FPS). Their formulations are given in Equations (36)–(40).

(36)

(37)

(38)

(39)

(40)

where Precision (P) quantifies the ratio of positive instances that are correctly identified by the model relative to all outputs labeled as positive, thereby indicating the trustworthiness of positive predictions. Recall (R) is described as the fraction of true positive samples that are successfully recognized among all annotated positive instances, which reflects the model’s ability to cover the target category. The F1-score is calculated as the harmonic mean of Precision and Recall, serving as a balanced metric that jointly reflects both measures. Mean Average Precision (mAP) comprehensively considers the performance of P and R under varying thresholds, functioning as a primary indicator for measuring the detector’s global recognition capability. This value is computed by combining the areas enclosed by the precision–recall curves across all object classes, serving as a comprehensive indicator of model stability and robustness in multi-class scenarios. In this study, a True Positive (TP) denotes a pest instance that is correctly recognized by the model, whereas a False Positive (FP) represents a non-pest region, such as background or other categories, that is incorrectly classified as a pest. A False Negative (FN) refers to a pest instance that exists in the image but is not correctly identified by the model. Here C denotes the number of pest categories under consideration, and and denote the precision and recall metrics of the i-th pest category, respectively. Frames per second (FPS) denotes the number of images that can be processed and detected by the network per second, which is adopted as the speed metric for object detection. Inference speed was evaluated using single-image inference (batch size = 1) on a single NVIDIA GeForce RTX 3090 GPU with an input imgsz of . To facilitate fair comparison with previous studies, the average inference latency (ms/image) is also reported.

4.3 Detection results of WSD-YOLO

In this experiment, we compared our method with the baseline models of the YOLOv11 series on the IPMaize dataset, as illustrated in Fig 8.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Performance comparison curves of WSD-YOLO and the YOLOv11 series.

(a) mAP@0.5. (b) mAP@0.5:0.95.

https://doi.org/10.1371/journal.pone.0357034.g008

The figure presents the complete convergence process of WSD-YOLO on this dataset. It can be clearly observed that the integration of the WMConv, SHLT, and DAWA modules demonstrated a comparatively strong positive performance. Notably, our model achieved significantly higher scores in both mAP@0.5 and mAP@0.5:0.95 compared to the YOLOv11 series models.

The Fig 9 presents the Precision–Recall (P–R) curves corresponding to YOLOv11n (Fig 9(a)) and WSD-YOLO (Fig 9(b)) on the IPMaize dataset, together with the AP@0.5 for each class and the overall mAP@0.5. Judging from the overall geometric distribution, almost all 13 P–R curves of WSD-YOLO are located above or in the upper-left region compared with the baseline curves. The mAP@0.5 improved from 72.6 % to 79.7 %, representing a relative gain of 7.1 %, which constitutes a significant improvement for agricultural pest-detection tasks.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Precision-Recall (PR) curve.

(a) Precision-Recall (PR) curve of YOLOv11n. (b) Precision-Recall (PR) curve of WSD-YOLO.

https://doi.org/10.1371/journal.pone.0357034.g009

The performance gains for small-sample and hard-sample categories are particularly pronounced: the AP for white-margined moth (n = 49) jumped from 0.301 to 0.766, an increase of 46.5 percentage points, with its P–R curve transitioning from a sharp drop to a gentle decline; the AP for large cutworm (n = 153) increased from 0.358 to 0.561, a gain of 20.3 percentage points. Absolute improvements in AP were also achieved for black cutworm, yellow cutworm, army worm, Potosia brevitarsis, and peach borer.

For categories with sufficient samples and a baseline AP > 0.80, WSD-YOLO further raised the upper limit of recall while maintaining high precision: the AP for both mole cricket and red spider reached 0.995, with their tail curves tending to level off, indicating that the SHLT module effectively suppresses backgrounds with repetitive textures.

The AP decreases observed for grub, wireworm, corn borer, and aphids are attributed to the high visual homogeneity between the pests’ body color/texture and the background (e.g., highlighted soil surfaces, leaf veins, leaf sheaths, tassels), resulting in weak edge discriminability. This leads to misclassification of background highlights, wrinkles, or intersecting leaf veins as pest bodies, consequently causing a reduction in AP values.

