Current Status and Future Prospects of Key Technologies in Variable-Rate Spray

Abstract

The traditional continuous, quantitative spraying technology ignores the severity of pests, diseases and grasses, spatial distribution and other differences, resulting in low effective utilization of pesticides, environmental pollution and other problems. Variable-rate spray technology has become an important development direction in the field of precision agriculture by dynamically sensing crop canopy morphology, pest and disease distribution, and environmental parameters, adjusting the application amount in real time, and significantly improving pesticide utilization. In this study, we systematically review the core progress of variable-rate spray technology; focus on the technical system of information detection, spray volume model, and control system; analyze the current bottlenecks; and propose an optimization path to adapt to the complex agricultural conditions. At the level of information perception, LiDAR, machine vision, and multi-source sensor fusion technology constitute the main perception architecture, and infrared and ultrasonic sensors assist target recognition in complex scenes. In the construction of the spray volume model, models based on canopy volume, leaf area density, etc., are used to realize dynamic application decision by fusing equipment operating parameters, pest and disease levels, meteorological conditions, and so on. The control system takes the solenoid valve + PID control as the core program, and improves the response speed through PWM regulation and closed-loop feedback. The current technical bottlenecks are mainly concentrated in the sensor dynamic detection accuracy, model environmental adaptability, and the reliability of the execution parts. In the future, it is necessary to further promote anti-jamming multi-source heterogeneous sensor data fusion, multi-factor adaptive spray model development, lightweight edge computing deployment, and solenoid valve structural parameter optimization and other technical research, with a view to promoting the application of variable-rate spray technology to the field on a large scale and providing a theoretical reference and technological support for the green transformation of agriculture.

1. Introduction

In 2022, the total global use of agricultural pesticides reached 3.7 million tons (based on active ingredients), an increase of 4% compared to 2021, with pesticide usage continuing to rise year by year. Traditional continuous quantitative spraying technology remains dominant globally (especially in developing countries). This spraying method ignores differences in the severity and spatial distribution of pests, diseases, and weeds, making it the primary cause of pesticide overuse and low utilization efficiency [1,2,3]. Currently, pesticide utilization rates in developing countries generally range from 35% to 45%. Globally, approximately 20% of pesticides are released into non-target environments annually, leading to soil degradation, reduced biodiversity, and water pollution [4,5,6].

In recent years, precision agriculture has integrated traditional agriculture with information technology to achieve intelligent management of agricultural operations. Among these technologies, variable-rate spray is one of the key technologies driving the development of precision agriculture [7,8,9]. It collects data on pest and disease information, growth status, and other parameters of target crops, and combines this with parameters such as the location, travel speed, and spray pressure of the application equipment to apply pesticides to target crops on an as-needed basis. As a green pest control method that significantly improves pesticide application efficiency, variable-rate spray technology effectively reduces pesticide residue risks, minimizes negative impacts on the ecological environment, and promotes pesticide reduction and efficiency enhancement. It has become a research hotspot in the field of agricultural engineering in recent years [10,11,12,13,14,15].

The complete workflow of variable-rate spray technology encompasses three key stages: target information sensing, spray decision generation, and precise execution. This paper systematically reviews the research progress in information sensing methods, spray volume models, and control execution systems. Beyond summarizing these technical advances, the overarching goal of this review is to identify the main technical bottlenecks—namely, limited dynamic sensing accuracy, insufficient model generalization ability, and inadequate actuator reliability—and to evaluate potential pathways for optimization. Based on this analysis, this paper proposes optimization suggestions tailored to the development of smart agriculture, aiming to provide theoretical support for the construction of high-precision, robust variable-rate spray systems and promote their large-scale application in the field.

2. Materials and Methods

This systematic review followed a structured methodological approach to maintain rigor and reduce potential biases. The procedure encompassed pinpointing pertinent literature, filtering records according to pre-established criteria, and performing a comprehensive full-text evaluation to decide upon final study inclusion. This methodology emphasized transparency and replicability at every stage of the review.

A systematic literature search was conducted using the Scopus database, covering the period from 1995 to 2025. A comprehensive set of keywords related to variable-rate spray technology was applied, and records were screened based on relevance to information detection, spray volume modeling, and control systems. After inclusion and exclusion criteria were applied, a representative set of studies was selected to form the basis of this review. The detailed search procedure and flowchart are provided in the Supplementary Material.

3. Information Detection

Accurate perception of target characteristics is a prerequisite for variable-rate spray, and its detection accuracy directly affects the quality of spraying operations. Modern sensors, with their high-resolution detection, fast response, and real-time data processing capabilities, have been widely used in variable-rate spray systems. The current mainstream sensors include laser scanning sensors, infrared sensors, ultrasonic sensors, radar, and machine vision. From a temporal perspective, the development of sensing technologies for variable-rate spraying has followed a clear evolutionary trend. Early studies in the late 1990s and early 2000s primarily relied on ultrasonic and infrared sensors due to their low cost and simplicity, although their accuracy was strongly influenced by environmental factors. With advances in computing power and laser technology during the 2010s, LiDAR and machine vision methods became increasingly prominent, offering richer canopy structure information and higher spatial resolution [16,17]. More recently, research has shifted toward multi-sensor fusion approaches that integrate LiDAR, vision, and other modalities to compensate for the weaknesses of single sensors and to improve robustness under complex field conditions. This thematic evolution underscores the trajectory from simple, low-cost sensing toward integrated, data-rich solutions that can better support precision agriculture.

3.1. Laser Scanning Sensor

Laser scanning sensors accurately obtain the distance and shape of objects by emitting laser beams and measuring the time it takes for them to be reflected back, thereby achieving high-precision spatial measurement. Table 1 summarizes representative studies on laser scanning sensors for canopy detection and volume estimation, showing their high precision but limited adaptability under dynamic and complex canopy conditions. Figure 1 shows some laser scanning sensor setups. Laser scanning sensors are mainly used in variable-rate spray systems for real-time detection of plant canopy structure and to obtain data such as target volume, density, and spatial distribution [18,19,20]. Currently, laser scanning sensors generally feature high resolution, fast detection cycles, and wide scanning ranges. However, they still face the following challenges: detection accuracy is significantly affected by distance, dynamic performance is limited, and adaptability to complex targets is insufficient [21,22].

Table 1. Application of laser scanning sensors in variable-rate spray systems.

Reference	Year	Model	Detection Performance
[18]	2012	LMS291 (SICK, Inc., Waldkirch, Germany), 0–180° scanning angle, angular resolution 0.5°, detection cycle 26 ms.	The model was used to detect the canopy structure of four tree species, including Douglas fir and western hemlock. When the machine is traveling at a speed of 0.89–1.78 m/s, the detection width is 0.023–0.046 m. Increasing the speed will result in a decrease in detection accuracy.
[19]	2016	UTM-30LX (Hokuyo Automatic Co., Ltd., Osaka, Japan), 0–270° scanning angle, angular resolution 0.25°, detection cycle 25 ms, maximum detection distance 30 m.	Within a detection range of 0.58–2.94 m, the average coefficient of variation for detecting spheres in the X/Y/Z directions is 6.1/8.6/5.3%, respectively, and the image similarity between the sensor and the camera is greater than 0.85. The detection error of the sensor in the Z direction increases with the detection distance, the resolution in the X direction is affected by the travel speed of the sprayer, and the sensor has a 90° blind spot on the ground.
[20]	2017	LMS511 (SICK, Inc., Germany), scanning angle 30–150°, angular resolution 0.5°, detection cycle 20 ms.	The maximum relative errors for detecting the cross-sectional volumes of rectangular prisms, triangular prisms, and cylinders were 3.3%, 7.9%, and 9.4%, respectively. The coefficients of variation for detecting the total volumes of pine trees and apple trees were 0.078 and 0.041, respectively. The sensor had a high detection error for irregular shapes (cylinders, triangular prisms) and was dependent on the preset tree row plane distance parameters.
[21]	2018	UST-10LX (Hokuyo Automatic Co., Ltd., Japan), 0–270° scanning angle, angular resolution 0.25°, detection cycle 25 ms, maximum detection distance 10 m.	The average root mean square error (RMSE) and coefficient of variation (CV) of the sensor in the X direction were 0.083 m/35.8% (toy ball), 0.081 m/44.4% (basketball), 0.057 m/25.3% (cardboard box), and 0.0606 m/51.5% (cylinder), respectively. The RMSEs of the two artificial plants (0.231 m and 0.254 m wide) in the X-direction were 0.075 m and 0.07 m, respectively; the average RMSEs in the Y-direction and Z-direction were 0.023 m/9.3% and 0.008 m/6.2% (toy ball), respectively. The measurement accuracy of the sensor in the X direction is generally low and decreases significantly with increasing target distance. Object color and texture do not significantly affect accuracy, but complex canopy structures may increase the difficulty of edge detection.
[22]	2019	UST-10LX (Hokuyo Automatic Co., Ltd., Japan), 0–270° scanning angle, angular resolution 0.25°, detection cycle 25 ms, maximum detection distance 10 m.	Used to detect objects of different shapes (toy balls, basketballs, rectangular boxes, cylinders) and two types of artificial plants with different canopy structures. The detection range is 0.033 to 0.083 m before reaching the target object and 0.013 to 0.084 m after passing the object. When moving at high speed, the spatial sampling density of the sensor decreases, affecting the recognition accuracy of small targets.

