Design of intelligent sprayer control for an autonomous farming drone using a multiclass support vector machine

Arjon Turnip; Mohammad Taufik

doi:10.1515/opag-2022-0375

Artikel Open Access

Design of intelligent sprayer control for an autonomous farming drone using a multiclass support vector machine

Arjon Turnip und Mohammad Taufik

Veröffentlicht/Copyright: 24. Dezember 2024

Veröffentlicht von

Veröffentlichen auch Sie bei De Gruyter Brill

Manuskript einreichen Informationen für Autor*innen Erkunden Sie dieses Fachgebiet

Aus der Zeitschrift Open Agriculture Band 9 Heft 1

Abstract

The increasing need for agricultural commodities poses a serious threat to the cultivating ability of agricultural supplies. The use of autonomous farming drones to support the cultivation process is a rapidly growing trend. In an effort to improve efficiency and accuracy in watering using drones, this research strived to propose a control design for the drones using the multiclass support vector machine (MSVM) method. The proposed system was an improvement compared to the common approach of using constant watering strength in all drone conditions. By obtaining suitable watering strength based on the drone’s altitude, wind speed, and speed sensor data as the input, the optimal solution between water usage and water efficiency was expected to be achieved. An experimental trial that consisted of 12 flights was conducted to acquire a data set with 3,750 data. The results of classification with MSVM obtained an accuracy of 90.82%. The efficiency of using the water resource and the accuracy of delivering the correct amount of water based on the drone condition were achieved. These results show that the proposed technology has great potential for using drones as an automatic watering system in agriculture.

Keywords: autonomous drone; intelligent sprayer; machine learning; multiclass support vector machines

1 Introduction

Agriculture is an important sector in the Indonesian economy, where most of the population depends on agricultural products for their livelihood [1]. Population growth in Indonesia has now reached 278.69 million people by mid-2023; this figure is up 1.05% from the previous year [2]. Thus, there will be an increase in the need for food supply and land to cultivate agricultural products. The prevalence of insufficiency in Indonesia’s food consumption in 2022 has a value of 10.21% or an increase of 1.72% from the previous year [3]. Moreover, the rate of conversion of agricultural land to other uses continues to increase, and opportunities to expand the area on conventional agricultural land are considered very limited [4]. Increased consumption of fertilizers and pesticides accompanied by intensification of agricultural activities may lead to future environmental challenges [5,6].

In recent years, the agricultural industry has seen a significant shift toward autonomous farming practices, driven by advancements in technology such as drones and artificial intelligence (AI). One critical aspect of autonomous farming is efficient and precise pesticide application to optimize crop yield while minimizing environmental impact. Traditional spraying methods often lack precision and can lead to the overuse of chemicals, resulting in increased costs and potential harm to the environment. To address these challenges, researchers are increasingly turning to intelligent sprayer control systems that utilize machine learning algorithms to accurately target areas in need of treatment. So an innovative approach based on modern technology is needed, which can increase production efficiency and affordable food quality, improve environmental sustainability, and improve the welfare of farmers and the sectors that support it [7,8]. Drones, also referred to as unmanned aerial vehicles (UAVs), are devices that can fly with a predetermined path on autopilot and GPS coordinates [9]. UAVs have undergone great development in recent decades. In the agricultural sector, drones have changed agricultural practices by providing significant cost savings, increased operational efficiency, and better profits for farmers. Over the past few decades, the topic of agricultural drones has attracted significant academic attention [5].

Traditional methods used for spraying pesticides and fertilizers take longer and are less effective [10]. So spraying using drones is applied in the field which increases coverage capabilities, increases the effectiveness of chemicals, and makes spraying work easier and faster [11,12]. However, in the application of the sprayer system on drones, determining the intensity of spraying becomes an important variable in providing efficient spraying results. A commonly used method is to classify drone altitude, wind speed, and speed data to determine watering strength [13]. The accuracy of the classification varies for each method used, so it is necessary to choose the right method for the nature of the dataset to be used.

