Intelligent Weed Recognition in Winter Wheat Fields through Deep Convolutional Neural Networks

Intelligent Weed Recognition in Winter Wheat Fields through Deep Convolutional Neural Networks | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Intelligent Weed Recognition in Winter Wheat Fields through Deep Convolutional Neural Networks Seyed Iman Saedi, Hassan Makarian This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6710451/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Improving wheat productivity is essential for maintaining food security in the coming years. A major challenge to wheat production is the presence of weeds, which can greatly reduce its yield. Therefore, precision weed management, utilising site-specific methods, will be vital for achieving sustainable wheat production. To pave the way for this goal, in the present study, we utilised colour images and deep learning techniques to distinguish winter wheat from four prevalent weeds, creating five distinct classes. A total of 542 images were captured in natural field conditions, and resized to 299 × 299 pixels. Four deep learning networks – Xception, EfficientNetB0, VGG19, and InceptionResNetV2 –were evaluated to serve as base networks to fulfil the classification requirements. These networks were pre-trained on ImageNet using transfer learning, then fine-tuned and enhanced with additional layers to improve performance on our dataset. The improved InceptionResNetV2 model demonstrated the highest performance among the four models, achieving an accuracy of 98.17% and a loss of 3.19 on the test data. Nevertheless, all models exhibited excellent performance in distinguishing plant classes, with F1-scores ranging from 93–100%, 69–98%, 82–100%, and 93–100% for models based on Xception, EfficientNetB0, VGG19, and InceptionResNetV2, respectively. Additionally, we analysed fifteen scenarios of weed presence in winter wheat fields, focusing on various weed types studied, to propose effective weed management strategies utilising any of the four models. The research findings provide a foundation for precision weed management that not only reduces herbicide usage and environmental impact but also enhances wheat yield and quality. Physical sciences/Engineering Biological sciences/Plant sciences Deep learning RGB image Weed recognition Winter wheat Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 1. Introduction As the global population continues to grow, wheat emerges as a vital crop in addressing the world's increasing nutritional demands (Sadeghi et al., 2022 ; S I Saedi et al., 2019 ; Shewry & Hey, 2015 ). Wheat plays a key role in global food security, supplying roughly 20% of the essential carbohydrates and proteins needed for human consumption (Langridge et al., 2022 ). It occupies about 30% of the world's cereal cropland, underscoring its critical importance in agricultural systems (FAO, 2021). In light of the increasing challenges posed by climate change and a rapidly growing global population, enhancing wheat productivity has emerged as a crucial strategy to uphold food security in the years ahead. One significant obstacle to wheat production is weeds, which reduce the growth and yield of wheat by competing for water, nutrients, and light (Singh et al., 2021 ). Weeds can significantly impact wheat productivity, with estimated yield losses of up to 30%, largely influenced by varying climatic conditions and differences in agricultural management practices (Montazeri, M., Zand, E., Baghestani, 2005 ). In many of Iran's key wheat-producing areas, broadleaved weeds have emerged as a prominent concern, accounting for a staggering 84% of the weed species infesting wheat fields and posing a significant threat to crop yields. Research on winter wheat has consistently demonstrated that the presence of weeds can have a profound impact on crop yields, resulting in losses of up to 20–30% in severe cases. Both broadleaf and narrow leaf weeds can infest wheat crops, collectively contributing to a global grain yield reduction of 13.1%. The extent of this yield loss, however, is highly variable and depends on factors such as the type and density of weeds, as well as local climate conditions and precipitation patterns (Gyawali et al., 2022 ). Manual removal of weeds in wheat fields is not a viable option due to the risk of crop damage and the high labour costs involved. Precise identification and differentiation of weeds from crop plants is a crucial component of effective site-specific weed management. By accurately distinguishing weeds from wheat, farmers can target herbicide applications to specific areas where weeds are present, thereby optimising weed control. By harnessing the power of advanced technologies such as computer vision and image processing, researchers can achieve high precision in weed detection, enabling the development of targeted management strategies that accurately distinguish between crops and weeds, ultimately optimising the effectiveness of site-specific weed control techniques (Juwono et al., 2023 ). Through providing two-dimensional information, colour images have been widely utilised in this domain, laying the groundwork for further advancements. Studies have shown that unique morphological characteristics, including leaf form, colour, and dimension, can serve as valuable features for object recognition and classification tasks (Susetyarini et al., 2020 ). Specifically, features like leaf area, length, perimeter, width, and colour can be employed to differentiate between plants and inform plant management strategies (Speck & Speck, 2021 ). Despite these challenges, significant progress has been made in developing object recognition algorithms for precision agriculture, with varying degrees of accuracy achieved in different applications. The application of deep learning techniques has demonstrated considerable potential in achieving these objectives. Over the past decade, deep learning has transformed the field of image processing, achieving unprecedented success and finding numerous practical applications in precision agriculture. While large datasets are essential for training and validating deep learning algorithms, advances in computational power and data availability have made them increasingly popular. As a subset of machine learning, deep learning leverages hierarchical representations of data through sequential layers, typically employing convolutional neural networks (CNNs) to extract complex, non-linear image features. This approach has been shown to exhibit exceptional generalisation capabilities, making it well-suited for classification tasks ( Seyed I Saedi & Rezaei, 2024 ). This task begins with the input of an image, which can vary in size, and undergoes preprocessing before entering the feature extraction stage. This phase involves a series of layers that progressively extract features from the image, with the complexity of the features increasing with each additional layer. Following feature extraction, the classification process begins, employing traditional fully connected layers commonly found in neural networks. The arrangement and configuration of these layers give rise to a diverse range of networks with distinct characteristics and capabilities (Chollet, 2017 ). Convolutional neural networks have been extensively explored for identification and classification tasks in agriculture. Research has demonstrated the effectiveness of these networks in various applications. For instance, Seyed Iman Saedi & Khosravi, ( 2020 ) developed an image-based CNN model to recognise six different types of on-branch fruits in orchards, achieving an impressive accuracy of 99.8%. Additionally, Khosravi et al., ( 2021 ) developed a CNN model to classify olive fruit at different cultivars and developmental phases, attaining 91% classification. Furthermore, researchers have used deep neural networks such as Xception (Salim et al., 2023 ) and lightweight CNNs (Seyed Iman Saedi et al., 2024 ), to recognise and classify fruits with high accuracy, demonstrating the potential of these models in image classification tasks. Deep learning algorithms have been successfully applied to weed detection and classification, demonstrating superior performance in distinguishing between crop plants and weeds (Li et al., 2024 ). For instance, Quan et al., ( 2019 ) employed the Faster R-CNN model to detect corn seedlings and weeds in complex environments, while Fan et al., ( 2023 ) improved this model to achieve an accuracy of over 98.43% in identifying weeds in cotton fields. Other studies have utilised deep learning-based models to identify common weeds in sugar beet cultivation (Nasiri et al., 2022 ), discriminate pepper from weeds (Subeesh et al., 2022 ), and detect weeds in sugar beet fields (Gao et al., 2020 ). Additionally, researchers have evaluated the performance of deep learning models such as Yolov8, Yolov9, and customised Yolov9 in real-field conditions for eight crop species (GC et al., 2024 ). Similarly, a modified Xception deep learning approach was successfully utilised to distinguish saffron from its broadleaf weeds, yielding promising results with F1 scores ranging from 91–96% (Makarian & Saedi, 2024 ). In a practical study, Rai & Sun, ( 2024 ) designed a single-stage deep learning architecture that accomplishes two tasks simultaneously: detecting weed presence through bounding boxes and achieving pixel-wise instance segmentation on unmanned aerial system (UAS) acquired remote sensing images. The model achieved the best detection and segmentation scores of 85.4% and 82.1%, respectively. The aforementioned works demonstrate the potential of deep learning algorithms in plant identification and classification, and highlight the need for further research in this area. This research aimed to address the lack of a thorough investigation into site-specific weed management for winter wheat. This work builds on the authors’ previous studies regarding weed-crop discrimination, aiming to develop a robust method for precision weed management in winter wheat fields. Here, advanced deep learning models were introduced to differentiate between winter wheat plants and four common weed species: annual bugloss, bastard cabbage, hoary cress, and flixweed. Utilising colour images, these models can accurately classify each plant category across various scenarios, paving the way for the creation of precise weed management systems for automated control in wheat fields. 2 Materials and Methods 2.1 Image preparation To address the requirements of the study, 542 images were collected from the five classes understudy: Winter Wheat ( Triticum aestivum L.), Annual Bugloss ( Anchusa arvensis L.), Bastard Cabbage ( Rapistrum rugosum L.), Hoary Cress ( Cardaria draba (L.) Desv.), and Flixweed ( Descurainia sophia L.). Figure 1 illustrates a selection of these images. The photos were taken with the 48-megapixel camera of an A-12 Samsung Galaxy (SM-A125F/DS). In the pie chart shown in Fig. 2 , the quantity and proportion of samples for each class are shown. The illustration highlights the balanced distribution of data used in this study. Having a balanced dataset is crucial to prevent bias, improve generalisation, enhance performance, promote stability, and avoid misinterpretation. This balance ensures that the model receives sufficient examples from each class, enabling effective learning and accurate predictions across all classes. The images captured underwent a review process in which inappropriate samples were excluded. An agronomist with expertise in weed and crop properties contributed to this selection process. Ultimately, 542 images were chosen to develop the model and train the network. The selected images were divided into three datasets: training, validation, and testing. 80% of the data was allocated for training, while the remaining 20% was split for testing. Within the training data, 15% was set aside for validation purposes. It is worth emphasising that the test set is reserved exclusively for evaluating the final model's performance and testing its accuracy on unseen data. This separation ensures a reliable assessment of the model's generalisation capabilities. 2.2 Image pre-processing Preprocessing input images is essential for improving the model's accuracy, mitigating overfitting, and enhancing its generalisation ability. 2.2.1 Resize and normalisation Initially, we standardised the image sizes to 299 × 299 pixels to ensure uniform resolution across all images. Subsequently, we normalised the resolutions to fall within the range [0,1] by utilising the Eq. 1 . $$\:{x}_{i}^{{\prime\:}}=\:\frac{{x}_{i}-\:{x}_{min}}{{x}_{max}-\:{x}_{min}}$$ 1 where \(\:{x}_{i}^{{\prime\:}}\) symbolises the normalised value within the dataset, \(\:{x}_{i}\) ​ represents the i-th value in the dataset, and \(\:\:{x}_{min}\) and \(\:{x}_{max}\:\) denote the minimum and maximum values in the dataset. This formula guarantees that after normalisation, the normalised value for the dataset's minimum value will consistently be 0, and the normalised value for the dataset's maximum value will invariably be 1. 2.2.2 Augmentation Data augmentation is a widely recognized method in computer vision. It seeks to create extra training data from existing samples by applying random transformations. This technique ensures that the generated images are realistic and that each specific image is unique (Chollet, 2017 ). Consequently, image augmentation can alleviate the challenges encountered by deep neural networks regarding their reliance on a substantial quantity of annotated data, leading to improved accuracy. The extensive dataset enhances the training procedure through allowing the network to acquire all essential parameters while reducing the likelihood of overfitting. In the present study, an automated data augmentation technique was utilized, incorporating random height and width shifts, random flipping, random contrast adjustments, and random rotations. Table 1 represents The specific data augmentation arguments for our task. Table 1 Augmentation parameters and values. Image augmentation parameters Values Random height translation 0.1 Random width translation 0.1 Random flip True Random contrast 0.15 Random rotation 0.15 2.3 Candidate models 2.3.1 Base networks The initial step in reaching our goal is to select a base network from several choices through transfer learning, which entails loading weights from the ImageNet dataset. To this end, we tested four innovative standard deep learning networks, including Xception, EfficientNetB0, VGG19, and InceptionResNetV2. 