Figure 1. Diagram of devices equipped with laser scanning sensors. (a) The variable-rate sprayer based on the LMS291 laser scanning sensor [18]; (b) the position of the laser sensor on the vehicle [19]; (c) a structure diagram of the laser-scanning system [20]; (d) the 270° radial range laser sensor mounted on a constant-speed track [21,22].

3.2. Infrared Sensor

Infrared sensors detect the presence of objects within their range by emitting and receiving high-intensity light pulses. They are inexpensive and fast-responding, and have been widely used for canopy detection, gap identification, and target positioning in orchard applications [23,24]. Table 2 presents applications of infrared sensors in orchard canopy detection and gap identification, highlighting their cost-effectiveness but also their sensitivity to environmental interference. Figure 2 presents examples of infrared sensor configurations. Infrared sensors are mainly used for fruit tree canopy detection, gap identification, and target positioning. However, their performance is significantly constrained by environmental factors, including differences in target reflectivity, light interference, and complex tree canopy shapes [25,26,27].

Table 2. Application of infrared sensors in variable-rate spray systems.

Reference	Year	Model	Detection Performance
[23]	2007	Composed of integrated circuits BA5104, BA5204, and integrated infrared receiver HS0038B.	The average detection distance was 6.15 m (white paper) and 4.33 m (simulated peach tree), and the average identification distance was 0.216 m. The detection distance was affected by the angle and density of the branches and leaves. The experiment was conducted under ideal laboratory conditions and has not been verified in an actual orchard environment.
[24]	2011	HS0038B, capable of emitting 38 kHz modulated infrared pulse signals. Maximum detection distance of 10 m, minimum detectable gap of 0.3 m.	The sensor can adapt to the detection of tree crowns with different row spacings, but detection accuracy depends on the stability of target reflectance, which is easily affected by complex tree crown shapes or environmental factors, and has limited adaptability to unstructured planting scenarios.
[25]	2012	The model consists of a receiving unit composed of a Wennerworth bridge sine wave oscillation circuit, a high-power LED emitting unit driven by a Darlington tube TIP142, and a silicon photovoltaic cell and CA3140 operational amplifier.	The detection distance for the Chinese scholar tree is 3.2 to 4.4 m, with an average detection interval of 0.22 m. However, the detection distance is greatly affected by lighting conditions, and the detection interval is greatly affected by the reflective properties of the target branches and leaves.
[26]	2015	E3F3-DS50N1 (Yueqing Gaode Electric Co., Ltd., Yueqing, China), maximum detection distance 70 cm, response time < 2.5 ms.	When detecting green leaves, the maximum distance can reach 0.65 m, and the target recognition rate is 100% at speeds of ≤ 1.5 m/s. However, there are significant differences in detection distances for different colored targets (white 0.70 m, gray 0.26 m, black 0.14 m).
[27]	2019	E3K80-DS5M1 (Shanghai Senzheng Electric Co., Ltd., Shanghai, China), maximum detection distance of 150 cm, response time ≤ 1 ms.	When the driving speed is ≤1.16 m/s, the cherry tree recognition rate is 100% and the detection error is <0.04 m. When the tree crown exceeds the 1.5 m detection radius, the sensor measures a width that is smaller than the actual tree crown width, with an error range of 3.8–7.9%.

Figure 2. Diagram of devices equipped with infrared sensors. (a) Automatic target detecting orchard sprayer with infrared sensors [24]; (b) automatic toward-target sprayer with infrared sensors [26].

3.3. Ultrasonic Sensor

Ultrasonic sensors consist of a transmitter and a receiver. The transmitter sends ultrasonic pulses toward the target object, while the receiver detects the reflected pulses. By measuring the time it takes for the pulses to return to the receiver, the sensor determines the distance and position of the object. Table 3 lists studies using ultrasonic sensors for canopy structure assessment, indicating their robustness in harsh conditions but accuracy challenges caused by multipath and scattering effects. Figure 3 shows two types of ultrasonic sensors with comparatively better performances. While ultrasonic sensors currently offer advantages for operating in harsh environments and can reduce overall tracking errors through multi-channel fusion and redundant measurements, their detection accuracy and stability remain limited by environmental sensitivity, physical beam characteristics, and target property dependencies [28,29]. Additionally, they struggle to fully overcome issues such as multipath reflections caused by complex canopy structures, non-uniform surface diffuse reflections, and signal attenuation resulting from random leaf and branch obstructions [30,31].

Table 3. Application of ultrasonic sensors in variable-rate spray systems.

Reference	Year	Model	Detection Performance
[28]	2011	LV-MaxSonar-WR1 (Maxbotix Inc., Brainerd, MN, USA), detection beam angle of approx. 10°, minimum detection distance 30.48 cm, resolution 0.32 cm, sampling frequency 20 Hz.	The sensor performs stably under harsh conditions such as low temperatures, dust, and strong winds, but the detection distance decreases by 0.05 m at high temperatures of 41.6 °C. Additionally, irregular tree crown shapes may cause ultrasonic multipath reflection errors. After artificial filtering, the root mean square error range decreases from 0.101 m (8.3% of the detection distance) to 0.194 m (15.9%) to 0.064 m (5.2%) to 0.079 m (6.5%).
[29]	2015	USS3 (Best Technology Ltd., Kyoto, Japan), output accuracy 1 cm, detection distance 0.04–10 m.	The surface of the pistachio tree canopy is composed of randomly distributed leaves, causing ultrasonic waves to produce diffuse reflection rather than specular reflection, which affects ranging accuracy. The sensor beam angle changes with the detection distance, resulting in reduced measurement consistency and stability.
[30]	2021	MB7060 (Maxbotix Inc., USA), detection distance 0.05–7 m.	The sensor’s accuracy is significantly affected by the target volume, density, and surface smoothness. The detection distance is generally greater than the actual closest distance, with a maximum error of 17.7%. In complex orchard environments, leaves and branches may obstruct the view and reduce the reliability of target identification.
[31]	2023	8-channel ultrasonic array, model unknown.	The overall average tracking error is reduced by 35%. Random variations in actual canopy morphology cause large fluctuations in single-measurement-angle values, requiring stabilization of data through the mean of four adjacent measurements. Complex canopy structures or high-density foliage affect the accuracy of ultrasonic echo signals.

Figure 3. Diagram of ultrasonic sensors. (A) LV-MaxSonar-WR1 ultrasonic sensor with protection components [28]; (B) USS3 ultrasonic sensor with connection board [29].