Research on UAVs in agricultural contexts has been done before. One of them, UAVs are used to map land suitable for agriculture [14]. Where this mapping is done using image recognition which is then processed using machine learning random forest (RF) and deep learning convolutional neural network (CNN). As a result, the RF method provides accuracy (≈85%) and CNN (≈90%) [15]. The use of drones allows for more efficient and flexible mapping compared to traditional methods. Then, another study also implemented a UAV with an AI-based sprayer system to measure liquid droplet deposition to eradicate pests on plants. The results showed that the artificial neural networks training variation was 0.003, so the model was stable and reliable. Field tests show that the error between predicted droplet deposition and actual droplet deposition is less than 20% [16]. These studies indicate the potential of AI to improve accuracy [17,18,19,20]. However, it should be noted that the use of AI in the study above requires complex sensors to obtain and process data. So a method is needed that can use common devices but with different processing.

Research using a drone sprayer that controls the water pump through the use of Arduino UNO and is connected directly to the pump motor is commonly carried out. The PIXHAWK flight controller provides data in the form of coordinates obtained from sensing in the field, and the Arduino UNO processes the data to determine the right time to activate the pump motor. However, by using only Arduino UNO, it is not possible to regulate the intensity of the flow of sprayed water [21]. Results from previous studies have shown that the use of drones in the agricultural sector can improve the efficiency and accuracy of watering crops. However, currently controlling the intensity of watering still cannot be done dynamically, resulting in a waste of resources. Therefore, it is necessary to develop a watering control system that is able to adapt dynamically. Machine learning methods are often used in data processing as feedback for watering systems [22,23]. However, the use of machine learning today still relies heavily on complex sensors such as cameras, so there are limitations between hardware and software.

As the global population continues to grow, there is an urgent need to enhance agricultural productivity and sustainability. Autonomous farming technologies, such as drones equipped with intelligent sprayer control systems, hold great promise in meeting this demand. However, to realize their full potential, these systems must be optimized for accuracy, efficiency, and adaptability. To overcome the bottleneck, a more efficient and practical watering control method is proposed by using only existing sensors in the form of altitude, wind speed, and drone speed sensors to determine the strength of watering. By utilizing data from these sensors, the watering control system can be set adaptively and dynamically, without relying on the limitations of complex sensors. The proposed method for processing such data is multiclass support vector machine (MSVM). By using MSVM, watering control can be done dynamically and adaptively based on data generated by existing sensors. MSVM can learn patterns and relationships from collected data to predict optimal watering needs. This enables saving water resources, reducing wastage, and increasing efficiency in the agricultural sector. The development of a reliable MSVM-based intelligent sprayer control for autonomous farming drones represents a critical step toward achieving these goals. Urgent research in this area is essential to accelerate the deployment of advanced agricultural technologies, mitigate environmental risks, and ensure food security for future generations.

This manuscript is organized as follows. First, the hardware used in constructing the smart sprayer system is presented. Then, the method to find the optimal solution of sprayer strength, MSVM, is elaborated. Furthermore, the implementation results are presented, emphasizing the obtained accuracy of the proposed method in determining the actuator setting for the sprayer system. The correlation among the inputs is first presented, followed by the final classification of the actuator setting by using MSVM. The manuscript is finalized with a conclusion.

2 Methods

This study was conducted at the Field of Basic Science Service Center, Padjadjaran University. Field trials involved creating an illustration resembling agricultural land. Data were collected from 12 drone flight tests, encompassing 4 distinct trajectory types and 3 altitudes. Throughout the test flights, five variables were gathered: altitude, wind speed, drone velocity, latitude, and longitude.

2.1 Smart sprayer system hardware

In this study, drones were equipped with an F450 frame and operated using the Pixhawk 2.4.6 Flight Controller, responsible for managing flight and landing operations. To track the drone’s location during flight, a GPS Module Radiolink M8N SE100 was installed onboard. The Raspberry Pi 4 served as the microcontroller, executing Python applications remotely. Additionally, DC motors were utilized for the watering system. The Pixhawk stored crucial flight data such as altitude, drone position, and waypoints. Calibration and configuration of the drones were facilitated through the Mission Planner application, which enabled users to define flight paths and waypoints. Figure 1 illustrates the hardware block diagram.

Figure 1

Intelligent sprayer system on autonomous drone hardware block diagram. Source: Own creation.

