A Detection Method for Sweet Potato Leaf Spot Disease and Leaf-Eating Pests

Abstract

Traditional sweet potato disease and pest detection methods have the limitations of low efficiency, poor accuracy and manual dependence, while deep learning-based target detection can achieve an efficient and accurate detection. This paper proposed an efficient sweet potato leaf disease and pest detection method SPLDPvB, as well as a low-complexity version SPLDPvT, to achieve accurate identification of sweet potato leaf spots and pests, such as hawk moth and wheat moth. First, a residual module containing three depthwise separable convolutional layers and a skip connection was proposed to effectively retain key feature information. Then, an efficient feature extraction module integrating the residual module and the attention mechanism was designed to significantly improve the feature extraction capability. Finally, in the model architecture, only the structure of the backbone network and the decoupling head combination was retained, and the traditional backbone network was replaced by an efficient feature extraction module, which greatly reduced the model complexity. The experimental results showed that the mAP0.5 and mAP0.5:0.95 of the proposed SPLDPvB model were 88.7% and 74.6%, respectively, and the number of parameters and the amount of calculation were 1.1 M and 7.7 G, respectively. Compared with YOLOv11S, mAP0.5 and mAP0.5:0.95 increased by 2.3% and 2.8%, respectively, and the number of parameters and the amount of calculation were reduced by 88.2% and 63.8%, respectively. The proposed model achieves higher detection accuracy with significantly reduced complexity, demonstrating excellent performance in detecting sweet potato leaf pests and diseases. This method realizes the automatic detection of sweet potato leaf pests and diseases and provides technical guidance for the accurate identification and spraying of pests and diseases.

1. Introduction

Sweet potato is an important economic crop, ranking fifth in tropical production and seventh in global food production, following wheat, rice, corn, potato, barley and cassava. Sweet potato is not only an important food source worldwide but is also widely used in animal feed and traditional medicine. All parts of the plant, including roots, vines and young leaves, have a wide range of application values [1]. However, sweet potato production faces serious problems with pests and diseases, which not only hinder the healthy growth of the crop but also severely impact its yield and quality. The initial symptoms of sweet potato leaf spot disease are small, irregular brown lesions on the edge or middle of the leaves. These spots gradually expand to a diameter of 5–25 mm, developing into gray-brown necrotic lesions with a light-yellow halo. Eventually, the diseased leaves age and fall off [2]. In addition, ground pests, such as hawk moths and wheat moths, also severely harm the growth of sweet potatoes. They eat the leaves, young shoots, tender stems and tender shoots of sweet potatoes, and even roll up the leaves or bite holes in them. In severe cases, only the veins may remain, resulting in large-scale damage to the leaves [3]. The damage caused by pests and diseases to sweet potato leaves reduces the photosynthesis of leaves and has a severe impact on the yield of sweet potatoes. For sweet potato crops produced on a large scale, how to efficiently and accurately detect sweet potato leaf spot and ground pests and take effective prevention and control measures in time is of great significance to reducing the loss of sweet potato production.

In the use of image processing technology and machine learning for crop pest and disease identification, researchers usually use traditional image processing methods to extract characteristic parameters, such as the color, texture and shape of plant diseases, and combine algorithms, such as principal component analysis [4], Bayesian discriminant [5], BP neural network [6] and support vector machine [7], to classify and identify plant diseases and pests. These methods have achieved good results to a certain extent. However, in practical applications, when faced with complex environments, it is difficult to ensure the stability and robustness of the recognition model [8].

With the development of deep learning technology, many achievements have been made in the research of plant pest and disease identification. Its core feature is the use of multi-layer neural networks for feature learning and pattern recognition. Through multi-level nonlinear transformation, deep learning technology can extract more advanced and abstract features from raw data, thereby better distinguishing different pest and disease samples and improving recognition accuracy and generalization ability. Researchers initially used a convolutional neural network (CNN) to identify plant pests and diseases [9,10,11], but training a CNN requires large data sets. Other researchers used transfer learning to adjust pre-trained CNNs [12,13] to solve the problem of insufficient data. With the continuous development of target detection technology, two-stage algorithms, such as Faster R-CNN [14] and Mask R-CNN [15], as well as one-stage target detection algorithms, such as SSD [16] and YOLO series [17], have gradually been applied to plant pest and disease detection, which has continuously improved the accuracy of plant pest and disease identification and classification.

Since single-stage target detection algorithms, especially the YOLO series algorithms, are widely used in various pest monitoring scenarios due to their real-time detection performance. Researchers have made improvements to these to obtain better actual detection results. Some researchers have introduced feature extraction and feature fusion networks, such as the sand clock feature extraction module (SCFEM) [18], multi-level feature fusion method [19], and bi-directional feature pyramid network (BiFPN) [20], to improve the model’s feature extraction ability and multi-scale feature fusion ability, thereby improving detection accuracy. Some researchers have enhanced the model’s ability to focus on areas of interest for pests and diseases by introducing attention mechanisms, the convolutional block attention module (CBAM) [21], efficient channel attention (ECA) [22], coordinate attention (CA) [23], transformer [24] and simple spatial attention (SSA) [25], at different positions in the network architecture. Some researchers have enhanced the feature extraction capability and improved the accuracy of crop disease and pest identification and detection by integrating convolution methods, switchable atrous convolution (SAConv) [26], Grouped Spatial Convolution (GSConv) [27], coordinate depthwise convolution (CDC) [28], spatial and channel reconstruction convolution (SCconv) [29] and omni-dimensional dynamic convolution (ODConv) [30]. Some researchers have reduced the number of model parameters and computational complexity by incorporating lightweight modules, ghost modules [31,32], or lightweight backbone networks, including MobileNet [33], GhostNet [34], ShuffleNet [35], and EfficientNet [36]. Additionally, network pruning has been used to reduce the number of parameters and computational complexity [37,38].

In terms of sweet potato pests and diseases, research in this field has mainly focused on the prevention and control of sweet potato virus diseases [1], because once a virus disease breaks out, it may cause crop failure. However, with the development of sweet potato virus-free technology, sweet potatoes planted with “virus-free” seedlings are almost virus-free, which can significantly reduce the incidence of field diseases [39]. However, sweet potato leaf spots and leaf-eating pests (hawk moths and wheat moths) still pose a threat to crops. Leaf spot [2] initially manifests as brown lesions on the edges or middle of the leaves [40]. As the lesions expand, they will reduce photosynthesis [41] and, in severe cases, cause leaf shedding. Pests feed on leaves, young shoots and tender stems, which can also cause leaf damage and reduce photosynthesis. Timely detection of these pests and diseases is crucial to increasing sweet potato yields. In terms of crop pest and disease research, researchers focus on crops, corn [42], rice [43], wheat [44] and fruits [45]. There are relatively few studies on sweet potato pest and disease detection. Initially, researchers used simple convolutional networks and SVM classifiers to detect common diseases of sweet potatoes [46,47,48,49,50]. Since the sweet potato planting pattern is one ridge and two rows, the occlusion between leaves is very serious. The performance of the detection algorithm based on machine learning and simple neural networks is limited, and the detection effect of sweet potato leaf diseases and pests is limited. Later, Rahman et al. [51] used YOLOv4 to detect sweet potato leaf spots and yellow spots but did not improve the model based on the characteristics of sweet potato diseases and pests, and the detection effect of sweet potato diseases and pests was also limited. The existing deep learning algorithm is not suitable for the detection of sweet potato leaf spot and leaf-feeding pests, and there are problems, low detection accuracy and poor robustness. Therefore, it is urgent to design a new sweet potato disease and pest detection model based on the characteristics of sweet potato leaf spot and leaf-feeding pests gnawing on leaves.

To address the problem of pests and diseases affecting sweet potato leaves, this study proposes a YOLO-based sweet potato leaf disease and pest (SPLDP) detection method, aiming to achieve timely and accurate monitoring of leaf spot disease and pests (hawk moths, wheat moths, and other above-ground insects) on sweet potato leaves. The main work of this paper includes the following aspects: (1) A local feature module (LF) consisting of three depth-separable convolutional layers and a jump connection is proposed, which can effectively retain and extract the key feature information of sweet potato leaves and avoid the loss of feature information when the network depth is deepened. (2) The LF module and the Enhanced Local Attention (ELA) attention mechanism are integrated to design a feature extraction structure based on the residual mechanism (FESRM) to improve the model feature extraction ability. (3) The structure of the traditional YOLO backbone network, neck network and decoupling head is changed, and only the structure of the backbone network and decoupling head is retained. The FESRM module is combined to replace the backbone network of YOLO, which greatly reduces the complexity of the model. (4) The designed network structure improves the detection accuracy of sweet potato leaf diseases and pests. (5) The designed model network structure has fewer parameters and calculations, and the network model is more compact.

