Kale, as a typical canopy-intensive crop, is an important cruciferous vegetable whose agronomic planting parameters are closely tied to the requirements of precision application technology. According to ISHS standards and actual production needs, the recommended planting spacing for kale is 40–60 cm. This spacing not only accommodates nodulation but also allows sufficient space for full leafball expansion during that stage, helping to prevent poor ventilation, reduced light transmission, and disease resulting from overly dense planting. Regarding application rates, due to the thick waxy layer of kale leaves and its dense canopy, the recommended spraying volume is 400–600 L/ha, which is 25%–33% higher than that of leafy vegetables such as cabbage (Zhang et al., 2017).

In this study, kale was cultivated according to standard agronomic practices with a plant spacing of 50 cm and a row spacing of 60 cm. The current YOLOv8 algorithm is widely adopted for high-density crop target detection due to its high efficiency, accuracy, and reliability, making it a popular choice in vegetable recognition applications. The complexity of kale field detection scenarios—such as small target scale and strict real-time requirements—is systematically evaluated. A comparative analysis was conducted on the migration learning performance of five commonly used models—Faster R-CNN, YOLOv8n, SSD, YOLOv5S, and YOLOv8—to determine the most suitable detection model for kale (Fu et al., 2022). While Faster R-CNN, a two-stage detection framework, excels in small target detection in complex scenes by generating candidate regions for feature extraction and classification, its time-consuming process and high deployment costs make it unsuitable for real-time kale detection. SSD, with single-stage detection and multi-scale feature mapping, underperforms on small targets due to insufficient shallow feature resolution. YOLO’s unique grid-based architecture enables efficient multi-category detection with fewer computational resources. YOLOv8 replaces the C3 structure in YOLOv5 with the C2f structure for improved gradient flow and modifies channel numbers for different model scales. As a result, even lightweight models like YOLOv8n perform well, particularly for lower-resolution images captured by mobile devices. In testing, YOLOv8n demonstrated the fastest detection speed and best performance among the lightweight models. Accordingly, YOLOv8n was selected in this study and further optimized to construct a recognition model capable of fast, accurate detection of field-grown kale under high-density, high-humidity conditions. Its architecture is illustrated in Figure 3 .

The constructed kale recognition model consists of three main parts: backbone, neck, and head. The backbone is responsible for extracting image feature information of kale and feeding it into the subsequent network. The Conv module contains convolutional layers, batch normalization, and activation functions for standard feature extraction. The DW-C2f module performs feature extraction and fusion using connections across different bottleneck layers. The Swin module enhances detection accuracy in edge regions of the kale image, thereby improving overall recognition precision for better variable spraying. The SPPF module fuses features that have undergone a maximum of three max-pooling operations. This sequence enables efficient extraction and fusion of rich feature information. The neck connects the backbone and head and facilitates deeper feature fusion. It integrates multiple Concat and DW-C2f modules to combine low- and high-level features after several convolution operations. The left and right sides of the head perform object detection in conjunction with binocular vision sensors, with a Bbox loss module for bounding box regression and a Cls module for classification.

The preliminary YOLOv8n detection model, while capable of identifying target crops in the field, exhibits inadequate adaptation to dense kale plantations. To improve this, the study uses larger convolutional kernels and depthwise convolutions to lower computational costs (Chen et al., 2024). Depthwise convolution decreases floating-point operations and parameters by applying each kernel to a single input image channel. Although dilated convolution can expand the receptive field, it might impact downstream tasks. To address this, the study employs a dilated convolution design with added shape biases to enhance model generalization and reduce overfitting risks (Ma et al., 2018). ConvFFN modules replace the original ones in the bottleneck to expand the model’s receptive field and improve feature extraction for target kale. The original YOLOv8n struggles with detecting edge regions in kale images due to insufficient convolutional operations and weak gradient updates for edge pixels. To solve this, the study integrates a Swin Transformer module with CNN. The Swin Transformer’s self-attention mechanism better extracts edge semantic information, overcoming CNN’s edge detection limitations and improving target semantic understanding (Nevavuori et al., 2019). In field scenes, kale is often occluded by weeds and other objects. A non-local attention mechanism is added to the algorithm to enhance feature extraction of kale in practical detection and strengthen the model’s ability to capture long-range dependencies. Channel or spatial attention mechanisms aim to generate more discriminative features and improve the distinctiveness of key features (Thenmozhi and Srinivasulu Reddy, 2019). However, channel-wise dimensionality reduction may affect the visual representation of target kale. Therefore, a spatial domain attention method is used. It transforms spatial information from the original image to another space while retaining key details, avoiding the adverse effects of channel attention mechanisms and improving target identification in high-density field scenarios.