Communication between the flight controller and microcontroller occurred via telemetry and the Micro Air Vehicle Link (Mavlink) protocol, allowing the Raspberry Pi to receive diverse drone data from the Pixhawk. Furthermore, the Raspberry Pi controlled the watering system’s power through the DC motor driver. Communication between the Raspberry Pi and the motor leveraged the RPI GPIO communication protocol, integrating the GPIO port within the Python program. Water distribution was achieved through hoses and nozzles connected to the water storage. Upon activation, water flowed through the nozzles for watering purposes. The Raspberry Pi regulated watering intensity by adjusting the voltage on the motor driver, utilizing altitude, wind speed, and drone speed measurements to control watering strength. Additionally, data regarding the drone’s waypoints influenced watering duration. Figure 2 provides an overview of the operational mechanism of the intelligent drone sprayer.

Figure 2

Intelligent sprayer system on autonomous drone flowchart.

2.2 MSVM: One-vs-all

The multiclass fine Gaussian support vector machine (SVM) method is a sophisticated approach employed for data classification in multiclass scenarios. This method combines SVM principles with the Gaussian kernel to generate precise classification models. Its efficacy has been demonstrated across diverse domains including pattern recognition, image processing, and bioinformatics. Essentially, SVM is a machine learning technique rooted in the concept of linear separation to categorize data. It identifies hyperplanes that optimize the separation between distinct classes within the feature space.

(1) f ( x ) = | w × x + b | ,

where f(x) is a decision function that classifies data x into either class (+1 or −1), x is the vector of the data feature you want to classify, w is the weight vector that is the normal vector of the hyperplane, and b is a hyperplane bias or shift. In binary SVM, we have two classes that we want to separate. However, in a multiclass problem, we have more than two classes. To solve the multiclass problem, the multiclass fine Gaussian SVM method adopts a one-vs-one or one-vs-all approach. The one-vs-one approach involves creating an SVM model for every possible pair of classes, while the one-vs-all approach involves creating an SVM model for each class against all other classes. In both approaches, each SVM model produces a decision function that classifies data into specific classes. In the one-vs-all method, multiple binary SVM models are trained by comparing each class to another. For example, if there are three classes, there will be three binary SVM models trained to classify each class against the other classes. The fine Gaussian kernel function is used to map data into higher dimensions of feature space, so SVM can better classify nonlinear data as in Figure 3.

Figure 3

Visualization of one-vs-all classification method (red square is class 1, blue triangle is class 2, and green circle is class 3).

Furthermore, the multiclass fine Gaussian SVM method uses the Gaussian kernel to transform data into a higher feature space. Gaussian kernels, also known as radial basis function kernels, map data into infinite-dimensional space using radial base functions. The Gaussian kernel describes the relationship between each data pair in a new feature space by using Gaussian functions. This allows SVM to handle non-linear data effectively. The function can be expressed by the following formula:

(2) f ( x ) = | ∑ ( a i × y i × K ( x i , x ) | + b ,

where a i is the Lagrange coefficient obtained during model training, y i is the class label of the training data, and K ( x i , x ) is the result of a fine Gaussian kernel between the training data x i and the test data x.

Mathematically, the binary SVM model in the one-vs-all approach can be defined whereby any ith class, the problem is divided into two classes: the ith class is considered a positive class (+1) and all other classes are considered a negative class (−1). Then, the binary SVM model is trained to classify the data into an ith class based on those settings (Figure 4). In SVM, the decision function for a binary SVM model can be expressed as follows:

(3) f i ( x ) = | w i × x + b i | ,

where f i ( x ) is the decision function for the ith class binary SVM model and w i is the weight vector.

Figure 4

MSVM algorithm.

3 Results

The dataset, comprising sensor measurements including altitude, wind speed, drone speed, and GPS error, consists of 3,750 entries. It is divided into 6 classes, each denoted by values 0 through 5, representing conditions ranging from “off” to “very high.” Each class contains an equal number of entries (625), ensuring data balance to prevent bias during model training. Imbalanced training data can skew the model toward classes with greater representation, leading to inaccurate predictions and reduced performance. Figure 5 illustrates the sensor measurements captured by the drone’s smart sprayer, while a detailed description of the dataset is provided in Table 1. Various methods exist for determining the classification range within a dataset. However, in this experiment, the box plot method was employed. Utilizing box plot enables visualization and adjustment of data ranges based on five statistical summaries: minimum value, first quartile, median, third quartile, and maximum value, ensuring greater accuracy in classification. The conversion of data into a box chart is presented in Figure 6, offering insight into the dataset’s distribution.