The rest of this paper is organized as follows. Section 2 lists the experimental materials used in the paper and introduces the designed SPLDP. Section 3 analyzes the detection effect of SPLDP and then verifies the performance of the model through ablation and comparative experiments. Section 4 discusses the limitations of the proposed method and data. Section 5 summarizes the detection performance of SPLDP.

2. Materials and Methods

2.1. Acquisition and Annotation of Images

This paper focuses on sweet potato leaf spot disease and ground pests (e.g., hawk moths and wheat moths) that feed on leaves. The research area is located in Haitou Town, Danzhou City, Hainan Province. A total of 983 sweet potato photos were collected with an image resolution of 4000 × 3000 pixels. The shooting height range was 50–80 cm, and the images were saved in JPG format. The data collection period was from 10 a.m. to 2 p.m. The sample images are shown in Figure 1, where Figure 1a is the original sweet potato image, Figure 1b shows the sweet potato leaves affected by pests, and Figure 1c shows the symptoms of leaf spot disease.

Since ground pests, such as hawk moths and wheat moths, that feed on sweet potato leaves have different foraging patterns, they may not always appear on the surface of the leaves and may not be detected. However, their common feature is the holes left after gnawing on the sweet potato leaves (as shown in Figure 1b), and their prevention and control methods are also similar. Therefore, this study labels all sweet potato leaves showing signs of gnawing as pest labels, while leaves with leaf spot disease are marked as disease labels.

This paper used LabelImg software (https://github.com/HumanSignal/labelImg/blob/master/LICENSE, accessed on 1 January 2025) to manually annotate the images. As shown in Figure 1d, the sweet potato leaves with leaf spot disease were annotated with a black frame, and the label category was represented by “1”. The leaves eaten by insects were annotated with a red frame, and the label category was represented by “0”. To ensure annotation effectiveness, sweet potato leaves obscured by more than two-thirds were excluded from annotation. After annotation, a txt file containing category and coordinate information was generated. These annotated data were divided into training, validation and test set in a ratio of 8:1:1. In order to expand and enhance the diversity of the dataset, we performed a variety of data augmentation operations on the original images, including cropping, rotation, translation, noise addition, cutout, and up and down, left and right flipping. Each image was randomly combined with the above six data enhancement methods, and each image was expanded to three images. The numbers of images in the expanded training set, validation set and test set were 4718, 590 and 590, respectively.

2.2. The Algorithm Principle of YOLO11

YOLOv11 [52] is the latest object detection algorithm developed by the Ultralytics team and is currently one of the best object detection models. As shown in Figure 2, the YOLOv11 network structure is mainly composed of a backbone network, a neck network, and a decoupling head. The YOLOv11 backbone network mainly includes C3k2, SPPF [53], and C2PSA modules. The C3k2 module improves the detection performance of multi-scale targets through a unique design. The C2PSA module integrates the attention mechanism, allowing the model to focus more on key features. The neck network primarily consists of C3K2 modules, which fuse low-level detailed features with high-level semantic features through upsampling and downsampling operations, resulting in a more comprehensive and enriched feature representation. Although YOLOv11 has made significant breakthroughs in many aspects, it may still encounter challenges, such as missed detections, incorrect detections, and duplicate detections in practical applications like sweet potato leaf pest detection. In response to these problems, further improvement and optimization of YOLOv11 can effectively improve the detection accuracy and achieve accurate identification of pest types.

2.3. The Algorithm Principle of SPLDP

Based on the structure of YOLOv11, an improved model SPLDP is proposed, as shown in Figure 3. The SPLDP model is mainly composed of a backbone network and a head. In the backbone network of SPLDP, the core modules include convolutional modules, FESRM and spatial pyramid pooling-fast (SPPF). For the task of pest and disease detection in sweet potatoes, a specific feature extraction module, FESRM, based on residual structure, is designed to extract the features from sweet potato leaves affected by pests or leaf spots. The structure of FESRM is concise and efficient, primarily consisting of two depthwise separable convolution layers, a 3 × 3 downsampling convolution layer, LF and ELA [54]. LF consists of a skip connection and three 3 × 3 depthwise separable convolutions, which can effectively extract local features while reducing computational complexity. ELA consists of a skip connection, a convolution module and average pooling, which accurately captures the location of the region of interest, enables the model to focus on the key area, and suppresses the interference of irrelevant background, thereby further improving the model’s accuracy in sweet potato pest and disease detection. In order to further improve the feature extraction performance of the network, the FESRM module is reused multiple times in the network, thereby enhancing the multi-level feature-learning ability of the model and improving overall detection performance. The designed SPLDP network has two versions, SPLDPvT and SPLDPvB, corresponding to the N version and S version of YOLOv11, respectively. N0, N1, N2, N3, and N4 in Figure 3 represent the number of repeated executions of the FESRM structure.

2.3.1. Feature Extraction Structure Based on the Residual Mechanism

The plug-and-play modules used by existing models, such as YOLO and Transform, can be summarized as the structure shown in Figure 4a. First, the number of channels is expanded through a convolution layer. Then different feature extraction networks are used to process the input data, and finally the number of channels is compressed through a 1 × 1 convolution layer. As shown in Figure 4b, the general structure is simplified to include only convolution operations. On this basis, further optimization measures are proposed. The first improvement is shown in Figure 4c. After the 1 × 1 convolution, a feature extraction module is added, then a 3 × 3 convolution layer is connected, and finally, the number of channels is compressed through a 1 × 1 convolution. This structure can extract local features more effectively and improve the expressiveness of the model. The second improvement is shown in Figure 4d. An attention mechanism is introduced after the 3 × 3 convolution and the 1 × 1 convolution. The attention mechanism can dynamically weight the input features so that the model pays more attention to the key areas and ignores the background noise, further improving the accuracy of feature extraction. The third improvement is shown in Figure 4e, which integrates the local feature extraction module and the attention mechanism. Finally, as shown in Figure 4f, the LF module is designed and the ELA attention mechanism is incorporated based on the above improvements. Through these improvements, FESRM combines the advantages of the LF module and the ELA attention mechanism, significantly improving the effect of feature extraction.

2.3.2. Local Feature Extraction Module

In order to avoid the loss of feature information due to the increase in the number of network layers, an improved structure LF is proposed based on the LocalSpatial block [55], as shown in Figure 5. The structure contains three depthwise separable convolutional layers and a skip connection, which aims to effectively retain important feature information through residual connections. The LF block mainly consists of two parts. The first part includes a feature extraction network, which consists of 3 × 3 depthwise separable convolution, normalization layer and activation function. The second part fuses the output of the first part with the original input feature map through skip connections. In the first part, the local spatial information in the input feature map is effectively extracted through depthwise separable convolution. Then, batch normalization is used to reduce overfitting and provide a certain regularization effect. Subsequently, depthwise separable convolution and the GELU activation function are used to better retain the distribution of information, enabling the model to learn more complex feature representations. The activation function uses nonlinear GELU to change the number of channels of the feature without changing the spatial dimension, thereby enhancing the representation of the feature. In the second part, the design of skip connections enables the network to retain more low-level information during the learning process while avoiding the loss of information when it is transmitted in multiple layers. Through these two parts, the LF block makes the model more layered and richer when extracting features, and can capture detailed information in the image while maintaining efficient computing performance. Especially when dealing with tasks, such as pest and disease detection, it can significantly improve the recognition accuracy and robustness of the model.

2.3.3. ELA Attention Mechanism

The attention mechanism can significantly enhance the network’s ability to process high-level semantic information and improve attention to different regions of interest, thereby greatly improving the accuracy of the model. To select a suitable attention mechanism, several mainstream attention algorithms were compared, including CA, CBAM, CCA [56], SE [57], SEAM [58], ECA, and ELA. After comparative experiments, ELA was finally selected as the attention mechanism module.

As shown in Figure 6, the convolution input of ELA is represented as $R^{H\times W\times C}$, where $H$, $W$, and $C$ represent the height, width, and number of channels of the feature map, respectively. In the ELA, each channel is first average pooled along the horizontal direction $(H,1)$ and along the vertical direction $(1,W)$, respectively. This produces two outputs: one is the representation of each channel in the vertical direction (height is $H$), and the other is the representation of each channel in the horizontal direction (width is $W$). Then, a one-dimensional convolution operation is applied to further enhance the horizontal and vertical position information, capture the spatial structure information in the input feature map, and enable the network to better understand the relationship between various regions in the image. Group Normalization (GN) is then used to process the enhanced position information to further improve the stability and consistency of the features, and obtain the attention representation of the position in the horizontal and vertical directions. Finally, the activation function Sigmoid is used to generate position attention predictions in two directions.