The performance of five mainstream detection models—Faster R-CNN, YOLOv8n, SSD, YOLOv5S, and YOLOv8—was tested on a kale dataset collected from a large field. The hardware platform for the test was the NVIDIA Jetson AGX Orin (32 GB video memory, CUDA 11.4), used to comprehensively evaluate the recognition accuracy of the models. Evaluation parameters such as mean average precision (mAP), image processing time, precision (P), and recall (R) (Wu et al., 2024) were adopted as the four criteria for model evaluation.

The response speed of the solenoid valve in the variable-rate spraying system directly influences flow control precision. In pest control operations for field kale, pesticide spray parameters must be dynamically adjusted based on crop growth and pest types. Because different pesticides cover different areas, constructing a theoretical mathematical flow decision model is crucial for achieving precise system-level flow control. The theoretical model for application rate per unit area in high-density kale fields (Zhang H. et al., 2024) is given in Equation 5.

where Q_area for the unit area of the amount of drug application, the unit is L/ha; C for the active ingredient concentration of the agent, the unit is %; N_plant unit area planting density, the unit is plant/ha.

There are significant differences in the canopy density of kale, and its canopy density and volume parameters directly affect the opening and closing control logic of the solenoid valve. Canopy density of kale refers to the number and distribution of kale leaves per unit area, serving as a vital measure of plant vigor and leaf coverage in the field. In the dynamic spraying management system for different kale canopy areas, a stereo visual sensor first captures the canopy’s volume and center position. The core processor then estimates pest and disease risks and analyzes the canopy density for each area. For the upper canopy edge, where less pest and disease occur due to more light, the system reduces the duty cycle via PWM signals to decrease pesticide application. The center area, with dense leaves and high pest/disease risk, is managed by the system module, which dynamically adjusts the electromagnetic valve’s duty cycle based on YOLOv8n’s visual feedback. This optimizes liquid distribution for precise spraying. For the lower canopy edge, which is near the ground and hard to spray, the system increases spray pressure using image data to enhance droplet penetration, ensuring effective coverage. In general, the higher the canopy density, the greater the corresponding required pesticide dosage. A three-dimensional canopy volume characterization was constructed based on canopy area and density (Manandhar et al., 2020). In this study, a pesticide flow decision model, Kcv, centered on canopy density is established for verifying the dynamic response characteristics of the PWM duty cycle flow model, and the model expression is shown in Equation 6.

where C_max is the maximum coverage area per unit area of the spray target in m²; NLWA is the normalized leaf weighted area in m²; NALLI is the normalized leaf anisotropic blade index; C_W pesticide maximum threshold.

Further, the PWM duty cycle–flow theoretical model Q(D, P) (Wen et al., 2018), whose expression is shown in Equation 7, is defined and embedded into the control unit of the variable-rate spraying system, controlled by adjusting the duty cycle of the solenoid valves’ opening and closing in real time.

where K = 0.694 is the flow coefficient, n = 0.583 is the pressure correction index, in line with ISO 16119–3 standard verification.

To verify the theoretical flow rate of the model and the control accuracy of the actual spraying unit, a spraying flow rate test was conducted to evaluate the opening and closing response characteristics of the solenoid valve under synergistic regulation with the PWM duty cycle signal. The test included four pressure gradients: 0.1 MPa, 0.2 MPa, 0.3 MPa, and 0.4 MPa, and duty cycle tests ranging from 10% to 90% under a 10 Hz control signal. Preliminary tests showed that within the low duty cycle range of 20%–40%, due to the imbalance in solenoid valve opening and closing time, there were significant fluctuations in flow output, and the terminal effective spray volume was insufficient to meet the pest prevention threshold requirements. In the 60%–80% duty cycle range, the system flow stability significantly improved, and pipeline pressure fluctuations were minimal. To quantify the control characteristics, the test was repeated using a flowmeter across 5 duty cycle gradient groups, each group maintained for 30 s for system stabilization. Data were collected and averaged over 5 repetitions to determine the effective output value.

