IntroductionA great deal has changed in human civilization. But agriculture is still crucial for every nation. Cotton, sugar, timber, and palm oil are just a few examples of the many raw commodities derived from agriculture. The pharmaceutical, diesel fuel, plastic, and many other key industries rely on these components. One of agriculture’s most crucial roles is providing food for people all around the globe1,2. The UN has projected that by the end of the year 2050, the population will have surpassed 10 billion. Hence, to account for this shift, it is must to quadruple the current food production. The next generation of farmers will face the formidable task of boosting both the amount and the quality of their harvest. Invasion by pests is one of the leading causes of crop damage. They make holes in stems, fruit, and roots to consume, and they can also devour leaves3. A crop can be infected with bacteria, viruses, or fungi by these pests. Farms and agricultural properties must have effective pest management. Chemical spray-like insecticides applied to agricultural areas during the larval stage can indirectly decrease the amount of pesticides required, and regular monitoring of these sites can detect early invasion. Stopping its spread is a top priority. In order to spot it early, farmers must check their maize crop every day. The process of physically identifying pests in the vast fields could take a considerable amount of time. More and better harvests have been achieved through the use of technology over the ages. Drones and intelligent robots will alter the face of agriculture by making farms more efficient. By eliminating the need for constant human oversight of crops, these technological advancements can greatly cut down on farmers’ workloads. Agriculture 5.0 is defined as a revolutionary phase in agriculture that integrates advanced technologies like robotics, extended reality (XR), 6G, AI etc. to improve the efficiency, productivity at the livestock level. Integrating AI and big data can enable the use of multimodal foundation models that can fuse diverse data sources to optimize farming practices. Adoption of Agriculture 5.0 technologies faces challenges related to farmer demographics, costs, data privacy concerns, and the need for behavioural shifts among stakeholders. Policies, training, collaborative approaches, and addressing these socio-technical factors are crucial for faster adoption 4. Based on the survey paper5, pests and disease management are some of the key challenges facing agriculture in the era of Industry 5.0 (I5.0) and remote sensing. Some of the key areas in Agriculture 5.0 are listed as the following:Early Detection of Pests Remote sensing technologies such as thermal imaging, multispectral imaging and hyperspectral imaging can help detect pathogens and plant diseases early before they become serious. This enables farmers to use targeted and timely strategies to control pests and diseases and improve crop health and yields6.Automated Pest Control I5.0 technologies such as robotics and automation enable targeted and precise application of pesticides and herbicides. Vision-guided robots can identify and spray affected areas, reducing waste and environmental impact compared to manual spraying7.Data-Driven Pest Management Advanced data analytics and machine learning can analyze data from remote sensing models and trends related to disease outbreaks. This can predict pests and take steps to prevent or reduce their impact on crops8.Pest classification in smart agriculture refers to the process of identifying and classifying different types of pests that can damage crops as shown in Fig. 1. The goal is to accurately identify and distinguish different types of insects, mites, rodents or other organisms that can be considered agricultural pests. It is important to note that the term ‘mites’ in this work refers specifically to pest mites known to damage crops, and does not include beneficial predatory mites. This treats pests such as ladybugs, mosquitoes, grasshoppers, etc. This involves dividing into taxonomic groups such as. This may include estimating population density or counting the number of pests per person found. Pest classification systems are designed to detect pests early before they affect major crops. Continuous monitoring and distribution allow farmers to intervene in time. Pest analysis data can be combined with other agricultural data (example weather, soil conditions) to enable targeted data-driven pest control strategies. This is based on the goal of precision agriculture and smart agriculture. Compared to pest monitoring by farmers, computerized pest classification and deep learning can improve speed, accuracy and efficiency. Pest classification is an important part of modern agricultural techniques used to measure, identify and monitor pests to maintain crop health and yields. The methods presented in this paper use deep learning tools to provide accurate and precise pest classification. Pest classification is important and necessary in the context of smart agriculture:Crop protection and food security Pest infestations can cause crop loss and serious damage, threatening food and agriculture. Timely and precise pest detection is crucial for effective intervention and management9.Economic impact Pests can have a significant financial impact on farmers and agriculture. Effective protection against pests and diseases helps reduce crop losses and maintain product quality10.Spread prevention The failure to apply pesticides to one limited area will lead to their fast dissemination among healthy regions and cross-border influences. Through pest identification and classification services managers can establish strategies to control and stop pest infestations11,12.Labour-intensive manual monitoring The task of pest identification conducted by farmers through manual methods becomes inefficient for big farms because it requires substantial labor effort. Automated pest classification technology shortens labor requirements by delivering instant identification services13.Precision agriculture and smart farming Extended agriculture challenges can be converted into smart farming objectives through joined efforts between pesticides and modern technologies such as drones and data analytics and robots. The technology fulfills the fundamental goal of smart agriculture to boost yield rates and maximize sustainability alongside operational excellence14.Fig. 1Types of pests.Full size imageModern smart agriculture depends on accurate distribution techniques which drive data-driven methods to achieve food safety combined with economic profitability and environmental safety in agriculture15. Deep learning techniques present the best solution for addressing these matters. The research gap involves the computational burden of using traditional deep learning and transfer learning models for pest classification. Previous approaches have several limitations:Computational Intensity of Transfer Learning Models Prior research heavily relied on pre-trained models which require significant computing resources. For example, InceptionResNetV2 has approximately 54.3 million parameters and requires 379.1 MB of memory.Resource Requirements for Farm Implementation Implementing these computationally intensive models in agricultural settings is challenging, especially in areas with limited technological infrastructure.Inefficient Feature Utilization Previous approaches processed all features from transfer learning models without discriminating between relevant and irrelevant features, creating unnecessary