Digital Twin System of Pest Management Driven by Data and Model Fusion

Abstract

Protecting crops from pests is a major issue in the current agricultural production system. The agricultural digital twin system, as an emerging product of modern agricultural development, can effectively achieve intelligent control of pest management systems. In response to the current problems of heavy use of pesticides in pest management and over-reliance on managers’ personal experience with pepper plants, this paper proposes a digital twin system that monitors changes in aphid populations, enabling timely and effective pest control interventions. The digital twin system is developed for pest management driven by data and model fusion. First, a digital twin framework is presented to manage insect pests in the whole process of crop growth. Then, a digital twin model is established to predict the number of pests based on the random forest algorithm optimized by the genetic algorithm; a pest control intervention based on a twin data search strategy is designed and the decision optimization of pest management is conducted. Finally, a case study is carried out to verify the feasibility of the system for the growth state of pepper and pepper pests. The experimental results show that the virtual and real interactive feedback of the pepper aphid management system is achieved. It can obtain prediction accuracy of 88.01% with the training set and prediction accuracy of 85.73% with the test set. The application of the prediction model to the decision-making objective function can improve economic efficiency by more than 20%. In addition, the proposed approach is superior to the manual regulatory method in pest management. This system prioritizes detecting population trends over precise species identification, providing a practical tool for integrated pest management (IPM).

1. Introduction

The factors influencing crop yield in modern agriculture are very complex, and one of the most important factors is insect pests, which, according to statistics, reduce the yield of fruits and vegetables by more than 20% worldwide [1]. Insect pests are one of the most important constraints to the development of modern agriculture. Due to their large population size, adaptability, and migratory nature, they can easily lead to large-scale pest outbreaks in a short period if not prevented and controlled in time, causing irreversible losses. Therefore, it is essential to establish IPM for the mitigation of pest infestations on vegetable crops [2].

Traditional pest management of crops relies heavily on the professional knowledge and personal experience of managers and the use of pesticides and other chemicals [3,4]. Due to the limitation of existing technologies, such as the high need for manual intervention, the inability to perform pre-process management, the high cost, or the technical complexity, it cannot provide efficient and intelligent management of crop pests. With the development of Internet of Things (IoT) technology, through sensors, camera devices, and operating mechanisms arranged on a farm, growers can monitor and perform operations remotely based on real-time digital information instead of direct observation and on-site manual operation [5,6]. In particular, digital twin technology combines artificial intelligence and data fusion based on the IoT architecture and uses modeling or simulation techniques to map the crop model in a virtual scene. It can simulate the impact of corrective and preventive measures on the crop. Growers are alerted if there is a pest problem with their crops and can check the condition of the greenhouse by viewing rich digital images of the plants or equipment in question, either on a PC or from a smartphone. Further, they can remotely execute the chosen intervention and can again use the digital representation to verify that the growing problem has been solved. The intelligent decision-making management system will become increasingly autonomous, without the need for human intervention by the grower.

In response to the current problems of heavy use of pesticides in pest management and over-reliance on managers’ personal experience on pepper plants, we have tried to employ digital twin technology to build a complete experimental framework of the pest management system. Through real-time situational awareness, ultra real-time virtual inference, and full-process interactive feedback, it will effectively achieve intelligent control of pest management systems. It will provide new ideas and methods for applying digital twin technology to agriculture. Therefore, the main purpose of our study is to develop a digital twin system that monitors changes in aphid populations on pepper plants, enabling timely and effective pest control interventions. This system prioritizes detecting population trends over precise species identification, providing a practical tool for IPM. The main contributions of this work are described below.

(1)

A digital twin system for pest management driven by data and model fusion is proposed. It includes environmental collection devices that collect crop environmental data through wireless transmission technology and a database in the form of a cloud platform.

(2)

A pest prediction model is established to predict the number of pests based on the random forest algorithm optimized by the genetic algorithm.

(3)

The proposed approach can achieve the virtual and real interactive feedback of the pepper pest management system, obtaining accurate prediction and management of pepper pests.

The remaining paper is structured as follows: Section 2 describes the literature review; Section 3 introduces the materials and methods; the proposed digital twin framework for pest management is presented in this Section; also, the detailed digital twin modeling approach in the framework is described; in Section 4 and Section 5, the experimental results and discussion are addressed, respectively; in Section 6, the main findings of this study are summarized. In addition, the limitations and suggestions for future work are presented.