Compared with other methods, the ELA attention mechanism is more flexible and accurate in processing local features, especially in complex scenes, and can effectively enhance the expression ability of the network. It can better capture spatial features, and through independent processing in two directions, it ensures the model’s accurate positioning and information integration of the area of interest.

2.4. The Indicators of Evaluation

In order to evaluate the performance of the sweet potato leaf pest and disease detection model proposed in this paper, recall rate (R), precision rate (P) and mean average precision (mAP) were selected as the main evaluation indicators. These indicators can effectively reflect the accuracy, comprehensiveness and overall performance of the model in the detection task. The calculation formula is as follows:

(1)$R={\displaystyle \frac{TP}{TP+FN}}\times 100\%$

(2)$P={\displaystyle \frac{TP}{TP+FP}}\times 100\%$

(3)$AP={\int }_0^{1}P\left(R\right)dR$

(4)$mAP={\displaystyle \frac{\sum _{i=1}^{n}A{P}_i}{n}}$

where, $TP$ stands for true positives, representing the number of pest-infested areas correctly detected by the model, and $FP$ stands for false positives, representing the number of non-pest-infested areas that the model misidentifies as pest-infested areas. $FN$ stands for false negatives, representing the number of actual pest-infested areas that the model fails to identify, $N$ is the total number of categories, and ${AP}_i$ is the average precision of the $i$-th category, which is expressed as the area under the P-R curve.

3. Experimental Design and Results Analysis

3.1. Experimental Environment and Parameter Setting

The SPLDP model and all comparison models were trained on the GPU server. The specific hardware configuration and software environment settings are shown in Table 1.

In this study, all the comparison models were tested on the constructed sweet potato leaf disease and insect pest dataset under the same hardware environment and trained from scratch. In order to ensure the stability and convergence of the training, we selected a smaller batch size and a lower learning rate for training according to the recommendations of the literature [59] and the computer hardware specifications. The specific training parameters were set as follows: the input image size was 640 × 640 pixels, the batch size was 16, the optimizer was stochastic gradient descent (SGD), the initial learning rate was 0.01, the weight decay was 0.0005, the number of data loader worker threads was 16, the momentum was 0.937, and the model was trained for 300 epochs. The early stopping strategy was applied during training. If the model’s validation accuracy did not improve significantly for 100 consecutive epochs, training would automatically stop to prevent overfitting and save computational resources. Additionally, all other training parameters were kept at their default settings.

The training results of the SPLDPvB and SPLDPvT models are shown in Figure 7. From the training curve, it can be observed that both versions of the model gradually converge after several rounds of training. As the training progresses, the loss function value of the model gradually decreases, and the precision and recall rate also steadily increase, indicating that the model is constantly optimizing and gradually reaching a relatively ideal performance level.

3.2. Ablation Experiment of FESRM Structure

To analyze the impact of different improved structures in Figure 4 on model performance, we compared the structures shown in Figure 4b–d,f to verify their effectiveness in enhancing model performance. In the experiment, YOLOv11S and SPLDPvB were selected as comparison models. The experimental results are shown in Table 2.

As shown in Table 2, compared with YOLOv11s, SPLDPvB-FESRM-Base exhibits a 0.3% and 1.7% reduction in mAP0.5 and mAP0.5:0.95, respectively, along with an 89.3% reduction in the number of parameters and a 71.8% reduction in computational cost. When the LF is used alone in the FESRM structure, compared with the YOLOv11s model, mAP0.5 and mAP0.5:0.95 are increased by 1.7% and 1.3%, respectively, the number of parameters is reduced by 89.2%, and the amount of calculation is reduced by 63.8%. When the ELA is used alone in the FESRM structure, compared with the YOLOv11s model, mAP0.5 and mAP0.5:0.95 are increased by 1.7% and 1.4%, respectively, the number of parameters is reduced by 89.3%, and the amount of calculation is reduced by 71.8%. When LF and ELA are used simultaneously in the FESRM structure, compared with the YOLOv11s model, mAP0.5 and mAP0.5:0.95 are improved by 2.3% and 2.8%, respectively, the number of parameters is reduced by 89.2%, and the amount of calculation is reduced by 63.8%. In addition, compared with the SPLDPvB-FESRM-Base, SPLDPvB-FESRM-LF and SPLDPvB-FESRM-ELA models, the final SPLDPvB model has significantly improved mAP0.5 and mAP0.5:0.95, respectively, while the number of parameters and the amount of calculation of the model have not changed significantly. The ablation experiment on FESRM demonstrates that the inclusion of LF and ELA significantly enhances the model’s detection accuracy. Although the amount of calculation is slightly increased, the number of parameters and the size of the model are greatly reduced, proving that our improvement is effective.

3.3. Comparative Experiments of Different Attention

To analyze the impact of different attention mechanisms on the model, we compared various mechanisms, including CA, CBAM, CCA, SE, SEAM, ECA, and ELA, to evaluate their effect on model performance. The experimental results are shown in Table 3, with SPLDPvB serving as the benchmark model for comparison.

As shown in Table 3, the model with the ELA attention mechanism has the highest mAP0.5 and mAP0.5:0.95 values, which are 88.7% and 74.6%, respectively. The model with the CBAM attention mechanism is second, with mAP0.5 and mAP0.5:0.95 values of 88.5% and 74.2%, respectively. However, the number of parameters of SPLDPvB-ELA is 51.5% less than that of SPLDPvB-CBAM. The model with the ELA attention mechanism has the fewest parameters, followed by the model with ECA. In terms of computational complexity, SPLDPvB-ELA and SPLDPvB-ECA both achieve the lowest complexity at 7.7. However, SPLDPvB-ECA does not perform well in improving model accuracy, and its parameter quantity and model size are higher than those of SPLDPvB-ELA. After comprehensively considering the average accuracy, parameter quantity and computational complexity, the more balanced ELA attention mechanism is selected.

3.4. Ablation Experiment

To verify the effects of LF, ELA and different necks on model performance, we conducted four groups of ablation experiments. As shown in Table 4, to demonstrate the improvement in network performance by removing the neck component, the models in the first, second, and third groups of experiments use different neck structures, while the model in the fourth group removes the neck component entirely. Model1, Model2, Model3 and Model4 in each group of experiments correspond to Figure 4b–d,f in the FESRM part. The first, second and third groups use the neck structures of YOLOv8, YOLOv10 and YOLOv11, respectively, and the fourth group does not use any neck structure.

As shown in Table 4, when the neck structure is different and when the LF and ELA structures are not used, the accuracy of the model without a neck is lower than that of the model with a neck. For instance, the mAP of Model1 in Group 4 is lower than that of Model1 in Groups 1, 2, and 3. When only LF or ELA or LF and ELA structures are used at the same time, each model in Groups 1, 2, and 3 is compared with the corresponding model in Group 4. The models in Group 4, except Model1 (because Model1 does not contain any LF or ELA modules), are better than the corresponding models in other groups in terms of mAP0.5 and mAP0.5:0.95. It can be clearly seen from the bar graph in Figure 8a that the height of the bar graph in Group 4 (except Model1) is higher than the bar graph of the corresponding models in other groups. In terms of model complexity, the number of parameters and the computational cost of models without a neck structure (Model 1, Model 2, Model 3, and SPLDPvB in the fourth group) are lower than those of models with any neck structure. As shown clearly in the bar graph in Figure 8b, the height of the bars representing the models in the fourth group is consistently lower than those of the corresponding models in the other groups. After removing the neck part, the proposed SPLDP model demonstrates better performance and higher accuracy, while maintaining a lower number of parameters and reduced computational cost.

When comparing experiments in the same group, that is, the structure of the neck part is the same, such as Group 1 in Table 4. Compared with Model1, Model2 (using LF alone in FESRM structure) has mAP0.5 and mAP0.5:0.95 increased by 1.7% and 2.3%, respectively, the number of parameters increased by only 4.3%, and the amount of calculation increased by 16.2%. Compared with Model1, Model3 (using ELA alone in FESRM structure) shows that mAP0.5 and mAP0.5:0.95 are improved by 1.7% and 1.7%, respectively, and the number of parameters and the amount of calculation are almost unchanged. Compared with Model1, Model4 (using LF and ELA in FESRM structure) shows that mAP0.5 and mAP0.5:0.95 are improved by 2.1% and 2.8%, respectively, and the number of parameters and the amount of calculation are only slightly increased. It can be clearly seen in Figure 8 that in the second, third and fourth groups of experiments, the Model4 and SPLDPvB using both LF and ELA have higher model accuracy than Model1, Model2 and Model3. The additional structure will inevitably increase the computational overhead of the model; however, the additional computational overhead is acceptable for the significant improvement in model accuracy.

