Plant pests and diseases are the main factors responsible for a significant reduction in crop production, including crop yield and quality (Godfray et al., 2016). Guo et al. reported that pests, weeds, and plant diseases accounted for 30% of global crop losses annually (Guo et al., 2019). Therefore, measures must be taken to reduce the enormous impact of pests and diseases. Currently, the spraying of chemical pesticides on crops is the most widely used method for preventing and controlling diseases and pests (Chen et al., 2021; Sparks and Bryant, 2021; Zhang et al., 2021). Various methods have been developed to improve the spraying efficiency and control the effect of pesticides, such as ground spraying, aerial spraying, air-assisted spraying, and knapsack spraying (Qin et al., 2016; Pan et al., 2017; Wang et al., 2019).

In the last decade, the price of pesticides for plant maintenance has soared (Wilson and Tisdell, 2001). Moreover, the use of pesticides in the past year reached an average of 5–7 liters/ha (Maria Travisi et al., 2006). However, the inappropriate use of pesticides results in a decrease in productivity owing to decreased soil nutrients. In addition, the indiscriminate use of pesticides affects crop yield because pesticides affect soil nutrients, interfering with plant growth and directly affecting crop yield (Tudi et al., 2021). Accordingly, the impact of indiscriminately used pesticides is felt on crops. In addition, the exposure of pests and diseases to large quantities of pesticides results in the development of resistance. In some plant pests, this resistance can be passed on to the 5th–7th generation, increasing difficulties in controlling these pests via chemical control in the next planting periods (Shi et al., 2011). This indicates that the correct application of pesticide-based plant treatments can improve the quality of the plants. In addition, this can affect the productivity and optimal growth of plants on fertile lands (Kalia and Gosal, 2011). This indicates the importance of the appropriate application of pesticides in terms of quantity and accuracy of the needs of plants.

Spraying systems that utilize aerial vehicles have a greater application range and can overcome the negative impacts of pesticide use, as most of these impacts are related to the ground rather than air (Li et al., 2022). Compared to conventional spraying methods, such as knapsack sprayers or ground vehicles, unmanned aerial spraying systems (UASS) exhibit a greater spread distance (Maheswaran et al., 2020). However, as this system operates by flying over the ground, it is susceptible to strong winds, which results in drift. In addition, owing to the maximum light payload and high energy output of the system, the power source, which is a battery, runs out quickly, resulting in a limited operational time. The use of a flexible platform makes it simpler to meet the intended needs. However, this platform encounters problems operating UASS for spraying fields in windy situations. Compared to conventional knapsack spraying techniques and ground vehicle plant protection, UASS exhibits significantly increased operational efficiency with reduced labor costs and pesticide exposure (Zhu et al., 2010). In addition, UASS exhibits numerous advantages, including a higher rate of pesticide penetration into the crop, owing to the ability of the rotors to overturn the leaves (Meng et al., 2019). To date, significant studies have been conducted on the use of plant protection techniques to reduce pests, such as the development of precision UASSs (Zhu et al., 2010; Li et al., 2022), autonomous ground vehicle plant treatment (Maheswaran et al., 2020), and in specific sectors, such as crop protection machinery for vegetables (Wang et al., 2019). Consequently, this has resulted in a significant increase in the use of plant-protection UASS.

The fundamental factor in agrochemical capacity control is the accuracy of the spray target in spray deposition pesticides. The use of a different platform slightly disrupts these fundamental plant protections; thus, it is essential to develop a new spraying system that will comply with applicable regulations. In addition, UASS manufacturers must consider variables that may emerge during operation (Chen et al., 2022; Huang et al., 2023). This is because ignoring these factors may reduce the effectiveness of UASS owing to malfunctions, such as drift (Lan and Chen, 2018; Liao et al., 2019; Hussain et al., 2022). Recently, UASSs have been widely employed in the agricultural sector and combined with cutting-edge technology to meet aerial spraying needs (Chen et al., 2022; Huang et al., 2023). In addition, research findings have been used as feature upgrades to develop new platforms. Typically, UASS developers do not produce upgraded components traded separately from the main platform, and they want to capitalize by increasing the selling value of their products by incorporating research features (Gregorio et al., 2014; Butler Ellis et al., 2017).