computational overhead.Scalability Issues As the number of pest classes increases (19 classes in this study), the computational burden of traditional methods becomes even more significant.Although deep learning techniques have been applied to pest detection, many models require substantial computational resources and lack generalizability across diverse pest classes. There is a pressing need for a computationally efficient and accurate pest classification system that can operate effectively across varied agricultural environments. The focus of this work is to develop an efficient and accurate pest classification system using deep learning and feature selection techniques to support automated pest management in smart agriculture. Motivated by the limitations of traditional manual inspection and the computational complexity of existing deep learning models, this study aims to enhance pest identification performance while reducing processing overhead. By leveraging transfer learning and a novel LDA-based feature selection method16,17, the proposed approach offers a lightweight yet powerful solution tailored for real-world agricultural applications. Fusion of features taken by diverse models could add advantage for results outcome18,19. The proposed approach addresses these limitations by:Feature Selection Over Model Retraining Instead of retraining or fine-tuning entire pre-trained models (which is computationally intensive), the authors apply LDA to select only the most relevant features.Dramatic Dimensionality Reduction The approach reduces thousands of features (e.g., 4608 from InceptionResNetV2) to just 18 discriminative features, significantly lowering computational requirements.Lightweight Classification Process By using a reduced feature set, the subsequent classification using DNN, LSTM, or Bi-LSTM becomes much more efficient.Improved Performance despite Lower Complexity The feature selection approach not only reduces computational needs but actually improves performance, achieving up to 99.99% accuracy compared to 96.56% with the best transfer learning model.Resource Efficiency for Practical Deployment The lightweight nature of this approach makes it more suitable for implementation in agricultural settings with limited computing resources.This research effectively addresses the gap between high-performance pest classification and the practical constraints of computational resources in agricultural applications, making automated pest classification more accessible and deployable in real-world farming environments. The integration of our lightweight pest classification system with Internet of Things (IoT) devices creates powerful opportunities for real-time agricultural monitoring 20. Smart cameras, sensor networks, and drones equipped with the proposed LDA-based classification model can process pest images locally with minimal computational resources, enabling edge computing that addresses rural bandwidth limitations.This IoT integration offers immediate detection and identification of pest outbreaks, allowing timely interventions before significant crop damage occurs. The system can simultaneously incorporate environmental sensors tracking conditions conducive to pest proliferation, enabling predictive models that forecast potential infestations based on changing environmental factors. The computational efficiency of our LDA-based feature selection method is particularly well-suited for resource-constrained IoT hardware typically deployed in agricultural settings. This addresses a significant barrier to implementing AI-based pest monitoring in farming environments where power, processing capability, and connectivity are limited. As Agriculture 5.0 emerges, such integrated pest monitoring systems represent a crucial component in the smart farming ecosystem, enabling sustainable agriculture through precise, data-driven pest management with minimal human intervention.Literature surveyIn recent times, many deep learning methodologies were implemented to classify pests and produce good results across various pest identification based projects21. The classification of insect invaders utilizing convolutional neural networks (CNN) and saliency techniques yielded approximate 92% accuracy for a reduced dataset21. To classify the defective wheat grains in a dataset of three hundred images, the researchers utilized optimized artificial neural networks, extreme learning machines, also an bee colony algorithm22. The performance of a DL framework for multi-class fruit/’s detection that incorporates images of fruits and augments data using Faster RCNN was assessed23. In addition to video-based performance metrics, a DL-based Faster RCNN was examined for identifying parasites and plant diseases in video/videos content 24. A survey paper was presented that examined recent advancements in digital image processing and enhancement techniques utilized for the computer based detection of leaf based “pests and diseases”25. Adao et al. compiled an image dataset of cotton fields, applied a deep residual design to the data, and utilized the Resnet 34 model to classify the pests with an F1-score of 0.9826. By applying CNN based frameworks incorporating attention mechanisms, feature pyramids, and fine-grained modeling techniques, an accuracy of approximately 73.99% was achieved on the “IP102” dataset27. Chen et al. utilized leaf images to implement a CNN model based on an AlexNet-modified architecture on a mobile application for the purpose of identifying tomato maladies. The achieved precision was 80.3%28. A very effective deep learning system identified pests for ten distinct pest classes with an average accuracy of approx. 69.99% utilizing the Yolov5-S model29. Although the model developed by Chen et al. achieved approximately 69.99% accuracy on ten pest classes, its generalizability remains a concern. The relatively modest accuracy suggests potential overfitting to a specific dataset or limitations in handling diverse pest images with high intra-class variability. For broader deployment in smart agriculture, models must demonstrate strong performance across heterogeneous environments, lighting conditions, and pest types. This highlights the need for approaches like ours, which aim to maximize accuracy while maintaining robustness and efficiency across diverse data. A comparative analysis of single-shot detector, KNN, SVM, Multilayer Perceptron, and Faster R-CNN classifiers was conducted to differentiate between Trialeurodes Vaporariorum embryo and Bemisia Tabacii embryo tomato insect classes30. For insect classification, K. Thenmozhi utilized three varieties of the dataset for forty classes and twenty-four classes, respectively—achieving an accuracy of 97% percent, 98%, and 96% with transfer learning and pre-trained DL techniques31. Nour et al. applied data augmentation to the AlexNet model in order to improve its accuracy for an eight-class insect pest on an IP102 dataset32. Detection of pests using a faster RCNN ResNet50 model, with eight-class insects obtaining an average accuracy of 93.99%33.Table 1 provides a details of key literature, including the research process employed, classification accuracy achieved, the number of pest classes considered, and the corresponding references. . After doing the literature survey and analysing the advantages and disadvantages of existing techniques, in this work, two Pests datasets were combined including 9 and 12 classes respectively. The combined dataset contains total 19 classes. In this work EfficientNetB3 is proven to give best results by giving 96.56% Accuracy, 98.90% Precision, 