2. Literature Review

Currently, to improve the operating efficiency and intelligence of IPM, researchers have focused on pest model prediction. They have investigated the relationship between pests and environmental factors like temperature, humidity, CO₂ concentration, wind speed, rainfall, and light intensity by using different sensing and detection methods. Bairwa et al. [7] pointed out that mung bean insect pests are highly influenced by meteorology, with a high positive correlation between maximum temperature and pests, a negative correlation between minimum temperature, wind speed, and light intensity, and a non-significant correlation between maximum humidity, minimum temperature, and rainfall. Zhang et al. [8] constructed a multi-sensor network system to collect small-scale environmental data from vegetable production sites for the study of early warning mechanisms for vegetable pests by collecting pest image data and information about four major vegetable pests in southern China. Machekano et al. [9] investigated the relationship between each environmental factor and moth population growth by data mining. They concluded that average temperature and maximum temperature were positively correlated with the rate of moth growth. Minimum temperature, minimum relative humidity, and maximum relative humidity were negatively correlated with the rate of moth growth, and the correlation functions derived from the study can be useful for pest early warning. Dan et al. [10] proposed an automated pest counting and environmental condition monitoring system, which can avoid laborious manual counting and can achieve timely assessment of pest and environmental information. The system provided an effective tool for long-term pest behavior observation and practical application of integrated pest management. Tonnang et al. [11] analyzed the effects of temperature and CO₂ concentration on pests. They concluded that temperature increases and CO₂ concentration changes can greatly affect the physiological habits of pests, accelerate their autogenous metabolic depletion, and ultimately lead to population density. Chen et al. [12] employed a long short-term memory network (LSTM)-based prediction of meteorological data and machine learning (ML) to identify correlations between pests and environmental factors, and their study confirmed that meteorological factors do affect pest occurrence. In addition to the above studies, other pest monitoring methods such as image classification [13], multispectral imaging [14], and thermal imaging [15] are gradually becoming new hot directions in pest non-destructive monitoring-related aspects. The development of the above-mentioned technologies can gradually improve the accuracy and applicability of pest prediction. However, these methods of pest management are only for detecting the infestation situation after it occurs and do not suggest specific management in the pre-infestation period.

With the development of IoT technology, growers can manage and control insect pests through sensors, camera devices, and operating mechanisms arranged on the farm. For instance, Azfar et al. [16] proposed an IoT framework to investigate the issue of smart pest detection and management of cotton plants through motion detection sensors. The result a modernization of pest management. Wang and Jannesari [17] developed a crop pest control system based on IoT and information management technologies to control insect pests and diseases, and they verified the effectiveness of the proposed system through implementation in a greenhouse facility. Ali et al. [18] addressed the implementation of an AI-enabled system based on IoT and advanced analytics technologies to detect, prevent, and control pests. The experimental results showed that the average design efficiency for pest management achieved 98.87%. Nadeem et al. [19] used IoT and data-driven machine learning to investigate stem borer attack prediction on sugarcane crops by directly sensing environmental conditions from the crop field. They aimed to support sustainable developments in agriculture by leveraging data-driven decisions. Crepon et al. [20] implemented IPM to manage insect infestations in the storage and handling chain by using IoT, near-infrared spectroscopy (NIRS), and hyperspectral imaging (HSI) and proved the effectiveness of the proposed method. Although these technologies have improved pest management to some extent, there are still significant limitations, and new solutions are needed.