2.3.2 Fine-tuning The objective of this fine-tuning process was to customise the candidate base networks for the particular dataset and task at hand, leveraging the previously learned features while adapting to the new data. This technique prevents the overfitting in pre-trained layers while facilitating the adaptation of the newly added layers to the new data. During the fine-tuning process, our goal was to selectively freeze certain layers (the earlier ones) while enabling the training of others (the later ones) when incorporating a pre-trained model into transfer learning within a deep learning framework. Initially, we disabled the trainable attribute for the first 20 layers, locking them during training to prevent their weights from being altered and preserving the valuable features they had already acquired from the original training data. Subsequently, we advanced to fine-tune the later layers of the base model. We then iterated through the layers starting from the 20th position and activated the trainable attribute for each of these layers. This adjustment allows these layers to undergo training and updates throughout the training phase. 2.3.3 Modifying block To develop a robust CNN model, we added a modifying block to the last layer of the proposed base networks. We tested various layer configurations for the modifying block, exploring different types, placements, and parameters to find the most efficient setup. This included examining combinations of 2D Convolution, Batch Normalization, Dropout, and Global Average Pooling layers. Parameters were selected through experimentation testing, though the specifics are not detailed here for the sake of brevity. Consequently, the optimized modifying block involved three sets of layers, including Convolution, Batch Normalization, Max Pooling, and Dropout, followed by the Fully Connected and Global Average Pooling layers. The final layer of the base candidate networks was supplemented with these additional layers to develop the proposed architectures in this study (Fig. 3 ). By following the steps outlined earlier, we developed four distinct models, labelled Model 1 through Model 4 (Table 2 ). Subsequently, we evaluated these models using images resized to 299×299 pixels, the SGD optimiser, a batch size of 32, and 200 epochs to identify the optimal model. During this process, we employed selection criteria that included assessing the model's stability during training, as well as analysing test accuracy and test loss. Table 2 Four proposed models for the purpose of this study. Model 1 Model 2 Model 3 Model 4 Modified Xception Modified EfficientNetB0 Modified VGG19 Modified InceptionResNetV2 The detailed outcomes of this step, including a comparison of the models' performance, are presented in Section 3.2. Figure 4 summarises the proposed model for the purpose of our study. Following the evaluation of the candidate standard networks, the Inception-ResNetV2 stood out as the top performer based on the specified selection criteria (section 3.1.). Therefore, we focus on the detailed structure of this network. Inception-ResNetV2, a cutting-edge deep learning architecture developed by Google, combines the Inception and ResNet modules to enhance performance in image recognition tasks. Renowned for its exceptional accuracy and efficiency in handling intricate visual data, this architecture integrates key elements such as the Inception module (GoogLeNet) for capturing multi-scale features and the ResNet-inspired residual connections that aid gradient flow during training. Additionally, Inception-ResNetV2 introduces Inception-ResNet blocks, merging Inception and ResNet principles to create powerful feature extraction mechanisms using convolutional layers of varying sizes and depths. By incorporating parallel paths within each module, the network explores diverse ways to represent input data, achieving a richer feature set and boosting overall recognition capabilities while maintaining efficient computation without compromising accuracy. This network consists of various types of layers, including Average Pooling, Dropout, Inception-ResNet-v2-A, Inception-ResNet-v2-B, Inception-ResNet-v2-C, Inception-ResNet-v2 Reduction-B, Reduction-A, and Softmax (Fig. 5 ). These layers serve different functions within the network, such as pooling operations, regularisation, image model blocks, and output functions. It has a total of 467 layers. This extensive depth in the network contributes to its ability to handle complex image classification tasks effectively. Figure 6 offers a comprehensive overview of the layers comprising the best model (Inception-ResNet-V2 – based network), including the output shapes and parameter count. The model includes approximately 24,000 (0.42%) non-trainable parameters from a total exceeding 59 million. Specific layers, such as Dropout, Max Pooling, and Global Average Pooling, do not add to the number of trainable parameters. This is because these layers do not possess any trainable parameters themselves. 2.4 Network training The training of the CNNs follows a standard neural network architecture, involving two primary stages: Feed-Forward and Back-Propagation. During the feed-forward phase, the network assesses its error by contrasting the output produced from the input image with the associated labelled output. In the backpropagation phase, gradients for the parameters are determined from this error, leading to subsequent updates of the weight matrices. This iterative process involves the input being fed into the network, propagated through various layers, and the resulting output being compared with the desired output to determine the error. This error is then utilised to adjust the network parameters, and the process is repeated multiple times, with each iteration referred to as an epoch. The error calculation involves various formulas and functions, and the optimisation process consists of incremental adjustments to the weights to minimise errors. To achieve this goal, we chose the appropriate optimiser from the options such as SGD, RMSprop, Nadam, and Adam. The performance of the model was evaluated using test accuracy and test loss. We additionally utilised the categorical cross-entropy loss function for the image data in this research. Table 3 represents the detailed properties of the optimizers. Table 3 The properties of the optimisers for the purpose of this study. Optimiser Properties RMSprop lr = 0.001 SGD lr = 0.001, momentum = 0.5, nesterov = FALSE Adam lr = 0.001, beta_1 = 0.9, beta_2 = 0.999 Nadam lr = 0.001, beta_1 = 0.9, beta_2 = 0.999 The training of the model entailed a systematic exploration of hyperparameters and components to determine the optimal configuration. During the training process, we carefully monitored the loss and accuracy trends for both the training and validation datasets at each epoch. This rigorous method enabled us to comprehensively evaluate the performance of the models and decide on their efficiency for our goal. The process of selecting the best model is depicted in the results and discussion section. To evaluate the potential models, we analysed several performance metrics, including accuracy, loss, and model stability. Accuracy measures the proportion of correct predictions made by the model, while loss indicates the average error for each instance. Generally, lower loss values suggest better model performance. However, it is essential to recognise that a model exhibiting high accuracy but significant loss may be at risk of overfitting, which can adversely affect its ability to generalise to unseen test data. So, we meticulously assessed the possibility of overfitting during our comparative analysis of the four candidate models. To guarantee the dependability and reproducibility of our findings, we conducted the training, validation, and testing phases using Python version 3.10.12 on the Google Colaboratory platform. The computational environment features an NVIDIA K80 GPU with 12 GB of memory, and we utilised Keras with TensorFlow as the backend (version 2.15.0), alongside libraries such as OpenCV to conduct the experiments. 3 RESULTS AND DISCUSSION The results of developing deep learning models for distinguishing between winter wheat and its four prevalent weeds, namely annual bugloss, bastard cabbage, hoary cress, and flixweed, using colour images is presented in this section. These models would be suitable for precision wheat weed management in various scenarios where any of these weeds are present. 3.1 Evaluation of models The first stage in developing an appropriate deep learning model entails the selection of a base network from a variety of standard options via transfer learning, where the initial weights are imported from ImageNet dataset. We assessed four cutting-edge standard deep learning networks: Xception, EfficientNetB0, VGG19, and InceptionResNetV2. Following fine-tuning, these networks underwent a modification phase by incorporating three blocks of CNN layers into the ultimate layer of the fine-tuned standard candidate models to establish four candidate architectures (Model 1 to Model 4) (Table 2 ). The models were trained using images resized to 299 × 299 pixels, with the SGD optimiser, a batch size of 32, and 200 epochs to determine the optimal network. During this process, we employed selection criteria that assessed model stability during training and analysed test accuracy and test loss. The results comparing the four candidate models based on maximum and minimum accuracy on the training, validation, and test datasets are presented in Table 4 . As observed in this table, Models 1 and 4 achieved 100% accuracy on the training dataset, while Model 3 attained 100% accuracy on the validation dataset. Notably, Model 3 produced the lowest loss values among all models on the training, validation, and test datasets. However, Model 4 achieved the highest accuracy on the test dataset, followed by Models 1 and 3, respectively. Model 2 yielded the lowest test accuracy. The accuracy and loss values of Models 1 and 4 are relatively close, which could be attributed to the presence of the Inception module in both models. Table 4 Comparison of the four deep learning models on training, validation, and test datasets. Model 1 Model 2 Model 3 Model 4 Maximum accuracy on train dataset 1.0000 (epoch: 161) 0.9837 (epoch: 179) 0.9810 (epoch: 190) 1.0000 (epoch: 179) Minimum loss on train dataset 3.1821 (epoch: 200) 3.2169 (epoch: 200) 2.9758 (epoch: 200) 3.1642 (epoch: 199) Maximum accuracy on validation dataset 0.9846 (epoch: 187) 0.9077 (epoch: 194) 1.0000 (epoch: 161) 0.9846 (epoch: 188) Minimum loss on validation dataset 3.2592 (epoch: 198) 3.4446 (epoch: 194) 2.9286 (epoch: 200) 3.2194 (epoch:197) Accuracy on test dataset 0.9725 0.8532 0.9358 0.9817 Loss on test dataset 3.2550 3.4555 2.9868 3.1933 To gain deeper insights into the training process of the four candidate models, we analysed the trends of accuracy and loss concerning the number of epochs during the training phase for both the training and validation datasets. The graphical representation of these trends for all models is depicted in Fig. 7. Accordingly, Model 2 exhibits slightly more pronounced fluctuations compared to the other three models, indicating a less stable training pattern. In contrast, Models 1, 3, and 4 display a consistent downward trend in losses and a steady increase in accuracies, with only minor variations. This behaviour indicates that Models 1, 3, and 4 exhibit greater resistance to instability and are likely to generalise more effectively to new, unseen data, rendering them more dependable options. Figure 8 compares the total parameters, trainable parameters, and non-trainable parameters for the four candidate models. This comparison provides valuable insights into the response time of the models. According to this figure, Model 1 and Model 4 exhibit comparable values for these parameters, which can be attributed to their Inception-based architectures. Model 2 offers the minimum number of trainable parameters, making it advantageous in terms of response time for real-time applications. However, it achieves the lowest accuracy and the highest loss on test dataset among all four models. Model 3, on the other hand, provides promising results in terms of accuracy and loss while maintaining a reasonable number of trainable parameters (Fig. 8), making it a compelling choice for our purposes. After carefully considering the evaluation criteria, we proposed Model 4 as the optimal choice for our specific purpose. This model utilises a modified and fine-tuned InceptionResNetV2 architecture with the SGD optimiser. Model 4 exhibited minimal fluctuations during the training process, indicating a high level of stability. Moreover, it achieved impressive accuracy and loss values on the test dataset (98.17% accuracy and a loss of 3.19, respectively), along with a reasonably fast response time. These factors suggest that Model 4 would perform exceptionally well in real-world scenarios, making it our preferred selection. 