The ablation experiment shows that the proposed SPLDP structure, when not using the neck structure, not only has lower model complexity but also has a significantly improved accuracy. After adding the LF and ELA structures, the feature extraction capability of the model is greatly improved, which is more conducive to the accurate detection of sweet potato pests and diseases in the field environment.

3.5. Comparison Experiment of the Mainstream One-Stage Algorithm

To demonstrate the model performance of SPLDP, it is compared with other SOTA models, including Faster R-CNN, SSD, YOLOV5, YOLOv7, YOLOv8, YOLOv9, YOLOV10, and YOLO11. The comparison results are shown in Table 5.

As shown in Table 5, the proposed SPLDPvT achieves the best performance in terms of parameter quantity, computational complexity, and model size. Its parameter quantity is 63.1% less than that of YOLOv9T, which has the smallest parameter quantity, and its computational complexity is 14.2% less than that of YOLOv11N, which has the lowest computational complexity. At the same time, SPLDPvT improves mAP0.5 by 9.4%, 14.2%, 1.1%, 1.2%, 3.8%, and 1.7%, respectively, compared to the Faster R-CNN, SSD, YOLOv5N, YOLOv8N, YOLOv10N, and YOLOv11N models. Similarly, it improves mAP0.5:0.95 by 27.4%, 26.7%, 0.8%, 0.8%, 2.9%, and 1.2%, respectively. Although it is 0.5% less than YOLOv9T’s mAP0.5:0.95, mAP0.5 has increased by 1%. As can be seen from Figure 9, the proposed SPLDPvT version has a lower model complexity.

The proposed SPLDPvB model outperforms other models in Table 5 in terms of precision, recall, mAP0.5, and mAP0.5:0.95. Compared with Faster R-CNN, SSD, YOLOv5S, YOLOv8S, YOLOv9S, YOLOv10S and YOLOv11S, the mAP0.5 increased by 12.2%, 17.0%, 2.2%, 2.6%, 1.7%, 2.1% and 2.3%, respectively, and the mAP0.5:0.95 increased by 32.9%, 32.2%, 2.1%, 2.7%, 1.8%, 2.7% and 2.8%, respectively. The number of parameters decreased by 99.1%, 95.8%, 87.9%, 90.0%, 84.5%, 86.2% and 88.2%, respectively, and the amount of calculation decreased by 97.9%, 87.7%, 67.6% and 72.8%, 71.1%, 68.4% and 63.8%, respectively. When the model accuracy is improved, the number of parameters and the amount of calculation are greatly reduced. Moreover, the parameters of SPLDPvB are reduced by 56.0%, 63.3%, 42.1%, 57.6% and 56.0%, respectively, compared with those of YOLOv5N, YOLOv8N, YOLOv9T, YOLOv10N and YOLOv11N models. It proves that the proposed SPLDPvB model has lower complexity and higher accuracy and is more suitable for sweet potato leaf pest detection.

In addition, compared with the most accurate models of other YOLO versions, SPLDP-B has mAP0.5 that is 2.4%, 2.6%, 1.7%, 1.6% and 0.9% higher than YOLOv5X, YOLOv8S, YOLOv9S, YOLOv10X and YOLOv11M, respectively, and mAP0.5:0.95 that is 0.9%, 1.5%, 1.7%, 0.4% and 0.3% higher than YOLOv5X, YOLOv8X, YOLOv9C, YOLOv10X and YOLOv11X, respectively. As shown in Figure 9a,b, the proposed model is more accurate than the best model of other YOLO series, while the number of parameters and calculations are greatly reduced, which greatly reduces the model’s complexity.

3.6. Performance Analysis of Algorithms in Complex Environments

In order to fully demonstrate the effect of the proposed SPLDP on sweet potato leaf pests and diseases detection, the detection effect comparison under typical complex scenes is shown in Figure 10. As shown in the figure, Faster R-CNN, SSD, YOLOv5, YOLOv8, YOLOv9, YOLOv10, YOLOv11, and SPLDP can accurately detect most sweet potato pests and diseases. However, when the leaves are mutually occluded or the target is small, these models, including SPLDPvT, encounter issues, such as missed detection, false detection, repeated recognition, and incomplete target recognition. Compared with other models, SPLDPvT and SPLDPvB have lower missed detection rates and wrong detection rates.

As shown in Figure 10b,c, there are many missed detections (marked by yellow ovals), wrong detections (marked by black ovals), and incomplete detections (marked by pink ovals) for the Faster R-CNN and SSD models, especially for the SSD model. In addition, as shown by the yellow ellipse marks in Figure 10d,e,h,j,l, the YOLOv5N, YOLOv5S YOLOv9T, YOLOv10N and YOLOv11N models all have some degree of missed detections. Except for YOLOv10N, the missed targets are mainly small targets. The YOLOv10N model has a poor detection effect on the diseases and pests of sweet potato leaves, and there are multiple missed detections. As shown in the black ellipses of Figure 10d–n, the YOLOv5, YOLOv8, YOLOv9, YOLOv10, YOLOv11 series and SPDPvT models identify sweet potato leaves without any diseases and pests as leaves with pests. At the same time, as shown in the orange ellipses of Figure 10d–n, they identify diseased leaves as leaves with pests, which has the problem of false detection. In addition, as shown in the purple ellipses of Figure 10b,l,n, they mistakenly detect two overlapping leaves as a leaf that is eaten by insects or has leaf spot disease. As shown in the red ellipses of Figure 10d,f,l, YOLOv5N, YOLOv8N and YOLOv11N repeatedly detect the same leaf as two leaf targets. In addition, as shown in the brown ellipses in Figure 10f,l, YOLOv8N and YOLOv11N identify a weed with similar color as a leaf, which results in the problem of incorrectly detecting the target. As shown in Figure 10n,o, the SPLDP model has relatively fewer problems of missed detection, false detection, and incomplete detection than the Faster RCNN, SSD, and YOLO series models.

In summary, by comprehensively comparing the detection effect and model complexity, the model we proposed has lower computational complexity and parameter volume. The SPLDPvT model has a lower missed detection rate and false detection rate, while the SPLDPvB model has higher detection performance and robustness.

3.7. Model Visualization Analysis

The interpretation and visualization of deep learning models are usually achieved through Grad-CAM [60]. The model outputs a heat map through Grad-CAM, which can highlight the model’s focus area on the image. The feature extraction ability of the model is analyzed by comparing the heat maps of the YOLOv11 and SPLDP algorithms.

As can be seen from Figure 11, the heat sources of the SPLDP and YOLOv11 models are concentrated on the target object, but the heat sources in Figure 11c,d are brighter than those in Figure 11a,b, indicating that the SPLDP we proposed is more sensitive to the characteristics of sweet potato leaves and pays more attention to the leaves of sweet potato. Although the heat sources in Figure 11a are also distributed on the sweet potato leaves, the heat sources are scattered, and some only cover part of the leaves, and the surrounding environment also slightly interferes with the model. The heat sources in Figure 11b are scarcer than those in Figure 11a, covering only a small part of the leaves, and have lower sensitivity to the characteristics of sweet potato leaves. The heat sources in Figure 11c basically cover all the sweet potato leaves, and obviously pay more attention to the leaf characteristics than the YOLOv11N model. The heat source in Figure 11d is evenly distributed on the sweet potato leaves. At the same time, the heat source is brighter than the heat source in Figure 11c, and the red part of the heat source accounts for a larger proportion, indicating that SPLDPvB has a stronger interest in sweet potato leaves and a greater ability to extract features from leaves.

As shown in Figure 11, the heat source of SPLDP basically covers all sweet potato leaf targets, indicating that SPLDP has a greater ability to recognize sweet potato leaf features and is more suitable for sweet potato disease and pest detection in natural environments.

4. Discussion

Pests and diseases have long plagued large-scale sweet potato planting, severely affecting the growth and yield of sweet potatoes. In order to achieve accurate identification of sweet potato pests and diseases, this study proposed the SPLDPvB and SPLDPvT models, which showed great potential in the detection of sweet potato pests and diseases in complex natural environments.