To further examine how WSD-YOLO differs from the baseline YOLOv11n in terms of performance, a comparison based on the confusion matrices of the two models was carried out, as shown in Fig 10. The confusion matrix, which visualizes the alignment between predicted categories and ground-truth categories, serves as an important tool for evaluating object detection models by intuitively reflecting the recognition accuracy and misclassification rates for each class.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 10. (a) Confusion matrix of YOLOv11n.

(b) Confusion matrix of WSD-YOLO.

https://doi.org/10.1371/journal.pone.0357034.g010

The normalized confusion matrix for YOLOv11n exhibits distinct characteristics, notably showing high values for pests being missed and classified as background. This indicates a relatively pronounced issue with missed detections during the actual detection process. In contrast, the confusion matrix of WSD-YOLO shows a decreased proportion of pests being missed and classified as background. Overall, the model demonstrates more accurate predictions for the majority of categories and a lower missed detection rate. Notably, the recall for the white-margined moth improved from 25 % to 50 %, representing one of the most groundbreaking improvements in the enhanced model.

4.4 Comparison with other classic models

To assess the effectiveness and robustness of the proposed approach, our method was evaluated against a range of representative advanced and traditional detection models, including the YOLOv5 [45], YOLOv8 [46], YOLOv10 [20], YOLOv11 [21], YOLOv12 [31],RT-DETR-l [47], Hyper-YOLO [48], Maize-YOLO model [49], YOLOv9-t [50], RTMDet-tiny [51], PP-YOLOE-s [52], D-FINE-n [53]. As shown in Table 5, the WSD-YOLO model contains 3.8 M parameters and requires 9.2 GFLOPs, achieving an mAP@0.5 of 79.7%. Its mAP surpasses all other classic detection models, while its computational cost is only 1.5 times higher than that of the most efficient model in the comparison. Simultaneously, our model achieves an appropriate balance among speed, accuracy, and computational cost.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Comparative experiments. Bold values indicate optimal results.

https://doi.org/10.1371/journal.pone.0357034.t005

To further evaluate deployment efficiency, Table 6 compares the model size, peak GPU memory consumption, and inference latency of YOLOv11n and WSD-YOLO. All measurements were obtained using the best-performing weights with an input resolution of and a batch size of 1 on a single NVIDIA GeForce RTX 3090 GPU.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Deployment efficiency comparison between YOLOv11n and WSD-YOLO.

https://doi.org/10.1371/journal.pone.0357034.t006

WSD-YOLO has a model size of 7.6 MB and requires 211.23 MB of peak GPU memory. Compared with YOLOv11n, the peak memory consumption increases by only 6.91%. Meanwhile, WSD-YOLO achieves an inference latency of 8.9 ms/image and a processing speed of 112 FPS, representing only a 4.71% latency increase over YOLOv11n. These results indicate that WSD-YOLO maintains a compact memory footprint and real-time inference capability, demonstrating its potential for agricultural pest monitoring on resource-constrained platforms.

4.5 Ablation and architectural analysis

To validate the impact of the proposed modules on maize pest detection performance and analyze the contribution of each component, ablation experiments were conducted on the IPMaize dataset. YOLOv11n was adopted as the baseline model, and the WMConv, SHLT, and DAWA modules were progressively incorporated into the network. Each configuration was independently trained three times using random seeds 42, 2026, and 3407 under identical experimental settings. Table 7 reports the mean sample standard deviation of the three runs, together with the corresponding model complexity (Params and GFLOPs).

Download:

• PNG
  larger image
• TIFF
  original image

Table 7. Ablation study results reported as mean standard deviation over three independent runs with random seeds 42, 2026, and 3407. Bold values indicate the optimal performance.

https://doi.org/10.1371/journal.pone.0357034.t007

4.5.3 Architectural configuration sensitivity analysis.

To further verify the rationality of the final network configuration, additional sensitivity experiments were conducted from two perspectives: the insertion positions of WMConv and the integration strategy of the Transformer. For each comparison, only the investigated architectural component was changed, while the remaining network structure and training settings were kept unchanged.

