The paper considered method workflow from PRISMA guidelines (Page et al., 2021) for insect detection and monitoring in agriculture based on NNs by investigating articles published between 2015 and 2022. This review article aims to provide an overview of the new trends and advancements in CNN research for insect pest detection in agriculture between 2015 and 2022. To select the papers for this review, the focus was primarily on papers that contribute to the development of CNN-based systems for insect pest detection in agriculture. Specifically, papers that propose novel CNN architectures, explore the use of transfer learning for insect pest detection, or apply CNNs to new insect pest detection tasks were prioritized. The selected papers demonstrate the power of CNNs in various applications for insect monitoring in modern agriculture, including object detection, segmentation, and recognition.

The research databases used in this review were: Web of Science, Scopus, and IEEE. Following the Prisma flow diagram ( Figure 2 ), several criteria were attached for searching and extracting articles of interest. Although there was an initially large number of papers identified for the topic of this review, the initial selection criteria extracted approximately 354 relevant studies in the first instance. Of all these, only 138 were chosen based on the final criteria related to new periods, new trends, attachment in top publications, and innovation. An initially large number of diverse research for the modern agricultural area and a considerable evolution in recent years are observed.

Searches for important terms and evolution as article numbers during the last years in the Web of Science, Scopus, and IEEE Xplore DBs between 2015 and 2021 with AND connector are presented in Figure 3 : A) (CNN) AND (agriculture) AND (image processing), B) (CNN) AND (agriculture) AND (insects), C) (CNN) AND (agriculture) AND (pest detection), D) (image processing) AND (pest detection), E) (CNN) AND (pest detection), and F) (CNN) AND (insects). The graphs highlight the strong increase in the number of research articles in the connected fields in recent years regarding the use of CNN.

A robust image database (DB) is crucial for DL classification and detection because it is the foundation upon which a model is trained (Ding & Taylor, 2016). The larger and more diverse the dataset is, the better the ML model’s performance will be. A robust image dataset allows a DL model to learn a general representation of the objects or classes it is supposed to recognize. The more diverse the dataset, the better the model will be at recognizing new images that it has not seen before. Also, it enables an ML model to achieve higher accuracy in classification and detection tasks. When the dataset is comprehensive and covers a wide range of scenarios, the model can learn more accurately how to identify objects and classify them.

Insect pest databases were used in agricultural monitoring applications to track and identify the presence of insect pests that can damage crops (Turkoglu et al., 2022). These databases are typically created by agricultural organizations, universities, and research institutions that have collected data on the life cycles, presence, behavior, and distribution of various insect pests. In this regard, analyzing the papers selected for this study, several ways to construct datasets in training and validating models used for insect detection and identification were observed. Several public databases have been used by researchers in their studies to measure the performance of the implemented architectures and to test the defined models against the obtained results. Table 1 presents a summary of the most known and frequently used databases for modern insect pest monitoring applications in agriculture.

Insect pest image databases often include images of insects at different life stages, including larvae and adult stages (Zhang S. et al., 2022). These images could be accompanied by additional information, such as the insect’s common name, scientific name, and the types of crops or plants that the insect pest is known to damage. This is an important aspect because the primary purpose of an insect image database is to provide a visual reference for identifying insect pests in the field. Insect pest image databases can be used as educational resources to help people learn about the different types of insect pests and their impact on agriculture and the environment (Shi et al., 2020).

One of the databases that is highlighted in the present study and that was used by the researchers in the selected papers is the IP102 DB. As presented in the acronym, it contains 102 classes of common insect pests with hierarchical taxonomy and broadly totals around 72,222 images (see Table 1 ). The database is regularly updated and maintained by a team of experts in the field of entomology. It covers a wide range of insect orders. Each entry in the IP102 Insect Database includes information on the insect’s scientific name, common name, description, habitat, diet, life cycle, behavior, and distribution, all being presented in high-quality images and illustrations, making it easy to identify different species.

The authors on IP102 DB note that existing image datasets primarily focus on everyday objects like flowers and dogs, limiting the applicability of advanced deep learning techniques in agriculture. To address this gap, they introduce a comprehensive dataset called IP102 for insect pest recognition. The authors conducted baseline experiments on the IP102 dataset using both handcrafted and deep feature-based classification methods. Their findings revealed that the dataset poses challenges related to inter-class and intra-class variance, as well as data imbalance. They anticipate that IP102 will serve as a valuable resource for future research in practical insect pest control, fine-grained visual classification, and addressing imbalanced learning challenges in this domain.