92.82% Recall and 41% Loss. Further, a novel feature selection technique was developed LDA which when applied on Deep Learning models giving better results than that of transfer learning models. The results obtained by this technique are 99.99% Accuracy, 100% validation, 99.99% Recall and 0% Loss. The main advantage of feature selection is that it makes the classification process lighter because it involves selecting relevant features from the existing dataset without training additional models, whereas transfer learning typically involves retraining or fine-tuning pre-existing models, which can be more computationally intensive.Table 1 Percentage accuracy given by different techniques for classification.Full size tableThe structure of this work includes an introduction to the problem, a review of related research, the proposed methodology, experimental results, and a concluding discussion. Each section builds upon the previous to present a complete deep learning-based pest classification framework.Methodological frameworkThe classification of pests into distinct categories plays a crucial role in agriculture and pest management. With advancements in computer vision, deep learning techniques have proven effective for image-based pattern recognition and are now increasingly applied in agricultural settings. This methodology outlines a feature selection-based deep learning framework for pest classification.Research designThis work presents a computer vision-based framework for classifying pests in agricultural environments. The proposed system leverages LDA for feature selection, facilitating both high classification performance and reduced computational complexity. The proposed method involves a two-stage process. First, transfer learning is used to extract deep features from pest images using (DenseNet201, EfficientNetB3, and InceptionResNetV2), which were chosen for their proven performance on image datasets such as ImageNet. Each model extracts a fixed number of features (1920, 1536, and 4608 respectively). LDA is then applied separately to each extracted feature set to reduce dimensionality while maximizing class separability. From each model, 18 top features are selected. Additionally, the features from all three models are concatenated to form a combined feature set (total 8064 features), and LDA is again applied to select 18 most discriminative features and then selected features trained on DNN, LSTM, and Bi-LSTM. Hyperparameters such as number of epochs, learning rate, and batch size were empirically optimized for each model through cross-validation to ensure robustness and fairness in comparison. This methodology ensures that both model-specific and combined feature sets are fully utilized and tuned for best performance on the combined pest dataset. Figure 2 illustrates the core steps involved in the research design.Fig. 2Research design of the proposed work.Full size imageData collectionThe Data in this work is a combination of two different Pests Datasets. The first dataset is The Agricultural Pest Image Dataset35 called as A which is a collection of images of 12 different types of agricultural pests, namely “Ants”, “Bees”, “Beetles”, “Caterpillars”, “Earthworms”, “Earwigs”, “Grasshoppers”, “Moths”, “Slugs”, “Snails”, “Wasps”, and “Weevils”. Although earwigs can act as beneficial insects by preying on smaller pests like aphids, in the context of this study they are considered pests due to their potential to damage soft fruits, flowers, and seedlings, particularly when their population density is high. There are 12 categories of pests; the data provide different images covering different shapes, colors and sizes; This makes it suitable in many cases for training and testing methods for pests identification and classification. The images in the dataset represent the real world and are not artificial. In addition, images are scaled down to a maximum width or height of 300 pixels, making the document lighter and easier to use. The second data file is the Pest database36 called as B. The dataset contains 9 types of pests including “aphids”, “armyworm”, “beetle”, “bollworm”, “grasshopper”, “mites”, “mosquito”, “sawfly”, “stem borer”. Some organisms such as ants and earthworms may not be harmful in all contexts—and are sometimes beneficial to soil health—they are included in the dataset as pests where they are known to negatively impact crop health. Similarly, mosquitoes, although primarily human health pests are considered here due to their potential to disrupt agricultural ecosystems. Dataset in this work is now generated by combining two datasets and deleting duplicate entries can be approached using mathematical set operations by combining both the datasets A and B to form a new dataset using (1)$${\text{C}} = {\text{A}} \cup {\text{B}}$$(1)Here, ∪ denotes the union operation, which merges the elements of both datasets while removing duplicates.After merging, and removing duplicates, the set difference operation is used as:$${\text{C}}_{{{\text{unique}}}} = {\text{C}} - \left\{ {\text{duplicates in C}} \right\}$$(2)Cunique now contains the combined datasets with duplicate entries removed. The overall descriptions of both the datasets are as shown in Table 2.Table 2 Overview of the datasets used in the study, detailing their sources, features, and key characteristics.Full size tableWorkflow of the implementationThe implementation of the pest classification system using feature selection involves a series of well-defined steps as shown in Fig. 3. The system implementation proceeds as follows:(1)Data Integration and Preprocessing:Combine datasets A and BResize and clean imagesEnsure class balance and apply 80/20 training/testing split(2)Feature Extraction via Transfer Learning:Use pre-trained CNNs: DenseNet201, EfficientNetB3, InceptionResNetV2Each model extracts features from image data:DenseNet201: 1920 featuresEfficientNetB3: 1536 featuresInceptionResNetV2: 4608 features(3)Dimensionality Reduction with LDA:Apply LDA individually to each model’s featuresReduce to 18 discriminative features per model (for 19 classes)Also apply LDA to the concatenated feature set (total 8064 features)(4)Classification:Use selected features to train using three classifiers:DNNLSTMBiLSTMEvaluate models using accuracy, precision, and recall(5)Performance Comparison:Fig. 3The process or workflow diagram of the proposed technique.Full size imageTechnologies appliedThe technologies incorporated into this work are described below.Pretrained model-based feature extractionThe study discussed in the document employs three pre-trained CNN models37 for transfer learning: DenseNet201, EfficientNetB3, and InceptionResNetV2. These models were initially trained on the ImageNet dataset and are known for their ability to extract rich, hierarchical visual features from input images38,39,40. The concept of transfer learning in machine learning involves repurposing a model that has been trained on one task for a different but related task. This approach is particularly beneficial when dealing with small datasets or limited computational resources, as it allows for efficient utilization of pre-existing knowledge. Once a suitable pre-trained model is chosen, feature extraction is performed. In this approach, the parameters of the pre-trained model are frozen, and it is used to extract features from fed dataset. These extracted features are then fed into a new classifier or task-specific layers, which are trained from scratch on the new data. Another approach to transfer learning is fine-tuning, where the pre-trained model’s architecture