3.2 Classification performance To evaluate the efficacy of the models in distinguishing the five classes, various parameters were considered. These parameters include True Positive (TP), representing the count of instances correctly identified as positive; True Negative (TN), denoting instances correctly identified as negative; False Positive (FP), indicating instances incorrectly identified as positive; and False Negative (FN), signifying instances incorrectly identified as negative. These parameters are utilised in the development of two classification metrics: the classification report and the confusion matrix. The classification report provides insights into the model's performance based on precision, recall, and F1-score, as outlined in Equations 2–4. Precision quantifies the model's accuracy in predicting positive cases. A high precision value suggests a lower ratio of incorrect to correct predictions, and conversely. Recall quantifies the ratio of accurately predicted positive instances to the total number of actual positive instances, highlighting the model's capability to correctly identify all positive instances. Furthermore, recall provides information on the model's capacity to minimise false negatives. The F1-score, calculated as the geometric mean of precision and recall, serves as a metric that balances these two parameters to achieve an optimal trade-off between precision and recall. \(\:Precision=\:\left(\frac{TP}{TP+FP}\right)\) (2) \(\:Recall=\:\left(\frac{TP}{TP+FN}\right)\) (3) \(\:f1-score=\:(2\times\:\frac{Precision\times\:Recall}{Precision+Recall})\) (4) The classification report describing the categorisation of winter wheat, annual bugloss, bastard cabbage, hoary cress, and flixweed classes by colour images through the four models is outlined in Table 5. Table 5 The classification report detailing the recognition performance of the five classes by the four candidate models. Annual Bugloss Bastard Cabbage Flixweed Hoary Cress Winter Wheat Micro Avg. Macro Avg. Weighted Avg. Samples Avg. Precision Model 1 0.95 0.95 1.00 0.95 1.00 0.97 0.97 0.97 0.96 Model 2 0.74 1.00 1.00 1.00 0.83 0.88 0.92 0.91 0.84 Model 3 1.00 0.80 1.00 1.00 1.00 0.95 0.96 0.96 0.93 Model 4 0.95 0.95 1.00 1.00 1.00 0.98 0.98 0.98 0.97 Recall Model 1 0.91 0.95 1.00 1.00 0.95 0.96 0.96 0.96 0.96 Model 2 1.00 0.52 0.96 0.71 1.00 0.84 0.84 0.84 0.84 Model 3 0.70 0.95 1.00 1.00 1.00 0.93 0.93 0.93 0.93 Model 4 0.91 1.00 1.00 1.00 0.95 0.97 0.97 0.97 0.97 f1-score Model 1 0.93 0.95 1.00 0.98 0.97 0.97 0.97 0.97 0.96 Model 2 0.85 0.69 0.98 0.83 0.91 0.86 0.85 0.86 0.84 Model 3 0.82 0.87 1.00 1.00 1.00 0.94 0.94 0.94 0.93 Model 4 0.93 0.98 1.00 1.00 0.97 0.98 0.98 0.98 0.97 Supports 23 21 24 21 20 109 109 109 109 Based on Table 5, the precision parameter analysis reveals that Model 1 was able to identify the flixweed and winter wheat classes with 100% accuracy. Using Model 2, three classes—bastard cabbage, flixweed, and hoary cress—were identified with 100% accuracy. Moreover, Model 3 successfully identified four classes—annual bugloss, flixweed, hoary cress, and winter wheat—with 100% accuracy, while Model 4 detected three classes—flixweed, hoary cress, and winter wheat—with 100% precision. On the other hand, if the objective is to identify a specific class (e.g., for selective treatment), it is observed that annual bugloss can be accurately identified using Model 3 and bastard cabbage using Model 2, both at 100% accuracy. Furthermore, Models 2, 3, and 4 can be utilised to identify hoary cress with 100% accuracy. It is worth noting that flixweed can be identified accurately with 100% precision by all four models. Conversely, if the goal is to differentiate weed species from winter wheat for herbicidal purposes, Models 1, 3, and 4 can be employed to identify winter wheat with 100% precision. Lastly, based on the aforementioned table, the lowest precision value was related to identifying annual bugloss with Model 2 (0.74). To evaluate the models' capacity to minimise false negatives, we can compare their recall values. The recall values in the table indicate that this parameter achieved 100% for the annual bugloss class using Model 2; for the bastard cabbage class using Model 4; for the flixweed and hoary cress classes using Models 1, 3, and 4; and for the winter wheat class using Models 2 and 3. Conversely, the lowest recall value was observed for Model 2 in relation to the bastard cabbage class (0.52). Furthermore, the F1-score parameter reached 100% for the flixweed class using Models 1, 3, and 4; for the hoary cress class using Models 3 and 4; and for the winter wheat class using Model 3. The lowest F1-score value was observed for the bastard cabbage class using Model 2 (0.69). Consequently, Model 4 demonstrated the best performance with the highest average precision, recall, and F1-score values. Figure 9 presents the confusion matrices for classifying five categories with four candidate models. The rows represent the true classes while the columns show the predicted classes. Through these matrices, we can see that there were only two annual bugloss instances mistakenly classified as bastard cabbage or hoary cress using Model 1. Additionally, Models 3 and 4 misclassified seven and one annual bugloss instances as bastard cabbage respectively, whereas Model 2 accurately identified this class without any errors. For the bastard cabbage class, one instance was misclassified as annual bugloss using Model 1; eight instances were misclassified as annual bugloss using Model 2; and one instance was misclassified as winter wheat also using Model 2. In contrast, Models 3 and 4 accurately classified this class with 100% precision. The flixweed class was accurately classified without error using Models 1, 3, and 4; however, Model 2 misclassified one instance as winter wheat. For the hoary cress class, Models 1, 3, and 4 accurately classified it without error; whereas Model 2 misclassified two instances as annual bugloss and four instances as winter wheat. Finally, for the winter wheat class, Models 1, 2, and 3 accurately classified it without error; whereas Model 4 misclassified only one instance as annual bugloss. 3.3 Winter wheat weed management scenarios It is essential to emphasise that the findings of this study can be effectively generalised to encompass various scenarios in winter wheat fields with respect to weed appearance. Given that four distinct weed species have been investigated, we can conclude that the proposed model(s) can be utilised as a tool for efficient weed management in 15 different weed appearance scenarios that may arise in winter wheat fields. These scenarios are detailed in Table 6, which accounts for one to four weeds that can coexist alongside winter wheat. Notably, scenario 15 represents the general condition that we have studied thus far in this investigation. Table 6 Fifteen scenarios for different weed appearance in winter wheat fields. Scenario number Weed(s) alongside winter wheat No. of classes Scenario 1 Flixweed 2 Scenario 2 Hoary Cress 2 Scenario 3 Annual Bugloss 2 Scenario 4 Bastard Cabbage 2 Scenario 5 Flixweed + Hoary Cress 3 Scenario 6 Annual Bugloss + Flixweed 3 Scenario 7 Bastard Cabbage + Flixweed 3 Scenario 8 Annual Bugloss + Hoary Cress 3 Scenario 9 Bastard Cabbage + Hoary Cress 3 Scenario 10 Annual Bugloss + Bastard Cabbage 3 Scenario 11 Annual Bugloss + Flixweed + Hoary Cress 4 Scenario 12 Bastard Cabbage + Flixweed + Hoary Cress 4 Scenario 13 Annual Bugloss + Bastard Cabbage + Flixweed 4 Scenario 14 Annual Bugloss + Bastard Cabbage + Hoary Cress 4 Scenario 15 Annual Bugloss + Bastard Cabbage + Flixweed + Hoary Cress 5 For this analysis, we used confusion matrices (Fig. 9), and for each scenario, we extracted a separate subset that must include winter wheat for each model. This analysis will involve creating 60 subsets of the confusion matrices (15 subsets for each model). The results of this analysis are presented in Table 7 and Fig. 10. Accordingly, it was concluded that Models 1, 3, and 4 are usable with 100% accuracy for scenarios 1, 2, 4, 5, 7, 9, and 12. Scenario 3 can be treated with 100% accuracy using Models 1, 2, and 3. Additionally, for scenario 6, Models 1 and 3 are suggested with 100% accuracy. For scenario 8, Model 3 will perform best with 100% accuracy. For scenario 11, Model 3 showed the best performance (100%). For scenarios 10, 13, and 14, while 100% accuracy was not achieved, Models 1 and 4 provided the highest accuracy of around 97% for scenario 10. The same two models can be used for scenario 13, with an accuracy of about 98%. Model 4 can be employed for scenario 14 with an accuracy of 98%. If we were to choose a comprehensive model for various scenarios, we would consider the average accuracy across different models and scenarios. Accordingly, based on Table 7, Model 4 has the highest average accuracy (98.92%) and the lowest standard deviation (0.0110), making it the most suitable model. Although Model 3 is noteworthy for achieving 100% accuracy in 11 different scenarios. It is important to reiterate that scenario 15 is the initial and overall scenario considered in this research, based on which the deep learning models have been designed. Table 7 Average and standard deviation of accuracies for the four models in all scenarios. Model 1 Model 2 Model 3 Model 4 Average accuracies in all scenarios 0.9890 0.9191 0.9776 0.9892 STD of accuracies in all scenarios 0.0161 0.0504 0.0380 0.0110 The ability to identify distinct weed and crop species from RGB images is a significant challenge, especially with a large number of species. The most promising research seems to focus on a broader range of species, achieving higher accuracy while de-emphasising specific environmental conditions. It is worth noting that a successful weed discrimination model for one crop may not perform adequately for others. However, considering various weed appearance scenarios for distinct crops and proposing models for each scenario could be advantageous, as it would cater to diverse conditions that may arise for those crops. This aspect could be a significant strength of our study compared to similar research. Shao et al., (2023) attempted to distinguish six weed species in paddy fields using deep learning methods under natural field conditions similar to our study but achieved lower accuracy. G C et al., (2024) applied five different deep learning frameworks to differentiate crop and weed species in a greenhouse that provides a roughly similar environment. Despite considering 14 crop and weed species, their accuracy was relatively low compared to our work. In contrast, our study achieved promising accuracy without such pre-processing steps. Finally, Makarian & Saedi, (2024) attained perfect accuracy in natural conditions; however, the small number of studied classes limits the model's generalisability. 4. CONCLUSION In this study, we developed four improved deep neural network models to differentiate winter wheat from four common weeds. We used colour images as input data to train and test the models. The results obtained from testing the models on unseen data (test data) demonstrated their ability to accurately identify these five plant species in photographs. Our findings show that all four proposed models have promising performance in distinguishing the plant classes, with InceptionResNetV2 exhibiting the highest accuracy. We also investigated fifteen distinct scenarios of weed occurrence in winter wheat fields, taking into account the number of weeds present alongside the wheat crop, to determine the most effective weed management strategy. The models reliably performed across these scenarios, considering winter wheat as a separate class. This approach also enables the targeting of non-wheat plants as weeds, facilitating a selective weed control strategy. Considering the significance of producing herbicide-free wheat and minimising environmental pollution, the findings of this research pave the way for precision weed management and the creation of automated weed removal solutions. This technology will not only reduce herbicide consumption and environmental impact but also contribute to increased wheat yield and quality. Declarations Acknowledgments The authors would like to acknowledge the financial support of Shahrood University of Technology for this research under project No: RP_S_219444. Author contributions S.I.S. contributed to software development, writing, formal analysis, and methodology. H. M. contributed to resources, data curation, conceptualization, and writing original draft. All authors reviewed the manuscript. Data availability statement The dataset used in the present study will be available upon reasonable request via [email protected] (S. I. Saedi). Competing Interests Statement The authors declare that they have no competing interests. S.I. Saedi: Software, Conceptualization, Writing – review and editing, Formal analysis, Writing – original draft, Supervision, Methodology H. Makarian: Resources, Data curation, Visualization, Conceptualization, Investigation, Validation References Chollet, F. (2017). Deep Learning with Python (1st ed.). Manning Publications Co. Fan, X., Chai, X., Zhou, J., & Sun, T. (2023). Deep learning based weed detection and target spraying robot system at seedling stage of cotton field. Computers and Electronics in Agriculture , 214 , 108317. https://doi.org/https://doi.org/10.1016/j.compag.2023.108317 Gao, J., French, A. P., Pound, M. P., He, Y., Pridmore, T. P., & Pieters, J. G. (2020). Deep convolutional neural networks for image-based Convolvulus sepium detection in sugar beet fields. Plant Methods , 16 (1), 29. https://doi.org/10.1186/s13007-020-00570-z GC, S., Zhang, Y., Howatt, K., Schumacher, L. G., & Sun, X. (2024). Multi-Species Weed and Crop Classification Comparison Using Five Different Deep Learning Network Architectures. Journal of the ASABE , 67 (2), 275–287. https://doi.org/https://doi.org/10.13031/ja.15590 Gyawali, A., Bhandari, R., Budhathoki, P., & Bhattarai, S. (2022). A REVIEW ON EFFECT OF WEEDS IN WHEAT (Triticum aestivum L.) AND THEIR MANAGEMENT PRACTICES . 34–40. https://doi.org/10.26480/faer.02.2022.34.40 Juwono, F. H., Wong, W. K., Verma, S., Shekhawat, N., Lease, B. A., & Apriono, C. (2023). Machine learning for weed–plant discrimination in agriculture 5.0: An in-depth review. Artificial Intelligence in Agriculture , 10 , 13–25. https://doi.org/https://doi.org/10.1016/j.aiia.2023.09.002 Khosravi, H., Saedi, S. I., & Rezaei, M. (2021). Real-time recognition of on-branch olive ripening stages by a deep convolutional neural network. Scientia Horticulturae , 287 , 110252. https://doi.org/https://doi.org/10.1016/j.scienta.2021.110252 Langridge, P., Alaux, M., Almeida, N., Ammar, K., Baum, M., Bekkaoui, F., Bentley, A., Beres, B., Berger, B., Braun, H.-J., Brown-Guedira, G., Burt, C., Caccamo, M., Cattivelli, L., Charmet, G., Civáň, P., Cloutier, S., Cohan, J.-P., Devaux, P., & Zhang, X. (2022). Meeting the Challenges Facing Wheat Production: The Strategic Research Agenda of the Global Wheat Initiative. Agronomy , 12 , 2767. https://doi.org/10.3390/agronomy12112767 Li, Z., Wang, D., Yan, Q., Zhao, M., Wu, X., & Liu, X. (2024). Winter wheat weed detection based on deep learning models. Computers and Electronics in Agriculture , 227 , 109448. https://doi.org/https://doi.org/10.1016/j.compag.2024.109448 Makarian, H., & Saedi, S. I. (2024). Automated classification of saffron and broadleaf weeds of flixweed and hoary cress using deep learning and color images. Crop Protection , 106750. https://doi.org/https://doi.org/10.1016/j.cropro.2024.106750 Montazeri, M., Zand, E., Baghestani, M. A. (2005). Weeds and their Control in Wheat Fields of Iran, first ed . Agricultural Research and Education Organization Press. Nasiri, A., Omid, M., Taheri-Garavand, A., & Jafari, A. (2022). Deep learning-based precision agriculture through weed recognition in sugar beet fields. Sustainable Computing: Informatics and Systems , 35 , 100759. https://doi.org/https://doi.org/10.1016/j.suscom.2022.100759 Nasiri, A., Taheri-Garavand, A., & Zhang, Y.