The Maryland Biodiversity Database (MBD) (Maryland Biodiversity Database, 2022) is another important database, and it has been used in various research works for the insect pest monitoring area. This database is a vast and valuable public resource that can serve as an important tool in researching various information for the insect pest area and querying it can create diverse datasets. The MBD database provides ecological information about species, including their habitats and interactions with other organisms. This can be useful for understanding the ecological context of insect pests, their host plants, and their natural predators. Its strengths lie in providing detailed species records and distribution data, facilitating ecological context for organisms, and supporting research on insect pests and native species. Researchers and conservationists benefit from its wealth of information to assess biodiversity impact and pest behavior. While not specialized in pest monitoring, MBD enhances pest management by offering a broader understanding of local ecosystems. This collaborative database stands as a crucial asset in safeguarding Maryland’s natural heritage and aiding scientific research. Scientists studying insect pests or conducting research on entomology can use the MBD to access data on insect species’ distributions and occurrences. Pest management strategies often require a comprehensive understanding of the local ecosystem. MBD can provide context by offering information on the diversity of species that may interact with or be affected by insect pests.

AgriPest introduces a domain-specific benchmark dataset for tiny wild pest detection in agriculture. This dataset contains over 49.7K images and 264.7K annotated pests, making it the largest of its kind. It aims to enhance the application of deep learning in agriculture by providing standardized data for pest detection research. AgriPest also defines sub-datasets, including challenges like pest detection and population counting, and validation subsets for various real-world scenarios. The authors build practical pest monitoring systems based on deep learning detectors and evaluate their performance using AgriPest. This dataset and associated code will be publicly available, facilitating further research in pest detection and precision agriculture.

Crafted to serve as a robust resource for training deep learning models in pest detection, Pest24 is another important DB which offers a vast repository of meticulously annotated images of agricultural pests. This paper addresses the challenges of real-time pest population monitoring in precision agriculture using AI technology. It introduces a large-scale standardized dataset called Pest24, comprising 25,378 annotated images of agricultural pests collected from automatic pest traps and imaging devices. The dataset covers 24 categories of common pests in China. On the other hand, the study applies various advanced deep learning detection methods, such as Faster RCNN, SSD, YOLOv3, and Cascade R-CNN, to detect these pests and achieves promising results for real-time field crop pest monitoring. The authors aim to advance accurate multi-pest monitoring in precision agriculture and provide a valuable object detection benchmark for the machine vision community.

The analysis of Pest24 highlights three key factors influencing pest detection accuracy: relative scale, number of instances, and object adhesion. Due to the scarcity of multi-target pest image big data, Pest24 holds great importance as a resource for advancing intelligent field crop pest monitoring. Characterized by its large-scale data, small relative object scales, high object similarity, and dense distribution, Pest24 presents unique challenges for deep learning-based object detection methods and is poised to drive progress in pest detection for precision agriculture while serving as a specialized benchmark for the computer vision community. Beyond its application in precision agriculture, Pest24 serves as an invaluable benchmark for the machine vision community, fostering advancements in specialized object detection. Future work aims to expand the dataset with more diverse multi-pest images from various practical.

Xie1 and Xie2 databases were other important resources in creating databases or testing and training the architectures defined in various works ( Table 1 ). Because these datasets are not large some authors have often resorted to augmentation techniques to increase the size of these datasets. The Xie2 dataset also called D0 contains 40 classes of insect pests represented in 4508 RGB images of 200 x 200px resolution.

Although there are several public databases illustrating and grouping various common classes of insect pests, most of the authors used their own datasets in solving the problems specific (Bhoi et al., 2021; Rajeena et al., 2022). Creating proprietary databases for insect pest detection or monitoring using NNs can help improve the accuracy and specificity of pest detection systems, while also providing flexibility and cost-effectiveness (Segalla et al., 2020; Hong et al., 2021). From the point of view of flexibility, creating its own database offers absolute control of the data that is attached to train the NNs for insect pest monitoring. This is about how the data set can be adjusted as needed to meet the changing need for insect pest families and environmental changes that may occur rapidly. Complete control of the specificity of pest populations was discussed in several works to describe the specificity zone (Khanramaki et al., 2021). By creating proprietary databases, specific insect pests can be tailored and described regarding each context and interest in pest recognition and monitoring. This can help ensure that the NNs are able to accurately identify and differentiate between the specific insect pests, rather than simply providing a general detection of any insect in the image (Liu and Wang, 2020; Xu et al., 2022).