3.4. Radar

Radar sensors are active sensors that emit electromagnetic waves (such as microwaves or millimeter waves) toward targets and calculate the target’s distance, velocity, and azimuth based on the time difference and Doppler shift of the echo signals. They are used for detection, tracking, imaging, and environmental sensing. Currently, the most widely used sensor in variable-rate spray systems is lidar. Table 4 summarizes research on 2D and 3D lidar techniques applied in variable-rate spray, illustrating their strong canopy reconstruction ability in addition to their real-time limitations, which are associated with algorithm and hardware constraints. Figure 4 shows two radar detection modes of horizontal detection and vertical detection. Two-dimensional lidar is primarily used in variable-rate spray systems for rapid scanning and modeling of plant two-dimensional geometric parameters (tree height, tree width), achieving real-time analysis of basic canopy contours through high angular resolution and millimeter-level ranging accuracy. However, its single-plane scanning mechanism has inherent limitations: “ shadow effects,” dynamic delays, and sparse point clouds at long distances can lead to volume calculation errors [32,33,34]. Three-dimensional lidar, on the other hand, supports three-dimensional plant model reconstruction through multi-dimensional scanning and high-precision point clouds. However, its performance is constrained by complex terrain interference, model distortion caused by reduced point cloud density at high speeds, and algorithm complexity. Additionally, vertical angle intervals and stitching errors may affect canopy volume estimation [35,36,37].

Table 4. Application of LIDAR in variable-rate spray systems.

Reference	Year	Model	Detection Performance
[32]	2020	2D laser radar UTM-30LX, 0–270° scanning angle, angular resolution 0.25°, scanning cycle 25 ms.	Within a detection distance of 1–2 m, the maximum relative errors for detecting the volumes of spheres, cylinders, and cuboids are 9.32%, 8.84%, and 8.53%, respectively. The accuracy of volume calculations decreases as the detection distance increases, mainly due to the accumulation of discretization calculation errors caused by sparse point clouds at long distances.
[33]	2020	UTM-30LX 2D laser radar.	The minimum identifiable target spacing is 0.08–0.18 m. When the sprayer vehicle speed exceeds 0.9 m/s, the minimum identifiable spacing exceeds 0.18 m, and the identification capability decreases linearly with increasing speed. There is a fixed response delay of 160 ms, and insufficient dynamic compensation at high speeds can easily lead to target deviation.
[34]	2024	SICK LMS5100 2D laser radar (SICK, Inc., Waldkirch, Germany), −5–185° scanning angle, 0.167–1° angular resolution, 25–100 Hz scanning frequency, 12 mm measurement accuracy.	The measurement errors for tree height and width were 2.22% and 4.11%, respectively, with a model fit > 0.9; however, the error in tree thickness increased with LAI (reaching 41.1% when LAI = 3.68), limited by the “shadow effect” caused by the characteristics of two-dimensional scanning (unable to penetrate a dense canopy).
[35]	2020	R-Fans-16 3D Laser Radar (Beijing Beike Tianhui Technology Co., Ltd., Beijing, China), 360° horizontal scanning and 30° vertical scanning angle (−15° to +15°, single-line vertical angle interval of 2°), maximum detection distance of 200 m, and ranging accuracy of 2 cm.	The maximum height identification error for sugarcane reached 8.42%, with an average error of 4.59%. Point cloud data processing relies on Python 3.8 programming to convert polar coordinates to rectangular coordinates, which is relatively complex. Its adaptability to low-growing crops or scenarios with significant differences in plant density has not been verified.
[36]	2022	RoboSense 16-line 3D laser radar (Sure Star Technology Co., Beijing, China), 360° horizontal scanning and 30° vertical scanning angle (−15° to +15°, single line vertical angle interval 2°), maximum detection distance 150 m, accuracy 2 cm.	It can achieve high-precision navigation (lateral deviation < 0.22 m, heading deviation < 4.02°), but data processing is complex, and canopy volume calculations are based on the linear assumption of leaf area index (LAI) and volume, which may introduce errors when fruit trees change during their growing season.
[37]	2024	R-Fans-16 3D Laser Radar (Beijing Beike Tianhui Technology Co., Ltd.).	A three-dimensional model of corn plants was successfully constructed, with a height prediction error of less than 10% and a leaf area prediction model R2 of over 0.9. Increasing the driving speed during dynamic scanning reduces point cloud density, affecting detection accuracy. Leaf area prediction accuracy decreases under complex canopy conditions (R2 as low as 0.74).

Figure 4. Diagram of some devices equipped with radar sensors. (a) System of target test [32]; (b) structure diagram of high-clearance sprayer [33].

3.5. Machine Vision

Machine vision is an effective method for identifying the location, size, shape, color, and texture information of vegetation and pests/diseases. It offers a variety of configurations, including monocular, binocular, and multi-camera systems, providing flexible solutions for different operational scenarios and requirements of variable-rate spray [38,39]. Table 5 compiles representative machine vision approaches for canopy recognition and feature extraction, showing high potential for detailed analysis but practical restrictions associated with illumination sensitivity and processing delays. Figure 5 shows some devices equipped with machine vision. However, practical applications still face the following limitations: light sensitivity, dynamic operation delays, and insufficient depth information. These systems rely on artificial environmental constraints and fixed threshold strategies, and their ability to analyze complex scenes is limited [40,41,42].

Table 5. Overview of machine vision-based variable-rate spray technology research.

Reference	Year	Model	Detection Performance
[38]	2018	UI-1220SE/C Digital Color Camera (IDS Imaging Development System Inc., Obersulm, Germany). 752 × 128-pixel resolution, 24-bit RGB color mode, a 3.5 mm focal length lens, and a 2 ms maximum auto exposure time.	At a height of 1.2 m, the field of view covers an area of 1.52 m × 0.28 m. In a mixed-vegetation canopy (e.g., when weeds and blueberries are similar in color), the resolution capability is limited and relies on texture analysis for assistance. Changes in lighting conditions (e.g., shadows or reflections) may lead to misjudgments.
[39]	2019	Camera model unknown. The image format grabbed by this camera is (TSCO, VGA (640 × 480), 30 fps, 10 megapixels).	In a laboratory environment, the average response delay was 0.55 s for start-up (corresponding to a spatial lag of 0.30–0.83 m) and 0.85 s for stop (corresponding to a lag of 0.55–1.32 m). Light sensitivity increases the risk of misidentification in complex backgrounds. A fixed threshold strategy is insufficiently adaptive to changes in olive tree canopy density. The spatial offset caused by mechanical delay increases linearly with speed.
[40]	2020	Logitech Pro 9000 webcam (Logitech Inc., San Jose, CA, USA), 640 × 480-pixel resolution.	Used to detect leaf lesions caused by rice blast disease. At an operating speed of 0.25 m/s, a processing delay of 800 ms can be achieved, but this relies on a green background cloth to eliminate light interference (limited field adaptability), and the trigger mechanism based on the travel sensor is easily affected by variations in plant spacing.
[41]	2021	XiaoMi Intelligent S1030-IR-120 Dual-Camera System (Slightech (Jiangsu) Co., Ltd., Jiangsu, Wuxi, China), pixel pitch 6.0 μm, focal length 2.1 mm, field of view 146° × 122° × 76°, 752 × 480-pixel resolution, detection range 0.8~5 m.	Used to detect canopy depth information and calculate canopy volume. After calibration compensation of 0.15 m, the linear correlation coefficient between the canopy volume detection results and manual measurement values reaches 0.933. In static tests, when the canopy volume exceeds 0.036 m3, the coefficient of determination between actual flow rate and theoretical flow rate is 0.990. Outdoor lighting conditions can interfere with the depth detection accuracy of the stereo camera, requiring calibration compensation values for correction. Additionally, the error characteristics at different detection distances must be calibrated separately.
[42]	2022	Microsoft LifeCam HD-3000 monocular camera (Microsoft, Redmond, WA, USA), 30 fps	Used to detect the leaf wall area of citrus tree canopies. The error in calculating LWA is only 0.5%, with a camera response time of 40 ms. The flow model constructed within a duty cycle range of 15% to 65% has a coefficient of determination R2 > 0.94. However, monocular vision lacks depth information, which limits the acquisition of three-dimensional canopy features.

Figure 5. Diagram of devices equipped with machine vision. (a) A smart sprayer showing the camera’s field of view for image acquisition [38]; (b) the position of the camera on the sprayer [39]; (c) the variable-rate pesticide sprayer prototype [40]; (d) the overall structure of the variable-rate sprayer [42].