-D. (2019). Image-based deep learning automated sorting of date fruit. Postharvest Biology and Technology , 153 , 133–141. https://doi.org/https://doi.org/10.1016/j.postharvbio.2019.04.003 Peng, H., Huang, B., Shao, Y., Li, Z., Zhang, C., Chen, Y., & Xiong, J. (2018). General improved SSD model for picking object recognition of multiple fruits in natural environment. Nongye Gongcheng Xuebao/Transactions of the Chinese Society of Agricultural Engineering , 34 , 155–162. https://doi.org/10.11975/j.issn.1002-6819.2018.16.020 Quan, L., Feng, H., Lv, Y., Wang, Q., Zhang, C., Liu, J., & Yuan, Z. (2019). Maize seedling detection under different growth stages and complex field environments based on an improved Faster R–CNN. Biosystems Engineering , 184 , 1–23. https://doi.org/https://doi.org/10.1016/j.biosystemseng.2019.05.002 Rai, N., & Sun, X. (2024). WeedVision: A single-stage deep learning architecture to perform weed detection and segmentation using drone-acquired images. Computers and Electronics in Agriculture , 219 , 108792. https://doi.org/https://doi.org/10.1016/j.compag.2024.108792 Sadeghi, S. H., Saedi, S. I., Peters, R. T., & Stöckle, C. (2022). Towards improving the global water application uniformity of centre pivots through lateral speed adjustment. Biosystems Engineering , 215 , 215–227. https://doi.org/https://doi.org/10.1016/j.biosystemseng.2022.01.012 Saedi, S I, Alimardani, R., Mousazadeh, H., & Salehi, R. (2019). Development and evaluation of an energy and water efficient intensive cropping system. INMATEH-Agricultural Engineering , 58 , 93–104. Saedi, Seyed I, & Rezaei, M. (2024). A Modified Xception Deep Learning Model for Automatic Sorting of Olives Based on Ripening Stages. In Inventions (Vol. 9, Issue 1). https://doi.org/10.3390/inventions9010006 Saedi, Seyed Iman, & Khosravi, H. (2020). A deep neural network approach towards real-time on-branch fruit recognition for precision horticulture. Expert Systems with Applications , 159 , 113594. https://doi.org/https://doi.org/10.1016/j.eswa.2020.113594 Saedi, Seyed Iman, Rezaei, M., & Khosravi, H. (2024). Dual-path lightweight convolutional neural network for automatic sorting of olive fruit based on cultivar and maturity. Postharvest Biology and Technology , 216 , 113054. https://doi.org/https://doi.org/10.1016/j.postharvbio.2024.113054 Salim, F., Saeed, F., Basurra, S., Qasem, S. N., & Al-Hadhrami, T. (2023). DenseNet-201 and Xception Pre-Trained Deep Learning Models for Fruit Recognition. In Electronics (Vol. 12, Issue 14). https://doi.org/10.3390/electronics12143132 Shao, Y., Guan, X., Xuan, G., Gao, F., Feng, W., Gao, G., Wang, Q., Huang, X., & Li, J. (2023). GTCBS-YOLOv5s: A lightweight model for weed species identification in paddy fields. Computers and Electronics in Agriculture , 215 , 108461. https://doi.org/https://doi.org/10.1016/j.compag.2023.108461 Shewry, P. R., & Hey, S. J. (2015). The contribution of wheat to human diet and health. Food and Energy Security , 4 (3), 178–202. https://doi.org/10.1002/fes3.64 Singh, M., Kukal, M. S., Irmak, S., & Jhala, A. J. (2021). Water Use Characteristics of Weeds: A Global Review, Best Practices, and Future Directions. In Frontiers in plant science (Vol. 12, p. 794090). https://doi.org/10.3389/fpls.2021.794090 Speck, O., & Speck, T. (2021). Functional morphology of plants - a key to biomimetic applications. The New Phytologist , 231 (3), 950–956. https://doi.org/10.1111/nph.17396 Subeesh, A., Bhole, S., Singh, K., Chandel, N. S., Rajwade, Y. A., Rao, K. V. R., Kumar, S. P., & Jat, D. (2022). Deep convolutional neural network models for weed detection in polyhouse grown bell peppers. Artificial Intelligence in Agriculture , 6 , 47–54. https://doi.org/https://doi.org/10.1016/j.aiia.2022.01.002 Susetyarini, E., Wahyono, P., Latifa, R., & Nurrohman, E. (2020). The Identification of Morphological and Anatomical Structures of Pluchea indica. Journal of Physics: Conference Series , 1539 (1), 12001. https://doi.org/10.1088/1742-6596/1539/1/012001 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6710451","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":471453598,"identity":"48e2f7f0-4051-47e5-9d20-0fd33857c642","order_by":0,"name":"Seyed Iman Saedi","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAuElEQVRIiWNgGAWjYBACNhDBY2ADJBkbD5CiJQ2kpYE4LWDAw3AYTBOnhY//8MMHbwrO261tPwy0pcYmmrDDJNKMDecY3E7ediYRqOVYWm4DYS0MZtI8QC1mB4BaGBsOE6GF//j33zwG55LNzj8kVgtDjhkzj8EBO7MbRNsikVMsOccgOcHsBtCWBGL8It9/fOOHN3/s7M3Opz988KHGhrAWGEgEq0wgVjkI2JOieBSMglEwCkYYAABaIULN9FWPTwAAAABJRU5ErkJggg==","orcid":"","institution":"Shahrood University of Technology","correspondingAuthor":true,"prefix":"","firstName":"Seyed","middleName":"Iman","lastName":"Saedi","suffix":""},{"id":471453599,"identity":"762041b0-7bfc-4a0f-8177-383a5707bbed","order_by":1,"name":"Hassan Makarian","email":"","orcid":"","institution":"Shahrood University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Hassan","middleName":"","lastName":"Makarian","suffix":""}],"badges":[],"createdAt":"2025-05-20 19:08:16","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6710451/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6710451/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":84778910,"identity":"72a7f6c0-d0be-4622-9135-cb01651c431c","added_by":"auto","created_at":"2025-06-17 09:19:59","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1991098,"visible":true,"origin":"","legend":"\u003cp\u003eSample images of winter wheat and weeds showcasing the five classes studied in fully natural condition within a field.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6710451/v1/11f37c45e36901787ce62e3b.png"},{"id":84778914,"identity":"2cf6e356-e9c6-4225-9d8a-cf5beaf8f31a","added_by":"auto","created_at":"2025-06-17 09:19:59","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":83343,"visible":true,"origin":"","legend":"\u003cp\u003eThe pie chart for the number and proportion of samples that showcases a balanced distribution of data used in this study.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6710451/v1/b9ef3ceda8cf97784a2f5299.png"},{"id":84778977,"identity":"ea72910c-008c-4313-afe8-2c7844d13a0f","added_by":"auto","created_at":"2025-06-17 09:20:00","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":135908,"visible":true,"origin":"","legend":"\u003cp\u003eModifying block appended to the final layer of the base networks.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6710451/v1/d3426b4f0612f8e15eff7b42.png"},{"id":84778916,"identity":"9864996b-912c-4f03-83bc-aedd2fd6d790","added_by":"auto","created_at":"2025-06-17 09:19:59","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":103322,"visible":true,"origin":"","legend":"\u003cp\u003eDeep learning structure for the classification of five plant species under study.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6710451/v1/a7bacce89303f1fb5e0d5e1e.png"},{"id":84778918,"identity":"18f067fa-efa7-4d41-a3f1-597a02e8186c","added_by":"auto","created_at":"2025-06-17 09:19:59","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":595530,"visible":true,"origin":"","legend":"\u003cp\u003eThe architecture of Inception-ResNet-V2 deep learning network.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6710451/v1/6eed948694d945f8a1e78d76.png"},{"id":84778978,"identity":"16195710-7052-4e34-9567-ee0fa754a021","added_by":"auto","created_at":"2025-06-17 09:20:00","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":424246,"visible":true,"origin":"","legend":"\u003cp\u003eLayer configuration of the improved InceptionResNetV2 deep learning model.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-6710451/v1/8426e2a9efa89dfb7ecd3d9f.png"},{"id":84778904,"identity":"360bdf44-ea36-4c5a-b04f-1fbf5ad8ebdb","added_by":"auto","created_at":"2025-06-17 09:19:59","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":467024,"visible":true,"origin":"","legend":"\u003cp\u003eAccuracy and loss trends on train and validation datasets for the four candidate model during training.\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-6710451/v1/bc2ade28c121ae1896893de7.png"},{"id":84778912,"identity":"478d481d-036d-43b5-a413-0c05dd4f0046","added_by":"auto","created_at":"2025-06-17 09:19:59","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":139296,"visible":true,"origin":"","legend":"\u003cp\u003eEvaluation of the number of parameters (total, trainable, and non-trainable) for the four candidate models.\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-6710451/v1/ed8f9e6895addbe179248f2d.png"},{"id":84780790,"identity":"e3b9d394-70c4-443c-b0ff-584f6e5bb3ec","added_by":"auto","created_at":"2025-06-17 09:28:00","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":271540,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrices illustrating the classification performance of the five classes using the four developed models.\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-6710451/v1/0193467aa448339d91e58baf.png"},{"id":84778973,"identity":"222dccce-2827-48cc-84e0-486bbf1d8653","added_by":"auto","created_at":"2025-06-17 09:20:00","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":218894,"visible":true,"origin":"","legend":"\u003cp\u003eThe accuracy of the four model candidates in different winter wheat weed appearance scenarios.\u003c/p\u003e","description":"","filename":"10.png","url":"https://assets-eu.researchsquare.com/files/rs-6710451/v1/7bcd540e5a37127cccbab51a.png"},{"id":88615380,"identity":"3d84f2e6-0f53-401f-9295-b4c5ddb90928","added_by":"auto","created_at":"2025-08-08 10:38:42","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":6090163,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6710451/v1/9f3b330a-3671-4c5a-a305-3861091ada46.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Intelligent Weed Recognition in Winter Wheat Fields through Deep Convolutional Neural Networks","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eAs the global population continues to grow, wheat emerges as a vital crop in addressing the world's increasing nutritional demands (Sadeghi et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; S I Saedi et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Shewry \u0026amp; Hey, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). Wheat plays a key role in global food security, supplying roughly 20% of the essential carbohydrates and proteins needed for human consumption (Langridge et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). It occupies about 30% of the world's cereal cropland, underscoring its critical importance in agricultural systems (FAO, 2021). In light of the increasing challenges posed by climate change and a rapidly growing global population, enhancing wheat productivity has emerged as a crucial strategy to uphold food security in the years ahead. One significant obstacle to wheat production is weeds, which reduce the growth and yield of wheat by competing for water, nutrients, and light (Singh et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Weeds can significantly impact wheat productivity, with estimated yield losses of up to 30%, largely influenced by varying climatic conditions and differences in agricultural management practices (Montazeri, M., Zand, E., Baghestani, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2005\u003c/span\u003e). In many of Iran's key wheat-producing areas, broadleaved weeds have emerged as a prominent concern, accounting for a staggering 84% of the weed species infesting wheat fields and posing a significant threat to crop yields. Research on winter wheat has consistently demonstrated that the presence of weeds can have a profound impact on crop yields, resulting in losses of up to 20\u0026ndash;30% in severe cases. Both broadleaf and narrow leaf weeds can infest wheat crops, collectively contributing to a global grain yield reduction of 13.1%. The extent of this yield loss, however, is highly variable and depends on factors such as the type and density of weeds, as well as local climate conditions and precipitation patterns (Gyawali et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Manual removal of weeds in wheat fields is not a viable option due to the risk of crop damage and the high labour costs involved. Precise identification and differentiation of weeds from crop plants is a crucial component of effective site-specific weed management. By accurately distinguishing weeds from wheat, farmers can target herbicide applications to specific areas where weeds are present, thereby optimising weed control.\u003c/p\u003e \u003cp\u003eBy harnessing the power of advanced technologies such as computer vision and image processing, researchers can achieve high precision in weed detection, enabling the development of targeted management strategies that accurately distinguish between crops and weeds, ultimately optimising the effectiveness of site-specific weed control techniques (Juwono et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Through providing two-dimensional information, colour images have been widely utilised in this domain, laying the groundwork for further advancements. Studies have shown that unique morphological characteristics, including leaf form, colour, and dimension, can serve as valuable features for object recognition and classification tasks (Susetyarini et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Specifically, features like leaf area, length, perimeter, width, and colour can be employed to differentiate between plants and inform plant management strategies (Speck \u0026amp; Speck, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Despite these challenges, significant progress has been made in developing object recognition algorithms for precision agriculture, with varying degrees of accuracy achieved in different applications. The application of deep learning techniques has demonstrated considerable potential in achieving these objectives. Over the past decade, deep learning has transformed the field of image processing, achieving unprecedented success and finding numerous practical applications in precision agriculture. While large datasets are essential for training and validating deep learning algorithms, advances in computational power and data availability have made them increasingly popular. As a subset of machine learning, deep learning leverages hierarchical representations of data through sequential layers, typically employing convolutional neural networks (CNNs) to extract complex, non-linear image features. This approach has been shown to exhibit exceptional generalisation capabilities, making it well-suited for classification tasks ( Seyed I Saedi \u0026amp; Rezaei, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). This task begins with the input of an image, which can vary in size, and undergoes preprocessing before entering the feature extraction stage. This phase involves a series of layers that progressively extract features from the image, with the complexity of the features increasing with each additional layer. Following feature extraction, the classification process begins, employing traditional fully connected layers commonly found in neural networks. The arrangement and configuration of these layers give rise to a diverse range of networks with distinct characteristics and capabilities (Chollet, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2017\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eConvolutional neural networks have been extensively explored for identification and classification tasks in agriculture. Research has demonstrated the effectiveness of these networks in various applications. For instance, Seyed Iman Saedi \u0026amp; Khosravi, (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) developed an image-based CNN model to recognise six different types of on-branch fruits in orchards, achieving an impressive accuracy of 99.8%. Additionally, Khosravi et al., (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) developed a CNN model to classify olive fruit at different cultivars and developmental phases, attaining 91% classification. Furthermore, researchers have used deep neural networks such as Xception (Salim et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) and lightweight CNNs (Seyed Iman Saedi et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), to recognise and classify fruits with high accuracy, demonstrating the potential of these models in image classification tasks.