is adapted to better fit the new dataset’s structure. In this method, certain components of the pre-trained model are removed, and the entire model is then trained using the new data. Fine-tuning helps to eliminate weaknesses in the pre-trained model and allows it to better adapt to the specific characteristics of the new dataset, typically resulting in improved performance, especially with larger datasets. Table 3 presents crucial information about the feature extractors and ImageNet weights utilized in the study for three pre-trained CNN models: DenseNet201, InceptionResNetV2, and EfficientNetB3. These models have been pre-trained on the ImageNet dataset, a large-scale dataset for image classification, and are employed in the study for transfer learning and feature extraction tasks. For the DenseNet201 model, the input shape expected by the model is (224, 224, 3), which corresponds to images with a height and width of 224 pixels and 3 color channels (RGB). The model extracts 1920 features from the Global Average Pooling 2D layer, which is a common technique for summarizing spatial information in CNNs. DenseNet201 has a substantial number of parameters, approximately 18.3 million, and requires around 327.5 MB of memory to store its weights and parameters. The InceptionResNetV2 model expects an input shape of (229, 229, 3), slightly larger than DenseNet201. It extracts 1536 features from the Global Average Pooling 2D layer, similar to DenseNet201. However, InceptionResNetV2 has a significantly higher number of parameters, around 54.3 million, which can contribute to its ability to capture more complex features. Consequently, it requires a larger memory footprint of approximately 379.1 MB to store its weights and parameters. The EfficientNetB3 model, part of the EfficientNet family of CNNs, expects an input shape of (300, 300, 3), which is larger than the other two models. It extracts 1536 features from the Dropout layer, which is a regularization technique commonly used in neural networks. EfficientNetB3 has a relatively smaller number of parameters, around 10.8 million, compared to the other two models, but still requires a substantial amount of memory, approximately 361.9 MB, to store its weights and parameters. The table highlights the trade-offs between model complexity, feature extraction capabilities, and memory requirements for the CNN models. While DenseNet201 and InceptionResNetV2 extract a similar number of features, InceptionResNetV2 has a significantly higher number of parameters, likely enabling it to capture more intricate features but at the cost of increased memory requirements. EfficientNetB3, on the other hand, strikes a balance between model complexity and memory usage while still maintaining a competitive feature extraction capability.Table 3 Feature extractors with pretrained weights utilized in current study.Full size tableLDAAs a prominent statistical tool LDA41 enables researchers to discover superior linear features that successfully distinguish between different data categories. The weight vector w obtained through LDA allows for the computation of linear combinations which both discover significant features while minimizing the number of dimensions. LDA functions effectively in practical applications although it needs some restrictive assumptions to function properly. The technique works with Gaussian data class distributions even though this assumption does not represent actual reality because it delivers acceptable results in practical uses. The decision boundaries from LDA depend on linear feature relationships and need equal covariance structures between all classes. LDA maintains widespread adoption because of its implementation efficiency and successful performance in class separation even though it might not match actual data patterns perfectly. Our system utilizes LDA to extract features which generates productive performance enhancements based on examination of imperfect data conditions. The research team recommends conducting future analyses which compare LDA with alternative techniques comprising Principal Component Analysis (PCA) paired with t-SNE. The feature extraction method of LDA implements the following process:LDA identifies a linear combination function as f(x)$${\text{f}}\left( {\text{x}} \right) = {\text{ w}}^{{\text{T}}} {\text{x}}$$(3)This function maximizes the ratio between class mean separation (μ0 − μ1) and within-class covariance Σ. The optimal weight vector is:$${\text{w}}^{*} = {\text{c}}\Sigma^{ - 1} \left( {\mu_{0} - { }\mu_{1} } \right)$$(4)where c represents an arbitrary constant. Equations (3) and (4) represent the standard formulation of LDA as described in41The magnitude of weights in w indicates feature importance, establishing a natural ranking of variables most critical for class discrimination.LDA assumes multivariate normal distribution within each class and shared covariance matrix Σ across all classes, enabling calculation of category averages and the combined variance–covariance structure derived from the learning dataset.The calculated coefficient array w* establishes a transformation that converts p-dimensional observations into single-dimensional classification values, successfully compressing the feature space.These discriminant scores enable classification of new observations based on threshold comparison.For K distinct classes, LDA extracts (K − 1) features.LDA was selected over other dimensionality reduction techniques like PCA primarily because of its supervised nature, which makes it more suitable for classification tasks42. Unlike PCA, which focuses on maximizing variance without considering class labels, LDA specifically maximizes the separation between different pest classes while minimizing within-class variance. For the pest classification problem, this class-aware approach provides several key advantages. LDA creates projections that explicitly maximize the distance between the means of different pest classes relative to the within-class variance, directly supporting the classification objective, whereas PCA might preserve variance that isn’t relevant for distinguishing between pest types. Additionally, LDA extracts precisely K − 1 features (where K is the number of classes), which for the 19-class pest dataset meant 18 features-significantly reducing the dimensionality from thousands of features extracted by the transfer learning models. The features selected by LDA are specifically those that best differentiate between the pest classes, focusing computational resources on the most classification-relevant information. LDA also naturally extends to multi-class problems like the 19-class pest dataset, providing an elegant mathematical framework for finding the optimal subspace for classification. By focusing only on class-discriminative features, LDA helps reduce the risk of overfitting that might occur when using the full feature set from transfer learning models.Deep learning techniquesDeep Neural Networks (DNN) are driving advancements in diverse fields such as computer vision, natural language processing, and speech recognition. As these technologies evolve, they may pave the way for even more groundbreaking innovations, potentially surpassing current capabilities in areas like genetic engineering and cloning, unlocking new possibilities for science and technology. As they develop machine learning and AI, they can analyze complex fixed-pattern data. A DNN classifier can be added to LDA feature sets. High-learning DNNs recognize pest species-target type feature set relationships. RNNs understand long dependences