To evaluate the variable control accuracy of the variable-rate spraying system in a field kale environment, a graded control experiment was designed, including a dynamically adjusted variable spraying group and a fixed-flow constant spraying group for comparison. Based on the morphological characteristics of the kale canopy (plant height 30–50 cm, row spacing 50 cm) (Wang Z. et al., 2024), the experimental target area was divided into three parts: the upper canopy edge (0–10 cm from the top), the target center (agronomic core control area), and the lower canopy edge (0–10 cm from the base). Flow data for each area were collected in real time using a Coriolis mass flowmeter (accuracy ±0.5%). During the test, the sprayer traveled at a constant speed of 1 m/s, the system working pressure was set to 0.3 MPa, and each test group was repeated three times with averaged results to reduce random error. For safety, water was used instead of pesticides, and line pressure stability was maintained using a pressure compensation valve. Different spraying methods were used to evaluate flow accuracy across kale target areas.

In agricultural plant protection operations, droplet deposition performance directly determines pesticide utilization and pest control efficiency. In this study, fog droplet deposition density, deposition amount, and coverage were used as core evaluation indicators. Field performance validation of the RealSense D455 binocular vision-based variable rate spray system was conducted. The experiment took place on December 3, 2024, at a vegetable plantation site in Junshan District, Yueyang City, Hunan Province (29.22°N, 112.88°E), under sunny, windless conditions with a temperature of 13.8 ± 0.5°C and relative humidity of 36%–43%. A control group (constant flow rate of 2.0 L/min) and a CV dynamic regulation treatment group were tested. Sampling points were arranged according to ISO 24253-2:2015 Plant Protection Machinery - Test for Arable Crops Deposit, as shown in the schematic diagram in Figure 5B . In the figure, each ridge of kale comprises 4 rows. In Figure 5B , A/B/C/D correspond to the respective rows of kale. Notably, the kale is planted in accordance with standard agronomic practices: a ridge spacing of 1.2 m, row spacing of 65 cm, and plant spacing of 60 cm. Based on the morphological characteristics of field kale, the target area was divided vertically into three gradients: upper canopy edge, target center, and lower canopy edge. For each kale group, 5–10 pieces of water-sensitive paper were placed in the same area, and cochineal solution was used in place of pesticide. Droplet coverage C d (%), deposition density ρ d (drops/cm²), and deposition per unit area V_d (uL/cm²) (Salcedo et al., 2021) were calculated according to Equations 8-10.

where A₁ indicates the droplet coverage area of the water-sensitive paper in cm²; A₂ indicates the total area of the water-sensitive paper region in cm²; ρ d is the droplet deposition density in drops/cm²; N_d indicates the total number of droplets in the water-sensitive paper region, which is automatically counted by the image analysis software.

where Ce₁ indicates the percentage of cochineal in the eluent, i.e., the corresponding concentration; V is the volume of eluent in mL; Ce₂ is the concentration of cochineal in the spray stock in ml/cm²; S indicates the area of water-sensitive paper in cm².

Water-sensitive paper (Chongqing Liuliushanxia Plant Protection Science and Technology Co., Ltd.) and cochineal reagent (Fuzhou FeiJing Plant Science and Technology Co., Ltd.) were used. The paper size was 110 mm × 35 mm, with an effective detection area of 385 cm². A 0.5 g/L cochineal solution was prepared with pure drinking water and detected at a wavelength of 507 nm using a Shimadzu UV-1240mini spectrophotometer, in accordance with GB 12475-2006. The kale planting parameters followed a monopoly planting pattern: four rows per monopoly, row spacing 65 cm, plant spacing 60 cm, monopoly width 1.2 m, in line with DB62/T 1978-2021. Water-sensitive paper was placed in three canopy zones: target center (stem–leaf junction), upper edge (0–15 cm from top), and lower edge (10–20 cm from ground), with ten sampling points per area, spaced at 60 cm in Figure 5A . A blank control group was included in a non-sprayed area. Spray pressure was 0.3 MPa, the sprayer moved in a straight line along the ridge at 1 m/s ( Figure 5D ), and water-sensitive paper was deployed 30 min prior to spraying to prevent environmental humidity interference. The variable-rate group adjusted flow in real time (1.5–3.0 L/min) using RealSense D455 data; the constant-rate group maintained 2.5 L/min at 0.3 MPa. Water-sensitive paper was recovered 5 min post-spray, dried in light, then scanned and analyzed using DepositScan software (Tewari et al., 2020). Each test group was repeated three times, outliers removed, and average values taken. Coverage and deposition density were calculated using Equations 8 and 9.