It is very important that the data set describes a real context to solve real problems with increased accuracy. What was observed in this regard as part of the present study in relation to the performances obtained by the authors in various works was a tendency to create robust datasets in increasing performances. The larger the database used for training and validating NNs, the higher the accuracy of the created models can be (Knyshov et al., 2021; Liu et al., 2022). The database used is determined by the precise study objectives and the sort of data required. Researchers interested in insect pest recognition, for example, may pick IP102 or Pest24, but those needing ecological context may prefer MBD. AgriPest is appropriate for precision agricultural research. Collectively, these databases help to advance pest detection and agricultural research. When paired with these different datasets, CNNs provide a very effective tool for insect pest study and control. They can help to increase pest detection accuracy, understand pest behavior in ecological contexts, and improve real-time monitoring and control tactics in precision agriculture. Researchers and practitioners may use these datasets to create more effective and efficient pest-related solutions in agriculture.

In another scenario, from a cost point of view, creating own database can be a cost-effective alternative. Test and training data creation solutions can capture data using low-cost methods like phones or digital cameras, which is a pretty good starting point. Where the data set is acquired using drones, high-fidelity cameras, robots, or specialized human resources, the cost of acquiring and creating the reference data set for pest monitoring can increase commensurately with the size and quality of data acquired (Xing et al., 2019; Tian H. et al., 2020; Genaev et al., 2022).

The organization of the data set represents another aspect noted by the authors in the development of models for harmful insect and pest detection in modern agriculture. In general, for training and evaluation using CNNs for pest detection and identification, the dataset division commonly includes training and validation sets or training, validation, and testing sets. The most common ratio observed in the last split was 70% for training, 20% for validation, and 10% for testing (Huang et al., 2022). The other ratio could include 80% for training and 20% for testing (Du et al., 2022; Zhang S. et al., 2022), or 70% with 30% respectively (Ahmad et al., 2022).

Regarding the dataset, the authors also followed techniques like data augmentation (Du et al., 2022; Zhang et al., 2023). Data augmentation in the context of CNNs is the process of producing additional training examples by applying various changes to existing pictures in the training dataset (Albanese et al., 2021). Geometric changes such as random rotation, horizontal and vertical flips, random cropping, and transformations such as brightness modifications or color jitter are examples of frequent transformations used for data augmentation in CNNs (Padmanabhuni and Gera, 2022). Adding random cropping can assist the model in learning to distinguish things that are not centered in the image (Genaev et al., 2022).

For data augmentation, some of the new trends include synthetic data generation to increase the number of samples if the number of representatives of a class is insufficient (Abbas et al., 2021). Using generative models to create synthetic images is one novel method of data augmentation. Augmentation through synthetic data generation is a novel technique of generating new training data using computer algorithms rather than gathering real-world data (Huang et al., 2022). The purpose of this method is to enhance the quantity and variety of the dataset, which can improve the performance of ML models (Divyanth et al., 2022). Synthetic data generation could address issues such as imbalanced datasets, lack of data privacy, and limited data availability (Lu et al., 2019). For the topic of agricultural pests, this can be done in a variety of ways. There are several methods for creating synthetic data for CNNs (Karam et al., 2022), including generative adversarial networks (GANs), deep learning picture synthesis, data augmentation, and data interpolation (Padmanabhuni and Gera, 2022). Conditional GAN was used by (Abbas et al., 2021) to generate synthetic images for tomato pests and to improve the performances. Another performance improvement was noted by (Divyanth et al., 2022) by creating an artificially generated dataset using GAN.

For the testing phase, the acquisition of digital images from real contexts can be noted. This was pursued by the authors to test the NN architectures they created and optimized against the real contexts, using pest images in the field (Brunelli et al., 2020). For modern agriculture, there are some ways of acquiring digital images, using various systems and techniques (Terentev et al., 2022). The present study identified four important directions that describe image acquisition vectors and were grouped and described in Table 2 . Based on the analyzed references, the performances obtained using the created databases were also noted. In this sense, satisfactory results are observed, and at the same time, it is important to note that the methods of image acquisition are done in an optimized framework and represent a strong point attached to the research areas in this field. For the modern agricultural area, the acquisition of data for the creation of models and automatic solutions in pest monitoring represents an extensive process that can include several resources (Nanni et al., 2022).

Table 2 summarizes the most common data gathering methods, with UAVs and pheromone traps emerging as the most popular options. This section will evaluate the benefits and drawbacks of different techniques. The integration of ML and deep learning DL for automated data processing, with a special focus on remote sensing and sensory data for complete area mapping, is an emerging research field. It is worth mentioning that remote sensing, as investigated by (Stefas et al., 2016; Ahmad et al., 2021), has several applications in fields such as agriculture and forestry.