1. 1. Sensitivity to WMConv Insertion Positions

WMConv was inserted only in the Backbone, only in the Neck, and simultaneously in both the Backbone and Neck. The results are summarized in Table 8.

Download:

• PNG
  larger image
• TIFF
  original image

Table 8. Sensitivity analysis of WMConv insertion positions. Bold values indicate the optimal performance.

https://doi.org/10.1371/journal.pone.0357034.t008

Compared with YOLOv11n, inserting WMConv only in the Backbone increased mAP@0.5 from 72.6% to 75.4% and recall from 71.6% to 76.1%, indicating that WMConv is particularly effective during backbone feature encoding. In contrast, inserting WMConv only in the Neck produced an mAP@0.5 of 72.5% and reduced recall to 64.9%, suggesting that applying WMConv solely during feature fusion cannot fully exploit its directional representation capability. When WMConv was deployed in both the Backbone and Neck, the model achieved the highest mAP@0.5 of 75.6% and an F1-score of 71%, while maintaining 2.5 M parameters and 6.3 GFLOPs. These results support the placement adopted in the proposed architecture.

1. 2. Comparison of Different Transformer Designs

To evaluate the single-scale design of SHLT, two alternative structures, namely SPPF + Transformer and a dual-scale Transformer, were implemented and compared with the original SPPF and SHLT. The results are presented in Table 9.

Download:

• PNG
  larger image
• TIFF
  original image

Table 9. Comparison of different transformer integration strategies. Bold values indicate the optimal performance.

https://doi.org/10.1371/journal.pone.0357034.t009

Adding a Transformer after SPPF improved mAP@0.5 from 72.6% to 73.7%, confirming the benefit of global contextual modeling. The dual-scale Transformer further increased mAP@0.5 to 73.9%, but raised the parameter count to 5.4 M and the computational cost to 7.4 GFLOPs. In comparison, SHLT achieved the highest mAP@0.5 of 74.6%, precision of 70.9%, recall of 73.7%, and F1-score of 72%, with only 3.2 M parameters and 6.6 GFLOPs. These results demonstrate that applying self-attention only to the deepest semantic feature layer provides a better balance between detection accuracy and computational efficiency than the alternative Transformer designs.

Overall, the sensitivity experiments provide quantitative support for the module placement and Transformer configuration adopted in WSD-YOLO.

4.5.4 Statistical validation and computational complexity analysis.

To further evaluate the reliability of the performance improvement and the computational cost introduced by the proposed modules, repeated experiments and complexity analysis were conducted. The baseline YOLOv11n and the proposed WSD-YOLO were independently trained three times using different random seeds (42, 2026, and 3407), while keeping all other training configurations unchanged. The experimental results are reported as mean standard deviation.

Since YOLOv11n and WSD-YOLO were evaluated under identical random seeds, a two-tailed paired Student’s t-test was employed to determine whether the observed improvements were statistically significant. A p-value smaller than 0.05 was considered statistically significant. The statistical significance test results are summarized in Table 10. The results demonstrate that WSD-YOLO consistently outperformed YOLOv11n across the major evaluation metrics. Specifically, the improvements in Recall (p = 0.0216), mAP@0.5 (p = 0.00016), mAP@0.5:0.95 (p = 0.0070), and F1-score (p = 0.0082) were statistically significant. In contrast, the improvement in Precision did not reach statistical significance (p = 0.2005). These results indicate that the proposed WSD-YOLO mainly enhances detection performance by improving feature representation and target retrieval capability.