\u003c/p\u003e \u003cp\u003eDeep learning algorithms have been successfully applied to weed detection and classification, demonstrating superior performance in distinguishing between crop plants and weeds (Li et al., \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). For instance, Quan et al., (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) employed the Faster R-CNN model to detect corn seedlings and weeds in complex environments, while Fan et al., (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) improved this model to achieve an accuracy of over 98.43% in identifying weeds in cotton fields. Other studies have utilised deep learning-based models to identify common weeds in sugar beet cultivation (Nasiri et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2022\u003c/span\u003e), discriminate pepper from weeds (Subeesh et al., \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2022\u003c/span\u003e), and detect weeds in sugar beet fields (Gao et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Additionally, researchers have evaluated the performance of deep learning models such as Yolov8, Yolov9, and customised Yolov9 in real-field conditions for eight crop species (GC et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Similarly, a modified Xception deep learning approach was successfully utilised to distinguish saffron from its broadleaf weeds, yielding promising results with F1 scores ranging from 91\u0026ndash;96% (Makarian \u0026amp; Saedi, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). In a practical study, Rai \u0026amp; Sun, (\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) designed a single-stage deep learning architecture that accomplishes two tasks simultaneously: detecting weed presence through bounding boxes and achieving pixel-wise instance segmentation on unmanned aerial system (UAS) acquired remote sensing images. The model achieved the best detection and segmentation scores of 85.4% and 82.1%, respectively.\u003c/p\u003e \u003cp\u003eThe aforementioned works demonstrate the potential of deep learning algorithms in plant identification and classification, and highlight the need for further research in this area. This research aimed to address the lack of a thorough investigation into site-specific weed management for winter wheat. This work builds on the authors\u0026rsquo; previous studies regarding weed-crop discrimination, aiming to develop a robust method for precision weed management in winter wheat fields. Here, advanced deep learning models were introduced to differentiate between winter wheat plants and four common weed species: annual bugloss, bastard cabbage, hoary cress, and flixweed. Utilising colour images, these models can accurately classify each plant category across various scenarios, paving the way for the creation of precise weed management systems for automated control in wheat fields.\u003c/p\u003e"},{"header":"2 Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Image preparation\u003c/h2\u003e \u003cp\u003eTo address the requirements of the study, 542 images were collected from the five classes understudy: Winter Wheat (\u003cem\u003eTriticum aestivum\u003c/em\u003e L.), Annual Bugloss (\u003cem\u003eAnchusa arvensis\u003c/em\u003e L.), Bastard Cabbage (\u003cem\u003eRapistrum rugosum\u003c/em\u003e L.), Hoary Cress (\u003cem\u003eCardaria draba\u003c/em\u003e (L.) Desv.), and Flixweed (\u003cem\u003eDescurainia sophia\u003c/em\u003e L.). Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e illustrates a selection of these images. The photos were taken with the 48-megapixel camera of an A-12 Samsung Galaxy (SM-A125F/DS).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn the pie chart shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, the quantity and proportion of samples for each class are shown. The illustration highlights the balanced distribution of data used in this study. Having a balanced dataset is crucial to prevent bias, improve generalisation, enhance performance, promote stability, and avoid misinterpretation. This balance ensures that the model receives sufficient examples from each class, enabling effective learning and accurate predictions across all classes.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe images captured underwent a review process in which inappropriate samples were excluded. An agronomist with expertise in weed and crop properties contributed to this selection process. Ultimately, 542 images were chosen to develop the model and train the network. The selected images were divided into three datasets: training, validation, and testing. 80% of the data was allocated for training, while the remaining 20% was split for testing. Within the training data, 15% was set aside for validation purposes. It is worth emphasising that the test set is reserved exclusively for evaluating the final model's performance and testing its accuracy on unseen data. This separation ensures a reliable assessment of the model's generalisation capabilities.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Image pre-processing\u003c/h2\u003e \u003cp\u003ePreprocessing input images is essential for improving the model's accuracy, mitigating overfitting, and enhancing its generalisation ability.\u003c/p\u003e \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e \u003ch2\u003e2.2.1 Resize and normalisation\u003c/h2\u003e \u003cp\u003eInitially, we standardised the image sizes to 299 \u0026times; 299 pixels to ensure uniform resolution across all images. Subsequently, we normalised the resolutions to fall within the range [0,1] by utilising the Eq.\u0026nbsp;\u003cspan refid=\"Equ1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:{x}_{i}^{{\\prime\\:}}=\\:\\frac{{x}_{i}-\\:{x}_{min}}{{x}_{max}-\\:{x}_{min}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{x}_{i}^{{\\prime\\:}}\\)\u003c/span\u003e\u003c/span\u003e symbolises the normalised value within the dataset, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{x}_{i}\\)\u003c/span\u003e\u003c/span\u003e​ represents the i-th value in the dataset, and\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\:{x}_{min}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{x}_{max}\\:\\)\u003c/span\u003e\u003c/span\u003edenote the minimum and maximum values in the dataset. This formula guarantees that after normalisation, the normalised value for the dataset's minimum value will consistently be 0, and the normalised value for the dataset's maximum value will invariably be 1.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e \u003ch2\u003e2.2.2 Augmentation\u003c/h2\u003e \u003cp\u003eData augmentation is a widely recognized method in computer vision. It seeks to create extra training data from existing samples by applying random transformations. This technique ensures that the generated images are realistic and that each specific image is unique (Chollet, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). Consequently, image augmentation can alleviate the challenges encountered by deep neural networks regarding their reliance on a substantial quantity of annotated data, leading to improved accuracy. The extensive dataset enhances the training procedure through allowing the network to acquire all essential parameters while reducing the likelihood of overfitting. In the present study, an automated data augmentation technique was utilized, incorporating random height and width shifts, random flipping, random contrast adjustments, and random rotations. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e represents The specific data augmentation arguments for our task.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eAugmentation parameters and values.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eImage augmentation parameters\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eValues\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom height translation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom width translation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom flip\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTrue\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom contrast\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.15\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom rotation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.15\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Candidate models\u003c/h2\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.3.1 Base networks\u003c/h2\u003e \u003cp\u003eThe initial step in reaching our goal is to select a base network from several choices through transfer learning, which entails loading weights from the ImageNet dataset. To this end, we tested four innovative standard deep learning networks, including Xception, EfficientNetB0, VGG19, and InceptionResNetV2.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003e2.3.2 Fine-tuning\u003c/h2\u003e \u003cp\u003eThe objective of this fine-tuning process was to customise the candidate base networks for the particular dataset and task at hand, leveraging the previously learned features while adapting to the new data. This technique prevents the overfitting in pre-trained layers while facilitating the adaptation of the newly added layers to the new data. During the fine-tuning process, our goal was to selectively freeze certain layers (the earlier ones) while enabling the training of others (the later ones) when incorporating a pre-trained model into transfer learning within a deep learning framework. Initially, we disabled the trainable attribute for the first 20 layers, locking them during training to prevent their weights from being altered and preserving the valuable features they had already acquired from the original training data. Subsequently, we advanced to fine-tune the later layers of the base model. We then iterated through the layers starting from the 20th position and activated the trainable attribute for each of these layers. This adjustment allows these layers to undergo training and updates throughout the training phase.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section3\"\u003e \u003ch2\u003e2.3.3 Modifying block\u003c/h2\u003e \u003cp\u003eTo develop a robust CNN model, we added a modifying block to the last layer of the proposed base networks. We tested various layer configurations for the modifying block, exploring different types, placements, and parameters to find the most efficient setup. This included examining combinations of 2D Convolution, Batch Normalization, Dropout, and Global Average Pooling layers. Parameters were selected through experimentation testing, though the specifics are not detailed here for the sake of brevity. Consequently, the optimized modifying block involved three sets of layers, including Convolution, Batch Normalization, Max Pooling, and Dropout, followed by the Fully Connected and Global Average Pooling layers. The final layer of the base candidate networks was supplemented with these additional layers to develop the proposed architectures in this study (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eBy following the steps outlined earlier, we developed four distinct models, labelled Model 1 through Model 4 (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Subsequently, we evaluated these models using images resized to 299\u0026times;299 pixels, the SGD optimiser, a batch size of 32, and 200 epochs to identify the optimal model. During this process, we employed selection criteria that included assessing the model's stability during training, as well as analysing test accuracy and test loss.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eFour proposed models for the purpose of this study.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel 1\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModel 2\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eModel 3\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eModel 4\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModified Xception\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModified EfficientNetB0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eModified VGG19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eModified InceptionResNetV2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe detailed outcomes of this step, including a comparison of the models' performance, are presented in Section 3.2. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e summarises the proposed model for the purpose of our study.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFollowing the evaluation of the candidate standard networks, the Inception-ResNetV2 stood out as the top performer based on the specified selection criteria (section 3.1.). Therefore, we focus on the detailed structure of this network.\u003c/p\u003e \u003cp\u003eInception-ResNetV2, a cutting-edge deep learning architecture developed by Google, combines the Inception and ResNet modules to enhance performance in image recognition tasks. Renowned for its exceptional accuracy and efficiency in handling intricate visual data, this architecture integrates key elements such as the Inception module (GoogLeNet) for capturing multi-scale features and the ResNet-inspired residual connections that aid gradient flow during training. Additionally, Inception-ResNetV2 introduces Inception-ResNet blocks, merging Inception and ResNet principles to create powerful feature extraction mechanisms using convolutional layers of varying sizes and depths. By incorporating parallel paths within each module, the network explores diverse ways to represent input data, achieving a richer feature set and boosting overall recognition capabilities while maintaining efficient computation without compromising accuracy. This network consists of various types of layers, including Average Pooling, Dropout, Inception-ResNet-v2-A, Inception-ResNet-v2-B, Inception-ResNet-v2-C, Inception-ResNet-v2 Reduction-B, Reduction-A, and Softmax (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). These layers serve different functions within the network, such as pooling operations, regularisation, image model blocks, and output functions. It has a total of 467 layers. This extensive depth in the network contributes to its ability to handle complex image classification tasks effectively.