with the help of Long Short-Term Memory (LSTM) and resist sequential data errors. Instead of memorizing everything, LSTM memory cells alter their values and selectively recall input information. For uncertain LSTM architecture, error flow can be continuous, keeping information over lengthy time sequences. LSTMs excel at context-overlong sequence problems like language modeling, translation, and speech recognition. LDA and feature classification by RNN LSTM classifies insect infestations. Long-term interdependence in sequential data is stressed in LSTM-based models, necessitating specific pest data pattern interpretation skills. Sequence modeling uses deep learning-specific RNN architecture BiLSTM. The second development is that this network analyzes input sequences like two-way communication to boost LSTM function. Most natural language processing tasks requiring past and future context use BiLSTMs.Results analysis and evaluationThis section describes the step-by-step experimental phases conducted to evaluate the effectiveness of the proposed pest classification framework. The experiments were designed to assess different combinations of feature extraction, dimensionality reduction, and classification techniques to determine the optimal configuration.Dataset preparationTwo datasets, referred to as A and B, were merged to form a comprehensive dataset comprising 19 distinct pest classes. To ensure class balance and prevent bias during model training, each class was standardized to contain 800 images. The images underwent input size resizing after preprocessing as per the specifications of the CNN networks for subsequent evaluation. Subsequently the dataset was partitioned into training (80%) and testing (20%) subsets for performance evaluation consistency.Feature extractionThe analysis used DenseNet201, EfficientNetB3 along with InceptionResNetV2 as pre-trained CNNs to conduct feature extraction. All images yielded feature outputs of 1920 by DenseNet201 while EfficientNetB3 generated 1536 features and InceptionResNetv2 generated 4608 features.The application of LDA methods operated on individual features from separate models resulting in decreased dimensions to 18 features for each model. All features obtained from the three models were combined into an 8064-dimensional vector before being reduced through LDA to 18 features.Three deep learning classifiers—DNN, LSTM, BiLSTM were trained and evaluated to assess their performance. The experiments were carried out using two types of input features: LDA features extracted from each individual CNN and LDA features derived from the concatenation of features obtained from multiple CNNs.Performance metricsEvaluation was conducted based on key performance metrics including accuracy, precision, and recall. Various classifiers and feature types were compared to determine the most effective configuration for the task.Final selectionThe best-performing combination (CNN+LDA+Classifier) was selected based on experimental results to be used as the final pest classification system. For taking the results, system configurations are present in Table 4.Table 4 configuration of the system used in this work.Full size tableEvaluation of baseline methods of transfer learningPest classification is applied using Transfer Learning models including DenseNet201, EfficientNetB3, and InceptionResNetV2. Figure 4: Baseline model performance shows (a) Training Accuracy, (b) Validation Accuracy, (c) Training Loss, (d) Validation Loss, (e) Training Precision, (f) Validation Precision, (g) Training Recall, (h) Validation Recall. Figure 4a,b shows the accuracy and validation accuracy performance comparisons, with InceptionResNetV2 consistently outperforming the other models across all epochs.Fig. 4The detailed comparison of baseline transfer learning techniques for pests classification.Full size imageThe validation accuracy values increase with the number of training epochs/iterations for all three models. The InceptionRestNetV2 model consistently outperforms the others in validation accuracy across all epochs. Figure 4c shows the comparison of the loss values during training for DenseNet201, EfficientNetB3, and InceptionRestNetV2. As expected, the loss values decrease as the number of training epochs/iterations increases for all models, with InceptionRestNetV2 achieving the lowest loss values, indicating it is the best-performing model in minimizing the loss function. Figure 4d shows the comparison of validation loss values during training. The validation loss values generally decrease for all three models, with EfficientNetB3 consistently achieving the lowest validation loss values, making it the best at minimizing the validation loss. Figure 4e compares the precision metric for the models. Precision values increase as the number of epochs/iterations rises, with InceptionRestNetV2 consistently achieving the highest precision values, indicating it is the best-performing model in terms of precision. Similarly, Fig. 4f compares validation precision values, where EfficientNetB3 performs the best in terms of validation precision, although InceptionRestNetV2 shows higher validation precision than DenseNet201 in most epochs. The difference in validation precision values becomes smaller toward the end of the training, suggesting convergence in performance. Figure 4g compares the recall metric for the models. Recall values increase with training epochs/iterations for all models, with InceptionRestNetV2 consistently achieving the highest recall values, indicating it is the best model for recall. Figure 4h compares validation recall values. EfficientNetB3 performs the best in terms of validation recall, while InceptionRestNetV2 outperforms DenseNet201 in most epochs. The difference in validation recall values becomes smaller toward the end of the training, suggesting the models are converging to similar performance levels.Table 5 shows the results of applying three different transfer learning models—DenseNet201, EfficientNetB3, and InceptionResNetV2—on the combined pest dataset with 19 classes. As shown in Table, EfficientNetB3 outperformed the other two models, achieving the highest accuracy of 96.56%, along with strong precision (98.90%) and recall (92.82%) metrics. All three models demonstrated good performance, with accuracy ranging from 94.86% (DenseNet201) to 96.56% (EfficientNetB3). The validation metrics (validation accuracy, validation precision, validation recall) were also high, indicating the models generalized well to the validation dataset. The loss values were relatively low, with EfficientNetB3 having the lowest training loss of 0.41 and validation loss of 0.53. Overall, the results show that transfer learning approaches, especially the EfficientNetB3 model, were able to effectively classify the diverse 19-class pest dataset with high accuracy and performance metrics. While these transfer learning models performed well, they can still be computationally heavy, leading them to explore a more lightweight feature selection approach as described in the later sections of the paper.Table 5 Evaluation of baseline methods of transfer learning for pest classification.Full size tableInvestigation of feature selection using pre-trained algorithms and DenseNet for classificationThis section presents a two-stage approach where deep features extracted from pre-trained models are optimized using LDA to enhance classification performance.To prevent overfitting in the proposed pest classification models, several strategies were employed. Data augmentation was applied to increase dataset diversity. Early