Unmanned Aerial Vehicles (UAVs) are gaining remarkable traction across diverse domains, with agriculture and environmental monitoring being prominent beneficiaries. One of their vital applications lies in the realm of pest detection and management within agricultural crops Mu et al., 2018. UAVs offer versatile data acquisition methods, including high-resolution imagery and sensory capabilities (Tian H. et al., 2020; Cochero et al., 2022). Equipped with high-resolution cameras, UAVs excel at capturing images and videos of crops, facilitating the identification of insect pests (Tian H. et al., 2020; Cochero et al., 2022). Subsequently, these images can undergo automated pest detection using ML algorithms (Preti et al., 2021). Moreover, UAVs can be equipped with sensors for detecting specific chemicals in the air or on plant surfaces, thus enabling pest identification, as well as treatment efficacy monitoring (Velusamy et al., 2022). To combat identified pests, certain UAVs are equipped with precision sprayers, targeting affected areas with minimal chemical usage and environmental impact (Iost Filho et al., 2019; Li C. et al., 2022). Thermal cameras mounted on UAVs provide valuable temperature data, aiding in pinpointing stressed or pest-infested crop areas due to temperature differences (Yuan & Choi, 2021). UAVs also use multispectral cameras, such as infrared and hyperspectral imaging, in addition to typical RGB images, which considerably improves the accuracy of pest detection models (Terentev et al., 2022). Another current technique employs lidar sensors to collect high-resolution 3D pictures of agricultural fields, allowing for the identification of pest-infested areas (Dong et al., 2018; López-Granados et al., 2019). Lidar imaging also provides information about crop dimensions, growth patterns, and prospective yield (Johansen et al, 2018; Ampatzidis et al., 2020).

Nonetheless, there are several drawbacks to the UAV-based strategy. UAVs, in general, have limited payload capacity, restricting their ability to carry large amounts of equipment and sensors. Furthermore, UAV flight durations are limited, often ranging from 20 to 30 minutes depending on the type and payload. As a result, covering large regions may demand numerous flights, which can be both time-consuming and costly (Dong et al., 2020). While UAVs excel in collecting high-resolution photographs of crops and insect pests, image analysis algorithms’ accuracy may be limited, necessitating professional analysis. Furthermore, the use of UAVs for data collecting is vulnerable to weather and legal limitations. These variables might limit the capacity to collect insect pest data during certain seasons or geographical locations. Many nations have tight UAV laws that include flying limitations as well as criteria for permissible equipment and sensors (Csillik et al., 2018).

Another way of data acquisition for insect monitoring in modern agriculture is based on pheromone traps ( Table 1 ). Data acquisition using pheromone traps is a useful tool for monitoring and controlling insect pests in agriculture and forestry. Pheromone traps are placed in strategic locations throughout a crop. The number and placement depend on the type of insect pest being targeted and the size of the area being monitored. Pheromone traps need to be checked regularly to ensure that they are working properly. Digital cameras can be attached to the pheromone traps to capture images of the trapped insects. The traps should be monitored regularly, typically every 1-2 weeks. During each monitoring visit, the traps are checked for trapped insects, and the digital cameras are checked to ensure they are functioning correctly.

While pheromone traps can be an effective tool for monitoring and managing insect pests, there are some disadvantages to their use. They are only effective against insect pests that are attracted to specific pheromones (Cardim Ferreira Lima et al., 2020). On the other hand, their effectiveness is limited to the area in which they are placed (Toscano-Miranda et al., 2022). Pheromone traps can attract not only the target insect pest but also non-target species that are attracted to the same pheromones. Additionally, pheromone traps can give an incomplete or inaccurate representation of the population of insect pests. This is because some individuals of the pest species may not respond to the pheromone lure or may be located outside the trapping area. This can lead to incorrect decisions about pest management strategies.

In agriculture, insect pest identification and monitoring are key parts of precision farming because they have a direct influence on crop production, quality, and overall agricultural sustainability. CNNs, in conjunction with specialist databases, provide significant promise for tackling the issues connected with insect pest control. Deep learning architectures, object identification, categorization and taxonomy, and real-time monitoring may be essential elements of neural networks applied to these datasets, as demonstrated in the selected research. With the use of CNNs and specialized datasets, insect pest identification and monitoring have entered a new age. This synergy has the potential to lead to more accurate, timely, and environmentally conscious pest management solutions. Continuous research, multidisciplinary cooperation, and an emphasis on practical application are required to fully achieve this promise. As technology advances, the future of insect pest identification and monitoring in agriculture remains bright, with the potential to greatly contribute to global food security and sustainable agriculture practices.

Automated monitoring systems use sensors and cameras to detect and identify insect pests (Amorim et al., 2019). These systems can be connected to the internet, allowing farmers to receive real-time information about pest populations. There are many solutions and methodologies based on image processing, DL, and NNs. CNNs are particularly well suited for tasks involving the detection of small objects, such as insects, within an image. In this scenario, CNNs are a powerful tool for pest detection and have been shown to achieve high accuracy in many applications. One key advantage of CNNs for pest detection is their ability to handle complex images. For example, a CNN can be trained to detect pests in images that contain multiple objects, different backgrounds, and varying lighting conditions (Fang et al., 2020). Additionally, CNNs can be trained on a large dataset of images, which can help improve the accuracy of the model. Another advantage of CNNs for monitoring crops for pest detection is their real-time ability (Ayan et al., 2020). On the other hand, one of the most significant advancements in this field is the development of transfer learning, where a pre-trained CNN model is fine-tuned on a smaller dataset of pest images. Some of the most used NNs for insect and pest detection and classification are presented in Figure 4 .