The residual structure, LF, we designed only contained a small number of depthwise separable convolutions, which made the LF module have extremely low computational complexity and avoided the loss of feature information as the network depth deepened. Based on this, we designed the FESRM structure, which combined LF, ELA and convolution. While ensuring the low module complexity, the feature extraction ability of the model was significantly improved. Therefore, the backbone network composed of FESRM was able to improve the detection performance of the model with lower parameters and computational complexity (as shown in Table 2). In order to further reduce the complexity of the overall model, we modified the traditional YOLO structure, removed the neck network, and retained only the backbone network and decoupling head. As shown in the ablation experiment results in Table 4, the SPLDP model containing only the backbone network and decoupling head not only improved detection accuracy but also greatly reduced the complexity of the model. For the SPLDPvT model, each FESRM structure is run only once in the backbone network, and its backbone network has fewer parameters and computational complexity than SPLDPvB. Therefore, SPLDPvT significantly reduces the complexity of the model while maintaining good accuracy. In addition, as shown in Figure 9, in the sweet potato pest and disease detection task, our proposed SPLDP model outperforms mainstream detection models, such as Faster R-CNN, SSD, YOLOv5, YOLOv8, YOLOv9, YOLOv10, and YOLOv11, showing its strong performance in complex environments.

In this study, we collected 983 images in the sweet potato field using image acquisition equipment and constructed a sweet potato dataset. Each image contained more than a dozen sweet potatoes, covering sweet potatoes at different growth stages, which posed a considerable challenge to identifying sweet potato leaf diseases and pests. In the study of sweet potato leaf diseases and pests, Goyal et al. [48] used a 7-layer convolutional network and SVM to establish a classification system for sweet potato leaf spot disease, with a detection accuracy of 96%. However, their dataset was collected in a laboratory environment with a simple background, a small number of sweet potatoes and similar growth conditions, so the model was difficult to adapt to complex field environments. Mehta et al. [49] combined federated learning with convolutional neural networks to detect the severity of sweet potato leaf diseases, with an average detection accuracy of 94.6%. However, the dataset only targeted a single leaf, and the detection effect dropped significantly when multiple leaves were blocked. Moreover, in actual sweet potato disease and pest detection, we could not only detect one leaf. The model we designed could detect all sweet potato leaf diseases and pests within a certain range, thus significantly improving the efficiency of sweet potato detection. Bayu et al. [50] used an Android-based mobile application for sweet potato disease and pest detection but did not describe its detection effect, making it difficult to evaluate its performance. Rahman and Marisa [51] used YOLOv4 for detection with an accuracy of 75%, which was 13.7% lower than our model. This showed that the SPLDP model we designed was more sensitive in detecting the characteristics of sweet potato leaf spots and leaf-feeding pests, with stronger feature extraction capabilities and higher detection accuracy. Overall, our study demonstrated that the proposed SPLDP model has a better detection effect and robustness in the detection of sweet potato leaf spots and leaf-feeding pests, providing strong technical support for sweet potato yield protection and precision agriculture.

In future research, we will expand the sweet potato pest and disease dataset as much as possible, not limited to leaf spot disease and pests, such as hawk moths and wheat moths that feed on leaves. At the same time, in order to adapt to more complex field environments, we will add data in different time periods, such as evening and night, to further improve the robustness and generalization of the model. In addition, we will continue to optimize the SPLDP model in terms of model structure, attention mechanism, model complexity and model pruning, continuously reduce the complexity of the model, and further improve the model detection accuracy.

5. Conclusions

In summary, our study demonstrates the detection effect of the proposed SPLDPvT and SPLDPvB, achieving accurate detection of sweet potato leaf spots and above-ground pests, such as hornworms and wheat moths. The mAP0.5 and mAP0.5:0.95 of the SPLDPvT model are 85.9% and 69.1, respectively, with only 0.7 M parameters and 5.4 G computational complexity. The mAP0.5 and mAP0.5:0.95 of the SPLDPvB model are 88.7% and 74.6%, respectively, with 1.1 M parameters and 7.7 G computational complexity. The experimental results confirm the feasibility and effectiveness of this method in detecting leaf spot disease and ground pests of sweet potato, effectively improving the detection accuracy and robustness of the model. This study is of great significance for the early detection and prevention of sweet potato pests and diseases. By accurately locating and detecting the location and type of the disease, it can significantly reduce the excessive use of pesticides and avoid environmental pollution caused by the abuse of pesticides.

Author Contributions

Methodology, K.X., Y.H. and W.S.; resources, K.X.; writing—original draft preparation, K.X., W.S., D.C. and J.X.; writing—review and editing, K.X., D.C., D.L. and R.Y.; supervision, J.X. and R.Y. All authors have read and agreed to the published version of the manuscript.

Institutional Review Board Statement

Not applicable.

Data Availability Statement

Dataset available on request from the authors.

Conflicts of Interest

The authors declare no conflicts of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1. Sweet potato leaf pests and diseases dataset. (a) Early sweet potato images collected in the field. (b) Leaves eaten by ground pests, hawk moths and wheat moths. (c) Sweet potato leaf spot. (d) Images annotated with LabelImg.

Figure 2. Structure of the YOLOv11 model.

Figure 3. Structure of the SPLDP model.

Figure 4. FESRM structural optimization process. (a) Common structure. (b) A general structure with 3 × 3 convolutional layers. (c) Structure of adding feature extraction module. (d) Structure with attention mechanism. (e) At the same time, feature extraction module and attention mechanism module are added. (f) The final structure of FESRM with LF and ELA added.

Figure 5. Structure of LF.

Figure 6. ELA attention mechanism.

Figure 7. Model training results. (a) Training results of the SPLDPvB model. (b) Training results of the SPLDPvT model.

Figure 7. Model training results. (a) Training results of the SPLDPvB model. (b) Training results of the SPLDPvT model.

Figure 8. Ablation experiment results. (a) Accuracy comparison of different structural combinations. (b) Comparison of model complexity of different combinations.

Figure 9. Comparison results with other models. (a) The accuracy of each model at mAP0.5. (b) The accuracy of each model at mAP0.5:0.95.

Figure 10. Detection effects of different models. (a) original image. (b) Faster RCNN. (c) SSD. (d) YOLOv5N. (e) YOLOv5S. (f) YOLOv8N. (g) YOLOv8S. (h) YOLOv9T. (i) YOLOv9S. (j) YOLOv10N. (k) YOLOv10S. (l) YOLOv11N. (m) YOLOv11S. (n) SPLDPvT. (o) SPLDPvB.

Figure 10. Detection effects of different models. (a) original image. (b) Faster RCNN. (c) SSD. (d) YOLOv5N. (e) YOLOv5S. (f) YOLOv8N. (g) YOLOv8S. (h) YOLOv9T. (i) YOLOv9S. (j) YOLOv10N. (k) YOLOv10S. (l) YOLOv11N. (m) YOLOv11S. (n) SPLDPvT. (o) SPLDPvB.

Figure 11. Heat map effects of different models. (a) YOLOv11N. (b) YOLOv11S. (c) SPLDPvT. (d) SPLDPvB.

Table 1

Hardware configuration and software environment.

Hardware Platform or Software Environment	Model Identity or Designation	Parametric or Version
CPU	Intel(R) Xeon(R) Platinum 8488C	Frequency: 3.80 GHz
GPU	NVIDIA GeForce RTX 4060	Memory: 24 GB
Computer system	Ubuntu 22	RAM: 512 GB
Deep learning framework	Pytorch	2.4.0
Deep learning framework	CUDA	11.7
Integrated development environment	Visual Studio Code	1.96.0
Programming language	Python	3.9.19

Table 2

Ablation experiment of FESRM structure.

Model	FESRM		mAP0.5 (%)	mAP0.5:0.95 (%)	Params (M)	FLOPs (G)	Size (M)
	LF	ELA
YOLOv11S(Figure 2)	×	×	86.4	71.8	9.4	21.3	19.2
SPLDPvB-FESRM-Base(Figure 4b)	×	×	86.1	70.1	1.0	6.0	2.3
SPLDPvB-FESRM-LF(Figure 4c)	√	×	88.1	73.3	1.1	7.7	2.5
SPLDPvB-FESRM-ELA(Figure 4d)	×	√	88.1	73.4	1.0	6.0	2.4
SPLDPvB (Figure 4f)	√	√	88.7	74.6	1.1	7.7	2.6

Table 3

Comparative experimental results of different attention mechanisms.