stopping was used during training by monitoring validation loss to halt training when performance plateaued. Dropout layers were incorporated in the DNN and LSTM models to reduce reliance on specific neurons, and batch normalization was used to stabilize training. Additionally, the use of LDA for feature selection reduced the dimensionality of input data, eliminating noise and focusing only on the most relevant features. These combined techniques contributed to improved generalization and robust model performance.Evaluation of DNN on extracted featuresThis subsection analyzes the performance of a DNN applied to LDA has selected features from the features attained by the applied pre-trained models.Figure 5 shows DNN model performance comparison with (a) Training Accuracy, (b) Validation Accuracy, (c) Training Loss, (d) Validation Loss, (e) Training Precision, (f) Validation Precision, (g) Training Recall, (h) Validation Recall. The data includes various performance metrics like accuracy, loss, precision, recall, and their validation counterparts over 50 epochs of training. All the models achieve very high accuracy (> 90%) and low validation loss by the end of training, indicating effective learning on the dataset. As shown in Figure, DenseNet classifier on MultiFeature selection method appears to be the best performing, reaching near-perfect accuracy and validation metrics by around epoch 10–15. The EfficientNetB3 feature selection model also performs very well, with high accuracy and reasonable validation metrics. The DenseNet201 and InceptionResNetV2 feature selection methods seem to lag behind the other two in terms of final performance. There are some fluctuations in the validation metrics during training, which is common and can be mitigated through techniques like early stopping. Overall, this data suggests that the Multi-Features feature selection with DenseNet classifier and EfficientNetB3 feature selection DenseNet classifiers were effective at learning the patterns in the dataset.Fig. 5Evaluation of DNN performance has been analyzed on the extracted features.Full size imageEvaluation of LSTM performance on extracted featuresThis section analyzes the performance of LSTM classifiers applied to LDA-selected features extracted from pre-trained deep learning models for pest classification.Figure 6 shows LSTM model performance with (a) Training Accuracy, (b) Validation Accuracy, (c) Training Loss, (d) Validation Loss, (e) Training Precision, (f) Validation Precision, (g) Training Recall, (h) Validation Recall. The metrics include accuracy, loss, precision, recall, validation accuracy, validation loss, validation precision, and validation recall for each model across 50 training epochs. The Multi-Features feature selection with LSTM classifier achieves the highest training and validation accuracy, reaching over 99% accuracy by the later epochs. This suggests it is the best performing model among the four. The DenseNet201 feature selection method with LSTM classifier, EfficientNetB3 feature selection with LSTM classifier, and InceptionResNetV2 feature selection with LSTM classifier, all achieve respectable training and validation performance, with validation accuracies in the range of 80–90% by the end of training. The precision and recall metrics provide additional insights into the models’ performance. For example, the Multi-Features feature selection with LSTM classifier has very high precision and recall, indicating it is able to accurately identify both positive and negative examples. Overall, the data suggests that this combination is the strongest performer among the four, though the other models also show promising results.Fig. 6LSTM classification performance evaluation on extracted features.Full size imageEvaluation of BiLSTM performance on extracted featuresThe detailed analysis for the effectiveness of BiLSTM classifiers on LDA-refined features extracted from pre-trained models has been discussed in this section.Figure 7 shows BiLSTM model results with (a) Training Accuracy, (b) Validation Accuracy, (c) Training Loss, (d) Validation Loss, (e) Training Precision, (f) Validation Precision, (g) Training Recall, (h) Validation Recall. The data covers 50 training epochs for each of the four models. The performance of the models varies across the different epochs. For example, the Multi-Features feature selection method with BiLSTM classifier achieves the highest accuracy and precision, while the InceptionResNetV2 feature selection with BiLSTM classifier, DenseNet201 feature selection with BiLSTM classifier, and EfficientNetB3 feature selection with BiLSTM classifier have lower, but still reasonably high, performance metrics.Fig. 7Analysis of BiLSTM on extracted features.Full size imageTable 6 presents the performance of various LDA-based feature selection and classification techniques on the pest dataset. The Multi-Features_BiLSTM model achieved the highest results with 99.99% accuracy, 100% validation accuracy, 99.99% precision and recall, and 0% loss, outperforming all other methods. The DNN and LSTM classifiers applied on the LDA-selected features also performed well, though not as impressively as the BiLSTM models. The key advantage of the LDA-based feature selection approach is that it is more lightweight compared to the computationally heavy transfer learning models, as it involves selecting a smaller set of relevant features rather than retraining or fine-tuning entire models.Table 6 Performance comparison of LDA-based feature selection with different deep learning classifiers on the pest dataset * Here feature selection methods are DenseNet201–DN201, EfficientNetB3–ENB3, InceptionResNetV2–INV2 and Multi-features–MF.Full size tableIn summary, this approach showcases the potential for developing effective and computationally efficient computer vision-based pest monitoring and management systems.Performance comparison analysis: accuracy and computation timeThis section discusses the comparison between baseline transfer learning models and our feature selection-based deep learning models concerning their classifying efficiency and processing capabilities. A system’s practical implementation as a pest classification tool depends heavily on efficiency due to resource limitations in particular environments. Our proposed approaches require dramatically shorter runtimes than the baseline transfer learning models for computing tasks according to the time analysis. Base transfer learning models required nearly three and a half hours to train InceptionResNetV2 and EfficientNetB3 and DenseNet201 at 13,931.17 s (3.87 h) and 12,262.54 s (3.41 h) respectively. The proposed approaches for feature selection performed the computations substantially faster than the baseline transfer learning models as shown in Table 7. When integrating EfficientNetB3 model with DNN classifier and applying the new feature selection process it cut training time by 92.7% down to 900.43 s (about 15 min) while achieving better classification performance.Table 7 Comparison of computation time between baseline models and proposed approaches.Full size tableThe Multi-Features with BiLSTM classifier consumed 3213.91 s (about 54 min) for completion which represented a 76.9% decrease in computation time when compared to the baseline InceptionResNetV2 model. Our proposed methodology provides an improved classification performance while needing dramatically less computation time during processing. Our proposed feature selection methods exist in an optimal section of the accuracy-time trade-off plot because they perform highly