3.6. Multi-Sensor Fusion

To combine the advantages of different types of sensors, multi-sensor fusion technology integrates data from various sensors to achieve more accurate results than could be achieved using a single sensor alone. Table 6 provides examples of multi-sensor fusion for improving detection accuracy and robustness, though challenges remain with regard to balancing algorithm complexity with real-time performance. Figure 6 illustrates some multi-sensor fusion strategies that integrate lidar, vision, and other sensors. Leveraging multi-modal data collaboration significantly enhances the accuracy of pest and disease identification, canopy structure analysis, and other applications [43,44]. However, these strategies still face multiple challenges with regard to their application, such as significant data heterogeneity due to sensor performance differences, and a prominent contradiction between algorithm complexity and real-time requirements, as well as the system’s reliance on manual experience for feature parameter selection and data annotation quality, with generalization capabilities limited by the diversity of training data [45,46,47].

Table 6. Application of multi-sensor fusion in variable-rate spray system.

Reference	Year	Model	Detection Performance
[43]	2021	RGB camera (LRCP Luoke USB, ShenZhen ZWAK LRCP Technology Co., LTD, Shenzhen, China, resolution 2592 × 1944, 3 channels), multispectral camera (XIMEA MQ022HG-IM-SM5 × 5 NIR, XIMEA GmbH, Am Mittelhafen, Münster, Germany, resolution 409 × 216, 25 bands) and thermal infrared camera (FILR Tau2-640, Flir Systems, Inc., Wilsonville, OR, USA, resolution 640 × 512, 3 channels).	Used to detect pests and diseases on grape leaves. The RGB model performed best in independent testing, with an overall accuracy rate of 93.77%. Through the proposed multi-source data fusion decision-making method, the accuracy rate was improved to 96.05%. However, thermal infrared images are easily affected by fluctuations in ambient temperature, and the high dimensionality of multispectral data leads to feature redundancy. During the fusion process, some correct samples may be misjudged due to the confidence threshold.
[44]	2022	Industrial camera (MER-131-210U3M/C, China Daheng Co., Ltd., Beijing, China, resolution 1280 × 1024 pixels, frame rate 210 frames/s), ultrasonic sensor (MB7155, MaxBotix Inc., Brainerd, MN, USA, detection range 20–765 cm, frequency 10 Hz, accuracy ±0.7 cm), pressure sensor (131-B, Beijing Aosheng Automation Technology Co., Ltd., Beijing, China, range 0–2.5 MPa) and flow sensor (HI2144, Yueqing Ponai Sensing Technology Co., Ltd., Yueqing, China, 1–30 L/min)	Used to detect field vegetables such as cabbage. The target error (mean absolute error ≤ 0.0363 m, root mean square error ≤ 0.0426 m) is significantly reduced compared to when not fused. However, industrial cameras have limited recognition stability in complex lighting or high-speed scenarios, and the low-frequency characteristics of ultrasonic sensors may cause data update delays during high-speed operations.
[45]	2023	LiDAR (SICK LMS291-S05, SICK, Inc., Waldkirch, Germany), stereo vision camera (Microsoft AzureKinectDK, Microsoft, Redmond, WA, USA, resolution 2560 × 1440, field of view 90° × 59°).	LiDAR was used to analyze the density distribution of the vertical layering of simulated fruit tree canopies, with a two-dimensional reconstruction accuracy error of 0.633% ± 0.0276%. The stereo camera calculates leaf wall area through reference calibration, with an area detection error of 1.525% compared to lidar. However, lidar point cloud accuracy and complex canopy segmentation algorithms still require optimization, and stereo vision faces the risk of mis-segmentation in scenarios with strong light interference and multiple overlapping trees.
[46]	2023	Visible-light and near-infrared camera, thermal imaging sensor, hyperspectral imaging sensor, model unknown	By integrating data through weighted averaging and principal component analysis algorithms, the model ultimately achieved an accuracy rate of 95%, a sensitivity of 96%, and a specificity of 94%, with detection accuracies of 92%, 96%, and 94% in the early, middle, and late stages of corn growth, respectively. The model’s performance is limited by environmental conditions and training data diversity, and insufficient data on specific diseases or regions may affect generalization capabilities.
[47]	2024	The DJI Phantom 4 Pro multispectral drone (SZ DJI Technology Co., Ltd., Shenzhen, China) is equipped with an RGB camera (red, green, and blue bands) and a multispectral camera (five bands: 450 nm, 560 nm, 650 nm, 730 nm, and 840 nm) for multi-source image fusion.	The five common pests and diseases identified in apple tree canopies are aphids, Alternaria leaf spot, mosaic virus, brown spot disease, and gray spot disease. For the self-built dataset, the model achieved a subset accuracy of 92.92%, a sample accuracy of 85.43%, and an F1 score of 86.21%, representing improvements of 8.93% and 10.9% over single-source RGB/multispectral models, respectively. The accuracy for single-disease identification exceeded 97%. However, data collection relies on clear, windless weather, and the model requires high-resolution images to distinguish between complex backgrounds and similar-symptom pests and diseases. Vegetation index calculations depend on manual experience-based screening, and model training requires a large amount of precisely annotated canopy region data.

Figure 6. Diagram of multi-source fusion sensor systems. (a) Image collection system: Multispectral (MS), RGB, and thermal infrared (TIR) cameras are connected to a portable computer through data lines. Additionally, the multispectral image (MSI), RGB image (RGBI), and thermal infrared image (TIRI) were collected by corresponding cameras controlled by the computer [43]. (b) Schematic diagram of sprayer structure equipped with multi-source sensors [44]; (c) diagram of detection system [45]; (d) visual representation of multi-sensor system [46].

Currently, information sensing technology for variable-rate spray systems is evolving toward multimodal fusion and intelligent collaboration. In recent years, lidar and machine vision technology have become the mainstream sensors for information sensing. Lidar, with its high-precision 3D point cloud modeling capabilities, plays a central role in scenarios such as crop canopy volume measurement and spatial analysis [48,49,50,51], while machine vision provides critical data support for pest and disease identification and crop health monitoring through color, texture, and spectral feature analysis [52,53]. Other types of sensors primarily serve as auxiliary tools to supplement data collection in specific scenarios: laser scanning sensors, with their millimeter-level modeling accuracy, are mainly used for precise canopy parameter extraction in orchards [54,55,56,57]; infrared sensors, due to their fast response speed and low cost, are commonly used in basic scenarios such as field crop and fruit tree boundary identification [58,59]; and ultrasonic sensors, due to their strong tolerance to harsh environments such as dust and humidity, are particularly suitable for mist droplet penetration detection in orchards and physical parameter collection in densely planted canopies [60,61].

Multi-sensor fusion technology significantly enhances the system’s overall detection capabilities by integrating the advantages of different sensors, particularly in complex operational environments. However, integrating different sensor types into variable-rate spray systems presents several challenges. First, data heterogeneity arises because sensors operate at different spatial resolutions, sampling rates, and noise levels, which complicates data alignment. Second, fusion algorithms such as Kalman filters, Bayesian inference, or deep learning models require substantial computational resources, making real-time deployment difficult under field conditions. Third, sensor occlusion or partial failure can degrade overall system performance if robust fault-tolerant fusion strategies are not in place. Finally, achieving a balance between algorithmic complexity and real-time applicability remains a central bottleneck, especially when aiming to deploy fusion methods on embedded hardware platforms. Addressing these issues will be essential for advancing multi-sensor fusion from experimental prototypes to practical field applications [62,63]. In the future, the development of information sensing technology for variable-rate spray systems will continue to follow the technical route of multi-source sensor collaboration, focusing on overcoming key technical bottlenecks such as optimizing anti-interference capabilities in complex environments, improving the accuracy of multi-modal data fusion, and enhancing real-time response speeds. Simultaneously, by reducing hardware costs and developing lightweight edge computing algorithms, the application of multi-source sensing systems in large-scale agricultural production will be promoted.

Building on accurate sensing of crop canopy morphology and pest/disease distribution, the next step is to transform this multidimensional information into decision-making parameters, which is achieved through spray volume models.