Model	P (%)	R (%)	mAP0.5 (%)	mAP0.5:0.95 (%)	Params (M)	FLOPs (G)	Size (M)
SPLDPvB-CA	81.6	80.9	88.3	74.0	1.4	7.8	3.3
SPLDPvB-CBAM	80.5	81.5	88.5	74.2	2.2	7.8	5
SPLDPvB-CCA	79.2	80.3	86.8	72.2	3.3	15.4	7.1
SPLDPvB-SE	80.6	80.8	88.0	73.5	1.8	7.8	4
SPLDPvB-SEAM	79.5	80.3	87.2	73.0	2.4	11.6	5.4
SPLDPvB-ECA	81.0	80.0	87.9	73.4	1.3	7.7	3.1
SPLDPvB-ELA	82.1	80.9	88.7	74.6	1.1	7.7	2.6

Table 4

Ablation experimental results of LF module, ELA module and different neck structures.

Groups	Model	FESRM		Neck			mAP0.5(%)	mAP0.5:0.95(%)	Params(M)	FLOPs(G)	Size(M)
		LF	ELA	v8	v10	v11
1	Model1	×	×	√	×	×	85.9	71.1	2.2	8.8	4.9
	Model2	√	×	√	×	×	87.6	73.4	2.3	10.5	5.1
	Model3	×	√	√	×	×	87.6	72.8	2.3	8.8	4.9
	Model4	√	√	√	×	×	88	73.9	2.4	10.5	5.2
2	Model1	×	×	×	√	×	86	71.7	1.9	8.6	4
	Model2	√	×	×	√	×	86.7	71.9	2.0	10.2	4.5
	Model3	×	√	×	√	×	87.4	72.7	1.9	8.6	4.3
	Model4	√	√	×	√	×	87.8	73.5	2.0	10.2	4.32
3	Model1	×	×	×	×	√	86.2	72.0	1.8	7.1	3.9
	Model2	√	×	×	×	√	86.7	72.9	1.8	8.8	4.1
	Model3	×	√	×	×	√	86.9	72.1	1.8	7.1	3.9
	Model4	√	√	×	×	√	88.1	73.8	1.9	8.8	3.99
4	Model1	×	×	×	×	×	86.1	70.1	1.0	6.0	2.3
	Model2	√	×	×	×	×	88.1	73.3	1.1	7.7	2.5
	Model3	×	√	×	×	×	88.1	73.4	1.0	6.0	2.4
	SPLDPvB	√	√	×	×	×	88.7	74.6	1.1	7.7	2.6

Table 5

Comparison experiment of the SPLDP with different detection algorithms.

Model	P (%)	R (%)	mAP0.5 (%)	mAP0.5:0.95 (%)	Params (M)	FLOPs (G)	Size (M)
Faster R-CNN	76.5	76.9	76.5	41.7	137.0	370	108
SSD	85.2	75.4	71.7	42.4	26.2	62.7	45.76
YOLOv5N	78.4	77.8	84.8	68.3	2.5	7.1	5.3
YOLOv5S	79.7	79.4	86.5	72.5	9.1	23.8	18.5
YOLOv5M	80.1	78.2	86	72.4	25.0	64.0	50.5
YOLOv5L	79.5	79.4	86.1	72.9	53.1	134.7	72.9
YOLOv5X	79.9	79.3	86.3	73.7	97.1	246.0	185.9
YOLOv8N	77.8	78.3	84.7	68.3	3.0	8.1	6.3
YOLOv8S	79.6	78.6	86.1	71.9	11.1	28.4	22.5
YOLOv8M	79.3	79.3	85.8	72.3	25.8	78.7	52
YOLOv8L	78.9	79.8	85.8	72.4	43.6	164.8	87.7
YOLOv8X	79.8	78.1	86	73.1	68.1	257.4	136.7
YOLOv9T	77.7	78.0	84.9	69.6	1.9	7.6	4.5
YOLOv9S	81.2	79.0	87.0	72.8	7.1	26.7	15.2
YOLOv9M	80.4	80.1	86.7	73.3	20.0	76.5	40.8
YOLOv9C	80.1	80.2	86.6	73.4	25.3	102.3	51.6
YOLOv10N	74.5	76.3	82.1	66.2	2.6	8.2	5.6
YOLOv10S	79.1	79.2	86.6	71.9	8.0	24.4	16.5
YOLOv10M	80.2	79.4	86.7	72.5	16.4	63.4	33.5
YOLOv10L	80.1	80.2	87.6	73.8	25.7	126.3	52.2
YOLOv10X	80.3	80.0	87.1	74.2	31.5	169.8	64.1
YOLOv11N	77.8	77.5	84.2	67.9	2.5	6.3	5.5
YOLOv11S	79.7	78.8	86.4	71.8	9.4	21.3	19.2
YOLOv11M	80.6	79.8	87.8	73.6	20.0	67.7	40.5
YOLOv11L	80.2	80.7	86.9	73.5	25.2	86.6	51.2
YOLOv11X	81.6	80.1	87.6	74.3	56.8	194.4	114.4
SPLDPvT	78.9	78.7	85.9	69.1	0.7	5.4	1.6
SPLDPvB	82.1	80.9	88.7	74.6	1.1	7.7	2.6

References

1. Zeng, F.; Ding, Z.; Song, Q.; Xiao, J.; Zheng, J.; Li, H.; Luo, Z.; Wang, Z.; Yue, X.; Huang, L. Feasibility of Detecting Sweet Potato (Ipomoea batatas) Virus Disease from High-Resolution Imagery in the Field Using a Deep Learning Framework. Agronomy; 2023; 13, 2801. [DOI: https://dx.doi.org/10.3390/agronomy13112801]

2. Chai, A.-L.; Du, G.-F.; Shi, Y.-X.; Xie, X.-W.; Li, B.-J. Leaf Spot on Sweet Potato (Ipomoea batatas) Caused by Stemphylium solani, a New Disease in China. J. Phytopathol.; 2015; 163, pp. 1046-1049. [DOI: https://dx.doi.org/10.1111/jph.12344]

3. Wang, T.; Niu, D.; Li, J.; Lin, G. Main sweet potato pests and diseases and integrated control measures in northern potato regions. Bull. Agric. Sci. Technol.; 2022; 5, pp. 215-217. [DOI: https://dx.doi.org/10.3969/j.issn.1000-6400.2022.05.063]

4. Chai, A.L.; Li, B.J.; Shi, Y.X.; Cen, Z.X.; Huang, H.Y.; Liu, J. Recognition of tomato foliage disease based on computer vision technology. Acta Hortic. Sin.; 2010; 37, pp. 1423-1430. [DOI: https://dx.doi.org/10.16420/j.issn.0513-353x.2010.09.030]

5. Guan, Z.; Tang, J.; Yang, B.; Zhou, Y.; Fan, D.; Yao, Q. Study on recognition method of rice disease based on image. Chin. J. Rice Sci.; 2010; 24, 497. [DOI: https://dx.doi.org/10.3969/j.issn.1001-7216.2010.05.009]

6. Li, Z.R.; He, D.J. Research on identify technologies of apple’s disease based on mobile photograph image analysis. Comput. Eng. Des.; 2010; 13, pp. 3051-3053. [DOI: https://dx.doi.org/10.16208/j.issn1000-7024.2010.13.027]

7. Dubey, S.R.; Jalal, A.S. Adapted Approach for Fruit Disease Identification using Images. Int. J. Comput. Vis. Image Process.; 2012; 2, pp. 44-58. [DOI: https://dx.doi.org/10.4018/ijcvip.2012070104]

8. Li, L.; Zhang, S.; Wang, B. Plant Disease Detection and Classification by Deep Learning—A Review. IEEE Access; 2021; 9, pp. 56683-56698. [DOI: https://dx.doi.org/10.1109/ACCESS.2021.3069646]

9. Kawasaki, Y.; Uga, H.; Kagiwada, S.; Iyatomi, H. Basic Study of Automated Diagnosis of Viral Plant Diseases Using Convolutional Neural Networks. Proceedings of the International Symposium on Visual Computing; Las Vegas, NV, USA, 14–16 December 2015; pp. 638-645. [DOI: https://dx.doi.org/10.1007/978-3-319-27863-6_59]

10. Mohanty, S.P.; Hughes, D.P.; Salathé, M. Using Deep Learning for Image-Based Plant Disease Detection. Front. Plant Sci.; 2016; 7, 1419. [DOI: https://dx.doi.org/10.3389/fpls.2016.01419]