accurate predictions efficiently. The Multi-Features with BiLSTM classifier stands out as the optimal model because it reaches an exceptional accuracy level of 99.99% while using only 3213.91 s of computation time. Experimental outcomes show that the deep learning model with feature selection achieves both better classification accuracy than baseline transfer learning methods and decreased computational durations. Our approach provides optimal performance combined with efficient computations which makes it appropriate for real-world agricultural pest classification deployments inharmonious with operational and resource-based needs.Limitations and challengesThe combination of LDA-based feature selection using deep learning classifiers particularly BiLSTM delivers better performance than traditional transfer learning approaches. The proposed approach overcomes challenges faced by EfficientNetB3 and InceptionResNetV2 which demand high memory capacity and retraining while it chooses select few relevant features to enhance accuracy up to 99.99%. LDA feature selection provides a fast training and inference process that enables the deployment of this model within agricultural fields. There are multiple advantages to the proposed approach yet its implementation might come with certain restrictions. Using LDA method presents a limitation because it requires linear class separation when working with certain types of pest image data that exhibit non-linear relationships. Classifiers show performance variability because the extracted features obtained from pre-trained models vary in quality. Introducing features from different models together results in performance degradation when such integration is not handled with care. The system should be evaluated for its performance in different real-world environments with respect to lighting conditions and background clutter along with object occlusion because generalizability matters. The proposed model presents a hopeful approach for large-scale quick pest classification functionality by finding the right balance between performance efficiency results.ConclusionThe creation of automated computer vision-based systems that identify agricultural insect pests stands as a vital operation to benefit agricultural economies together with the sustainability of farmer livelihoods. A properly developed system helps create better methods for controlling pests and diseases in agricultural fields. A lightweight insect pest classification system for agriculture served as the main proposition within this research. A combined dataset of 19 agricultural pest classes emerged from the authors who merged two pest classifications. They initially applied transfer learning models with EfficientNetB3 giving the best results at 96.56% accuracy. However, the authors note that transfer learning models can be computationally heavy, so they proposed a novel feature selection technique using LDA. Applying LDA-based feature selection and then using DNN, LSTM, and Bi-LSTM classifiers, yielded even better results, with up to 99.99% accuracy. The proposed LDA-based feature selection approach, combined with BiLSTM, achieved exceptional classification accuracy of 99.99% while significantly reducing computational complexity. This demonstrates the model’s potential for real-time, scalable pest detection systems that support sustainable and precision agriculture. The main advantage of the feature selection approach is that it makes the classification process more lightweight compared to retraining or fine-tuning entire transfer learning models. In conclusion, this work presents an effective and efficient computer vision-based framework for classifying a wide range of agricultural pests, which can be valuable for sustainable crop management in farming communities. The novel feature selection technique demonstrates the potential for improving pest classification performance while reducing computational complexity.The potential of deep learning in pest classification requires constant research investment to achieve its maximum utility through collaborative work because this development would revolutionize agriculture and strengthen food safety while reducing pesticide evaluation’s environmental impacts. Moving forward this research will examine how to develop the LDA-based classification system to process streaming video data coming from actual field-deployed IoT devices for continuous pest activity monitoring. Further analysis of ensemble approaches for joining multiple feature selection methods would boost robustness while preserving numerical computing speed.Data availabilityThe dataset utilized during the current study is publicly available at “https://www.kaggle.com/datasets/vencerlanz09/agricultural-pests-image-dataset”.ReferencesKaur, H. J. & Himansh, H. The role of internet of things in agriculture. In: Proceedings–International Conference on Smart Electronics and Communication, ICOSEC 2020, 667–675 (2020).Anand, A., Trivedi, N. K., Gautam, V., Tiwari, R. G., Witarsyah, D. & Misra, A. Applications of Internet of Things(IoT) in agriculture: The need and implementation. In: Proceedings–International Conference Advancement in Data Science, E-Learning and Information Systems, ICADEIS 2022, 1–5 (2022).Upadhyay, N. & Gupta, N. A survey on diseases detection for agriculture crops using artificial intelligence. In: 2021 5th International Conference on Information Systems and Computer Networks (ISCON), 1–8 (2021).Fountas, S. et al. Agriculture 5.0: Cutting-edge technologies, trends, and challenges. IT Prof. 26, 40–47 (2024).Google Scholar Victor, N. et al. Remote sensing for agriculture in the era of industry 5.0 & #x2013; a survey. IEEE J. Sel. Top Appl. Earth Obs. Remote Sens. 17, 5920–5945 (2024).ADS Google Scholar Bütüner, A. K., Şahin, Y. S., Erdinç, A., Erdoğan, H. & Lewis, E. Enhancing pest detection: Assessing Tuta absoluta (Lepidoptera: Gelechiidae) damage intensity in field images through advanced machine learning. Tarim Bilimleri Dergisi 30, 99–107 (2024).Google Scholar Kohlberg, A. B., Myers, C. R. & Figueroa, L. L. From buzzes to bytes: A systematic review of automated bioacoustics models used to detect, classify and monitor insects. J. Appl. Ecol. 61, 1–13 (2024).Google Scholar Godara, S. et al. Data-driven insights for agricultural extension services in Rajasthan: A study of Kisan call center queries. Indian J. Ext. Educ. 60, 53–58 (2024).Google Scholar Poudel, A. Interplay of plant pathogens and host defenses: Unveiling the mechanisms and strategies for crop protection. Arch. Agric. Environ. Sci. 9, 93–101. https://doi.org/10.26832/24566632.2024.0901014 (2024).Article Google Scholar Tiwari, A. K. Insect pests in agriculture identifying and overcoming challenges through IPM. Arch. Curr. Res. Int. 24, 124–130 (2024).Google Scholar Mourya, P. K., Singh, J. K., Chaudhary, P., Chaudhary, A. K. & Upadhayay, V. Role of drone technology in insect pest management. 