To quantify droplet deposition on kale canopy, water-sensitive paper from the test ( Figure 5C ) was analyzed. To guarantee the safety of operators, carmine solution (which is made of biological reagents) was used in the experiment instead of pesticides. The cochineal solution was calibrated via UV spectrophotometry with the UV-1240mini, and eluent absorbance (ABS) was measured. A linear regression equation was established (Equation 11) through standard concentration gradient testing. Droplet spatial distribution uniformity and liquid solution permeability were found to be kinetically coupled in agricultural spray deposition. When conducting field operations, the spray droplets landing on the water-sensitive papers placed in the kale field cause a physical diffusion phenomenon. To prevent the liquid solution’s physical diffusion from affecting the spray deposition results, the water-sensitive papers should be photographed promptly during the experiment. Based on plant surface interface science, the coefficient of variation (Dai et al., 2019) was used to evaluate uniformity (Equations 12-14), calculated as follows:

where ABS is the absorbance value of the measured solution; C_e is the concentration of the measured cochineal solution in g/L.

Format:

where CV is the coefficient of variation; δ is the sample standard deviation; u is the sample mean; u_i is the sample observation; and n is the number of samples in the data set.

To validate the target flow accuracy and dynamic response of a binocular vision-guided variable-rate spraying system, this study focused on the coupling between solenoid valve switching frequency and spray volume. Operation efficiency was optimized by adjusting sprayer speed. Experiments were conducted per JB/T 9782–2014 Test Methods for Plant Protection Machinery (Wang et al., 2023). Based on field agronomy and anticipated variable spray response delay, solenoid valve flow at different start/stop frequencies was tested to identify discrepancies between actual and theoretical spray volumes. Sprayer speed was adjusted across four groups: 0.5 m/s (low), 1 m/s (benchmark), 1.5 m/s (medium), and 2 m/s (high), tested with different duty cycles. Spray volume during a single valve cycle and over 1 min was measured using a standard 1 mL measuring cup, repeated 10 times per frequency group. The experiment was conducted using standard flat - fan nozzles. Spray nozzle models FPV110-015, FPV110-02, FPV110-03, and FPV110–04 were used, with spray pressures set to 0.2–0.35 MPa according to nozzle parameters. The experiment was repeated 10 times for averaging. Results confirmed the effects of rapid solenoid valve switching on flow regulation, providing a basis for determining the optimal operating speed of the variable-rate system and improving kale disease control efficiency in large fields.

A systematic experimental design was used to analyze the relationship between PWM duty cycle and spray flow under fan-shaped anti-drift nozzles at pressure levels from 0.1–0.4 MPa. Flow test data were processed via linear regression analysis using Origin 2021. The results of solenoid valve flow response to different duty cycle signals are shown in Figure 9 , based on the average of five test groups. The theoretical and actual flow rates closely matched. The regression coefficients (R²) for the spraying models under four pressures were all above 0.9, indicating strong linearity. Additionally, in similar studies, Han et al. (2024) achieved significant success using an energy - saving PWM - based electromagnetic - valve control approach. This underscores the effectiveness of PWM methods in delivering precise electromagnetic - valve control. The composite model average was R² = 0.9958, meaning the PWM duty cycle signal explained 99.58% of the flow rate variation. The maximum relative error was 4.1%, meeting the ISO 5628-1:2017 Plant Protection Machinery-Spraying Equipment technical standard for precision spraying operations. This error threshold is below 5%, aligning with the requirements for accurate application in high-density crops like kale. As system pressure increased from 0.1 MPa to 0.4 MPa, flow rates also increased, consistent with fluid dynamics theory. When the duty cycle was below 20%, the flow dropped below 0.01 L/min due to solenoid valve mechanical delay and the minimum pressure threshold. In the 40%–80% duty cycle range, the coefficient of variation (CV) was less than 2%, confirming suitability for variable spraying applications, with minimal influence from signal changes.

In the performance test of kale spraying across different canopy areas, spray volume per test kale and canopy area were recorded. Results are shown in Table 3 , comparing continuous and variable spraying. The average spray volume under variable spraying in three replicated groups was significantly higher than under continuous spraying, and the CV decreased by 0.2%. For kale canopy edge zones, spray volume was reduced by 42.31% at the upper edge and 52% at the lower edge, significantly improving pesticide targeting at the bullseye. However, environmental airflow caused some spray drift, and leaf interference affected droplet penetration. Overall, the variable-rate spraying system demonstrated clear advantages over constant spraying in targeting kale pest control. The results showed that spray distribution was significantly improved under variable spraying. As seen in Figure 10 , the spray volume in the bullseye region increased by an average of 10.7% compared to constant spraying, closely aligning with control requirements for pest management (Seol et al., 2022). The similar variable and constant spraying results in different kale groups may stem from spray drift during the test (Li et al., 2021). This has minimal impact on overall pest control in extensive kale fields.