11. Ma, J.; Du, K.; Zheng, F.; Zhang, L.; Gong, Z.; Sun, Z. A recognition method for cucumber diseases using leaf symptom images based on deep convolutional neural network. Comp. Electron. Agric.; 2018; 154, pp. 18-24. [DOI: https://dx.doi.org/10.1016/j.compag.2018.08.048]

12. Wang, D.; Deng, L.; Ni, J.; Gao, J.; Zhu, H.; Han, Z. Recognition pest by image-based transfer learning. J. Sci. Food Agric.; 2019; 99, pp. 4524-4531. [DOI: https://dx.doi.org/10.1002/jsfa.9689]

13. Mallick, T.; Biswas, S.; Das, A.K.; Saha, H.N.; Chakrabarti, A.; Deb, N. Deep learning based automated disease detection and pest classification in Indian mung bean. Multimed. Tools Appl.; 2023; 82, pp. 12017-12041. [DOI: https://dx.doi.org/10.1007/s11042-022-13673-7]

14. Hou, J.; Yang, C.; He, Y.; Hou, B. Detecting diseases in apple tree leaves using FPN–ISResNet–Faster RCNN. Eur. J. Remote. Sens.; 2023; 56, 2186955. [DOI: https://dx.doi.org/10.1080/22797254.2023.2186955]

15. Kukreja, V.; Kumar, D.; Bansal, A.; Solanki, V. Recognizing Wheat Aphid Disease Using a Novel Parallel Real-Time Technique Based on Mask Scoring RCNN. Proceedings of the 2022 2nd International Conference on Advance Computing and Innovative Technologies in Engineering (ICACITE); Greater Noida, India, 28–29 April 2022; pp. 1372-1377. [DOI: https://dx.doi.org/10.1109/ICACITE53722.2022.9823459]

16. Wang, L.; Shi, W.; Tang, Y.; Liu, Z.; He, X.; Xiao, H.; Yang, Y. Transfer Learning-Based Lightweight SSD Model for Detection of Pests in Citrus. Agronomy; 2023; 13, 1710. [DOI: https://dx.doi.org/10.3390/agronomy13071710]

17. Tian, Y.; Wang, S.; Li, E.; Yang, G.; Liang, Z.; Tan, M. MD-YOLO: Multi-scale Dense YOLO for small target pest detection. Comput. Electron. Agric.; 2023; 213, 8233. [DOI: https://dx.doi.org/10.1016/j.compag.2023.108233]

18. Li, D.; Ahmed, F.; Wu, N.; Sethi, A.I. YOLO-JD: A Deep Learning Network for Jute Diseases and Pests Detection from Images. Plants; 2022; 11, 937. [DOI: https://dx.doi.org/10.3390/plants11070937]

19. Zhu, X.; Jia, B.; Huang, B.; Li, H.; Liu, X.; Seah, W.K. Pest-YOLO: A Lightweight Pest Detection Model Based on Multi-level Feature Fusion. Proceedings of the International Conference on Intelligent Computing; Tianjin, China, 5–8 August 2024; pp. 137-148. [DOI: https://dx.doi.org/10.1007/978-981-97-5591-2_12]

20. Lin, J.; Bai, D.; Xu, R.; Lin, H. TSBA-YOLO: An Improved Tea Diseases Detection Model Based on Attention Mechanisms and Feature Fusion. Forests; 2023; 14, 619. [DOI: https://dx.doi.org/10.3390/f14030619]

21. Sangaiah, A.K.; Yu, F.-N.; Lin, Y.-B.; Shen, W.-C.; Sharma, A. UAV T-YOLO-Rice: An Enhanced Tiny Yolo Networks for Rice Leaves Diseases Detection in Paddy Agronomy. IEEE Trans. Netw. Sci. Eng.; 2024; 11, pp. 5201-5216. [DOI: https://dx.doi.org/10.1109/TNSE.2024.3350640]

22. Zhang, D.-Y.; Zhang, W.; Cheng, T.; Zhou, X.-G.; Yan, Z.; Wu, Y.; Zhang, G.; Yang, X. Detection of wheat scab fungus spores utilizing the Yolov5-ECA-ASFF network structure. Comput. Electron. Agric.; 2023; 210, 107953. [DOI: https://dx.doi.org/10.1016/j.compag.2023.107953]

23. Jia, L.; Wang, T.; Chen, Y.; Zang, Y.; Li, X.; Shi, H.; Gao, L. MobileNet-CA-YOLO: An Improved YOLOv7 Based on the MobileNetV3 and Attention Mechanism for Rice Pests and Diseases Detection. Agriculture; 2023; 13, 1285. [DOI: https://dx.doi.org/10.3390/agriculture13071285]

24. Liu, Y.; Yu, Q.; Geng, S. Real-time and lightweight detection of grape diseases based on Fusion Transformer YOLO. Front. Plant Sci.; 2024; 15, 1269423. [DOI: https://dx.doi.org/10.3389/fpls.2024.1269423] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/38463562]

25. Zhang, D.-Y.; Luo, H.-S.; Cheng, T.; Li, W.-F.; Zhou, X.-G.; Guo, W.; Gu, C.-Y.; Diao, Z. Enhancing wheat Fusarium head blight detection using rotation Yolo wheat detection network and simple spatial attention network. Comput. Electron. Agric.; 2023; 211, 107968. [DOI: https://dx.doi.org/10.1016/j.compag.2023.107968]

26. Wang, S.; Li, Q.; Yang, T.; Li, Z.; Bai, D.; Tang, C.; Pu, H. LSD-YOLO: Enhanced YOLOv8n Algorithm for Efficient Detection of Lemon Surface Diseases. Plants; 2024; 13, 2069. [DOI: https://dx.doi.org/10.3390/plants13152069] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/39124187]

27. Li, H.; Li, J.; Wei, H.; Liu, Z.; Zhan, Z.; Ren, Q. Slim-neck by GSConv: A better design paradigm of detector architectures for autonomous vehicles. arXiv; 2022; [DOI: https://dx.doi.org/10.48550/arXiv.2206.02424] arXiv: 2206.02424

28. Yang, J.; Zhu, W.; Liu, G.; Dai, W.; Xu, Z.; Wan, L.; Zhou, G. ICPNet: Advanced Maize Leaf Disease Detection with Multidimensional Attention and Coordinate Depthwise Convolution. Plants; 2024; 13, 2277. [DOI: https://dx.doi.org/10.3390/plants13162277]

29. Yang, Z.; Jin, A.; Wang, J.; Xie, H.; Zhan, X.; Li, Q. Wheat Disease Recognition Based on a Fusion Model with Improved Spatial and Channel Reconstruction Convolution Combined with Contextual Transformer. Proceedings of the 2024 5th International Conference on Computer Vision, Image and Deep Learning (CVIDL); Zhuhai, China, 19–21 April 2024; pp. 339-343. [DOI: https://dx.doi.org/10.1109/CVIDL62147.2024.10604082]

30. Luo, D.; Xue, Y.; Deng, X.; Yang, B.; Chen, H.; Mo, Z. Citrus Diseases and Pests Detection Model Based on Self-Attention YOLOV. IEEE Access; 2023; 11, pp. 139872-139881. [DOI: https://dx.doi.org/10.1109/ACCESS.2023.3340148]

31. Wang, Y.; Wang, Y.; Zhao, J. MGA-YOLO: A lightweight one-stage network for apple leaf disease detection. Front. Plant Sci.; 2022; 13, 927424. [DOI: https://dx.doi.org/10.3389/fpls.2022.927424]

32. Pan, P.; Guo, W.; Zheng, X.; Hu, L.; Zhou, G.; Zhang, J. Xoo-YOLO: A detection method for wild rice bacterial blight in the field from the perspective of unmanned aerial vehicles. Front. Plant Sci.; 2023; 14, 1256545. [DOI: https://dx.doi.org/10.3389/fpls.2023.1256545]

33. Xu, W.; Wang, R. ALAD-YOLO: An lightweight and accurate detector for apple leaf diseases. Front. Plant Sci.; 2023; 14, 1204569. [DOI: https://dx.doi.org/10.3389/fpls.2023.1204569]

34. Li, R.; Li, Y.; Qin, W.; Abbas, A.; Li, S.; Ji, R.; Wu, Y.; He, Y.; Yang, J. Lightweight Network for Corn Leaf Disease Identification Based on Improved YOLO v8s. Agriculture; 2024; 14, 220. [DOI: https://dx.doi.org/10.3390/agriculture14020220]