Download:

• PNG
  larger image
• TIFF
  original image

Table 10. Statistical significance analysis between YOLOv11n and WSD-YOLO using paired Student’s t-test.

https://doi.org/10.1371/journal.pone.0357034.t010

Furthermore, the repeated experiments show that WSD-YOLO maintains stable performance under different random initializations. The standard deviations of WSD-YOLO for Precision, Recall, mAP@0.5, mAP@0.5:0.95, and F1-score were 0.26%, 1.51%, 0.83%, 0.31%, and 0.58%, respectively. The relatively small variations among three independent runs demonstrate the reproducibility and robustness of the proposed method.

The computational complexity analysis further illustrates the trade-off between detection accuracy and model efficiency. Compared with YOLOv11n, the complete WSD-YOLO increased the number of parameters from 2.6 M to 3.8 M and GFLOPs from 6.3 to 9.2, corresponding to increases of approximately 46.2% and 46.0%, respectively. Meanwhile, the mean mAP@0.5 and mAP@0.5:0.95 were improved by 6.97 and 4.46 percentage points, respectively. These results indicate that the additional computational overhead introduced by WMConv, SHLT, and DAWA provides substantial accuracy improvements while maintaining the lightweight characteristics of the YOLOv11n framework.

Among the proposed modules, WMConv achieves the most favorable accuracy–complexity balance. It improves the mean mAP@0.5 by 2.83 percentage points while reducing the parameter number and maintaining the same computational complexity. SHLT introduces global contextual modeling with only a minor increase in computational cost, whereas DAWA provides stronger feature fusion capability at the expense of additional parameters and GFLOPs. Overall, WSD-YOLO achieves an effective balance between detection accuracy and computational efficiency, demonstrating its potential for practical maize pest detection applications.

4.5.5 Comparison with representative modules.

To further verify the effectiveness of the proposed modules compared with representative feature enhancement methods, additional comparative experiments were conducted. Specifically, WMConv was compared with ACB and WTConv, SHLT was compared with LSKA and Mona, and DAWA was compared with CAFM and CGAFusion. All experiments were performed under identical training settings while replacing only the corresponding modules.

As shown in Table 11, the proposed WMConv, SHLT, and DAWA achieved better detection performance than their corresponding representative alternatives. Compared with ACB and WTConv, WMConv obtained higher accuracy, demonstrating that directional spatial modeling is more suitable for capturing irregular pest textures than conventional asymmetric convolution or frequency-domain feature extraction. SHLT achieved a better accuracy-efficiency trade-off than LSKA and Mona by introducing global contextual modeling only on the deepest semantic feature layer. Meanwhile, DAWA outperformed CAFM and CGAFusion, indicating that adaptive same-scale aggregation of complementary backbone and neck features is beneficial for maize pest detection.

Download:

• PNG
  larger image
• TIFF
  original image

Table 11. Comparison with representative modules.

https://doi.org/10.1371/journal.pone.0357034.t011

Overall, WMConv, SHLT, and DAWA improve feature representation from three complementary perspectives, namely local structural perception, global contextual modeling, and adaptive feature aggregation. Together with the conceptual comparison presented in Table 2, the controlled module-replacement experiments provide empirical support for the effectiveness of the proposed design choices relative to the selected representative methods. These findings further support the effectiveness and design rationale of the proposed WSD-YOLO architecture for maize pest detection.

4.6 Visualization of results

To present a clearer comparison between the detection results of our proposed model and the baseline model YOLOv11n, we selected five representative sets of detection results from the IPMaize dataset for visualization, as shown in Fig 11.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 11. Visualization results on the IPMaize dataset.

(a) Original images. (b) Detections by YOLOv11n. (c) Detections by WSD-YOLO.

https://doi.org/10.1371/journal.pone.0357034.g011

As clearly shown in Fig 11, YOLOv11n still exhibits significant false detections for easily confusable categories. Comparing b(1), b(2) with c(1), c(2), it can be seen that the baseline YOLOv11n model misdetected category 3 (White margined moth) as category 4 (Black cutworm). However, WSD-YOLO not only corrected these misdetections but also achieved a relative increase in target confidence scores.