VGG Net (Visual Geometry Group) (Simonyan & Zisserman, 2014) is a key architecture used in insect detection and monitoring, especially VGG-16 and VGG-19. The architecture is widely used in computer vision applications such as object detection and image segmentation (Popescu et al., 2022). The architecture for VGG-16 (Ramadhan & Baykara, 2022) is shown in Figure 4A , and it was the most used for insect detection and classification tasks. The convolutional layers are responsible for extracting features from the input image, while the pooling layers reduce the spatial dimensions of the feature maps to reduce computation time. The fully connected layers are used to classify the features extracted by the convolutional and pooling layers. The most used, VGG-16 model has a total of 16 layers and the VGG-19 has 19 layers being a modified version of VGG-16 with the addition of the new three convolutional layers.

Residual Network (ResNet) is another CNN family used in insect monitoring for modern agriculture. ResNet-18, ResNet-34, ResNet-50, ResNet-101, and ResNet-152 are variants of the CNN architecture that was introduced in 2015 by researchers at Microsoft (He et al., 2016). The key innovation of ResNet is the use of “residual connections,” or shortcut connections, that allow the network to learn identity mapping and make it easier to train very deep networks. This is shown in Figure 4B as a residual block example. According to the investigated papers, ResNet-50 was the most used for insect detection and classification tasks, and the basic architecture is shown in Figure 4C .

The R-CNN (Region-based CNN) architecture is a type of object detection model that uses a combination of CNNs and region proposal algorithms to detect objects within an image (Ren et al., 2015) and was also used for insect monitoring. It is a two-stage process that first generates a set of region proposals and then uses a CNN to classify and refine the proposals. The first stage of the R-CNN architecture is the region proposal algorithm, which generates a set of regions or “proposals” that may contain an object of interest. These regions are then passed to the second stage of the R-CNN architecture, which is the CNN. This is used to classify and refine the regions generated by the region proposal algorithm and it is done by extracting features from each region and passing them through a series of convolutional and fully connected layers. In this context, another architecture often used for insect detection tasks was Faster R-CNN. This is a type of object detection model that uses a CNN to extract features from an image and then uses a region proposal network (RPN) to propose regions that may contain objects. The feature extractor typically uses a pre-trained CNN, such as VGG or ResNet, to extract features from the input image. The main advantage of Faster R-CNN over other object detection models is its efficiency, as it shares computation between the RPN and the classifier. The Faster R-CNN architecture, adapted from (Ren et al., 2015) is presented in Figure 4E .

The Inception CNN architecture (Szegedy et al., 2015) is also representative of insect classification and detection. This deep CNN architecture utilizes a combination of convolutional, pooling, and inception modules to efficiently learn hierarchical representations of visual data. The novel aspect is that it includes a series of components named Inception modules that apply a combination of convolutional and pooling layers at different scales, allowing the network to efficiently capture and learn both the high-level and low-level features of the image. This review highlighted that the most used architecture from this family, for insect detection and classification was the InceptionV3 (Szegedy et al., 2016). Following the structure and features presented previously, the basic scheme of Inception V3 can be viewed in Figure 4F (adapted from Szegedy et al., 2016).

Dense Convolutional Network (DenseNet) is another CNN family used for insect detection and classification. This neural network architecture is characterized by dense layers Huang et al., 2020. Each layer is connected to every other layer in the network (Huang et al., 2017). This creates a dense network of connections, which allows for a more efficient flow of information and a greater capacity for learning. A dense block is shown in Figure 4D , adapted from (Huang et al., 2017). One of the main advantages of DenseNet architecture is its ability to effectively handle large amounts of data and complex patterns (Huang et al., 2017).

YOLO (You Only Look Once) is another state-of-the-art family that is widely used in the modern agricultural sector for real-time insect detection and monitoring. YOLO (Redmon et al., 2016) is an object detection algorithm that uses a single stage to perform object detection. Unlike other object detection algorithms that rely on region proposals, YOLO uses a grid of cells to divide the image into smaller regions and predicts the object class and location for each cell. The algorithm is trained on large datasets, such as the COCO (Common Objects in Context) or ImageNet, and has been designed to be fast and accurate. Different variants from this family were used: YOLOv2 (Redmon & Farhadi, 2017), YOLOv3 (Redmon & Farhadi, 2018), YOLOv4 (Bochkovskiy et al., 2020), YOLOv5s, YOLOv5m, and YOLOv5l (Ultralytics, 2020).