A digital twin (DT) is a dynamic virtual representation of a physical object or system, usually across multiple stages of its lifecycle, that uses real-world data, simulation, or machine learning models combined with data analysis to enable understanding, learning, and reasoning [21,22,23]. The DT, as a unification of virtual and physical assets in product lifecycle management [24], has been successfully applied to manufacturing, aerospace, and medical industries [25,26,27]. However, research on DTs in agriculture is in the initial development stage. Digital twin technology can bridge the barrier between physical entities and virtual models, providing new methods and ideas for crop pest management. Li et al. [28] presented a deep learning-based method for single-view leaf reconstruction in plant-growth digital twin systems, providing important ideas and methods for single-view leaf recovery in plant-growth digital twin systems. Juan et al. [29] studied the monitoring digital twin of controlled environment agricultural services (CEAs), which can collect and analyze CEA equipment data to build a database. It can be used in the future by predictive and prescriptive algorithms to suggest actions that can improve the productivity and sustainability of the process. Naftali et al. [30] addressed a digital greenhouse twin to track crop growth with possible longer user interaction in a virtual environment to explore the application of virtual reality technology in future farms. Rafael et al. [31] developed a digital twin system for water-saving irrigation management, which consists of a FIWARE-based IoT platform and a discrete event simulation model in Siemens plant simulation software. The system can collect, aggregate, and process soil, weather, and crop data to calculate daily irrigation prescriptions and can simulate the behavior of the irrigation system as defined in the application scenario. In addition, Christos et al. [32] pointed out that most documented agricultural applications were still at the conceptual level. DTs are powerful technology, demonstrating their potential to accomplish digitization and replication of complex systems across a variety of domains in agriculture. However, their key role in enabling sustainability in agriculture is often absent in research [33].

In summary, DTs have begun to be applied in the field of agriculture in recent years. A DT combines artificial intelligence (AI) and data fusion based on the architecture of the Internet of Things and uses modeling and simulation techniques to map the crop model in a virtual scene. It can simulate the impact of corrective and preventive measures on the crop [34]. It can be seen that the DT has the ability of real-time simulation, prediction, interactivity, interoperability, and visualization compared with traditional technologies. Thus, growers are alerted if there is a pest problem with their crops and can check the condition of the greenhouse by viewing rich digital images of the plants or equipment in question, either on a personal computer (PC) or a tablet PC. Further, they can remotely execute the chosen intervention and can again use the digital representation to verify that the growing problem has been solved. The intelligent decision-making management system will become increasingly autonomous, without the need for human intervention by the grower. However, the overall research carried out is in the exploratory stage; the theoretical framework is not complete and the experiments stay in the simulation stage, which cannot specifically solve the practical problems of pest management. It is challenging for us to investigate the practical effects of digital twin technology applied in agriculture. Many aspects should be taken into account; for example, what kind of technical problems can be solved by digital twin technology applied in agriculture? And what are the specific differences between digital twin technology applied in agriculture and other fields like industrial production?

3. Materials and Methods

3.1. Data Collection

3.1.1. Experimental Environment Setting

The experimental crop was selected from the Plant Tissue Culture Technology Laboratory of Haidian District, Beijing, China. The germinated peppers [Capsicum annuum L. (Solanales: Solanaceae)] were “Sea Flower No.3” that came from Lixia River Research Institute, Jiangsu. Pepper seeds were sown in cavity trays, and when the seedlings reached 6–7 leaves, the seedlings with uniform growth and robustness were selected and transplanted in pots; the test aphids are green peach aphids [Myzus persicae Sulzer (Hemiptera: aphididae)] that came from the laboratory of the School of Horticulture and Plant Protection, Yangzhou University, and were propagated with pepper (Haihua 3) for more than 3 generations in the insect-rearing room. The experiment was carried out from March to May 2023.

The pepper cultivation environment was a climatic chamber with an STM32F103C8T6 microcontroller, a GY-302 light intensity sensor, a DHT11 temperature and humidity sensor, and an MG-812 carbon dioxide concentration sensor. We can adjust the light, temperature, humidity, and carbon dioxide required by the pepper in a timely manner according to the monitoring terminal. The monitoring terminal includes the monitoring terminal host and monitoring terminal sub-node. LoRa was used as a wireless communication means to transmit data to the microcontroller unit (MCU). A one-to-one communication mode was employed to synchronize different sensing information to the OneNet cloud platform through ESP8266 WiFi modules. Thus, it enabled users to realize remote monitoring of the real-time environmental conditions of the climate chamber on the PC side, which facilitates further data mining, processing, and modeling. The experimental environment was set up as shown in Figure 1.