Pesticides are typically applied uniformly across the land. Even if not all areas of the agricultural land are infected and require treatment, the treatment requirements of disease-infected plants determine the pesticide dose. As the distribution of pests and diseases determines treatment, the amount of pesticides sprayed is unevenly distributed. In contrast, a plant-targeted protection treatment with the right dose will enable the rapid completion of the process before the disease spreads throughout the land (Bottrell and Schoenly, 2018).

Previous studies have reported several limitations of each UASS category. Operation parameters are the main factors that are still an operational consideration. Rotor rotation is an advantage not exhibited by other sprayer systems (Meng et al., 2020), but should be considered in the operation of the UASS (Lan and Chen, 2018). Some UASS manufacturers advise that they should be operated in a conducive environment. This is because an unstable condition affects the effectiveness of the flexible platform; particularly, the use of the UASS during inclement weather may result in the uncontrollable loss of pesticides. Manufacturers have addressed this weakness by upgrading the system via the development of a new platform rather than the addition of feature-enhancing components, such as new control systems. To overcome the limitations of UASSs, their features must be improved to meet the needs of their users. For optimal performance, not only the quantity of pesticides required but target location requiring a specific treatment should be considered, and these must be fulfilled under any operating conditions or environment, indicating that the system must overcome all negative parameters for the UASS to complete the target as soon as possible before the spread of pests and diseases (Hanif et al., 2022).

The quality of effective spray width and overlap identifies the effectiveness of a UASS spray, as shown in the method in ISO 5682-1 (ISO (International Organization for Standardization), 2017a), and another indication is uniformity. The uniform distribution of pesticides on plants indicates a good spraying distribution and the safety of pesticide penetration on plants (Lv et al., 2019). The spray’ uniformity across the nozzle’s working width demonstrates an even distribution. This context has been demonstrated by the low coefficient of variation (CV), which has been reported that the maximum acceptable level of CV is 30% (Parkin and Wyatt, 1982; Richardson et al., 2004; Richardson et al., 2020). CV value obtained from overlapping between spray lines in the operation of spray distribution (Griesang et al., 2022). Nevertheless, a lower CV indicates a uniform and even distribution in agricultural spraying. The output is expected to meet the goal of spraying as uniformly as possible while using a minimum amount of pesticides.

The most common method for assessing these characteristics is by recording droplet deposition using water-sensitive paper (WSP) and conventional optical techniques to assess the droplet images of the WSP. The WSP can rapidly and easily calculate droplet deposition, coverage, and distribution (Zhu et al., 2010; Cerruto et al., 2019; Brandoli et al., 2021). However, it only qualitatively displays the occurrence of deposition, approximates the droplet size distribution portion, and cannot be used to investigate the dynamic. Cunha et al. evaluated the capability of the imaging systems of WSPs and found that most imaging systems cannot precisely measure the coverage density of droplets when the coverage rate exceeds 17%. Consequently, with the growing concern about the adoption and growth of UASS, it is essential to develop a specific standard method or equipment that can determine the spraying deposition pattern of UASS, either in target or off-target regions (Cunha et al., 2012).

This study conducted UASS tests in an indoor simulator under varied controlled conditions, such as wind effect and UASS operation conditions, focusing on developing a spraying control system for row crops. Data collection as a modeling dataset in the indoor simulator had also been adapted to the layout of row crops in the field, with the same planting distance and planting rows in accordance with the standard planting of row crops. The main objective was to develop a machine learning-based model for spraying distribution and characteristic data using the indoor simulator analysis results. Most studies used indoor simulators to analyze the spray distribution characteristics of the UASS type, nozzle type, or spraying scheme. The specific objectives of this study were: 1) analyzing UASS operational conditions, such as flight speed, flight height, rotor rotation, and even spraying pattern, and environmental factors, such as wind condition, as the parameters that need to be considered during UASS operation and assign their values as condition variations, 2) conducting simulations using indoor flight simulators under various conditions to analyze the spray distribution characteristics, and 3) modeling the control system logic using the dataset generated by the simulation in the mission to obtain a uniform spraying distribution.