4. Spray Volume Model

The precise acquisition of core parameters such as the canopy morphology of target crops and the distribution characteristics of pests and diseases is essential for constructing a scientifically sound spray volume model, which is a critical step in achieving on-demand spraying. By analyzing multi-dimensional data such as the canopy density and leaf area index, the optimal spray dosage can be dynamically calculated to ensure effective coverage of the target area while minimizing pesticide waste and negative environmental impacts, thereby balancing economic benefits with ecological sustainability.

Spray volume models for fruit tree crops have developed into a relatively complete technical system, primarily including classic models based on ground area (GA), tree row volume (TRV), leaf wall area (LWA), and leaf wall height (LWH) [64]. Although these models are designed for fruit trees, their parameterization principles can be extended to other crops. Table 7 summarizes spray volume models in variable-rate spray, reflecting diverse parameterization strategies but also limited adaptability across crops and environments. Currently, spray volume models have formed a multi-dimensional parameterization application system in variable-rate spray systems. By integrating crop morphological parameters (such as canopy volume and leaf area index), equipment characteristics (nozzle flow rate and flight parameters), and pest and disease severity data, combined with fuzzy logic and neural network algorithms, spray decisions are made [65,66,67,68,69]. However, the following core issues remain: Parameter acquisition relies on offline databases or manual input; model assumptions are simplified, and missing environmental factors lead to insufficient adaptability in complex scenarios; weak cross-crop generalization capability, as existing models are primarily developed for single crops and lack a universal parameter system [70,71,72,73]. Future spray volume models should focus on establishing intelligent models with dynamic regulation and multi-factor integration [74], primarily with regard to the following three aspects: dynamic multi-factor integrated modeling, i.e., the development of adaptive models that integrate multiple environmental factors such as wind speed, temperature, and humidity, and strengthen dynamic compensation mechanisms for environmental factors such as wind speed disturbances and temperature and humidity changes; crop growth stage adaptation optimization, which will involve a spray parameter adjustment mechanism for different crop growth stages, focusing on analyzing the correlation between leaf distribution density changes, pest and disease occurrence cycles, and canopy structure evolution characteristics, and will enhance cross-crop generalization capability. A universal parameter system will also be constructed; it will cover major crops, develop a modular model architecture, and support rapid adaptation to different crop types through parameter configuration.

Table 7. Spray volume modeling studies in variable-rate spray systems.

Reference	Year	Model Introduction	Model Effects
[65]	2009	Based on multidimensional data such as crop parameters, pesticide characteristics, equipment parameters, and environmental conditions, the system quantifies efficiency factors to correct for losses during the spraying process (such as canopy obstruction and meteorological interference), and ultimately outputs the optimized spray volume.	The model can reduce traditional spray volume by approximately 30%, but it is highly parameter-dependent (requiring precise input of data such as LAI and canopy volume that is difficult to obtain in real time). Simplified assumptions affect accuracy (such as the assumption of uniform droplet coverage ignoring heterogeneity of deposition within the canopy), efficiency factor weights lack dynamic verification (some sub-factor scores rely on empirical values), and its universality is limited (mainly suitable for vineyard scenarios).
[66]	2010	Based on canopy width, height, tractor travel speed, and a preset unit volume spray coefficient, the nozzle flow rate is dynamically adjusted to achieve a linear match between canopy volume and spray volume.	The model achieved an average reduction of 58% in pesticide usage across three grape varieties and different growth stages, while normalized leaf deposition and leaf recovery rates were comparable to or better than those of traditional spraying, with no significant decrease in deposition uniformity. However, when the canopy volume was low, the system’s minimum pressure limit caused the spray coefficient to deviate from the theoretical value, resulting in localized over-spraying. Additionally, the model does not incorporate parameters such as leaf density or leaf area index, which may affect its adaptability to complex canopy structures.
[67]	2011	The optimal spray volume is calculated based on crop canopy characteristics, spray equipment parameters, environmental conditions, and pest characteristics. By setting the target droplet density and droplet volume, the spray volume is dynamically adjusted in combination with a leaf area database, and an efficiency factor is introduced to quantify spray loss.	The model reduced spray volume by an average of 39.9% and pesticide use by 53%, with leaf surface deposition uniformity superior to that of traditional methods. Its efficacy in controlling diseases such as powdery mildew and gray mold was comparable to that of traditional high-dose treatments. However, the model relies on a pre-established crop parameter database and does not account for synergistic effects of multiple pesticides. Under high-density canopies, traditional methods achieve higher recovery rates.
[68]	2012	The model is based on a fuzzy neural network and includes parameters such as crop area, target distance, and pest and disease severity. It adopts a five-layer fuzzy neural network structure, which uses Gaussian membership functions to fuzzy the input parameters, combines 18 fuzzy rules for inference, and utilizes the BP algorithm to dynamically adjust the weights, membership function centers, and width parameters to achieve self-learning.	After 191 training sessions, the model achieved a target error of 0.001. The decision error of the test samples was significantly lower than that of traditional fuzzy systems, and the model demonstrated stronger generalization capabilities for samples that were not included in the training, enabling real-time response to uncertain greenhouse environments. However, the model did not explicitly consider dynamic interference factors such as environmental parameters and robot movement speed, and model verification relied on laboratory simulation data.
[69]	2013	Based on fuzzy neural networks, input parameters include the area of the spraying target, target distance, and pest and disease severity. The model learns from operational experience data through offline training and outputs pesticide flow rates.	The model has a small average decision error on the test set, with an error reduction of approximately 30–40% compared to traditional fuzzy decision systems. However, the model’s pest and disease severity ratings depend on manual input and do not yet support automated real-time detection. The model does not consider the potential impact of other environmental factors such as temperature and humidity on pesticide application rates, and the experimental data was collected in a specific greenhouse scenario.
[70]	2017	The model is based on the canopy grid volume algorithm, with key parameters including canopy grid height, width, depth, single-point height, sprayer travel speed, vertical distance from the sensor to the tree row, and nozzle flow coefficient. The model calculates the grid volume and converts it into PWM duty cycle to control the nozzle flow, achieving variable-rate spray that matches the canopy volume and speed.	The model shows a small range of spray position distance errors at different speeds (0–0.14 m), with the lowest error occurring at fine grid settings, validating the effectiveness of the delay control method. However, the model’s accuracy is limited by the grid width setting; as the grid size increases, the error significantly increases. Additionally, the model does not account for the potential impact of dynamic factors such as environmental wind speed.
[71]	2020	The model mainly integrates four parameters: flight altitude, flight speed, nozzle voltage, and solenoid valve opening angle. It achieves stable spray volume per unit area through dynamic adjustment of the solenoid valve opening angle.	In single-factor tests, the model control deviation range was 0.11–9.79%, and in multi-factor orthogonal tests, the maximum deviation was 9.46%, both of which met agricultural aviation spraying requirements (<10%). However, the droplet movement model only considers gravity and ignores complex factors such as air resistance, and the spray width prediction is based on an approximate sine curve for a specific centrifugal spray nozzle (Auz900), which may affect the model’s universality.
[31]	2023	The core parameters of the model include crown unit volume, leaf area density, and spray volume per unit volume, with maximum theoretical leaf area density as the benchmark parameter. Spray flow is dynamically adjusted based on crown volume and leaf area density parameters.	The spray variability coefficient of the contour variable-rate spray was reduced by 25.9% and 21.9% in the inner and outer layers of the tree canopy, respectively, with spray uniformity significantly superior to traditional methods. While achieving a 32.1% reduction in flow rate, key indicators such as droplet coverage and volume median diameter (VMD) showed no significant difference from traditional spraying methods. However, ground deposition coverage decreased from 33.2% to 13.0%. However, the complexity of canopy morphology causes significant fluctuations in single contour angle calculations, necessitating smoothing through the average of four measurements. Leaf area density estimation relies on ultrasonic echo characteristics, which are susceptible to environmental noise and extreme conditions of branch and leaf density.
[72]	2024	Based on the dynamic adjustment of spray flow according to cotton canopy volume, the main parameters include drone flight parameters, canopy volume, nozzle duty cycle, and geographic coordinate information. The model establishes a linear relationship between flight parameters, canopy volume, and duty cycle, and combines polygon prescription maps to decode geographic coordinates and flow information in real time, thereby achieving variable-rate spray.	The model-calculated canopy volume showed a coefficient of determination (R2) of 0.89 compared to manually measured values, confirming a high correlation. The variable-rate spray system saved 43.37% of the pesticide compared to the constant-rate spray system. The deposition efficiency (average 0.31 μL/cm2) and coverage rate (7.78%) met operational requirements, and no over-spraying occurred. However, the model is limited by the PWM duty cycle adjustment range at low flow rates, and the onboard processor’s ability to process large-scale point cloud data in real time may affect the dynamic response speed under complex canopy conditions.
[73]	2024	The model is based on the rubber tree powdery mildew control standard and integrates parameters such as leaf area index (LAI), individual tree height, canopy volume, powdery mildew severity, and the distance from the powder spray nozzle to the canopy center. The model dynamically adjusts the application rate of sulfur powder based on LAI values, calculates the initial wind speed using jet decay theory, and adjusts the spray angle in real time based on the canopy center height.	The accuracy of the model-controlled actuator exceeds 95.9%. Compared to the quantitative mode, the coefficient of variation (CV) of canopy deposition in the variable mode is reduced to 35.38–36.90%, and the effective utilization rate of sulfur powder is improved by 20.03%. However, the model wind speed is fixed at 2 m/s and is not dynamically optimized based on spray volume, resulting in an actual deposition growth trend below the theoretical value at high spray volumes. Additionally, the model does not account for the influence of natural wind on airflow, limiting its applicability under high-wind conditions, and further integration of meteorological parameters is required.