Figure 5

Data from sensor measurements on autonomous drones.

Table 1

Description of smart spraying drone dataset

	Altitude (m)	Wind speed (m/s)	Velocity (m/s)	GPS error (%)
Count	3,750	3,750	3,750	3,750
Mean	3.8	3.73	3.01	3.74
Std	2.24	2.21	1.74	3.3
min	0.004	0.002	0.003	0.002
25%	1.87	1.88	1.5	1.5
50%	3.75	3.64	3.02	2.97
75%	5.6	5.5	4.54	4.45
Max	8	8	6	15

Figure 6

Transforming data plot into a box chart.

Each variable in the dataset correlates with other variables. Correlations between variables can be classified into two types, namely negative correlation when the value is less than 0, and positive correlation when the value is more than 0. If the correlation value between two variables is 0, then the variables are not correlated. The level of correlation can be seen through the level of darkness or brightness on the heatmap. The correlation between these variables is shown in the heatmap in Figure 7. In the smart sprayer algorithm that will be used in this system, rules have been compiled based on a dataset that has been prepared beforehand. This dataset includes 250 different conditions that will produce an output of when the watering force will be in a predefined category, i.e., “off,” “very low,” “low,” “medium,” “high,” or “very high.” This rule aims to correlate the variables in the dataset with the appropriate watering strength category. Rules are created based on parameters such as altitude, wind speed, velocity, and GPS error. The rules are created based on the exact linguistic values or categories for each parameter. Table 2 lists the rules that relate the combination of parameter values with the desired watering strength category. These rules provide clear guidance for the smart sprayer system to determine the optimal watering strength based on the variables measured.

Figure 7

Correlation between variables using heatmap.

Table 2

Rules for every parameter

No	Altitude	And	Windspeed	And	Velocity	And	Error	Then	Watering power
1	Very_low	&	Very_low	&	Very_low	&	Low	=	out1mf1
2	Very_low	&	Very_low	&	Very_low	&	High	=	out1mf2
3	Very_low	&	Very_low	&	Low	&	Low	=	out1mf3
4	Very_low	&	Very_low	&	Low	&	High	=	out1mf4
5	Very_low	&	Very_low	&	Medium	&	Low	=	out1mf5
6	Very_low	&	Very_low	&	Medium	&	High	=	out1mf6
…	…		…		…		…		…
245	Very_high	&	Very_high	&	Medium	&	Low	=	out1mf245
246	Very_high	&	Very_high	&	Medium	&	High	=	out1mf246
247	Very_high	&	Very_high	&	High	&	Low	=	out1mf247
248	Very_high	&	Very_high	&	High	&	High	=	out1mf248
249	Very_high	&	Very_high	&	Very_high	&	Low	=	out1mf249
250	Very_high	&	Very_high	&	Very_high	&	High	=	out1mf250

In this study, an MSVM algorithm was created to process the dataset that had been prepared. The features of the dataset as well as the prediction targets are extracted and stored in different variables. Furthermore, the data are divided into training and testing sets in an 80:20 ratio, where 80% of the data is used to train the model and 20% of the data is used to test the performance of the model. In the SVM algorithm, adjusted hyperparameters are used to achieve optimal results. One important hyperparameter is the selection of kernel functions. In this case, a fine Gaussian kernel function is used to generate an SVM model. Fine Gaussian kernel functions can help map data into higher feature spaces, allowing for better separation between existing classes. In addition, in this SVM model, a multiclass one-vs-all mode is also used which allows classifying each subset of data against other subsets. Next is a scatter plot which is a type of data visualization used to show the relationship between two numerical variables. In a scatter plot, each point is plotted on a graph that represents the value of both variables. The scatter plot in Figure 8 depicts the data before it is classified where each dot of a different color describes the six classes.