3.1.2. Data Collection Methods

The 80 seedlings of late-stage “Sea Flower No.3” pepper were selected and transplanted into 8 pots (marked as A, B, C, D, E, F, G, H). Two pots were considered as a group, and four groups were placed in the incubator of the Mechanical Engineering College, Yangzhou. We inoculated aphids on pepper leaves to create an aphid-infested pepper growth environment. Four comparable groups were set with different environmental parameters, and the environmental parameters and population change data of each comparable group were recorded three days apart. In addition, yellow trap boards were placed in the center of the pepper incubator to trap aphids, and the data on the number of pest insects were obtained automatically by the monitoring equipment. We collected experimental information on pest management based on the digital twin method, including daytime temperature (8 am), night temperature (8 pm), average humidity (8 am–6 pm), daytime light intensity (8 am–6 pm), carbon dioxide content (12 am), ultraviolet (UV) exposure time (10 am), initial aphid population, and changes in aphid population. Finally, a total of 130 sets of data were collected for the experiment (see Table S1), and the testing dataset is shown in Table S2.

In addition, we captured more than 1000 original pepper leaf aphid images at different growth stages of peppers by using an RGB camera to support pest prediction [12]. To ensure comprehensive coverage of aphid infestations, the images of pepper leaves infested with green peach aphids were taken of both the upper and lower surfaces of the leaves. To avoid the overfitting of an algorithm model, data augmentation is an essential means of balancing the number of samples and expanding the amount of data in deep learning technology. We further employed data augmentation methods like random rotation, random cropping, color enhancement, and noise addition to obtain more images of pepper pests for both training and validation of the model [35]. In addition, the changes in damaged leaves in different pots with the progress of the cycle can be observed in

Table 1. The experimental changes in damaged leaves in different pots.

Plant Number	1	2	3	4	5
A
B					None
C
D
E
F
G
H

, where the plant number represents the plant to which the leaves belong and the leaf infected with aphid infestation is selected as a representative of different plants. Images were taken of the same leaf at the end of five consecutive cycles of different environmental variables, but the leaf of plant B fell off after the 4th cycle; thus, the image at the end of the 5th cycle is missing.

3.2. Digital Twin Framework for Pest Management

Applying the emerging digital twin technology to pest management can realize the data mapping from physical entities to digital twin models, which can assist production managers in supervising crop pests and also self-generate decisions to intelligently manage pests. To establish a digital twin system for pest management, the environmental data of crops and pest infestation in greenhouses are mapped into the twin model. Then, an all-round, multi-level, and multi-domain data integration and sharing module is established to fully integrate state perception, virtual reality, data analysis, AI, model fusion, and other related technologies to drive the continuous iterative update of the twin mechanism model for pest management. Hence, it can sense the state of the physical objects in the greenhouse. The twin mechanism model of pest management is driven by continuous iterative updates to realize the state perception of the real plant. We present a digital twin framework model (M) applicable to IPM, which refers to the five-dimensional model of a digital twin proposed by Tao et al. [22].

M = (PEC, VEC, TDC, SC, DC)

(1)

PEC is the physical entity component of the pest management system, and it represents the physical attributes of data processing devices in climatic chambers; VEC is the virtual entity component of the pest management system, and it represents the digital mapping model and function of the physical entity; TDC is the twin data component of the pest management system, and it represents fusion information based on physical and virtual space; SC is the service component of the pest management system, and it represents application software for supporting pest management; DC is the data connection among the components of the pest management system, and it achieves the information interconnection and interaction between PEC, VEC, TDC, SC, and data. The system realizes the virtual–real fusion and dynamic update of the digital twin of pest management, and the specific implementation framework is shown in Figure 2.

3.2.1. Physical Entity Component (PEC)

The physical entity is the basis for building a digital twin system for pest management, which consists of climatic chambers, crops, pests, and equipment. To realize the monitoring of pest management, it is necessary to collect various sensing data of the crop and the environmental information in real time or dynamically. These data are digitized to establish digital twins of corresponding entities. Here, the sensing devices consist of light-sensitive panels, carbon dioxide sensors, and temperature and humidity sensors. They constitute the physical entities of the pest management system.

3.2.2. Virtual Entity Component (VEC)

The virtual entity for pest management includes virtual, functional, and rule models, which provide systematic descriptions of physical entities in different dimensions. It provides a virtual environment for IPM. Through crop environmental parameters and pest infestation data, machine learning models are constructed to mine the hidden mathematical and theoretical relationships among crop physiological states, environmental parameters, and pest infestation levels. Then, specific implementation suggestions are made for crop regulation, and the control strategy of the virtual entity is formulated by establishing the rule model of the virtual entity. Finally, the physical entity data are updated after executing the decision in the database to achieve the effect of iterative updates of the mathematical model, together realizing crop pest management with interactive functions.