During field tests, droplet deposition data from different spraying methods were obtained via DepositScan software analysis of water-sensitive paper. As per NY/T 650–2013 agricultural standards for kale pest control, spray deposition density should exceed 85 drops/cm², and coverage should surpass 20%. As shown in Table 4 , compared to constant spraying, the CV-based variable-rate method achieved a droplet deposition density increase of 28.45% and deposition volume increase of 0.87 µL/cm²—an improvement, though not drastic. These results generally meet on-target spraying standards. In the constant mode, coverage was 22.93%, with some fluctuation likely due to spray drift, which caused non-target deposition and droplet coalescence, reducing the number of distinct droplets. The reason for these results is similar to Zhang’s findings on PWM - based variable spraying (Zhang C. et al., 2024). This further shows that spray drift can affect the system’s deposition performance. In the CV-based variable-rate mode, the rapid response of the vision module ensured accurate flow control. The solenoid valve, controlled via varying PWM duty cycles, switched frequently and generated a water hammer effect (Liu et al., 2014), producing smaller droplets. Consequently, droplet count increased in variable spraying, with a variation of 0.33 µL/cm².

At the end of the test, atomized deposition on water-sensitive paper was scanned using the Aficio MP 7502 scanner (Deep Xiangyu Technology Co., Ltd., Ricoh Group, RICOH) at a resolution of 600 × 600 ppi. DepositScan was used to process the grayscale images and extract droplet coverage and deposition density to evaluate the spray effect test indices. Based on the droplet deposition distribution data, Origin 2021 software was used to generate distribution maps of the mean droplet coverage and deposition density across the kale canopy area in the field. The droplet deposition density under different spray modes is shown in Figures 11A, B . Compared with constant spraying, the deposition density in the canopy target center under variable-rate spraying was significantly higher, with an average increase of 16.1%, indicating that the variable-rate system better achieves targeted spraying. This contributes to improved pest and disease control in the plant core. In contrast, constant spraying resulted in high deposition in canopy edge areas (non-target zones), with wide-area coverage leading to pharmaceutical waste. In both modes, droplet deposition density exceeded 16 drops/cm², meeting GB/T 8231, Guidelines for the Rational Use of Pesticides in China. Furthermore, the significantly higher deposition density in the target center under variable-rate spraying is attributed to the binocular vision sensor’s rapid target acquisition and the solenoid valve’s frequent switching, which increases local nozzle pressure. The system can thus adjust the flow rate according to variations in the canopy volume grid at different locations, enabling precise application. Regarding droplet coverage under both modes ( Figures 11C, D ), results from the variable spraying test show that average droplet coverage at the target center reached 34.42%, representing a 17.72% increase compared to constant spraying. Although the water hammer effect from frequent solenoid valve switching may affect results, variable-rate spraying still showed good droplet uniformity and stronger penetration.

Testing the system’s performance at different travel speeds revealed that at speeds over 2 m/s, the binocular vision sensor could still identify kale targets; however, the solenoid valve’s frequent switching hindered accurate measurement of spray volume per pass. Therefore, in this test, the valve’s opening time was set to 0.5 s and closing time to 1 s to ensure accurate measurement at 2 m/s. To minimize error, each test was repeated five times under the same conditions, and the average values were taken, as shown in Figure 12 . As indicated in the figure, the system’s theoretical spray volume aligns closely with the actual spray volume. Here, “theoretical” refers to the system’s theoretical flow rate, while “actual” denotes the measured value. This indicates the electromagnetic valve’s flow reliability is effective. The corresponding single and cumulative spray volumes for four travel speeds over 1 min are shown in Table 5 . Results show that increased speed significantly impacts spray volume. At 2 m/s, the FPV110–04 nozzle showed a single-pass error of 1.79%, while the FPV110–03 at 1.5 m/s had an error of 4.64%, indicating a strong influence of travel speed on actual volume. The influence of pressure on spray volume aligns with the findings of Dai et al. (2019) on electromagnetic valve - controlled spray flow characteristics under dynamic conditions. Thus, it’s essential to conduct tests on the actual flow values of variable - rate spraying systems. Higher spray pressure increased atomization, enabling droplets to better penetrate dense canopies at high speeds and reducing under-coverage. However, at low pressure (e.g., FPV110–015 at 0.20 MPa), fog droplets lacked kinetic energy and were prone to rebounding, contaminating non-target areas and diverging from theoretical values. This suggests that the system performs best at a medium travel speed of 1 m/s for effective kale pest and disease control.