A comparison between b(3), b(4), b(5) and c(3), c(4), c(5) reveals that the confidence scores for targets detected by YOLOv11n were only 0.82, 0.48, and 0.85, respectively. In contrast, WSD-YOLO increased these confidence scores to 0.92, 0.89, and 0.91. This indicates that the WMConv convolution, the SHLT module, and the DAWA detail fusion module in WSD-YOLO effectively provide more discriminative textural features, enabling correct identification of the corresponding categories with significantly higher confidence. These improvements offer a more reliable decision-making basis for subsequent variable-rate pesticide application and pest infestation warnings in complex field environments.

4.7 Detection error analysis of AP-decreased categories

Although WSD-YOLO achieves overall performance improvements on the IPMaize dataset, slight AP degradation is observed for several categories, including Grub, Wireworm, Corn borer, and Aphids. To further investigate the underlying reasons for these performance variations, representative failure cases were visualized, and different error types were quantitatively counted.

Fig 12 presents representative detection error cases of YOLOv11n and WSD-YOLO for the AP-decreased categories. The first column shows the original images, while the second and third columns present the detection results of YOLOv11n and WSD-YOLO, respectively. Red dashed boxes indicate background false positives (BFP), where non-target background regions are incorrectly detected as pest targets. Green dashed boxes indicate false negatives (FN), where existing pest instances are missed by the detector.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 12. Visualization results on the IPMaize dataset.

(a) Original images. (b) Detections by YOLOv11n. (c) Detections by WSD-YOLO. Red dashed boxes indicate background false positives, and green dashed boxes indicate false negatives.

https://doi.org/10.1371/journal.pone.0357034.g012

To further quantify the sources of detection errors, three major error types were considered, including false negatives, background false positives, and category confusion. The statistical results are summarized in Table 12.

Download:

• PNG
  larger image
• TIFF
  original image

Table 12. Statistical analysis of different error types for AP-decreased pest categories.

https://doi.org/10.1371/journal.pone.0357034.t012

As shown in Table 12, background false positives represent the most frequent error type among the AP-decreased categories. For Grub, Wireworm, Corn borer, and Aphids, 7, 6, 6, and 38 background false positive cases were observed, respectively. These errors mainly result from the visual similarity between pest targets and complex field backgrounds. Specifically, background structures such as soil textures, leaf veins, and crop residues may exhibit similar color or texture patterns to pest bodies, causing incorrect detections.

In addition, category confusion remains an important factor for some pest categories. Wireworm and Corn borer show relatively higher category confusion cases, with 4 and 5 errors, respectively. This indicates that morphological similarities among pest categories may introduce feature ambiguity during classification. Although WSD-YOLO improves feature representation through enhanced feature extraction and fusion mechanisms, distinguishing visually similar pest categories remains challenging under natural field conditions.

Aphids exhibit the highest detection difficulty, with 22 false negatives and 38 background false positives. This is mainly attributed to their extremely small size, weak boundaries, and frequent occurrence on leaf surfaces with complex textures. These characteristics make it difficult for the detector to extract sufficiently discriminative features and accurately localize pest instances.

Overall, the error analysis demonstrates that the AP degradation of several categories is mainly caused by intrinsic challenges in field-based pest detection, including weak target-background discrimination, small object characteristics, and inter-class morphological similarity. These findings further explain the remaining limitations of WSD-YOLO and provide guidance for future improvements in more challenging agricultural scenarios.

4.8 Validation of model generality

To assess the robustness and generalization of the proposed model, experiments were carried out using the Tomato Pest&Diseases dataset [54] and the IP102 dataset.

4.8.1 Tomato Pest&Diseases Dataset.

The Tomato Pest&Disease Dataset contains a total of 2574 images representing 15 different types of tomato pests and diseases, as shown in the Fig 13. The high similarity among different pest and disease categories in the dataset poses a great challenge for accurate detection. The performance of WSD-YOLO and other models are summarized in Table 13. When compared to other existing models, the proposed WSD-YOLO model demonstrated superior detection performance, achieving improvements of 3.9% and 2.0% in mAP@0.5 and mAP@0.5:0.95, respectively, over the baseline model YOLOv11n.