A synthetic presentation of NNs used for insect and pest detection and classification in agricultural applications is given in Table 3 . Based on the information from Table 3 , the graph in Figure 5 describes the evolution over the last three years of the most used neural networks for insect monitoring in modern agriculture.

This study primarily centers its focus on exploring emerging trends in CNNs for insect pest detection and monitoring through the innovative application of new combinations, while also acknowledging classic CNN models as reference points. This approach aligns with the prevalent practice in the field, where most studies strive to strike a balance between pioneering CNN architectures and established, foundational models. This dual perspective, embracing innovation while respecting tradition, mirrors a common practice observed in the contemporary studies within the deep learning community. Researchers understand that leveraging the strengths of both new and classic CNN models can yield comprehensive insights and solutions, ultimately driving the field forward.

Regarding key trends and advancements, CNNs continued to be a popular choice for image-based insect pest detection. Researchers were developing and fine-tuning CNN architectures to achieve higher accuracy in recognizing and classifying pests from images. Transfer learning techniques were becoming increasingly important in this domain. Researchers were pre-training CNN models on large datasets and then fine-tuning them for insect pest detection tasks. This approach helped in achieving better results even with limited labeled data for specific pests. For object detection and localization, object detection models like Faster R-CNN, YOLO and SSD were adapted for insect pest monitoring. These models not only classified pests but also provided bounding box coordinates, which is crucial for precise pest localization. As a new trend, researchers were experimenting with advanced data augmentation techniques to improve model robustness. Techniques like GANs were used to create synthetic pest images to augment the training dataset. Next, focusing on network architectures, capsule networks, which aim to address the limitations of traditional CNNs in handling hierarchical features, have been explored for insect pest recognition (Xu et al., 2022; Zhang S. et al., 2022). They can capture the spatial hierarchies of pest body parts for improved classification. Some researchers have proposed hybrid architectures that combine the strengths of CNNs for image processing and recurrent neural networks (RNNs) for sequential data processing. This is particularly useful when tracking pests’ movements over time (Butera et al., 2021; Alsanea et al., 2022; Du et al., 2022). To make pest detection systems more transparent and interpretable, explainable AI in architectures techniques have been integrated into neural network architectures. This allows users to understand why a particular pest detection decision was made. Researchers often choose or design architectures based on the unique characteristics and challenges of the pests they are targeting and the monitoring environment. Advancements in neural network architectures for insect pest detection and monitoring are ongoing, so staying up to date with the latest research papers and developments in the field is essential for the most current insights.

Looking at the area of impact and innovation, the new trends stand out with high-performance indices in relation to the area of pest identification. Attaching these was done to create a comparison area. Since the research was based on deep learning models, the indicators most used as evaluation methods of these models were highlighted as part of this study, being represented by accuracy, precision, sensitivity, specificity, F1 score, Jaccard index, mean average precision (mAP), and sometimes R2. Names and calculation formulas are attached in Table 4 . The most used performance indicators were mAP, accuracy, and F1 score. Representative indices were also extracted from the creation of the confusion matrix (Ahmad et al., 2022) where the values for TP – True Positive, TN – True Negative, FP – False Positive, and FN – False Negative are indicated.

This study underlines the need of tracking the software used in NNs ( Table 5 ). This is especially important given the fast developments in NNs and the advent of new software and approaches. Nevertheless, various software programs may yield somewhat different results due to differences in implementation and optimization strategies. Knowing what software was used allows others to replicate and validate the results. This is especially significant for improving the area and expanding on previous studies. Furthermore, knowing the software utilized helps enhance collaboration in the fields of NNs and PA. It allows academics to share code and data, enabling the flow of ideas and speeding up research and development.

As can be observed from Table 5 , Tensorflow in combination with Keras is the most popular choice for software development in pest detection or identification systems using CNNs (Fang et al., 2020). The second popular way of software implementation, showing increasingly high and modern adoption, is represented by PyTorch with the attached torch and torch-vision libraries.

TensorFlow is an open-source software library developed by Google for building and training ML models (Abadi et al., 2016). It is a popular and powerful DL framework that provides a wide range of tools and APIs for building and training models (Wang et al., 2022). TensorFlow is a library for numerical computation that is particularly well-suited to the computation of large-scale linear algebra operations, which are a common component of many ML algorithms. It provides a wide range of tools for building and training DL models, including CNNs and recurrent NNs. It also includes support for distributed training and deployment on different hardware platforms. For the task of detection of harmful insects and pests in modern agriculture, it was a popular choice ( Table 5 ).