3.2.3. Twin Data Component (TDC)

The twin data mainly include perceptual data from physical space, virtual model data, and service system data in the pest management system. They involve the collection and analysis of plant growth data and pest data to provide data support for decision-making through real-time monitoring of plant health. A large amount of data generated during the pest management process is collected through the database for unified data processing to form a structured dataset. Meanwhile, dynamic updating of the virtual model can be realized based on the collected real-time data. With the continuous accumulation of pest data and the improvement of the database, the feasibility of the established pest prediction and regulation model gradually increases.

3.2.4. Service Component (SC)

The service system is a package of individual models and algorithms in the service component that realizes crop environmental parameters and pest information display, pest prediction, and decision-making. The SC is the core of the system and integrates models and algorithms for supporting pest management. It is responsible for processing and integrating various data from PEC and VEC and provides decision support services, such as pest risk assessment and control strategy recommendations.

3.2.5. Data Connectivity (DC)

Data connection realizes the information interaction between various parts of the digital twin, including the connection between the service system, twin data, physical entities, and virtual models. The physical space in the pest management system can be input into the digital space through the data transfer protocol from the real entity to the virtual entity; the optimization and intelligent decisions of the twin mechanism model are carried out in the digital space, and the model is migrated to the physical space from the virtual entity to the real entity. Through continuous iterations and optimization of the above process, the connection and dynamic interaction of real-time data between physical space and digital space can be realized.

3.3. Digital Twin Modeling Approach

According to the digital twin framework of pest management, the digital twin-based pest management model is created as follows (see Figure 3): (1) we arrange environmental collection devices (such as RGB camera, hyperspectral camera, and different types of sensors) in a controlled agricultural environment through IoT architecture in the perceptual layer; (2) we collect crop environmental data through wireless transmission technology (such as LoRa, and 5G) and the database of the OneNet cloud platform in the transport layer; (3) we construct a virtual model for pest prediction through AI and machine learning technology based on the collected data and carry out the IPM decision-making based on the analysis result of AI in the application layer.

3.3.1. Pest Prediction Model

Machine learning algorithms serve the purpose of human–computer interaction and act by mining the potential laws implied in the data and establishing the mathematical relationship between input and output on this basis, to achieve the approximation of the correlation between input variables and target variables in physical space and achieve the full cycle of analysis, control and prediction. Since the influencing factors affecting crop pests are very complex, this paper combines previous research based on the selection of light, humidity, morning and evening temperature, and carbon dioxide concentration as external variables affecting pests, with the addition of UV lamps for more effective pest control [36,37,38]. To improve the model prediction accuracy as well as convergence speed, a random forest prediction regression algorithm based on a genetic algorithm (GA) was employed to train the crop environment data collected by LoRa.

Random forest (RF) is a parallel integrated learning algorithm with a decision tree as the base learner [39], which enhances the generalization ability of the pest prediction model by achieving the randomness of pest sample selection and pest feature selection through the idea of Bagging (Bootstrap aggregating) and Random Subspace Method (RSM) compared to the base learner alone. According to the Bagging idea, as shown in Figure 4, the RF model can obtain pest-sampled training sets M with the same capacity as the original training set by M independent random sampling with put-back. We use these pest-sampled training sets for training to obtain the corresponding M base learners. Due to the independence of sampling, the sampled training set obtained from each sampling is different from the original training set and other sampled training sets, which can effectively avoid the emergence of locally optimal solutions from the perspective of training sample selection and also ensure low correlation among each decision subtree [40].

In addition, GA is an efficient stochastic search and optimization method based on the genetic mechanism of nature and biological evolution theory [41]. GA has global optimization performance and can find the most suitable value of the parameters of the algorithm. Hence, the algorithm could reach the optimal configuration to predict the insect population more accurately. The error function of RF is considered the fitness function of GA, and the goal of the optimization is to minimize the value of the error function. The flowchart of GA to optimize the parameters of the RF model is shown in Figure 5.

3.3.2. Twin Data Search Strategy

The general pest prediction model is based on the prediction results obtained at a single time point. However, the number of insect pests varies over time; no specific opinion can be given for pest management. It is necessary to introduce twin data and AI to optimize the environmental parameters and combine the objective function for the dynamic management of pests. In this paper, we propose a pest-finding model combining an objective function and artificial fish swarm algorithm (AFSA) to conduct the decision optimization of pest management.