While spray volume models provide the theoretical basis for determining pesticide dosage, their effectiveness ultimately depends on real-time execution, which is realized through the control system that drives actuators such as solenoid valves.

5. Control System

From sensing to modeling and then to control, these three components form a closed-loop technical chain: information detection supplies real-time input, spray volume models convert input into optimized dosage, and the control system ensures precise implementation under dynamic field conditions. Table 8 outlines control strategies and actuators used in variable-rate spray, showing the dominance of PID-based methods with PWM solenoid valves, while intelligent optimization approaches are still at an early stage of field application. Figure 7 shows some control systems with different algorithms. Currently, variable-rate spray control systems mainly differ in the selection of control algorithms (PID, fuzzy logic, neural networks, and fusion algorithms, etc.) and actuators (solenoid valves, pumps, electric actuators, etc.) [75]. The choice of algorithms and actuators should be made according to the spraying objectives and the expected deposition outcomes, because spray pressure and PWM duty cycle strongly influence droplet size (VMD), drift, and canopy penetration. Higher pressure generally produces finer droplets with better canopy penetration but higher drift risk, while lower pressure generates coarser droplets that reduce drift but limit inner-canopy deposition [76,77,78]. Likewise, a higher PWM duty cycle helps improve spray distribution uniformity and deposition density, especially in dynamic spray systems, where it can significantly reduce the coefficient of variation (CV); whereas a low duty cycle can easily lead to uneven spray coverage and decreased deposition density [79,80,81]. These relationships imply that control algorithms must dynamically adjust parameters to stabilize droplet size under variable field conditions, and actuator performance (e.g., high-frequency response of solenoid valves) determines whether such adjustments can be realized effectively.

Table 8. Control system study in variable-rate spray system.