Figure 8

Scatter plot data after pre-processing.

In the performance evaluation of the model, the Confusion matrix is used. This confusion matrix measures model performance using several metrics, including precision or positive predictive value, recall or true positive rate (TPR), F1 score or harmonic mean of precision and recall, support, false positive rate (FPR), and false discovery rate (FDR). Precision describes the extent to which the model can correctly predict a positive class. Recall measures the model’s ability to identify all positive samples. F1 score is a metric that combines precision (precision) and recall. Precision measures the accuracy of a model’s predictions on a positive class, while recall measures the model’s ability to identify all positive samples. F1 score provides an overall measurement of model accuracy that balances precision and recall.

Support refers to the volume of data tested within a specific class. FPR denotes the ratio of negative actual cases incorrectly identified as positive by the model. Meanwhile, FDR represents the proportion of false positive predictions to the total positive predictions made by the model. In our model, multiclass SVM exhibits an impressive accuracy of 90.82%. The comprehensive breakdown of results, including the confusion matrix (Figure 9) and other metric measurements (Table 2), offers insights into the model’s performance across various criteria. Table 2 presents a range of metrics associated with the classification model’s performance across multiple classes. Notably, overall accuracy is notably high, ranging between 95.05 and 99.83%. This suggests the model effectively categorizes instances across all classes. However, relying solely on accuracy may not provide a complete understanding, particularly with imbalanced datasets. Precision measures the accuracy of positive predictions made by the model, calculated as the ratio of true positives to the sum of true positives and false positives. Precision values range from 0.83 to 0.99, indicating generally accurate positive predictions by the model (Table 3).

Figure 9

The confusion matrix output of multiclass SVM.

Table 3

Evaluation matrix calculation results based on confusion matrix

Class	n (truth)	n (classified)	Accuracy (%)	Precision	Recall	F1 score
1	97	98	99.83	0.99	1.0	0.99
2	92	98	97.62	0.90	0.96	0.93
3	101	98	97.11	0.93	0.90	0.91
4	101	98	97.11	0.87	0.84	0.85
5	99	98	95.05	0.83	0.82	0.82
6	98	98	97.96	0.94	0.94	0.94

Recall, also known as sensitivity or TPR, assesses the model’s ability to capture all relevant instances of a class. It is the ratio of true positives to the sum of true positives and false negatives. Recall values range from 0.82 to 1.0, with some classes achieving perfect recall (1.0), indicating the model’s effectiveness in identifying instances. The F1 score, a harmonic mean of precision and recall, offers a balanced measure considering both false positives and false negatives. With F1 score values ranging from 0.82 to 0.99, the model demonstrates a good balance between precision and recall across most classes. Overall, the model exhibits strong performance, characterized by high accuracy and balanced precision and recall. Notably, class 1 stands out with exceptionally high precision, recall, and F1 score, indicative of very accurate and comprehensive predictions. Although class 5 shows slightly lower precision and recall, it remains within an acceptable range. Classes 2–4 demonstrate solid overall performance with high accuracy and balanced precision and recall. Further evaluation of the model should consider the specific context of the classification problem and any unique requirements or constraints. Additionally, analyzing the confusion matrix and evaluating the potential impact of false positives and false negatives on the application can offer valuable insights for refinement and optimization.

4 Conclusion

The development of an intelligent sprayer on autonomous farming drones utilizing MSVM: one-vs-all for predicting watering forces has been successfully achieved. Through experimentation with a dataset comprising 3,750 entries divided into 6 classes of 625 entries each, a comprehensive analysis of metrics including accuracy, precision, recall, and F1 score has been conducted. The findings indicate that the multiclass SVM model exhibits strong performance in classifying instances across diverse classes, with overall accuracy ranging impressively from 95.05 to 99.83%. However, acknowledging the limitations of accuracy, particularly in imbalanced datasets, the examination of precision, recall, and F1 score provides a nuanced evaluation of the model’s efficacy. Precision values ranging from 0.83 to 0.99 underscore the model’s accuracy in predicting positive classes, while recall values spanning from 0.82 to 1.0 highlight its effectiveness in capturing relevant instances across all classes. Furthermore, the F1 score, representing a harmonized measure of precision and recall, reinforces the model’s commendable performance, with values ranging between 0.82 and 0.99. A class-specific analysis reveals notable strengths, notably the exceptional precision, recall, and F1 score of class 1, indicative of accurate and comprehensive predictions. Although class 5 exhibits slightly lower precision and recall, its performance remains within an acceptable range. Moreover, classes 2–4 demonstrate robust overall performance, characterized by high accuracy and balanced precision and recall. To gain deeper insights into the model’s practical applicability, it is essential to consider the specific context of the classification problem and the potential implications of false positives and false negatives. By doing so, further optimization and refinement of the model can be pursued, ensuring its efficacy and reliability in real-world autonomous farming scenarios.