Nozzle opening control systems to support the requirements of precision and uniform pesticide spraying in UASS were modeled based on the occurring parameters and are most influential in pesticide spraying operations. The determination of parameters was based on literature studies and UASS operating standard settings, both from nozzle type, pump pressure, pesticide dilution ratio, and spraying interval to UASS operating properties, such as flight speed, spray height, spraying pattern used, and fluid dynamics that occur, to produce sprays that must also meet standards, such as uniformity and precise quantity (Zhang et al., 2021). These conditions, particularly the UASS operating properties, cannot be determined based on user preferences. However, crop treatment operations must be adapted to the current needs in the field. In some cases, as reported by Martin et al. in 2019, the operation of UASS is highly influenced by weather conditions. They also analyzed the environmental conditions, particularly wind properties, that affect the results of the spray patterns and droplet spectra from UASS; thus, some are recommended to be kept from being operated in windy weather conditions (Martin et al., 2019), which are similar problems and phenomena that prompted this research.

The previously mentioned parameters were used as a reference in determining the parameters used in the simulation so that the environmental conditions can be adjusted according to the values of the predetermined parameters when using machine learning, where a dataset must be formed and used in the calibration and validation process. Obtaining CV values in this simulation requires a long data processing series and particular circumstances must also be discussed to validate this study. The rest of this chapter will go over the phenomena that occur during project development until there are variables that affect the output data. Furthermore, the discussion continues with the practical application of the results of this study to enable the evaluation of the practical application of the model to ensure its application in agriculture.

Various observations were made during the simulation, as shown in Figure 8D, where one of the simulations using the XR series nozzle produced an abnormal coverage value. A similar thing can happen in simulation when the operational conditions of the simulator are unstable. This specific kind of data can still be used because there is an R² value in modeling that indicates the accuracy of the prediction when the dataset is used, as well as the accuracy of other machine learning methods (Wen et al., 2019) and this type of phenomenon contributes to the development of the model while considering a situation that may occur in actual use.

The ANOVA results revealed that all the parameters in calculating the CV value of each spraying pattern and interval range were valid with a P-value of less than 0.05 (Wang et al., 2023), indicating that the parameters significantly affected the output results (i.e., CV value). Few parameters, whether used as a single value or combined, have a P-value above 0.05. This begs the question of whether such parameters should still be used or whether they can be eliminated. RPM and wind direction are the only parameters with a P-value greater than the default. First, the parameters are discussed to discover the facts in the field. Almost all parameter values, except the RPM value, can be set precisely in the simulator. The use of optical sensors in calculating the number of rotor rotations does help in monitoring. Still, the rotor rotation was controlled manually, so the possibility of error is considerably high. Moreover, the wind direction parameter, which has a high P-value, indicates that this parameter does not really affect the output value because wind direction is directly affected by the wind speed value. If the wind speed value is zero, any predetermined wind direction value will not affect the output CV value. According to the orthogonal matrix used for the simulation, there are three levels of wind speed parameter, one of which is 0 m/s. This value will have no effect regardless of the wind direction value, which directly reduces the influence of the parameter on the CV.

Initially, the modeling in this study was performed separately for both nozzle types and spraying patterns used in the processing coverage data. However, these two classifications can be combined by assigning identities with numbers, increasing the data used in modeling. Recognizing that the number of datasets used influences the accuracy of the resulting model, this merging step was performed to improve the accuracy. Furthermore, as various investigations were performed, several modeling methods are still used as a comparative aspect where nonlinear or machine learning techniques have performed better than linear methods. The regression equation generated from four nonlinear methods also exhibited different performances, so random forest regression was selected as the modeling method with the best performance and was used for the development of the system control.