Keras is an open-source software library written in Python that provides a high-level interface for building and training DL models (Chollet et al., 2015). It is built on top of other popular DL frameworks, including TensorFlow, and provides a simple and intuitive API for defining and training models. Keras was designed with the goal of making DL accessible to a wider audience, including researchers, students, and developers with limited ML experience.

For insect monitoring tasks, another software used was PyTorch. Based on the analyzed papers, a strong adoption of the framework in pest detection tasks is observed in the last three years, especially in 2022. Facebook (actual Meta) team created PyTorch as an open-source machine learning framework (Paszke et al., 2019). It is a well-known and sophisticated DL framework that offers a variety of tools and APIs for developing and training ML models. Torch is a scientific computing framework that enables efficient tensor operations and automated differentiation. PyTorch is built on top of the Torch library and improves these capabilities by including a dynamic computational graph, allowing for more flexible and intuitive model creation, and debugging. PyTorch includes a variety of tools and APIs for developing and training DL models such as CNNs, recurrent NNs, and others.

The MATLAB programming and numerical computing platform do not have the same characteristics as the libraries and deep learning frameworks like Tensorflow + Keras or PyTorch, based on the performance and flexibility associated with the Python programming language in which they are implemented. MATLAB (matrix laboratory) is a programming environment and a programming language used primarily for numerical computing and scientific computing (MathWorks Matlab 22). MathWorks MATLAB provides a wide range of built-in functions and tools specifically designed for image processing and computer vision applications (Nagar and Sharma, 2021). MATLAB’s Image Processing Toolbox provides a comprehensive set of tools for image analysis, filtering, segmentation, feature extraction, and object recognition (Divyanth et al., 2022). The toolbox includes functions for common image processing tasks such as image smoothing, noise reduction, edge detection, and morphological operations. MATLAB also provides support for deep learning and machine learning, which can be used for image classification and object recognition tasks.

In this sense, although there are various software solutions, the Python programming language remains a solid basis to build such deep-learning systems based on artificial neural networks in the detection, identification or even monitoring of insect pest. Cloud computing services capable of providing modules, APIs or even software platforms as a service in the development of deep-learning solutions for pest detection have also been noted. The main characteristic in their case is represented by the availability and flexibility in accessing these types of cloud resources, being therefore part of the new trends.

Another software used for insect monitoring in precision and modern agriculture was Fastai. It is a high-level open-source DL library built on top of PyTorch (Howard & Gugger, 2020). It is designed to make it easier to train state-of-the-art DL models with as little code as possible. The library provides a simple and consistent API for quickly training deep NNs on a wide range of tasks, such as image classification, object detection, text classification, and natural language processing. One of the unique features of Fastai is its approach to transfer learning, which involves leveraging pre-trained models and fine-tuning them for specific tasks (Cochero et al., 2022).

Another modern software that was used for insect monitoring in agriculture was Imagga Cloud API which is a cloud-based image recognition platform that provides a suite of APIs for developers to build image-related applications (Imagga, 2020). Imagga API was used for rice pest detection (Bhoi et al., 2021), integrating IoT and UAV systems. The Imagga Cloud API provides a range of image analysis and recognition services, including image tagging, content-based image search, color extraction, cropping, and ML algorithms that can identify objects, scenes, colors, and other attributes within an image.

CNNs have become increasingly popular in image-processing applications for modern agriculture following their ability to identify insects and features in images. According to this study, one of the applications of CNNs in the field of modern and precision agriculture is harmful insect detection. The identification of harmful insects is crucial for the protection of crops and the prevention of plant diseases (Lv et al., 2022). CNNs can be an effective tool for harmful insect detection in images (Guo et al., 2021; Alsanea et al., 2022). By training the network on a large and diverse dataset, CNN can learn to identify a wide range of harmful insects. However, the issues of class imbalance and transferability need to be addressed to ensure that CNN performs well in real-world applications. For effective detection of harmful insects, the first step is to collect and label a dataset of digital images containing both harmful and non-harmful insects (Cochero et al., 2022; Wang et al., 2022). In this case, the dataset should be large and diverse to ensure the great performance of the CNN model and to ensure that the CNN can learn to recognize a wide range of harmful insects. Because this detection uses CNN models that learn different features of an image through convolutional operations, the second step is the preprocess the images in the dataset created to ensure that they are in a format that can be fed into the CNN. This may involve resizing the images, converting them to grayscale, or normalizing the pixel values (Alsanea et al., 2022; Zhang Y. et al., 2022). Following this scenario, the next step is to train the chosen CNN model using the dataset prepared (Malathi and Gopinath, 2021; Nagar and Sharma, 2021; Liu et al., 2022). This involves feeding the network the labeled images and adjusting the weights of the neurons through backpropagation to minimize the error between the predicted and actual labels. Transfer learning applied on a custom insect pest dataset can be used and hyperparameter tuning to speed up the process in this topic. Related to this aspect, most of the papers analyzed for this study include such methodology (Ayan et al., 2020). After the CNN was trained, it can be used to classify new images of insects as either harmful or non-harmful. To do this, the new image is fed into the CNN, and the output is a probability score indicating the likelihood that the insect in the image is harmful. A threshold value can be set, and if the probability score is above this value, the insect is classified as harmful.