Acknowledgments

This research was funded by Riset Percepatan Lektor Kepala (RPLK) No.: 1549/UN6.3.1/PT.00/2023 and supported by the Department of Electrical Engineering, Universitas Padjajaran, Indonesia.

Author contributions: All authors have accepted responsibility for the entire content of this manuscript and consented to its submission to the journal, reviewed all the results and approved the final version of the manuscript. AT: conceptualization, formal analysis, investigation, methodology, software, supervision, validation, visualization, writing – original draft, writing – review and editing with the contribution of all co-authors. MT: data curation and methodology, resources, and project administration.
Conflict of interest: Authors state no conflict of interest.
Data availability statement: Data will be made available on request.

References

[1] Noor FA, Nasution DD. BPS: Agricultural sector growth still below pre-pandemic levels. 7 February 2023. https://ekonomi.republika.co.id.Suche in Google Scholar

[2] Central Bureau of Statistics. Mid-year population (Thousand People), 2021–2023. Central Bureau of Statistics, 2023. https://www.bps.go.id/indicator/12/1975/1/jumlah-penduduk-pertengahan-tahun.html [Accessed 23 October 2023].Suche in Google Scholar

[3] Central Bureau of Statistics. Prevalence of food consumption insufficiency (Percent), 2020–2022. 2022. https://www.bps.go.id/indicator/23/1473/1/prevalensi-ketidakcukupan-konsumsi-pangan.html [Accessed 23 October 2023].Suche in Google Scholar

[4] Purnomo D. Food needs, agricultural land availability, and plant potential. Semarang, Indonesia: Sebelas Maret University Library; April 6, 2016. https://library.uns.ac.id/kebutuhan-panganketersediaan-lahan-pertanian-dan-potensi-tanaman/ [Accessed 23 10 2023].Suche in Google Scholar

[5] Rejeb A, Abdollahi A, Rejeb K, Treiblmaier H. Drones in agriculture: A review and bibliometric analysis. Computers Electron Agric. 2022;198(107017):1–19.10.1016/j.compag.2022.107017Suche in Google Scholar

[6] Apushev AK, Yussupov B, Salybekova NN, Mamadaliev A. Biomorphological analysis of tulip varieties on substrates in covered ground in Turkestan. J Human Earth Future. 2023;4(2):207–20.10.28991/HEF-2023-04-02-06Suche in Google Scholar

[7] Coordinating Ministry of Economic Midwives of the Republic of Indonesia. The government encourages the improvement of the food and agriculture sector for the welfare of the Indonesian people. Coordinating Ministry of Economic Midwives of the Republic of Indonesia, November 16, 2020. https://www.ekon.go.id/publikasi/detail/647/pemerintah-dorong-peningkatan-sektor-pangan-dan-pertanian-untuk-kesejahteraan-masyarakat-indonesia. [Accessed 23 October 2023].Suche in Google Scholar

[8] Cornejo J, Barrera S, Ruiz CH, Gutiérrez F, Casasnovas MO, Kot L, et al. Industrial, collaborative and mobile robotics in Latin America: Review of mechatronic technologies for advanced automation. Emerg Sci J. 2023;7(4):1430–58.10.28991/ESJ-2023-07-04-025Suche in Google Scholar

[9] Ahirwar S, Swarnkar R, Bhukya S, Namwade G. Application of drone in agriculture. Int J Curr Microbiol Appl Sci. 2019;8(1):2500–5.10.20546/ijcmas.2019.801.264Suche in Google Scholar