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e offers a comprehensive overview of the layers comprising the best model (Inception-ResNet-V2 \u0026ndash; based network), including the output shapes and parameter count. The model includes approximately 24,000 (0.42%) non-trainable parameters from a total exceeding 59\u0026nbsp;million. Specific layers, such as Dropout, Max Pooling, and Global Average Pooling, do not add to the number of trainable parameters. This is because these layers do not possess any trainable parameters themselves.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Network training\u003c/h2\u003e \u003cp\u003eThe training of the CNNs follows a standard neural network architecture, involving two primary stages: Feed-Forward and Back-Propagation. During the feed-forward phase, the network assesses its error by contrasting the output produced from the input image with the associated labelled output. In the backpropagation phase, gradients for the parameters are determined from this error, leading to subsequent updates of the weight matrices. This iterative process involves the input being fed into the network, propagated through various layers, and the resulting output being compared with the desired output to determine the error. This error is then utilised to adjust the network parameters, and the process is repeated multiple times, with each iteration referred to as an epoch.\u003c/p\u003e \u003cp\u003eThe error calculation involves various formulas and functions, and the optimisation process consists of incremental adjustments to the weights to minimise errors. To achieve this goal, we chose the appropriate optimiser from the options such as SGD, RMSprop, Nadam, and Adam. The performance of the model was evaluated using test accuracy and test loss. We additionally utilised the categorical cross-entropy loss function for the image data in this research. Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e represents the detailed properties of the optimizers.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eThe properties of the optimisers for the purpose of this study.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOptimiser\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eProperties\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRMSprop\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003elr\u0026thinsp;=\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSGD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003elr\u0026thinsp;=\u0026thinsp;0.001, momentum\u0026thinsp;=\u0026thinsp;0.5, nesterov = FALSE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAdam\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003elr\u0026thinsp;=\u0026thinsp;0.001, beta_1\u0026thinsp;=\u0026thinsp;0.9, beta_2\u0026thinsp;=\u0026thinsp;0.999\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNadam\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003elr\u0026thinsp;=\u0026thinsp;0.001, beta_1\u0026thinsp;=\u0026thinsp;0.9, beta_2\u0026thinsp;=\u0026thinsp;0.999\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe training of the model entailed a systematic exploration of hyperparameters and components to determine the optimal configuration. During the training process, we carefully monitored the loss and accuracy trends for both the training and validation datasets at each epoch. This rigorous method enabled us to comprehensively evaluate the performance of the models and decide on their efficiency for our goal.\u003c/p\u003e \u003cp\u003eThe process of selecting the best model is depicted in the results and discussion section. To evaluate the potential models, we analysed several performance metrics, including accuracy, loss, and model stability. Accuracy measures the proportion of correct predictions made by the model, while loss indicates the average error for each instance. Generally, lower loss values suggest better model performance. However, it is essential to recognise that a model exhibiting high accuracy but significant loss may be at risk of overfitting, which can adversely affect its ability to generalise to unseen test data. So, we meticulously assessed the possibility of overfitting during our comparative analysis of the four candidate models.\u003c/p\u003e \u003cp\u003eTo guarantee the dependability and reproducibility of our findings, we conducted the training, validation, and testing phases using Python version 3.10.12 on the Google Colaboratory platform. The computational environment features an NVIDIA K80 GPU with 12 GB of memory, and we utilised Keras with TensorFlow as the backend (version 2.15.0), alongside libraries such as OpenCV to conduct the experiments.\u003c/p\u003e \u003c/div\u003e"},{"header":"3 RESULTS AND DISCUSSION","content":"\u003cp\u003eThe results of developing deep learning models for distinguishing between winter wheat and its four prevalent weeds, namely annual bugloss, bastard cabbage, hoary cress, and flixweed, using colour images is presented in this section. These models would be suitable for precision wheat weed management in various scenarios where any of these weeds are present.\u003c/p\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Evaluation of models\u003c/h2\u003e \u003cp\u003eThe first stage in developing an appropriate deep learning model entails the selection of a base network from a variety of standard options via transfer learning, where the initial weights are imported from ImageNet dataset. We assessed four cutting-edge standard deep learning networks: Xception, EfficientNetB0, VGG19, and InceptionResNetV2. Following fine-tuning, these networks underwent a modification phase by incorporating three blocks of CNN layers into the ultimate layer of the fine-tuned standard candidate models to establish four candidate architectures (Model 1 to Model 4) (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). The models were trained using images resized to 299 \u0026times; 299 pixels, with the SGD optimiser, a batch size of 32, and 200 epochs to determine the optimal network. During this process, we employed selection criteria that assessed model stability during training and analysed test accuracy and test loss.\u003c/p\u003e \u003cp\u003eThe results comparing the four candidate models based on maximum and minimum accuracy on the training, validation, and test datasets are presented in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. As observed in this table, Models 1 and 4 achieved 100% accuracy on the training dataset, while Model 3 attained 100% accuracy on the validation dataset. Notably, Model 3 produced the lowest loss values among all models on the training, validation, and test datasets. However, Model 4 achieved the highest accuracy on the test dataset, followed by Models 1 and 3, respectively. Model 2 yielded the lowest test accuracy. The accuracy and loss values of Models 1 and 4 are relatively close, which could be attributed to the presence of the Inception module in both models.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison of the four deep learning models on training, validation, and test datasets.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModel 1\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eModel 2\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eModel 3\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eModel 4\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMaximum accuracy on train dataset\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.0000\u003c/p\u003e \u003cp\u003e(epoch: 161)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.9837\u003c/p\u003e \u003cp\u003e(epoch: 179)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.9810 (epoch: 190)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.0000\u003c/p\u003e \u003cp\u003e(epoch: 179)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMinimum loss on\u003c/p\u003e \u003cp\u003etrain dataset\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.1821\u003c/p\u003e \u003cp\u003e(epoch: 200)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.2169\u003c/p\u003e \u003cp\u003e(epoch: 200)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.9758 (epoch: 200)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3.1642\u003c/p\u003e \u003cp\u003e(epoch: 199)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMaximum accuracy on validation dataset\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.9846\u003c/p\u003e \u003cp\u003e(epoch: 187)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.9077\u003c/p\u003e \u003cp\u003e(epoch: 194)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.0000 (epoch: 161)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.9846\u003c/p\u003e \u003cp\u003e(epoch: 188)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMinimum loss on validation dataset\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.2592\u003c/p\u003e \u003cp\u003e(epoch: 198)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.4446\u003c/p\u003e \u003cp\u003e(epoch: 194)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.9286 (epoch: 200)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3.2194\u003c/p\u003e \u003cp\u003e(epoch:197)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAccuracy on test dataset\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.9725\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8532\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.9358\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.9817\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLoss on test dataset\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.2550\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.4555\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.9868\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e3.1933\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eTo gain deeper insights into the training process of the four candidate models, we analysed the trends of accuracy and loss concerning the number of epochs during the training phase for both the training and validation datasets. The graphical representation of these trends for all models is depicted in Fig.\u0026nbsp;7. Accordingly, Model 2 exhibits slightly more pronounced fluctuations compared to the other three models, indicating a less stable training pattern. In contrast, Models 1, 3, and 4 display a consistent downward trend in losses and a steady increase in accuracies, with only minor variations. This behaviour indicates that Models 1, 3, and 4 exhibit greater resistance to instability and are likely to generalise more effectively to new, unseen data, rendering them more dependable options.\u003c/p\u003e \u003cp\u003eFigure 8 compares the total parameters, trainable parameters, and non-trainable parameters for the four candidate models. This comparison provides valuable insights into the response time of the models. According to this figure, Model 1 and Model 4 exhibit comparable values for these parameters, which can be attributed to their Inception-based architectures. Model 2 offers the minimum number of trainable parameters, making it advantageous in terms of response time for real-time applications. However, it achieves the lowest accuracy and the highest loss on test dataset among all four models. Model 3, on the other hand, provides promising results in terms of accuracy and loss while maintaining a reasonable number of trainable parameters (Fig. 8), making it a compelling choice for our purposes.\u003c/p\u003e\n\u003cp\u003eAfter carefully considering the evaluation criteria, we proposed Model 4 as the optimal choice for our specific purpose. This model utilises a modified and fine-tuned InceptionResNetV2 architecture with the SGD optimiser. Model 4 exhibited minimal fluctuations during the training process, indicating a high level of stability. Moreover, it achieved impressive accuracy and loss values on the test dataset (98.17% accuracy and a loss of 3.19, respectively), along with a reasonably fast response time. These factors suggest that Model 4 would perform exceptionally well in real-world scenarios, making it our preferred selection.\u003c/p\u003e\n\u003cdiv id=\"Sec14\"\u003e\n \u003ch2\u003e3.2 Classification performance\u003c/h2\u003e\n \u003cp\u003eTo evaluate the efficacy of the models in distinguishing the five classes, various parameters were considered. These parameters include True Positive (TP), representing the count of instances correctly identified as positive; True Negative (TN), denoting instances correctly identified as negative; False Positive (FP), indicating instances incorrectly identified as positive; and False Negative (FN), signifying instances incorrectly identified as negative. These parameters are utilised in the development of two classification metrics: the classification report and the confusion matrix. The classification report provides insights into the model\u0026apos;s performance based on precision, recall, and F1-score, as outlined in Equations 2\u0026ndash;4. Precision quantifies the model\u0026apos;s accuracy in predicting positive cases. A high precision value suggests a lower ratio of incorrect to correct predictions, and conversely. Recall quantifies the ratio of accurately predicted positive instances to the total number of actual positive instances, highlighting the model\u0026apos;s capability to correctly identify all positive instances. Furthermore, recall provides information on the model\u0026apos;s capacity to minimise false negatives. The F1-score, calculated as the geometric mean of precision and recall, serves as a metric that balances these two parameters to achieve an optimal trade-off between precision and recall.\u0026nbsp;\u003c/p\u003e\n \u003ctable id=\"Tabb\" border=\"1\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\\(\\:Precision=\\:\\left(\\frac{TP}{TP+FP}\\right)\\)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e(2)\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\\(\\:Recall=\\:\\left(\\frac{TP}{TP+FN}\\right)\\)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e(3)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\\(\\:f1-score=\\:(2\\times\\:\\frac{Precision\\times\\:Recall}{Precision+Recall})\\)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e(4)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cp\u003eThe classification report describing the categorisation of winter wheat, annual bugloss, bastard cabbage, hoary cress, and flixweed classes by colour images through the four models is outlined in Table 5.