Reference	Year	System Principle	Control Performance
[82]	2015	Comprising an electronic linear actuator, a variable orifice nozzle (VariTarget), a pressure sensor, an electromagnetic flowmeter, and a data acquisition module, the control algorithm employs proportional control (variable gain P control) based on dynamic adjustment of the system pressure gain, with gain adaptation achieved through a pressure polynomial equation.	Within the pressure range of 138–414 kPa, the step response settling time is <1 s (as fast as 0.63 s), overshoot is ≤7.2%, and the steady-state error is <2%. The ramp response has an average absolute error of <5% at an input of 0.07 Hz, a hysteresis time of <0.38 s, and a maximum flow rate change rate of 6.39 mL/s2 (414 kPa). However, system gain requires adjustment based on pressure experience, and laboratory conditions do not account for field vibrations/temperature effects.
[83]	2020	The system consists of a multi-sensor data acquisition module (including speed, flow, pressure, and liquid level sensors), an STC12C5A60S2 microcontroller, and an electrically controlled regulating valve actuator. The system employs a PID control algorithm tuned using a BP neural network, leveraging the neural network’s self-learning capabilities to dynamically adjust PID parameters, thereby achieving precise control of the drug solution return flow.	When the vehicle speed randomly varies between 1.11 and 3.06 m/s, the spray volume control error remains stable within ±3%, with an average adjustment time of 0.72 s, an overshoot of only 2.1%, a coefficient of variation for longitudinal spray uniformity below 6%, and a standard deviation of droplet deposition density < 1.4 per cm2, significantly outperforming traditional PID and fuzzy PID control. However, the system still has limitations such as the manual pressure regulating valve affecting automation levels, pressure regulation potentially disrupting spray atomization stability, and the complexity of neural network parameter training.
[82]	2021	The system consists of an STM32 processor, electric valve control circuit, flow sensor, pressure sensor, liquid level sensor, and actuators (electric control valve and anti-drip nozzle). The system uses an improved genetic PID (IGA-PID) control algorithm, which optimizes population selection through mean selection operators, and combines adaptive crossover and mutation operators to improve parameter optimization efficiency.	The algorithm outperforms traditional PID, fuzzy PID, and standard genetic PID in terms of overshoot (1.25%), steady-state error (1.21%), and settling time (0.157 s). The field test response time is 3.84 s, with a spray volume error of 2.59%. However, the system relies on electric valves to regulate flow, and pressure changes may affect atomization efficiency.
[83]	2022	The system consists of an STM32H743VIT6 processor, a diaphragm pump drive module, a flow/pressure sensor, a PWM drive module, and a wireless communication module. It uses a single-neuron PID control algorithm based on the Hebb learning rule to achieve precise control of the spray flow by dynamically adjusting the PID parameters.	Single-neuron PID control demonstrates significantly higher accuracy than traditional PID control. Simulation results show that the rise time is reduced to 0.054 s with no steady-state error. In indoor tests, the maximum overshoot is 3.88%, and the average response time is 0.85 s. In field tests, the flow error is less than 5%, and the droplet deposition rate is positively correlated with the flow rate. However, overshoot still requires optimization during sudden changes in system flow (reaching a maximum of 9.1% in field tests), and single-neuron parameters (such as learning rate and proportional coefficient) still require manual tuning.
[84]	2023	The system consists of a microcontroller (ATmega328), pressure sensor, PWM controller, precision potentiometer, buck module, and diaphragm pump. The system employs a PWM-based open-loop control strategy, where the target pressure value is set via the potentiometer. The microcontroller continuously monitors the pressure sensor signal and adjusts the PWM duty cycle to control the pump speed.	The system control accuracy can reach ±0.1 kg/cm2 (±9.81 kPa) of the target pressure value, significantly outperforming the ±0.3 kg/cm2 fluctuation of the traditional pressure relief valve + gate valve combination. Experiments show that this system keeps the coefficient of variation (CV) of nozzle flow below 1%, while traditional methods often exceed 1%. However, this system has only been validated in a laboratory single-nozzle scenario and has not yet been adapted to the complex operating conditions of multi-nozzle spray bar sprayers.
[85]	2024	The S7-200 PLC serves as the central controller, equipped with Hall effect speed sensors, flow sensors, and pressure sensors to collect real-time operational parameters. Flow is regulated via proportional electric control valves, and the power source uses a diaphragm pump to drive the delivery of the solution. The system employs an ABC-optimized PID control algorithm (ABC-PID), utilizing swarm intelligence to dynamically adjust PID parameters.	In the simulation, the overshoot of the ABC-PID control was only 1.7% (compared to 24.5% for conventional PID), the rise time was reduced to 0.17 s (compared to 0.3 s for conventional PID), the average error in spray volume during the indoor constant-speed test was 1.3% (compared to 3.1% for fuzzy PID), and the error in the variable test was 2.1%; it outperformed the comparison algorithms in all these areas. However, the experiments were conducted solely in a laboratory environment and have not been validated in actual field operation scenarios. The number of iterations and population size of the ABC algorithm have not been optimized, which may increase the computational burden in real-time applications.
[86]	2011	The system is composed of an 8-channel variable controller, solenoid valve, LandManager II controller, flow valve, and nozzles. The control algorithm is based on real-time threshold judgment. When the height of weeds exceeds the preset threshold, the variable controller immediately activates the corresponding nozzles using open-loop control logic.	The system response time is 0.05 ± 0.003 s, with a flow control error of ≤2.5%. A t-test of water-sensitive paper coverage in field trials showed no significant difference between variable-rate spray (10.01–81.22%) and uniform spray (5.39–72.67%). However, the system is only applicable for weed detection between 0.3 and 0.55 m in height and is ineffective for short weeds (<0.3 m) and early growth stages. Spray coverage is significantly dependent on weed height, with coverage decreasing for weeds over 0.55 m in height, necessitating spraying during specific growth stages.
[87]	2012	The system consists of a PWM solenoid valve nozzle assembly, a microcontroller (PIC18F4523), and a speed sensor. Each nozzle is independently controlled by a PWM solenoid valve to regulate flow. The microcontroller adjusts the PWM duty cycle of each nozzle based on sensor data and travel speed, and uses a pre-trigger mechanism to ensure that the spraying timing is synchronized with the canopy position.	At a speed of 0.89 m/s, the system achieves an average response time of 296 ms, with a trigger error of 0.045–0.125 m. The coefficient of variation for spray coverage and deposition rate ranges from 44.5% to 59.2%. However, there were no significant differences in average coverage (12.0–14.7%) and deposition rate (0.72–0.90 μL/cm2) across different speeds (0.89–2.22 m/s). Nevertheless, the simplification of the canopy volume using a rectangular assumption may affect the accuracy of detecting complex shapes, and the sensor is susceptible to interference from path deviations and mechanical vibrations.
[88]	2020	The system uses three pressure sensors and two flow sensors to monitor the main pipeline pressure and spray flow in real time. The actuator includes an electric control valve and a return flow control valve. The control algorithm uses a composite control strategy based on chaos optimization: the outer loop flow control combines Bang-Bang relay control with adaptive fuzzy nonlinear PID, while the inner loop pressure control uses variable domain fuzzy control, and the controller parameters are globally optimized through a chaos algorithm.	The simulation results show that the optimized controller’s rise time (0.9 s), settling time (1.5 s), and overshoot (2.67%) are all better than the traditional PID (2.2 s, 5 s, and 6%). Experimental verification shows that the effective droplet rate reaches 89.4% at a pressure of 300 kPa and a flow rate of 0.08 m3/h, which is an 8.1% improvement over the PID. However, the system’s reliance on multiple sensors increases complexity, and the significant computational load of chaotic optimization may impact real-time performance.
[89]	2016	The system consists of a closed-loop PID controller, a single-phase AC chopper voltage regulator, an electric impeller pump, a pressure transmitter, a PLC controller, solid-state relays, solenoid valves, and hollow conical nozzles. Pressure sensors provide real-time feedback on pipeline pressure, which is used by the PID algorithm to calculate deviation signals. The single-phase AC chopper then dynamically adjusts the speed of the pump motor to achieve closed-loop pressure control.	Within the set pressure range of 100–400 kPa, the maximum relative error of closed-loop control is 8.36% (400 kPa/three-nozzle condition), while the error of open-loop control is as high as 40%. The system performs optimally in the medium-pressure range of 200–300 kPa, with a spray angle deviation of less than 0.87°. However, the error increases at high pressure (400 kPa) due to the limitations of the pump’s maximum head. The AC chopping voltage regulation technology used has a narrower speed control range and lower precision compared to the variable-frequency speed control scheme.
[90]	2018	The system consists of a DSP controller (TMS320F2812 chip), a multi-nozzle solenoid valve drive circuit, a network of flow/level/pressure sensors, a WiFi communication module, and an alarm circuit. The system uses an improved particle swarm optimization algorithm (IPSO) for weed image segmentation, combined with a lateral histogram algorithm to quickly extract weed distribution maps, and implements multi-nozzle combination control based on weed density through the DSP.	The system has high control accuracy, with a spray response time of 1.562 s and a flow rate adjustment range of 84–674.4 L/hm2, which is eight times greater than that of traditional single-nozzle systems. However, at low pressures (200 kPa), small flow rates cannot be detected by the sensor, and the system has only been verified during the early stages of corn seedling growth (3–4 leaf stage). Its applicability under complex canopy conditions has not been tested.
[91]	2019	The system consists of a core control board (based on the NXPKinetisK60 microcontroller), a miniature diaphragm pump, a flow sensor, a pump motor speed controller, and a spray nozzle. The system employs a closed-loop PID control algorithm, which continuously collects real-time flight speed and altitude parameters. These are combined with pre-set nozzle spacing, target spray volume, and other parameters to dynamically calculate the target flow rate and adjust the PWM duty cycle of the diaphragm pump. Additionally, feedback from the flow sensor is utilized to form a closed-loop correction mechanism.	The system’s average control response time is 0.18 s, with a flow stabilization time of 0.75 s, enabling rapid tracking of changes in flight status to achieve flow regulation. However, insufficient GPS positioning accuracy may lead to speed calculation errors; the acceleration integration method for speed measurement carries the risk of cumulative errors; the mechanical inertia of the diaphragm pump causes startup/shutdown delays (average stabilization time of 0.54 s), affecting precision in high-frequency speed-changing scenarios; and parameter settings rely on manual input (nozzle type, spacing, etc.) and lack intelligent adaptation capabilities.
[92]	2024	The Xinjie XD3-16RT-C PLC is used to implement pulse width modulation control of the solenoid valve opening through a time relay, which essentially belongs to an open-loop time proportional control algorithm.	Test results indicate that the system achieves a maximum variable capacity of 9.91 times within the pressure range of 100–200 kPa. The coefficient of determination for the exponential/polynomial fit between the solenoid valve opening and spray volume exceeds 0.99, with relative error controlled within 4%, demonstrating high-control precision. However, the system lacks a closed-loop feedback mechanism, relies on a constant-pressure pipeline environment, and requires manual intervention for pressure adjustments. Additionally, the system does not integrate a flow sensor for real-time correction, which may result in cumulative errors during long-term operation.

Figure 7. Diagram of some control systems with different compositions. (a) Diagram of variable-rate herbicide spraying controller [75]; (b) diagram of variable-rate spray on double closed-loop [82]; (c) schematic principle of constant-pressure control system [83].

In laboratory environments, control systems based on intelligent optimization algorithms (such as chaotic optimization PID and improved genetic algorithms) have achieved high-precision dynamic regulation [82,83,84,85,86,87]; however, in actual field applications, traditional PID control combined with electromagnetic valve execution remains the primary method, limited by insufficient sensor interference resistance, mechanical transmission delays, and edge computing power constraints, resulting in control errors that are generally higher than in laboratory scenarios [88,89,90,91,92,93,94]. Real-time decision-making is particularly critical under rapidly changing field conditions, such as fluctuating wind speed, irregular canopy structures, or varying travel velocities of the sprayer. Traditional fixed-parameter controllers often struggle to adapt in these scenarios. Closed-loop strategies that incorporate feedback from flow, pressure, or canopy sensors, combined with edge computing and lightweight learning models, can help ensure timely adjustments and improve spray robustness in practice. Future development directions should focus on edge intelligence deployment: by compressing lightweight models and optimizing embedded hardware, the real-time performance of complex algorithms in the field can be improved. Additionally, the reliability of actuators should be enhanced: optimizing the structural parameters of solenoid valves and using wear-resistant materials can enhance stability under high-frequency variable operating conditions. Recent advances also illustrate how emerging digital technologies can be integrated into control systems, which highlights the potential of IoT-enabled control architectures to enhance both technical precision and practical feasibility in variable-rate spraying [95].