35. Li, Y.; Li, A.; Li, X.; Liang, D. Detection and Identification of Peach Leaf Diseases based on YOLO v5 Improved Model. Proceedings of the ICCCV 2022: 2022 The 5th International Conference on Control and Computer Vision; Xiamen, China, 19–21 August 2022; pp. 79-84. [DOI: https://dx.doi.org/10.1145/3561613.3561626]

36. Li, K.; Zhu, J.; Li, N. Lightweight Automatic Identification and Location Detection Model of Farmland Pests. Wirel. Commun. Mob. Comput.; 2021; 2021, 993. [DOI: https://dx.doi.org/10.1155/2021/9937038]

37. Chen, R.; Qi, H.; Liang, Y.; Yang, M. Identification of plant leaf diseases by deep learning based on channel attention and channel pruning. Front. Plant Sci.; 2022; 13, 1023515. [DOI: https://dx.doi.org/10.3389/fpls.2022.1023515] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36438120]

38. Liu, Y.; Liu, J.; Cheng, W.; Chen, Z.; Zhou, J.; Cheng, H.; Lv, C. A High-Precision Plant Disease Detection Method Based on a Dynamic Pruning Gate Friendly to Low-Computing Platforms. Plants; 2023; 12, 2073. [DOI: https://dx.doi.org/10.3390/plants12112073] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37299053]

39. Xu, Y.; Wang, L.-X.; Chen, C.; Ma, S.-S.; Zhou, R.; Xiong, A.-S. Virus-Free Sweet Potato Industry: Development Status and Production Suggestions. Horticulturae; 2024; 10, 979. [DOI: https://dx.doi.org/10.3390/horticulturae10090979]

40. Ilondu, E. Etiology and assessment of leafspot disease of sweet potato (Ipomoea batatas (L.) Lam) in selected farms in Delta State, Nigeria. Agric. Biol. J. N. Am.; 2013; 4, pp. 476-484. [DOI: https://dx.doi.org/10.5251/abjna.2013.4.4.476.484]

41. Halder, B.; Sultana, S.; Akter, T.; Begum, S. Life cycle, feeding behavior and nature of damage of sweet potato leaf moth, Agrius cingulata (Fabricius) and Agrius convolvuli (Linnaeus) (Lepidoptera: Sphingidae). Dhaka Univ. J. Biol. Sci.; 2018; 27, pp. 125-134. [DOI: https://dx.doi.org/10.3329/dujbs.v27i2.46461]

42. Rashid, R.; Aslam, W.; Aziz, R.; Aldehim, G. An Early and Smart Detection of Corn Plant Leaf Diseases Using IoT and Deep Learning Multi-Models. IEEE Access; 2024; 12, pp. 23149-23162. [DOI: https://dx.doi.org/10.1109/ACCESS.2024.3357099]

43. Daniya, T.; Vigneshwari, S. Rider Water Wave-enabled deep learning for disease detection in rice plant. Adv. Eng. Softw.; 2023; 182, 103472. [DOI: https://dx.doi.org/10.1016/j.advengsoft.2023.103472]

44. Xu, L.; Cao, B.; Zhao, F.; Ning, S.; Xu, P.; Zhang, W.; Hou, X. Wheat leaf disease identification based on deep learning algorithms. Physiol. Mol. Plant Pathol.; 2023; 123, 101940. [DOI: https://dx.doi.org/10.1016/j.pmpp.2022.101940]

45. Gupta, S.; Tripathi, A.K. Fruit and vegetable disease detection and classification: Recent trends, challenges, and future opportunities. Eng. Appl. Artif. Intell.; 2024; 133, 108260. [DOI: https://dx.doi.org/10.1016/j.engappai.2024.108260]

46. Onyejegbu, L.N. AI and IoT-Based Model for the Detection and Treatment of Sweet Potato Pests and Diseases in Precision Farming. In Deep Learning Indaba 2023, Accra, Ghana. 3 September 2023; Available online: https://openreview.net/forum?id=5xqyt2vs3S (accessed on 31 July 2023).

47. Banerjee, D.; Rajora, R.; Garg, M. A Hybrid CNN–SVM Model for Early Detection and Classification of Sweet Potato Leaf Diseases. Latest Trends in Engineering and Technology; Singh, S.; Kaur, S. CRC Press: London, UK, 2024; pp. 81-88. [DOI: https://dx.doi.org/10.1201/9781032665443]

48. Goyal, D.; Banerjee, D.; Chauhan, R.; Devliyal, S.; Gill, K.S. Advanced Techniques for Sweet Potato Leaf Disease Detection: A CNN-SVM Hybrid Approach. Proceedings of the 2024 3rd International Conference for Innovation in Technology (INOCON); Bangalore, India, 1–3 March 2024; pp. 1-6. [DOI: https://dx.doi.org/10.1109/INOCON60754.2024.10512114]

49. Mehta, S.; Kukreja, V.; Sharma, V.; Aeri, M. Cutting-edge Analysis of Sweet Potato Leaf Diseases: Leveraging Federated Learning and CNNs for Severity Evaluation. Proceedings of the 2023 14th International Conference on Computing Communication and Networking Technologies (ICCCNT); Delhi, India, 6–8 July 2023; pp. 1-6.

50. I Bayu, M.S.Y.; Prayogo, Y.; Indiati, S.W.; Sumartini,; Indriyani, F.C.; A Susanto, G.W.; Yuniar, E. Trizelia Early detection of sweet potato weevil (Cylas formicarius) and scab diseases (Sphaceloma batatas) outbreak based on android application. Proceedings of the IOP Conference. Series: Earth and Environmental Science; Malang, Indonesia, 17–18 October 2023; 12044. [DOI: https://dx.doi.org/10.1088/1755-1315/1312/1/012044]

51. Nisti, M.; Rahman, A.Y.; Marisa, F. Detection of Diseases and Pests on the Leaves of Sweet Potato Plants sing Yolov4. Buana Inf. Technol. Comput. Sci.; 2024; 5, pp. 39-44. [DOI: https://dx.doi.org/10.36805/bit-cs.v5i1.6065]

52. Khanam, R.; Hussain, M. Yolov11: An overview of the key architectural enhancements. arXiv; 2024; [DOI: https://dx.doi.org/10.48550/arXiv.2410.17725] arXiv: 2410.17725

53. Sun, W.; Xu, M.; Xu, K.; Chen, D.; Wang, J.; Yang, R.; Chen, Q.; Yang, S. CSGD-YOLO: A Corn Seed Germination Status Detection Model Based on YOLOv8n. Agronomy; 2025; 15, 128. [DOI: https://dx.doi.org/10.3390/agronomy15010128]

54. Xu, W.; Wan, Y. ELA: Efficient Local Attention for Deep Convolutional Neural Networks. arXiv; 2024; [DOI: https://dx.doi.org/10.48550/arXiv.2403.01123] arXiv: 2403.01123

55. Wang, Z.; Li, C.; Xu, H.; Zhu, X. Mamba YOLO: SSMs-Based YOLO For Object Detection. arXiv; 2024; [DOI: https://dx.doi.org/10.48550/arXiv.2406.05835] arXiv: 2406.05835

56. Wang, B.; Teng, Y.; Lau, V.; Han, Z. CCA-Net: A Lightweight Network Using Criss-Cross Attention for CSI Feedback. IEEE Commun. Lett.; 2023; 27, pp. 1879-1883. [DOI: https://dx.doi.org/10.1109/LCOMM.2023.3267461]

57. Hu, J.; Shen, L.; Sun, G. Squeeze-and-excitation networks. arXiv; 2018; [DOI: https://dx.doi.org/10.48550/arXiv.1709.01507] arXiv: 1709.01507

58. Wang, Y.; Zhang, J.; Kan, M.; Shan, S.; Chen, X. Self-Supervised Equivariant Attention Mechanism for Weakly Supervised Semantic Segmentation. arXiv; 2020; [DOI: https://dx.doi.org/10.48550/arXiv.2004.04581] arXiv: 2004.04581

59. Kandel, I.; Castelli, M. The effect of batch size on the generalizability of the convolutional neural networks on a histopathology dataset. ICT Express; 2020; 6, pp. 312-315. [DOI: https://dx.doi.org/10.1016/j.icte.2020.04.010]

60. Selvaraju, R.R.; Cogswell, M.; Das, A.; Vedantam, R.; Parikh, D.; Batra, D. Grad-CAM: Visual explanations from deep networks via gradient-based localization. Int. J. Comput. Vis.; 2020; 128, pp. 336-359. [DOI: https://dx.doi.org/10.1007/s11263-019-01228-7]