\u003c/p\u003e\n \u003cdiv\u003e\n \u003ctable id=\"Tab5\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 5\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eThe classification report detailing the recognition performance of the five classes by the four candidate models.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAnnual Bugloss\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eBastard Cabbage\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eFlixweed\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eHoary Cress\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eWinter Wheat\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMicro Avg.\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMacro Avg.\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eWeighted Avg.\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSamples\u003c/p\u003e\n \u003cp\u003eAvg.\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" rowspan=\"4\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eModel 1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eModel 2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.74\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eModel 3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eModel 4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" rowspan=\"4\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eModel 1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eModel 2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eModel 3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eModel 4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" rowspan=\"4\"\u003e\n \u003cp\u003ef1-score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eModel 1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eModel 2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.69\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.86\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.86\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eModel 3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.87\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eModel 4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eSupports\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e23\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e21\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e24\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e21\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e109\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e109\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e109\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e109\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eBased on Table 5, the precision parameter analysis reveals that Model 1 was able to identify the flixweed and winter wheat classes with 100% accuracy. Using Model 2, three classes\u0026mdash;bastard cabbage, flixweed, and hoary cress\u0026mdash;were identified with 100% accuracy. Moreover, Model 3 successfully identified four classes\u0026mdash;annual bugloss, flixweed, hoary cress, and winter wheat\u0026mdash;with 100% accuracy, while Model 4 detected three classes\u0026mdash;flixweed, hoary cress, and winter wheat\u0026mdash;with 100% precision. On the other hand, if the objective is to identify a specific class (e.g., for selective treatment), it is observed that annual bugloss can be accurately identified using Model 3 and bastard cabbage using Model 2, both at 100% accuracy. Furthermore, Models 2, 3, and 4 can be utilised to identify hoary cress with 100% accuracy. It is worth noting that flixweed can be identified accurately with 100% precision by all four models. Conversely, if the goal is to differentiate weed species from winter wheat for herbicidal purposes, Models 1, 3, and 4 can be employed to identify winter wheat with 100% precision. Lastly, based on the aforementioned table, the lowest precision value was related to identifying annual bugloss with Model 2 (0.74). To evaluate the models\u0026apos; capacity to minimise false negatives, we can compare their recall values. The recall values in the table indicate that this parameter achieved 100% for the annual bugloss class using Model 2; for the bastard cabbage class using Model 4; for the flixweed and hoary cress classes using Models 1, 3, and 4; and for the winter wheat class using Models 2 and 3. Conversely, the lowest recall value was observed for Model 2 in relation to the bastard cabbage class (0.52). Furthermore, the F1-score parameter reached 100% for the flixweed class using Models 1, 3, and 4; for the hoary cress class using Models 3 and 4; and for the winter wheat class using Model 3. The lowest F1-score value was observed for the bastard cabbage class using Model 2 (0.69). Consequently, Model 4 demonstrated the best performance with the highest average precision, recall, and F1-score values.\u003c/p\u003e\n \u003cp\u003eFigure 9 presents the confusion matrices for classifying five categories with four candidate models. The rows represent the true classes while the columns show the predicted classes. Through these matrices, we can see that there were only two annual bugloss instances mistakenly classified as bastard cabbage or hoary cress using Model 1. Additionally, Models 3 and 4 misclassified seven and one annual bugloss instances as bastard cabbage respectively, whereas Model 2 accurately identified this class without any errors. For the bastard cabbage class, one instance was misclassified as annual bugloss using Model 1; eight instances were misclassified as annual bugloss using Model 2; and one instance was misclassified as winter wheat also using Model 2. In contrast, Models 3 and 4 accurately classified this class with 100% precision. The flixweed class was accurately classified without error using Models 1, 3, and 4; however, Model 2 misclassified one instance as winter wheat. For the hoary cress class, Models 1, 3, and 4 accurately classified it without error; whereas Model 2 misclassified two instances as annual bugloss and four instances as winter wheat. Finally, for the winter wheat class, Models 1, 2, and 3 accurately classified it without error; whereas Model 4 misclassified only one instance as annual bugloss.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec15\"\u003e\n \u003ch2\u003e3.3 Winter wheat weed management scenarios\u003c/h2\u003e\n \u003cp\u003eIt is essential to emphasise that the findings of this study can be effectively generalised to encompass various scenarios in winter wheat fields with respect to weed appearance. Given that four distinct weed species have been investigated, we can conclude that the proposed model(s) can be utilised as a tool for efficient weed management in 15 different weed appearance scenarios that may arise in winter wheat fields. These scenarios are detailed in Table 6, which accounts for one to four weeds that can coexist alongside winter wheat. Notably, scenario 15 represents the general condition that we have studied thus far in this investigation.\u003c/p\u003e\n \u003cdiv\u003e\n \u003ctable id=\"Tab6\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 6\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eFifteen scenarios for different weed appearance in winter wheat fields.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eScenario number\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eWeed(s) alongside winter wheat\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eNo. of classes\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eScenario 1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFlixweed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eScenario 2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHoary Cress\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eScenario 3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAnnual Bugloss\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eScenario 4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBastard Cabbage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eScenario 5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFlixweed\u0026thinsp;+\u0026thinsp;Hoary Cress\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eScenario 6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAnnual Bugloss\u0026thinsp;+\u0026thinsp;Flixweed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eScenario 7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBastard Cabbage\u0026thinsp;+\u0026thinsp;Flixweed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eScenario 8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAnnual Bugloss\u0026thinsp;+\u0026thinsp;Hoary Cress\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eScenario 9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBastard Cabbage\u0026thinsp;+\u0026thinsp;Hoary Cress\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eScenario 10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAnnual Bugloss\u0026thinsp;+\u0026thinsp;Bastard Cabbage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eScenario 11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAnnual Bugloss\u0026thinsp;+\u0026thinsp;Flixweed\u0026thinsp;+\u0026thinsp;Hoary Cress\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eScenario 12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBastard Cabbage\u0026thinsp;+\u0026thinsp;Flixweed\u0026thinsp;+\u0026thinsp;Hoary Cress\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eScenario 13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAnnual Bugloss\u0026thinsp;+\u0026thinsp;Bastard Cabbage\u0026thinsp;+\u0026thinsp;Flixweed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eScenario 14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAnnual Bugloss\u0026thinsp;+\u0026thinsp;Bastard Cabbage\u0026thinsp;+\u0026thinsp;Hoary Cress\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eScenario 15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAnnual Bugloss\u0026thinsp;+\u0026thinsp;Bastard Cabbage\u0026thinsp;+\u0026thinsp;Flixweed\u0026thinsp;+\u0026thinsp;Hoary Cress\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eFor this analysis, we used confusion matrices (Fig. 9), and for each scenario, we extracted a separate subset that must include winter wheat for each model. This analysis will involve creating 60 subsets of the confusion matrices (15 subsets for each model). The results of this analysis are presented in Table 7 and Fig. 10. Accordingly, it was concluded that Models 1, 3, and 4 are usable with 100% accuracy for scenarios 1, 2, 4, 5, 7, 9, and 12. Scenario 3 can be treated with 100% accuracy using Models 1, 2, and 3. Additionally, for scenario 6, Models 1 and 3 are suggested with 100% accuracy. For scenario 8, Model 3 will perform best with 100% accuracy. For scenario 11, Model 3 showed the best performance (100%). For scenarios 10, 13, and 14, while 100% accuracy was not achieved, Models 1 and 4 provided the highest accuracy of around 97% for scenario 10. The same two models can be used for scenario 13, with an accuracy of about 98%. Model 4 can be employed for scenario 14 with an accuracy of 98%. If we were to choose a comprehensive model for various scenarios, we would consider the average accuracy across different models and scenarios. Accordingly, based on Table 7, Model 4 has the highest average accuracy (98.92%) and the lowest standard deviation (0.0110), making it the most suitable model. Although Model 3 is noteworthy for achieving 100% accuracy in 11 different scenarios. It is important to reiterate that scenario 15 is the initial and overall scenario considered in this research, based on which the deep learning models have been designed.\u003c/p\u003e\n \u003cdiv\u003e\n \u003ctable id=\"Tab7\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 7\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eAverage and standard deviation of accuracies for the four models in all scenarios.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eModel 1\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eModel 2\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eModel 3\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eModel 4\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAverage accuracies in all scenarios\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.9890\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.9191\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.9776\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.9892\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSTD of accuracies in all scenarios\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.0161\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.0504\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.0380\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.0110\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eThe ability to identify distinct weed and crop species from RGB images is a significant challenge, especially with a large number of species. The most promising research seems to focus on a broader range of species, achieving higher accuracy while de-emphasising specific environmental conditions. It is worth noting that a successful weed discrimination model for one crop may not perform adequately for others. However, considering various weed appearance scenarios for distinct crops and proposing models for each scenario could be advantageous, as it would cater to diverse conditions that may arise for those crops. This aspect could be a significant strength of our study compared to similar research. Shao et al., (2023) attempted to distinguish six weed species in paddy fields using deep learning methods under natural field conditions similar to our study but achieved lower accuracy. G C et al., (2024) applied five different deep learning frameworks to differentiate crop and weed species in a greenhouse that provides a roughly similar environment. Despite considering 14 crop and weed species, their accuracy was relatively low compared to our work. In contrast, our study achieved promising accuracy without such pre-processing steps. Finally, Makarian \u0026amp; Saedi, (2024) attained perfect accuracy in natural conditions; however, the small number of studied classes limits the model\u0026apos;s generalisability.