(1) Optimization-seeking objective function

Most crops are used as cash crops, and one of the key indicators evaluating the economic value of crops is yield. Different degrees of pest infestation on a pepper crop can directly affect the yield of pepper. In this paper, we mainly consider the impact of aphid infestation on the yield of pepper, inoculating the number of aphids on a pepper crop to represent the pest infestation degree of the crop. In the objective function, it is very critical to evaluate the implementation of the decision on the degree of pest infestation on the crop yield. The residual percentage function F(L) of the predicted economic benefits is designed to represent the implementation of the decision on the degree of pest infestation on the pepper yield. F(L) is obtained by multiplying the control action (n–n’) of pest change and the sensitivity factor L. L is the sensitivity coefficient and it can be artificially adjusted to meet the actual production needs according to a specific pest infestation in different growth stages of peppers. n represents the number of insect pests before the decision execution of IPM; n’ represents the predicted number of insect pests after the decision execution of IPM. We substitute various environmental parameters of the decision-making strategy into the prediction model based on RF mentioned above. Thus, we can obtain pest prediction data directly without having to count the actual pest population indirectly. The specific formula is as follows:

$$F(L)=\left(n-n^{\prime }\right)\ast L\%$$

(2)

To obtain the highest economic value, it is also necessary to consider the energy involved in each decision strategy. Different decision strategies need to regulate corresponding environmental parameters like temperature, humidity, CO₂, and light in the pepper crop cultivation environment. These parameters are controlled by various sensors and devices, which consume a certain amount of electric energy. Thus, for a given decision strategy, the economic cost of the energy consumption involved can be represented by ${\displaystyle \sum _{i=1}^n{C}_i}{X}_i$. X_i (i = 1, 2, 3, 4, 5) represents the energy cost per unit value for each environmental variable. C_i (i = 1, 2, 3, 4, 5) represents the value of environmental variables that can be adjusted; more precisely, C₁ represents the daily average temperature in the incubator, C₂ represents the daily average humidity in the incubator, C₃ represents the simulated light intensity in the incubator (12 h of light per day), C₄ represents the daily average carbon dioxide concentration in the incubator, and C₅ represents the UV irradiation time in a day. X_i represents the economic value required to adjust the value of each unit of the corresponding environmental parameter. While the adjustable range of different environmental parameters is not quite the same, the specific constraints are as follows:

$$\mathrm{s}.\mathrm{t}.\{\begin{array}{l}{K}_i-a_i<C_i<K_i+b_i, i=1,2,3,4\\ 0<C_i<10, i=5\end{array}$$

(3)

where K_i (i = 1, 2, 3, 4) represents the initial value of the environmental parameter, a_i (i = 1, 2, 3, 4) represents the search range in which the environmental parameter decreases, and b_i (i = 1, 2, 3, 4) represents the search range in which the environmental parameter can increase.

Therefore, the optimal function Z for the highest economic value of regulatory decisions for pepper crops in the region can be established and is presented as follows:

$$\max Z=P\ast Q\ast F(L)-{\displaystyle \sum _{i=1}^n{C}_i}{X}_i$$

(4)

where P represents the economic price of the crop market and Q represents the total yield in the region.

(2) Optimization-seeking algorithm

To obtain the optimal solution of the objective function, different optimization algorithms such as particle swarm optimization, the ant colony algorithm, and the firefly algorithm can be employed to solve the problem. In particular, AFSA is a new swarm intelligence optimization algorithm in which the place with the highest number of individual fish is generally the place with the highest amount of nutrient-rich material in this water. Based on this phenomenon and the characteristics of fish, it imitates the behavior of fish such as feeding, clustering, and tail-chasing to achieve global optimization. In this study, we take into account AFSA due to its powerful global search capability, dynamic adaptability, and parallel processing capability [42]. It avoids prematurity by modeling the diversity and randomness of fish populations while coping with complex environments by learning and adapting its strategies. In addition, the probability-directed search of AFSA improves the likelihood of finding a globally optimal solution, which makes it show unique advantages in decision search problems [43]. In the digital twin system, the strategy search for pest management decisions requires a strong global search capability as well as decision adaptation for different pest situations. Hence, AFSA is selected for decision optimization for pest management.