5, 104–107 (2024).Upadhyay, N. & Gupta, N. Potato leaves disease detection with data augmentation using deep learning approach. In Information and Communication Technology for Competitive Strategies (ICTCS 2022) (eds Joshi, A. et al.) 589–599 (Springer Nature, Singapore, 2023).Google Scholar Kakarla, U. S., Venkataramana, A., Gopinath, S., Shivaram, K. & Yugendhar, K. Revolutionizing Application Agricultural Harvesting with IoT Introduction, 01014:1–10 (2024).Hashmi, A. U. H. et al. Effects of IoT communication protocols for precision agriculture in outdoor environments. IEEE Access 12, 46410–46421 (2024).Google Scholar Upadhyay, N. & Gupta, N. Diagnosis of fungi affected apple crop disease using improved ResNeXt deep learning model. Multimed. Tools Appl. 83, 64879–64898 (2024).Google Scholar Alahmadi, A., Hussain, M. & Aboalsamh, H. LDA-CNN: Linear discriminant analysis convolution neural network for periocular recognition in the wild. Mathematics https://doi.org/10.3390/math10234604 (2022).Article Google Scholar Belattar, K., Mostefai, S. & Draa, A. A hybrid GA-LDA scheme for feature selection in content-based image retrieval. Int. J. Appl. Metaheuristic Comput. (IJAMC) 9, 48–71 (2018).Google Scholar Chen, J., Pan, S., Peng, W. & Xu, W. Bilinear spatiotemporal fusion network: An efficient approach for traffic flow prediction. Neural Netw. 187, 107382 (2025).PubMed Google Scholar Chen, J., Ye, H., Ying, Z., Sun, Y. & Xu, W. Dynamic trend fusion module for traffic flow prediction. Appl. Soft Comput. 174, 112979 (2025).Google Scholar Luo, J., Zhao, C., Chen, Q. & Li, G. Using deep belief network to construct the agricultural information system based on Internet of Things. J. Supercomput. 78, 379–405 (2022).Google Scholar Nanni, L., Maguolo, G. & Pancino, F. Insect pest image detection and recognition based on bio-inspired methods. Ecol. Inform. https://doi.org/10.1016/j.ecoinf.2020.101089 (2020).Article Google Scholar Sabanci, K. Detection of sunn pest-damaged wheat grains using artificial bee colony optimization-based artificial intelligence techniques. J. Sci. Food Agric. 100, 817–824 (2020).CAS PubMed Google Scholar Wan, S. & Goudos, S. Faster R-CNN for multi-class fruit detection using a robotic vision system. Comput. Netw. 168, 107036 (2020).Google Scholar Li, D. et al. A recognition method for rice plant diseases and pests video detection based on deep convolutional neural network. Sensors (Switzerland) https://doi.org/10.3390/s20030578 (2020).Article PubMed Central Google Scholar Ngugi, L. C., Abelwahab, M. & Abo-Zahhad, M. Recent advances in image processing techniques for automated leaf pest and disease recognition–a review. Inf. Process. Agric. 8, 27–51 (2021).Google Scholar Alves, A. N., Souza, W. S. R. & Borges, D. L. Cotton pests classification in field-based images using deep residual networks. Comput. Electron. Agric. 174, 105488 (2020).Google Scholar Ung, H. T., Ung, H. Q., Nguyen, T. T. & Nguyen, B. T. An efficient insect pest classification using multiple convolutional neural network based models. Front. Artif. Intell. Appl. 355, 584–595 (2022).Google Scholar Chen, H. C. et al. AlexNet convolutional neural network for disease detection and classification of tomato leaf. Electronics (Switzerland) 11, 1–18 (2022).Google Scholar Doan, T. N. An efficient system for real-time mobile smart device-based insect detection. Int. J. Adv. Comput. Sci. Appl. 13, 30–36 (2022).Google Scholar Gutierrez, A. et al. A benchmarking of learning strategies for pest detection and identification on tomato plants for autonomous scouting robots using internal databases. J. Sens. https://doi.org/10.1155/2019/5219471 (2019).Article Google Scholar Thenmozhi, K. & Srinivasulu Reddy, U. Crop pest classification based on deep convolutional neural network and transfer learning. Comput. Electron. Agric. 164, 104906 (2019).Google Scholar Khalifa, N. E. M., Loey, M. & Taha, M. H. N. Insect pests recognition based on deep transfer learning models. J. Theor. Appl. Inf. Technol. 98, 60–68 (2020).Google Scholar Ramalingam, B. et al. Remote insects trap monitoring system using deep learning framework and IoT. Sensors (Switzerland) 20, 1–17 (2020).Google Scholar Nanni, L., Manfè, A., Maguolo, G., Lumini, A. & Brahnam, S. High performing ensemble of convolutional neural networks for insect pest image detection. Ecol. Inform. https://doi.org/10.1016/j.ecoinf.2021.101515 (2022).Article Google Scholar Agricultural Pests Image Dataset. https://www.kaggle.com/datasets/vencerlanz09/agricultural-pests-image-dataset. Accessed 15 Apr 2024.Pest Dataset. https://www.kaggle.com/datasets/simranvolunesia/pest-dataset. Accessed 15 Apr 2024.Shaha, M. & Pawar, M. Transfer learning for image classification. In: Proceedings of the 2nd International Conference on Electronics, Communication and Aerospace Technology, ICECA 2018, 656–660 (2018).Zhao, C., Shuai, R., Ma, L., Liu, W. & Di Hu, Wu. M. Dermoscopy image classification based on StyleGAN and DenseNet201. IEEE Access 9, 8659–8679 (2021).Google Scholar Batool, A. & Byun, Y. C. Lightweight EfficientNetB3 model based on depthwise separable convolutions for enhancing classification of leukemia white blood cell images. IEEE Access 11, 37203–37215 (2023).Google Scholar Ferreira, C. A., Melo, T., Sousa, P., Meyer, M. I., Shakibapour, E., Costa, P. & Campilho, A. Classification of breast cancer histology images through transfer learning using a pre-trained inception Resnet V2. In: Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 10882 LNCS, 763–770 (2018).Graf, R., Zeldovich, M. & Friedrich, S. Comparing linear discriminant analysis and supervised learning algorithms for binary classification—a method comparison study. Biom. J. 66, 1–12 (2024).MathSciNet Google Scholar Jimenez-Narvaez, A. D. et al. Predictive modeling for fetal health: A comparative study of PCA, LDA and KPCA for dimensionality reduction. IEEE Access 13, 59687–59703 (2025).Google Scholar Download referencesAcknowledgementsThe authors would like to thank Deanship of Scientific Research at Majmaah University for supporting this work under Project Number R-2025-1951.Author informationAuthors and AffiliationsChitkara University Institute of Engineering and Technology, Chitkara University, Rajpura, Punjab, IndiaVikas Khullar & Isha KansalDepartment of Computer Science and Engineering, Faculty of Science, Engineering and Technology, University of Science and Technology Chittagong (USTC), Chittagong, BangladeshShyama Barna BhattacharjeeDepartment of Allied Science (Mathematics), Faculty of Science, Engineering and Technology (FSET), University of Science and Technology Chittagong (USTC), Chittagong, BangladeshZarin TasneemDepartment of Computer Science and Engineering, School of Engineering and Technology, Central University of Haryana, Mahendergarh, Haryana, IndiaNitin GoyalDepartment of Computer Science, College of Computer and Information Sciences, Majmaah University, Al Majmaah, Saudi ArabiaShirina SamreenDepartment of Electronics and Communication Engineering, Central University of Jammu, Samba, Jammu, (UT of J&K), 181143, IndiaSachin Kumar GuptaAmity School of Engineering & Technology, Amity University, Gurgaon, Haryana, IndiaShubham MahajanAuthorsVikas KhullarView author publicationsSearch author on:PubMed Google ScholarIsha KansalView author publicationsSearch author on:PubMed Google ScholarShyama Barna BhattacharjeeView author publicationsSearch author on:PubMed Google ScholarZarin TasneemView author publicationsSearch author on:PubMed Google ScholarNitin GoyalView author publicationsSearch author on:PubMed Google ScholarShirina SamreenView author publicationsSearch author on:PubMed Google ScholarSachin Kumar GuptaView author publicationsSearch author on:PubMed Google ScholarShubham MahajanView author publicationsSearch author on:PubMed Google ScholarContributionsAll authors have contributed equally.Corresponding authorsCorrespondence to Shirina Samreen or Shubham Mahajan.Ethics declarationsCompeting interestsThe authors declare no competing interests.Additional informationPublisher’s noteSpringer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.Rights and permissionsOpen Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.Reprints and permissionsAbout this article