Using the simulator as a data collection platform had many advantages for this study. The predetermined parameters could be achieved with considerable ease and precisely. On the other hand, testing directly in the field would result in random parameter values that are difficult to establish. Developing an independent UASS spraying control system using machine learning is also an exciting advancement in precision agriculture. Achieving uniform spray distribution is one of the most critical factors in effective aerial spraying. The fact that the system processes data of CV values below 30% indicates that the machine learning algorithm can detect and correct uneven spray distribution. In addition, the amount of pesticide sprayed is another priority for developing precision spraying. Both scenarios improve the effectiveness of spraying and reduce the waste of chemicals, thus saving costs for farmers.

Another significant advantage of using machine learning for spraying control is the ability to target all UASS types in the future. This indicates that the system will be adaptable to a wide range of UASS, making it accessible to more farmers and agricultural businesses. Additionally, the ability to adapt and learn from new data means that the system can be continually improved and evolved over time, ultimately leading to even better results. However, there are also some potential challenges and considerations that should be addressed. One concern is the need for consistent and accurate data input, as the performance of the machine learning algorithm will depend heavily on the quality and accuracy of the data it receives. There also might be regulatory and ethical considerations regarding using autonomous UASS for spraying, particularly regarding safety and potential environmental impacts. Overall, developing an independent UASS spraying control system using machine learning is a promising advancement in agriculture technology, and it has the potential to improve spraying effectiveness, reduce waste, and increase cost savings for farmers. As the system is completed and ready to be fully examined, field tests will be conducted to determine the control system’s performance in addressing environmental conditions and the operation of the UASS. However, carefully considering potential challenges and ethical implications will be important moving forward.

The overall objective of this study was answered during the experiments. The essential results of this study are briefly summarized in the following conclusion:

Spraying distribution data was collected under the simulator environment conditions with 72 simulations. Using the spraying pattern and interval from 1–9 m, 648 kinds of CV datasets were obtained for each pump opening. AI nozzle produced a better ability to overcome drift at 1–7 m intervals under several simulation conditions and exhibited a CV value of<30%. The TeeJet nozzle type XR exhibited a narrower interval width with a CV of<30%, which ranges from 1–3 m in some simulation conditions. The CV values obtained in the spraying pattern race track tend to be greater, ranging from 4 to 26% at a spray interval of 1 m using the AI series nozzle and 7 to 57% at the XR series TeeJet nozzle. In contrast, in the back-and-forth spraying pattern, the CV values obtained ranged from 1 to 24% at a spray interval of 1 m while using the AI series nozzle and 3 to 54% at a spray interval of 1 m while using the XR series TeeJet nozzle. The control system was modeled using machine learning and linear regression methods based on the CV datasets for each pump opening. Using nine levels of interval parameters and two spraying pattern types also helps increase the number of datasets. In this study, the random forest regression model achieved the best accuracy and intercept error value compared to other models: it exhibited an R² of 0.96 and RMSE of 0.04% for pump openings A and AB. This model could predict the CV value under parameters outside the modeling boundary. As confirmed in the simulation, the model can predict the CV value and make decisions on pump operation. The findings of this study confirmed the ability of random forest regression to develop a model for a functioning control system to establish an independent precision spraying control system.

The original contributions presented in the study are included in the article/Supplementary Material. Further inquiries can be directed to the corresponding author.

Written informed consent was obtained from the individual(s) for the publication of any potentially identifiable images or data included in this article.

Conceptualization, AH, and XH. Methodology, AH, XH, S-HY, CH, and SB. Analysis, AH, S-HY, CH, and SB. Visualization, AH, S-HY, CH, and SB. Supervision, XH. Reviewing the manuscript, XH, S-HY, CH, SB, C-GL, D-HL, and YK. All authors contributed to the article and approved the submitted version.