3.3.3. Execution of Decisions and Feedback Iterations

The prediction model based on machine learning and the economic efficiency objective function based on the optimization algorithm are integrated into the decision-making of the pest management system. It can obtain the optimal decision based on the current crop pest situation and environmental parameters, adjusting the environmental parameters in the incubator, collecting the data at the end of each cycle to correct the prediction model error, and carrying out feedback iteration. In addition, to improve the real-time accuracy of the prediction method, the pepper data within the experiment using LoRa wireless transmission technology are uploaded to the cloud platform and imported into the algorithmic database to give appropriate weights to correct the prediction error.

4. Experimental Results

4.1. Pest Prediction Accuracy

To evaluate the accuracy of pest prediction, Matlab2022a was used as the simulation platform. The population size of the genetic algorithm was set to 100, the maximum number of reproduction generations was set to 60, the probability of chromosome crossover was 80%, the probability of chromosome variation was the inverse of the chromosome length, and the fitness value function was set to the accuracy of the pest data in the RF model. The evaluation metric is the R-square, also known as goodness of fit. It is commonly used to describe how well the data fit the model, which can be described as

$$R^2=1-\frac{{\displaystyle \sum _i{\left({\widehat{y}}_i-y_i\right)}^2}}{{\displaystyle \sum _i{\left({\overline{y}}_i-y_i\right)}^2}}$$

(5)

where i represents the data serial number, ${\widehat{y}}_i$ represents the predicted value, and $y_i$ represents the real pest value.

According to the 130 groups collected from the climatic chamber, the pepper aphid pest dataset was divided into the training set and test set. GA was employed to find the hyper-parameter combination of the RF regression model; the experimental results are shown in Figure 6 and Figure 7. It can be observed that the result of R-square prediction accuracy is 0.8801 in the training set and 0.8573 in the test set. According to the application of AI techniques in plant pest detection, a prediction accuracy of more than 0.8 meets the pest prediction accuracy requirement [44]. Hence, the pest data-driven machine learning model could be used for the digital twin supervision of pest management in the experimental environment of the climatic chamber.

4.2. Digital Twin Supervision Results

First, AFSA was employed to obtain the best solution of environmental parameter variables in the climate chamber. The population size of the fish swarming algorithm was set to 100, the maximum number of iterations was 70, the maximum number of experiments for artificial fish was 30, the perceptual distance was 0.5, the crowding factor was 0.618, and the swimming step was 0.03. The goal of the optimization strategy was the economic optimal strategy consisting of energy consumption and yield. Here, the consumption of electrical energy in the incubator was considered as the energy consumption. The power of the air conditioner was 450 W, the humidifier was 100 W, fluorescent lamps were stepped 20 W pairs to provide 20% of light intensity up to five pairs, the UV lamp was 10 W, and the ventilation fan was 80 W. The artificial fish population was randomly generated data within reasonable limits of environmental parameters to find the optimal environmental parameter variables with the goal of economic efficiency. The iteration diagram of AFSA is shown in Figure 8.

In Figure 8, it can be found that AFSA quickly reaches 5.516 RMB from the initial value of 4.88 RMB at the initial stage of the algorithm, an increase of more than 20%, for the economically optimal objective function Z. It tends to converge to the stability value of the algorithmic model when it is iterated about 50 times, which indicates that the proposed algorithm is feasible and effective. Hence, the execution decision is carried out according to the experiment result of the fish swarm algorithm.

Further, three experiments were set up for different pest management methods. Experiment 1 was supervised in a normal greenhouse environment for pepper growth, Experiment 2 was manually supervised by traditional manual experience to judge the pepper environment intervention strategy, and Experiment 3 was carried out based on the proposed digital twin supervision strategy. The pest control situation and the energy consumption of pest management are shown in Figure 9 and Figure 10.

Through continuous tracking of the strategy cycles, in terms of pest management, Experiment 1 shows a rapid increase in the number of aphids, and the spread of the pest situation is rapid, while Experiment 3 and Experiment 2 are able to control the spread of the pest situation in Figure 10. Hence, Experiment 2 and Experiment 3 can control the spread of the pest situation effectively, but the proposed method in Experiment 3 is superior to the empirical method in Experiment 2. In addition, in terms of energy consumption, Experiment 1 does not need to consume electricity because of no environmental parameter intervention, while Experiment 2 consumes more electricity than Experiment 3 in Figure 10. To sum up, Experiment 3 presented a more effective pest management effect and consumed less electricity, which illustrated the effectiveness and superiority of the data model-driven pest supervision method based on a digital twin.