One of the main challenges that were identified in applications for harmful insect detection using CNNs is the issue of class imbalance (Du et al., 2022). Harmful insects may be rare in the dataset, which can lead to the CNN being biased towards non-harmful insects. To overcome this, techniques such as over-sampling or under-sampling can be used to balance the dataset. Another challenge identified is the issue of transferability. CNNs trained on one dataset may not perform well on a different dataset due to differences in the types of insects or the background images. To address this, transfer learning can be used, which involves using a pre-trained CNN as a starting point and fine-tuning the network on the new dataset, as mentioned earlier (Butera et al., 2021; Li W. et al., 2022; Popkov et al., 2022).

CNNs are a powerful tool for identifying insect-infected crops. They can be trained to learn patterns and features in images that are indicative of insect damage and provide predictions on whether the crops are healthy or infected (Turkoglu et al., 2022; Zhang S. et al., 2022). The use of CNNs in agriculture can improve crop yields and help farmers prevent and manage insect infestations more effectively (Espinoza et al., 2016; Sharma et al., 2020; Bereciartua-Pérez et al., 2022). Infected crops by insects can have a significant impact on the agricultural industry, leading to the loss of crops and revenue (Xu et al., 2022). With the increasing advancements in computer vision, for modern agriculture, our study highlights that the CNNs became an effective tool for identifying and detecting insect infestations in crops.

CNNs are commonly utilized in applications such as image classification, object identification, and segmentation. CNNs may be taught to recognize patterns and characteristics in images that are indicative of insect damage in the context of recognizing insect infestations in crops. Similarly, to the insect detection tasks discussed, a huge collection of images of healthy and infected crops must be developed for applications used to target diseased crops. The images are then annotated with whether the crops are healthy or sick, as well as the species of bug inflicting the harm. The CNN models are then trained by giving them tagged images, allowing them to understand the patterns and characteristics associated with insect-infested crops.

On the other hand, the CNN model can also provide information about the type of insect causing the damage, enabling farmers to take appropriate measures to prevent further damage. In addition to identifying insect-infected crops, CNNs can also be used for segmentation tasks (Zhang & Chen, 2020). Segmentation involves dividing an image into different regions or objects. In the context of identifying insect infestations, segmentation can be used to identify and evaluate the areas of the crop that are infected. This can provide more detailed information to farmers and enable them to target their treatment strategies more effectively.

The third area of applications using models based on DL, respectively on CNNs, is crop monitoring. This area of application has a major impact when considering pest insect populations and managing the effects of their presence. In this sense, there have been several studies that have included this direction of development. Crop monitoring refers to the process of keeping track of the growth, development, and health of crops. Crop monitoring can be done using a variety of methods, including satellite imagery, drone imagery, ground-based sensors, and visual inspections. However, the traditional methods of crop monitoring can be time-consuming, expensive, and require a significant number of resources. With the advent of AI and ML, the use of CNNs for crop monitoring has become increasingly popular. In crop monitoring, CNNs can be used to analyze images of crops and provide insights into their growth, development, and health. The process of crop monitoring using CNNs typically involves several steps including data collection, data preprocessing, training NNs, and evaluating performances in a specific area of interest (Dai et al., 2021). Applications of crop monitoring using CNNs have a wide range of applications in modern agriculture, including disease and pest detection or even yield estimation (Zhong et al., 2018; Tian G. et al., 2020). CNNs can be used to detect the presence of diseases in crops by analyzing the images of the leaves and other parts of the plant. This can help farmers to take timely action to prevent the spread of diseases and minimize crop losses. On the other hand, this process can be automated by introducing real-time monitoring modules, based on hardware systems and software modules optimized for mobile platforms, used in the field. Important to note, crop monitoring using CNNs has the potential to revolutionize agriculture by providing farmers with real-time insights into the growth, development, and health of their crops. CNNs can analyze images of crops quickly, accurately, and at a fraction of the cost of traditional methods (Vanegas et al., 2018). By using CNNs for crop monitoring, farmers can make informed decisions about crop management, minimize losses due to meteorological conditions, diseases, and pests, and optimize their yields.