6. Discussion

While this review systematically reviewed the core technical components of variable-rate spray technology—sensing, modeling, and control—a broader discussion is essential, as it will contextualize these technologies within the diverse agricultural landscapes where they are being developed and deployed. The global adoption of variable-rate spray technology is not uniform; it is influenced by a wide variety of factors, including regional economic drivers, farm structures, regulatory pressures, technological priorities, and so on.

In developed countries, the push for variable-rate spray technology is largely driven by the pursuit of high precision, labor efficiency, and stringent environmental regulations, although the technological pathways vary. One major avenue of research focuses on optimizing sophisticated ground-based systems for high-value crops. For example, studies in the USA have demonstrated that laser-guided systems can reduce spray volume by over 80% in apple orchards while maintaining efficacy [96]. This finding is consistent with the results reviewed in Section 3.1, where laser scanning sensors were highlighted for their high precision. Similar outcomes have been reported in European perennial cropping systems; research in Italy confirmed that a variable-rate spray technology sprayer in vineyards reduced spray volume by an average of 35%, achieving significant savings in pesticides, water, and fuel while maintaining canopy deposition comparable to uniform rates and avoiding overspray [97]. The emphasis on optimization extends to highly automated, per-target spraying, as seen in Israeli research on adjustable, vision-guided devices [98].

Another significant pathway in developed nations is the focus on Unmanned Aerial Vehicles (UAVs), which offer a scalable platform for variable-rate spray technology. Research from South Korea exemplifies this trend, with studies refining both the hardware control and mission planning aspects of UAV spraying. At the control system level, innovations such as Kalman filter algorithms to reduce nozzle pressure pulsation and response delays are critical for ensuring spray uniformity from lightweight aerial platforms [99]. This aligns with the intelligent optimization approaches identified in Section 5 of this review. In addition, predictive models based on machine learning have been developed to estimate spray volumes from operational parameters such as flight speed and PWM settings, thereby optimizing UAV missions without relying solely on complex, real-time sensor data [100]. These advances demonstrate the increasing importance of integrating spatial variability into UAV-based spraying strategies.

Beyond technical achievements, adoption ultimately depends on user acceptance. Research from Spain highlights the critical need for manufacturer-independent prescription map formats to lower barriers to entry for growers [101]. This is echoed in studies from the United Kingdom, which show that while growers are motivated by economic returns and environmental sustainability, adoption hinges on clear demonstrations of the technology’s viability, ease of use, and a manageable upfront investment relative to farm size [102].

In emerging agricultural economies and developing countries, research and development often prioritize cost-effectiveness, accessibility, and adaptation of existing infrastructure. In Brazil, for instance, studies have validated the use of relatively low-cost electronic flow controllers in cotton production, demonstrating that significant benefits can be achieved without investing in high-end systems [103]. Similarly, research in Pakistan has shown that retrofitting a conventional sprayer into a real-time variable-rate spray technology system with infrared sensors and an Arduino microcontroller improved spray coverage and vertical distribution while reducing costs to a fraction of those of imported commercial sprayers [104]. Such approaches directly address the cost barriers that hinder adoption and provide practical models for expanding access to precision spraying in capital-constrained regions.

This divergence in technological pathways also highlights a universal challenge: the lack of standardized protocols for data collection and model validation. Currently, variable-rate spray technology studies often adopt heterogeneous experimental setups, sampling resolutions, and performance metrics, which makes cross-comparison across crops, regions, and equipment types difficult. Developing unified guidelines for sensor calibration, environmental factor measurement, and spray deposition assessment, together with benchmark datasets for model testing, would significantly improve reproducibility and comparability. Such efforts would provide a more solid basis for evaluating and integrating sensing, modeling, and control technologies in practical agricultural systems.

Beyond technical standardization, the economic implications of variable-rate spray technology technologies are also noteworthy. Although this review primarily focuses on technical aspects, prior studies suggest that variable-rate spray technology can reduce pesticide use by 10–30% compared with conventional practices, which may offset the higher initial investment in sensing devices and control systems. However, systematic cost–benefit analyses remain limited, especially across different farm scales and crop types. Future research should therefore examine the trade-offs among precision, robustness, and cost, and integrate these evaluations with economic analysis to better support decision-making. For small-scale farmers, high equipment costs, limited access to technical support, and complex system integration remain major barriers. Simplified hardware designs, modular sensor packages, and user-friendly control interfaces could improve accessibility and reduce operational burdens. Moreover, low-cost, lightweight solutions tailored to smallholder farming systems will be essential to ensure that the benefits of variable-rate spray technology can be realized across diverse agricultural contexts. In particular, in South Asian countries dominated by smallholders with limited resources, strategies for affordable and easy-to-deploy variable-rate spray technology systems are urgently needed [105]. Future research should prioritize field validation in these regions to assess performance under local cropping practices, climatic conditions, and operational constraints.

7. Conclusions and Prospect

Variable-rate spray technology has evolved as a cornerstone of precision chemical management, yet its implementation pathways diverge significantly across technical architectures and operational contexts. Our analysis reveals that no single approach universally dominates; rather, each technology aligns with distinct agricultural scenarios based on inherent trade-offs between precision and robustness. While economic cost is also a critical factor influencing adoption, detailed cost analyses are beyond the scope of this review and should be systematically addressed in future studies. While this review summarizes the major advances in sensing, modeling, and control, it is important to note that reported performance metrics are not always directly comparable. Studies often differ in experimental design, crop types, and environmental conditions, which makes it difficult to establish a uniform benchmark. Even under apparently similar conditions, some approaches (e.g., ultrasonic vs. LiDAR sensing) have shown varying levels of accuracy and robustness; moreover, it should be acknowledged that this review may also be subject to publication bias, as studies reporting positive outcomes are more likely to be published than those reporting negative or inconclusive findings, suggesting that the results should be interpreted with caution.

(1) Information detection forms the foundation of variable-rate spray systems, which utilize sensors to collect real-time data on crop morphology, pest and disease conditions, and environmental information. Currently employed sensors include laser scanning sensors, infrared sensors, ultrasonic sensors, lidar, machine vision, and multi-source fusion. The primary issues that need to be addressed in the future are as follows: Existing sensors have weak interference resistance in complex field environments, leading to reduced data accuracy. Signal processing algorithms need to be improved to eliminate background noise. Additionally, some sensors lack sufficient dynamic detection capabilities to match the high-speed movement of spray operations. To overcome the limitation of sensor dynamic detection accuracy, future research should focus on adaptive calibration methods and motion compensation algorithms. For example, integrating inertial measurement units (IMUs) with optical or ultrasonic sensors can help correct for vibration and movement during field operations. In addition, machine learning-based filtering techniques could be employed to improve signal-to-noise ratios under rapidly changing conditions.

(2) Spray volume models are the core foundation for converting sensor detection data into precise spraying strategies. Common models include those based on canopy volume and leaf area density. Different models demonstrate good performance in adjusting spray volume and addressing crop-specific characteristics, effectively reducing pesticide usage. The primary issues to be addressed in the future are as follows: existing models lack adaptability to complex meteorological conditions, necessitating the integration of more environmental variables to develop multi-factor optimization models; Additionally, most models lack dynamic optimization capabilities and cannot adjust parameters in real time based on crop growth stages. It is necessary to combine machine learning methods to build adaptive models and achieve dynamic matching of spray parameters. In the future, hybrid approaches that combine mechanistic models with data-driven corrections show promise, as they can incorporate real-tie weather data (e.g., wind speed, humidity, and temperature) to adjust predictions dynamically. Cross-crop training datasets and transfer learning strategies may further enhance model generalization across different plant species and canopy structures.

(3) Variable-rate spray control systems regulate spray flow and pressure through actuators. The electromagnetic valve + PID control algorithm is the current mainstream control solution, achieving significant improvements in response speed and flow control accuracy. The primary issues to be addressed in the future are as follows: The opening and closing speed of electromagnetic valves affects the response speed of the control system, and prolonged high-frequency operation can lead to wear and tear, resulting in reduced control accuracy. It is necessary to optimize material and structural design to enhance response speed and service life. Additionally, the control system exhibits poor stability in complex scenarios, necessitating the introduction of deep learning algorithms to optimize control strategies and enhance interference resistance in dynamic environments.