5. Discussion

In response to the current problems of heavy use of pesticides in pest management and over-reliance on managers’ personal experience with pepper plants, we designed a digital twin system for pest management driven by data and model fusion. Based on the population trends observed in our study, we recommend initiating pest control measures when aphid populations reach a certain threshold, which the system approximately detects a specific number of days after initial infestation. Thus, a detailed analysis of the optimal result for pest control interventions based on the population dynamics detected by the system can be discussed below.

In the laboratory environment, a climate chamber platform was built for pepper crops, in which most of the environmental parameters could be adjusted. According to Table 1, we observed that these leaf image data indicate that the pest severity will be affected to a certain extent through the changes in environmental parameters in different cycles, and they also indicate that the pest can be regulated to a certain extent through the changes in environmental parameters.

In the experiment section, on the one hand, Experiment 1 was in a natural environment and was easily influenced by the environment parameters. Experiment 2 was supervised manually and had great uncertainty, relying too much on manual experience. From the experimental results, it can be seen that the spread of insect pests is rapid in the unsupervised experiment, while the spread of insect pests is controlled by manual intervention. But for resource conservation and early intervention, Experiment 1 and Experiment 2 are inferior to Experiment 3. Hence, IPM based on the digital twin method is feasible and effective. On the other hand, in Experiment 3, algorithms are used based on the historical data between crops and pests as a training set, instead of manually acquiring knowledge of pest management from practice. An optimization algorithm is used to find environmental regulation decisions, which can be specified as the optimization goal. Then, integrating LORA communication equipment, machine learning algorithms, and optimization algorithms could work together to generate a complete closed-loop system for the supervision and management of pepper insect pests [44,45,46,47]. It could collect the pest data, predict the pest situation, and make decisions on regulation. Every cycle, the decision strategy is generated and the latest data are uploaded to the cloud for storage and used as feedback to correct machine learning algorithm errors. In Experiment 3, the system initiates pest control measures after initial infestation when aphid populations reach a certain threshold from the second week. The optimal timing for pest control interventions based on the population dynamics is detected by the system. Then, the spread of the pest infestation is significantly controlled by implementing the system strategy. Thus, Experiment 3 could provide a feasible and effective method for pest management using digital twin technology.

6. Conclusions and Future Works

Pest control interventions have always been the focus of research in crop planting. Digital twins are considered to be an effective means of pest control interventions. They can realize the interactive integration of the agricultural information world and the physical world. In this paper, we leverage a digital twin system for pest management driven by data and model fusion in a controlled agricultural environment. Specifically, we construct a digital twin framework to manage insect pests by deploying environmental collection devices. Various environmental parameters closely related to pests in crop growth are fused to investigate crop entities at multiple levels. Then, we establish a digital twin model to predict the number of pests based on the random forest algorithm optimized by the genetic algorithm. Based on the proposed prediction model, we propose a pest-finding strategy combining objective function and AFSA to conduct the decision optimization of pest management. Finally, the aphid pest is investigated in the growth process of peppers as an entry point, tracking and recording the pepper crop data, using the proposed method for pepper crop supervision of pest management. The system can achieve virtual and real interactive feedback of the pepper pest management system, obtaining accurate prediction and management of pepper pests. The experimental results prove the effectiveness of the digital twin system in monitoring aphid population dynamics and informing pest control strategies. The system provides a practical tool for IPM in improving intelligent agricultural practices and reducing pesticide reliance.

While the proposed DT system represents a significant advancement in improving intelligent agricultural practices and reducing pesticide reliance, it is important to acknowledge its limitations. Expanding the dataset and enhancing model precision will further improve the system’s robustness and practical applicability. It is worth integrating the system with other agricultural technologies. In future work, we will focus on validating the digital twin system in various field conditions, considering factors such as complex weather patterns, crop nutrition, and natural pest predators. It is necessary to collect a large amount of crop data in the actual field environment and comprehensively consider various variables. Furthermore, how to generate feasible decisions, how to intelligently regulate these variables, or how to adjust decisions when some variables are uncontrollable will be the focus of future research.