[10] Mrema G, Soni P, Rolle RS. A regional strategy for. Bangkok: Food and Agriculture Organization of the United Nations; 2014.Suche in Google Scholar

[11] Hafeez A, Husain MA, Singh S, Chauhan A, Khan MT, Kumar N, et al. Implementation of drone technology for farm. Inf Process Agric. 2023;10:192–203.10.1016/j.inpa.2022.02.002Suche in Google Scholar

[12] Umeda S, Yoshikawa N, Seo Y. Cost and workload assessment of agricultural drone sprayer. Sustainability. 2022;14(10850):2–11.10.3390/su141710850Suche in Google Scholar

[13] Shaikh TA, Rasool T, Lone FR. Towards leveraging the role of machine learning and artificial intelligence in precision agriculture and smart farming. Computers Electron Agric. 2022;198(107119):2–29.10.1016/j.compag.2022.107119Suche in Google Scholar

[14] Jia X, Cao Y, Zhu J, Tsang DC, Zou B, Hou D. Mapping soil pollution by using drone image recognition and machine. Environ Pollut. 2021;270(116281):2–10.10.1016/j.envpol.2020.116281Suche in Google Scholar PubMed

[15] Bhatnagar S, Gill L, Ghosh B. Drone image segmentation using machine and deep. Remote Sens. 2020;12(2602):2–26.10.3390/rs12162602Suche in Google Scholar

[16] Wen S, Zhang Q, Yin X, Lan Y, Zhang J, Ge Y. Design of plant protection UAV variable spray. Sensors. 2019;19(1112):2–23.10.3390/s19051112Suche in Google Scholar PubMed PubMed Central

[17] Ponnusamy V, Natarajan S. Precision agriculture using advanced technology of IoT, Unmanned aerial vehicle, augmented reality, and machine learning. Dalam smart sensors for industrial Internet of Things. Cham: Springer; 2021. p. 207–29.10.1007/978-3-030-52624-5_14Suche in Google Scholar

[18] Suhaeni S, Wulandari E, Turnip A, Deliana Y. Factors influencing green, environmentally-friendly consumer behaviour. Open Agric. 2024;9(1):20220269.10.1515/opag-2022-0269Suche in Google Scholar

[19] Turnip A, Pebriansyah FR, Simarmata T, Sihombing P, Joelianto E. Design of smart farming communication and web interface using MQTT and Node.js. Open Agric. 2022;8(1):1–12. 10.1515/opag-2022-0159.Suche in Google Scholar

[20] Andayani SA, Umyati S, Dinar, Tampubolon GM, Ismail AY, Dani U, et al. Prediction model for agro-tourism development using adaptive neuro-fuzzy inference system method. Open Agric. 2022;7:644–55.10.1515/opag-2022-0086Suche in Google Scholar

[21] Duttaroy S, Nemade MU, Devgonde S, Parmar V, Gounder D. Efficient Field Monitoring Autonomous Drone (FMAD) with pesticide. Int Res J Eng Technol. 2021;8(4):2741–6.Suche in Google Scholar

[22] Benos L, Tagarakis AC, Dolias G, Berruto R, Kateris D, Bochtis D. Machine learning in agriculture: A comprehensive updated review. Sensors. 2021;21(11):3758.10.3390/s21113758Suche in Google Scholar PubMed PubMed Central

[23] Khan S, Tufail M, Khan MT, Khan ZA, Iqbal J, Wasim A. Real-time recognition of spraying area for UAV sprayers using a deep learning approach. PLOS One. 2021;16(4):1–17.10.1371/journal.pone.0249436Suche in Google Scholar PubMed PubMed Central

Received: 2024-01-13

Revised: 2024-09-29

Accepted: 2024-10-03

Published Online: 2024-12-24

This work is licensed under the Creative Commons Attribution 4.0 International License.

Artikel in diesem Heft

https://doi.org/10.1515/opag-2022-0375

Schlagwörter für diesen Artikel

autonomous drone; intelligent sprayer; machine learning; multiclass support vector machines

Creative Commons

BY 4.0