Download:

• PNG
  larger image
• TIFF
  original image

Table 13. Comparative evaluation on the Tomato Pest&Diseases Dataset. Bold values indicate optimal results.

https://doi.org/10.1371/journal.pone.0357034.t013

Download:

• PNG
  larger image
• TIFF
  original image

Fig 13. Distribution of 15 object categories.

https://doi.org/10.1371/journal.pone.0357034.g013

To provide a clear comparison between the performance of our model and the baseline YOLOv11n, three representative detection results from the Tomato Pest&Diseases dataset were selected, as shown in Fig 14. The results clearly demonstrate that WSD-YOLO outperforms YOLOv11n in this dataset. This comparison highlights the superior effectiveness and generalization ability of the proposed WSD-YOLO model on the Tomato Pest&Diseases dataset.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 14. Visualization results on the Tomato Pest&Diseases dataset.

(a) Original images. (b) Detections by YOLOv11n. (c) Detections by WSD-YOLO.

https://doi.org/10.1371/journal.pone.0357034.g014

4.8.2 IP102 dataset.

The IP102 dataset is a widely used pest image dataset, designed to provide a standardized platform for agricultural pest detection and recognition. It covers 102 unique pest categories, including a variety of common pests found in different crops and environments. The entire dataset consists of 18,975 images. The wide variety of pest categories and their high similarity in the dataset present substantial challenges for the accurate detection of the proposed model. When compared to other existing models, as shown in Table 14, the proposed WSD-YOLO model demonstrated superior detection performance, achieving improvements of 4.2% and 3.3% in mAP@0.5 and mAP@0.5:0.95, respectively, over the baseline model YOLOv11n.

Download:

• PNG
  larger image
• TIFF
  original image

Table 14. Comparative evaluation on the IP102 dataset. Bold values indicate optimal results.

https://doi.org/10.1371/journal.pone.0357034.t014

Three representative detection results on the IP102 dataset are illustrated in Fig 15. In the last group, YOLOv11n caused misdetection. Experimental results on the IP102 dataset demonstrate that our proposed WSD-YOLO model achieves favorable detection performance and strong generalization ability.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 15. Visualization results on the IP102 dataset.

(a) Original images. (b) Detections by YOLOv11n. (c) Detections by WSD-YOLO.

https://doi.org/10.1371/journal.pone.0357034.g015

4.8.3 Quantitative analysis of cross-dataset performance degradation.

Although WSD-YOLO was evaluated on the Tomato Pest&Diseases and IP102 datasets, its detection performance varies across datasets because of domain shift. Therefore, the relative performance degradation (PD) is introduced to quantitatively evaluate the performance difference between the source-domain IPMaize dataset and each target dataset. All results in this subsection are obtained using the same experimental configuration as the preceding cross-dataset experiments.

(41)

where and denote the mAP@0.5 values obtained on IPMaize and the corresponding target dataset, respectively. A lower PD indicates better generalization under the evaluated target-domain conditions.

As shown in Table 15, the PD of WSD-YOLO is 56.84% on the Tomato Pest&Diseases dataset and 11.79% on the IP102 dataset. The larger degradation on Tomato Pest&Diseases can be attributed to its substantial differences from IPMaize in crop species, object categories, imaging conditions, background characteristics, and target appearance. In contrast, IP102 contains numerous agricultural pest categories with visual characteristics similar to those in IPMaize, resulting in a relatively smaller distribution discrepancy and lower performance degradation.

Download:

• PNG
  larger image
• TIFF
  original image

Table 15. Quantitative analysis of cross-dataset performance degradation between YOLOv11n and WSD-YOLO.

https://doi.org/10.1371/journal.pone.0357034.t015

Despite domain-induced performance degradation, WSD-YOLO consistently achieves higher absolute detection accuracy than YOLOv11n on both target datasets. Its PD is slightly lower on the Tomato Pest&Diseases dataset, whereas a higher PD is observed on IP102 due to its substantially improved source-domain performance. These results suggest that WSD-YOLO maintains strong cross-dataset performance, although the relative performance retention varies across target domains.