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"4. CONCLUSION","content":"\u003cp\u003eIn this study, we developed four improved deep neural network models to differentiate winter wheat from four common weeds. We used colour images as input data to train and test the models. The results obtained from testing the models on unseen data (test data) demonstrated their ability to accurately identify these five plant species in photographs. Our findings show that all four proposed models have promising performance in distinguishing the plant classes, with InceptionResNetV2 exhibiting the highest accuracy. We also investigated fifteen distinct scenarios of weed occurrence in winter wheat fields, taking into account the number of weeds present alongside the wheat crop, to determine the most effective weed management strategy. The models reliably performed across these scenarios, considering winter wheat as a separate class. This approach also enables the targeting of non-wheat plants as weeds, facilitating a selective weed control strategy. Considering the significance of producing herbicide-free wheat and minimising environmental pollution, the findings of this research pave the way for precision weed management and the creation of automated weed removal solutions. This technology will not only reduce herbicide consumption and environmental impact but also contribute to increased wheat yield and quality.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eAcknowledgments\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe authors would like to acknowledge the financial support of Shahrood University of Technology for this research under project No: RP_S_219444.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eS.I.S. contributed to software development, writing, formal analysis, and methodology. H. M. contributed to resources, data curation, conceptualization, and writing original draft. All authors reviewed the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eData availability statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe dataset used in the present study will be available upon reasonable request via [email protected] (S. I. Saedi).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eCompeting Interests Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\u003cp\u003eS.I. Saedi:\u003c/p\u003e\n\u003cp\u003eSoftware, Conceptualization, Writing \u0026ndash; review and editing, Formal analysis, Writing \u0026ndash; original draft, Supervision, Methodology\u003c/p\u003e\n\u003cp\u003eH. Makarian:\u003c/p\u003e\n\u003cp\u003eResources, Data curation, Visualization, Conceptualization, Investigation, Validation\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eChollet, F. (2017). \u003cem\u003eDeep Learning with Python\u003c/em\u003e (1st ed.). Manning Publications Co.\u003c/li\u003e\n\u003cli\u003eFan, X., Chai, X., Zhou, J., \u0026amp; Sun, T. (2023). Deep learning based weed detection and target spraying robot system at seedling stage of cotton field. \u003cem\u003eComputers and Electronics in Agriculture\u003c/em\u003e, \u003cem\u003e214\u003c/em\u003e, 108317. https://doi.org/https://doi.org/10.1016/j.compag.2023.108317\u003c/li\u003e\n\u003cli\u003eGao, J., French, A. P., Pound, M. P., He, Y., Pridmore, T. P., \u0026amp; Pieters, J. G. (2020). Deep convolutional neural networks for image-based Convolvulus sepium detection in sugar beet fields. \u003cem\u003ePlant Methods\u003c/em\u003e, \u003cem\u003e16\u003c/em\u003e(1), 29. https://doi.org/10.1186/s13007-020-00570-z\u003c/li\u003e\n\u003cli\u003eGC, S., Zhang, Y., Howatt, K., Schumacher, L. G., \u0026amp; Sun, X. (2024). Multi-Species Weed and Crop Classification Comparison Using Five Different Deep Learning Network Architectures. \u003cem\u003eJournal of the ASABE\u003c/em\u003e, \u003cem\u003e67\u003c/em\u003e(2), 275\u0026ndash;287. https://doi.org/https://doi.org/10.13031/ja.15590\u003c/li\u003e\n\u003cli\u003eGyawali, A., Bhandari, R., Budhathoki, P., \u0026amp; Bhattarai, S. (2022). \u003cem\u003eA REVIEW ON EFFECT OF WEEDS IN WHEAT (Triticum aestivum L.) AND THEIR MANAGEMENT PRACTICES\u003c/em\u003e. 34\u0026ndash;40. https://doi.org/10.26480/faer.02.2022.34.40\u003c/li\u003e\n\u003cli\u003eJuwono, F. H., Wong, W. K., Verma, S., Shekhawat, N., Lease, B. A., \u0026amp; Apriono, C. (2023). Machine learning for weed\u0026ndash;plant discrimination in agriculture 5.0: An in-depth review. \u003cem\u003eArtificial Intelligence in Agriculture\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e, 13\u0026ndash;25. https://doi.org/https://doi.org/10.1016/j.aiia.2023.09.002\u003c/li\u003e\n\u003cli\u003eKhosravi, H., Saedi, S. I., \u0026amp; Rezaei, M. (2021). Real-time recognition of on-branch olive ripening stages by a deep convolutional neural network. \u003cem\u003eScientia Horticulturae\u003c/em\u003e, \u003cem\u003e287\u003c/em\u003e, 110252. https://doi.org/https://doi.org/10.1016/j.scienta.2021.110252\u003c/li\u003e\n\u003cli\u003eLangridge, P., Alaux, M., Almeida, N., Ammar, K., Baum, M., Bekkaoui, F., Bentley, A., Beres, B., Berger, B., Braun, H.-J., Brown-Guedira, G., Burt, C., Caccamo, M., Cattivelli, L., Charmet, G., Civ\u0026aacute;ň, P., Cloutier, S., Cohan, J.-P., Devaux, P., \u0026amp; Zhang, X. (2022). Meeting the Challenges Facing Wheat Production: The Strategic Research Agenda of the Global Wheat Initiative. \u003cem\u003eAgronomy\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e, 2767. https://doi.org/10.3390/agronomy12112767\u003c/li\u003e\n\u003cli\u003eLi, Z., Wang, D., Yan, Q., Zhao, M., Wu, X., \u0026amp; Liu, X. (2024). Winter wheat weed detection based on deep learning models. \u003cem\u003eComputers and Electronics in Agriculture\u003c/em\u003e, \u003cem\u003e227\u003c/em\u003e, 109448. https://doi.org/https://doi.org/10.1016/j.compag.2024.109448\u003c/li\u003e\n\u003cli\u003eMakarian, H., \u0026amp; Saedi, S. I. (2024). Automated classification of saffron and broadleaf weeds of flixweed and hoary cress using deep learning and color images. \u003cem\u003eCrop Protection\u003c/em\u003e, 106750. https://doi.org/https://doi.org/10.1016/j.cropro.2024.106750\u003c/li\u003e\n\u003cli\u003eMontazeri, M., Zand, E., Baghestani, M. A. (2005). \u003cem\u003eWeeds and their Control in Wheat Fields of Iran, first ed\u003c/em\u003e. Agricultural Research and Education Organization Press.\u003c/li\u003e\n\u003cli\u003eNasiri, A., Omid, M., Taheri-Garavand, A., \u0026amp; Jafari, A. (2022). Deep learning-based precision agriculture through weed recognition in sugar beet fields. \u003cem\u003eSustainable Computing: Informatics and Systems\u003c/em\u003e, \u003cem\u003e35\u003c/em\u003e, 100759. https://doi.org/https://doi.org/10.1016/j.suscom.2022.100759\u003c/li\u003e\n\u003cli\u003eNasiri, A., Taheri-Garavand, A., \u0026amp; Zhang, Y.-D. (2019). Image-based deep learning automated sorting of date fruit. \u003cem\u003ePostharvest Biology and Technology\u003c/em\u003e, \u003cem\u003e153\u003c/em\u003e, 133\u0026ndash;141. https://doi.org/https://doi.org/10.1016/j.postharvbio.2019.04.003\u003c/li\u003e\n\u003cli\u003ePeng, H., Huang, B., Shao, Y., Li, Z., Zhang, C., Chen, Y., \u0026amp; Xiong, J. (2018). General improved SSD model for picking object recognition of multiple fruits in natural environment. \u003cem\u003eNongye Gongcheng Xuebao/Transactions of the Chinese Society of Agricultural Engineering\u003c/em\u003e, \u003cem\u003e34\u003c/em\u003e, 155\u0026ndash;162. https://doi.org/10.11975/j.issn.1002-6819.2018.16.020\u003c/li\u003e\n\u003cli\u003eQuan, L., Feng, H., Lv, Y., Wang, Q., Zhang, C., Liu, J., \u0026amp; Yuan, Z. (2019). Maize seedling detection under different growth stages and complex field environments based on an improved Faster R\u0026ndash;CNN. \u003cem\u003eBiosystems Engineering\u003c/em\u003e, \u003cem\u003e184\u003c/em\u003e, 1\u0026ndash;23. https://doi.org/https://doi.org/10.1016/j.biosystemseng.2019.05.002\u003c/li\u003e\n\u003cli\u003eRai, N., \u0026amp; Sun, X. (2024). WeedVision: A single-stage deep learning architecture to perform weed detection and segmentation using drone-acquired images. \u003cem\u003eComputers and Electronics in Agriculture\u003c/em\u003e, \u003cem\u003e219\u003c/em\u003e, 108792. https://doi.org/https://doi.org/10.1016/j.compag.2024.108792\u003c/li\u003e\n\u003cli\u003eSadeghi, S. H., Saedi, S. I., Peters, R. T., \u0026amp; St\u0026ouml;ckle, C. (2022). Towards improving the global water application uniformity of centre pivots through lateral speed adjustment. \u003cem\u003eBiosystems Engineering\u003c/em\u003e, \u003cem\u003e215\u003c/em\u003e, 215\u0026ndash;227. https://doi.org/https://doi.org/10.1016/j.biosystemseng.2022.01.012\u003c/li\u003e\n\u003cli\u003eSaedi, S I, Alimardani, R., Mousazadeh, H., \u0026amp; Salehi, R. (2019). Development and evaluation of an energy and water efficient intensive cropping system. \u003cem\u003eINMATEH-Agricultural Engineering\u003c/em\u003e, \u003cem\u003e58\u003c/em\u003e, 93\u0026ndash;104.\u003c/li\u003e\n\u003cli\u003eSaedi, Seyed I, \u0026amp; Rezaei, M. (2024). A Modified Xception Deep Learning Model for Automatic Sorting of Olives Based on Ripening Stages. In \u003cem\u003eInventions\u003c/em\u003e (Vol. 9, Issue 1). https://doi.org/10.3390/inventions9010006\u003c/li\u003e\n\u003cli\u003eSaedi, Seyed Iman, \u0026amp; Khosravi, H. (2020). A deep neural network approach towards real-time on-branch fruit recognition for precision horticulture. \u003cem\u003eExpert Systems with Applications\u003c/em\u003e, \u003cem\u003e159\u003c/em\u003e, 113594. https://doi.org/https://doi.org/10.1016/j.eswa.2020.113594\u003c/li\u003e\n\u003cli\u003eSaedi, Seyed Iman, Rezaei, M., \u0026amp; Khosravi, H. (2024). Dual-path lightweight convolutional neural network for automatic sorting of olive fruit based on cultivar and maturity. \u003cem\u003ePostharvest Biology and Technology\u003c/em\u003e, \u003cem\u003e216\u003c/em\u003e, 113054. https://doi.org/https://doi.org/10.1016/j.postharvbio.2024.113054\u003c/li\u003e\n\u003cli\u003eSalim, F., Saeed, F., Basurra, S., Qasem, S. N., \u0026amp; Al-Hadhrami, T. (2023). DenseNet-201 and Xception Pre-Trained Deep Learning Models for Fruit Recognition. In \u003cem\u003eElectronics\u003c/em\u003e (Vol. 12, Issue 14). https://doi.org/10.3390/electronics12143132\u003c/li\u003e\n\u003cli\u003eShao, Y., Guan, X., Xuan, G., Gao, F., Feng, W., Gao, G., Wang, Q., Huang, X., \u0026amp; Li, J. (2023). GTCBS-YOLOv5s: A lightweight model for weed species identification in paddy fields. \u003cem\u003eComputers and Electronics in Agriculture\u003c/em\u003e, \u003cem\u003e215\u003c/em\u003e, 108461. https://doi.org/https://doi.org/10.1016/j.compag.2023.108461\u003c/li\u003e\n\u003cli\u003eShewry, P. R., \u0026amp; Hey, S. J. (2015). The contribution of wheat to human diet and health. \u003cem\u003eFood and Energy Security\u003c/em\u003e, \u003cem\u003e4\u003c/em\u003e(3), 178\u0026ndash;202. https://doi.org/10.1002/fes3.64\u003c/li\u003e\n\u003cli\u003eSingh, M., Kukal, M. S., Irmak, S., \u0026amp; Jhala, A. J. (2021). Water Use Characteristics of Weeds: A Global Review, Best Practices, and Future Directions. In \u003cem\u003eFrontiers in plant science\u003c/em\u003e (Vol. 12, p. 794090). https://doi.org/10.3389/fpls.2021.794090\u003c/li\u003e\n\u003cli\u003eSpeck, O., \u0026amp; Speck, T. (2021). Functional morphology of plants - a key to biomimetic applications. \u003cem\u003eThe New Phytologist\u003c/em\u003e, \u003cem\u003e231\u003c/em\u003e(3), 950\u0026ndash;956. https://doi.org/10.1111/nph.17396\u003c/li\u003e\n\u003cli\u003eSubeesh, A., Bhole, S., Singh, K., Chandel, N. S., Rajwade, Y. A., Rao, K. V. R., Kumar, S. P., \u0026amp; Jat, D. (2022). Deep convolutional neural network models for weed detection in polyhouse grown bell peppers. \u003cem\u003eArtificial Intelligence in Agriculture\u003c/em\u003e, \u003cem\u003e6\u003c/em\u003e, 47\u0026ndash;54. https://doi.org/https://doi.org/10.1016/j.aiia.2022.01.002\u003c/li\u003e\n\u003cli\u003eSusetyarini, E., Wahyono, P., Latifa, R., \u0026amp; Nurrohman, E. (2020). The Identification of Morphological and Anatomical Structures of Pluchea indica. \u003cem\u003eJournal of Physics: Conference Series\u003c/em\u003e, \u003cem\u003e1539\u003c/em\u003e(1), 12001. https://doi.org/10.1088/1742-6596/1539/1/012001\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Deep learning, RGB image, Weed recognition, Winter wheat","lastPublishedDoi":"10.21203/rs.3.rs-6710451/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6710451/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eImproving wheat productivity is essential for maintaining food security in the coming years. A major challenge to wheat production is the presence of weeds, which can greatly reduce its yield. Therefore, precision weed management, utilising site-specific methods, will be vital for achieving sustainable wheat production. To pave the way for this goal, in the present study, we utilised colour images and deep learning techniques to distinguish winter wheat from four prevalent weeds, creating five distinct classes. A total of 542 images were captured in natural field conditions, and resized to 299 \u0026times; 299 pixels. Four deep learning networks \u0026ndash; Xception, EfficientNetB0, VGG19, and InceptionResNetV2 \u0026ndash;were evaluated to serve as base networks to fulfil the classification requirements. These networks were pre-trained on ImageNet using transfer learning, then fine-tuned and enhanced with additional layers to improve performance on our dataset. The improved InceptionResNetV2 model demonstrated the highest performance among the four models, achieving an accuracy of 98.17% and a loss of 3.19 on the test data. Nevertheless, all models exhibited excellent performance in distinguishing plant classes, with F1-scores ranging from 93\u0026ndash;100%, 69\u0026ndash;98%, 82\u0026ndash;100%, and 93\u0026ndash;100% for models based on Xception, EfficientNetB0, VGG19, and InceptionResNetV2, respectively. Additionally, we analysed fifteen scenarios of weed presence in winter wheat fields, focusing on various weed types studied, to propose effective weed management strategies utilising any of the four models. The research findings provide a foundation for precision weed management that not only reduces herbicide usage and environmental impact but also enhances wheat yield and quality.\u003c/p\u003e","manuscriptTitle":"Intelligent Weed Recognition in Winter Wheat Fields through Deep Convolutional Neural Networks","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-17 09:19:53","doi":"10.21203/rs.3.rs-6710451/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"41c19c4d-7d78-4856-b494-769795ff7617","owner":[],"postedDate":"June 17th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":50065506,"name":"Physical sciences/Engineering"},{"id":50065507,"name":"Biological sciences/Plant sciences"}],"tags":[],"updatedAt":"2025-08-08T10:38:12+00:00","versionOfRecord":[],"versionCreatedAt":"2025-06-17 09:19:53","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6710451","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6710451","identity":"rs-6710451","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}