LiDAR-Assisted UAV Variable-Rate Spraying System

Abstract

In wheat pest and disease control methods, pesticide application occupies a dominant position, and the use of UAVs for precise pesticide application is a key technology in precision agriculture. However, it is difficult for existing UAV spraying systems to accurately achieve variable spraying according to crop growth conditions, resulting in pesticide waste and environmental pollution. To address this issue, this paper proposes a LiDAR-assisted UAV variable-speed spraying system. Firstly, a biomass estimation model based on LiDAR data and RGB data is constructed, LiDAR point cloud data and RGB data are extracted from the target farmland, and, after preprocessing, key parameters including LiDAR feature variables, canopy cover, and visible-light vegetation indices are extracted from the two types of data. Using these key parameters as model inputs, multiple machine learning methods are employed to build a wheat biomass estimation model, and a variable spraying prescription map is generated based on the spatial distribution of biomass. Secondly, the variable-speed spraying system is constructed, which integrates a prescription map interpretation module and a PWM control module. Under the guidance of the variable spraying prescription map, the spraying rate is adjusted to achieve real-time variable spraying. Finally, a comparative experiment is designed, and the results show that the LiDAR-assisted UAV variable spraying system designed in this study performs better than the traditional constant-rate spraying system; while maintaining equivalent spraying effects, the usage of chemical agents is significantly reduced by 30.1%, providing a new technical path for reducing pesticide pollution and lowering grain production costs.

1. Introduction

Wheat, as a globally cultivated staple crop, plays a critical role in worldwide food production [1,2]. It stands as one of China’s three primary grain commodities, alongside rice and maize [3]. Ensuring sustainable high-yield wheat production constitutes a critical strategic priority for China’s economic development and food security [4,5]. Currently, chemical-based crop protection inputs dominate pest management strategies, exhibiting significantly higher adoption rates than alternative control measures [6]. Pesticide application predominantly relies on empirical decision-making, resulting in suboptimal efficacy and excessive chemical loading [7]. Compared to conventional ground-based application methods, agricultural UAVs demonstrate significant advantages in operational efficiency, terrain adaptability, and whole-growth-cycle applicability [8]. These capabilities enhance application rationality, reducing chemical usage and environmental contamination risks [9]. The efficacy is particularly pronounced in heterogeneous canopies like wheat fields, where spatially adaptive spraying strategies are essential.

UAV spray dosage determination involves multivariate factors including crop biomass, flight parameters and spray characteristics [10]. Conventional constant-rate systems fail to respond to real-time biomass variations, resulting in chemical overuse in sparse zones. The existing UAV variable spraying systems generate prescription maps by using simple image recognition technologies; without considering the biomass information of crops, there is still the phenomenon of liquid medicine waste [11].

Currently, there are two types of mainstream sensors used for UAV-based biomass estimation methods: RGB cameras and multispectral cameras [12,13,14,15]. UAV remote sensing based on RGB data was first used for crop biomass monitoring [16,17,18]. Jiang et al. [19] developed a multisource integration approach combining spectral vegetation indices, RGB-derived structural indices, and meteorological parameters, achieving a 0.92 correlation coefficient for rice biomass estimation using random forest regression. However, significant limitations exist: when acquiring crop canopy information via UAV-mounted imaging systems, the substantial sensor-to-canopy distance causes measurable accuracy degradation due to signal attenuation and structural data loss. UAV-based multispectral remote sensing employs drone-mounted multispectral cameras to acquire high-frequency multispectral imagery. This technology exhibits two distinctive advantages: seamless image mosaicking capability and customizable spectral band configurations that satisfy general crop monitoring requirements [20,21]. When integrated with texture analysis, biomass inversion models combining spectral indices and textural features demonstrate robust performance across all growth stages [22]. Li et al. [23] established that vegetation indices derived from specific spectral bands effectively estimate crop height. Despite the utility of textural parameters for biomass estimation, spectral saturation effects under high canopy coverage remain a significant constraint for multispectral-based biomass prediction [24]. Additionally, similar to RGB cameras, multispectral cameras cannot directly acquire crop structural information, resulting in the low accuracy of biomass estimation models. It should be noted that compared to professional multispectral imaging equipment, RGB-based vegetation indices have certain limitations: for instance, they have limited spectral information, are highly sensitive to imaging conditions, require calibration and standardization, and even with identical RGB camera settings, vegetation indices calculated directly from images acquired by different devices may lack direct comparability. It should also be noted that multispectral cameras are expensive, typically costing several times more than RGB cameras.

Recent studies have explored the potential of UAV light detection and ranging (LiDAR) in biomass estimation for winter wheat [25,26]. Unlike traditional optical remote sensing [27], LiDAR-collected point cloud data possesses unique penetration capabilities that reveal three-dimensional canopy architecture [28,29,30]. Compared to optical methods, LiDAR operates independently of lighting conditions and weather variations, enabling all-weather stable performance. Furthermore, the high-resolution nature of LiDAR data facilitates a granular analysis of crop growth dynamics, significantly enhancing biomass estimation accuracy [31,32]. Fareed et al. [33] established that multi-temporal LiDAR achieved R² = 0.83–1.00 for tall crops versus R² = 0.59–0.75 for short crops across phenological stages. Harkel et al. [26] employed UAV-mounted VUX-SYS LiDAR to acquire canopy data for biomass and height estimation in potato, sugar beet, and winter wheat. Sofonia et al. [34] demonstrated UAV LiDAR’s superiority over photogrammetric point clouds for sugarcane biomass estimation despite comparable height measurement accuracy.

Researchers have attempted to use LiDAR point cloud information to guide the variable-rate spraying of UAVs [35]. For example, Shaoyong et al. [36] used airborne LiDAR to estimate the canopy volume of cotton and generated variable-rate spraying prescription maps based on the volume, ultimately achieving a 43.37% reduction in spray volume. The above studies fully demonstrate the application potential of LiDAR in the field of UAV variable-rate spraying.

In response to the shortcomings of existing research, this paper proposes a LiDAR-assisted UAV variable-rate spraying system; the method introduces LiDAR and RGB imagery to construct a wheat biomass estimation model, using LiDAR-derived parameters and canopy coverage as structural descriptors for wheat and combining visible-light vegetation indices acquired by RGB cameras as model input parameters. After obtaining the biomass of each plot through the biomass estimation model, spraying prescription maps are generated, and a variable-speed spraying system is established. The system architecture integrates a prescription map parsing module and a PWM (Pulse Width Modulation) control module, enabling real-time variable-rate application according to spatial prescription maps. The innovation of this paper lies in proposing an integrated wheat biomass estimation model that fuses LiDAR and RGB data; by combining the three-dimensional structural parameters obtained from LiDAR with the vegetation indices extracted from RGB imagery, a high-precision biomass estimation model is constructed. Furthermore, the model dynamically guides variable spraying operations based on biomass estimation values, achieving precise matching between spray volume and crop growth conditions. This not only significantly improves the accuracy of variable spraying operations but also effectively reduces pesticide waste and environmental pollution risks.

2. Materials and Methods

2.1. Experimental Site

First, we collected ground data and UAV data to construct a biomass inversion model. Field experiments were conducted at the Qinfeng Seed Industry Biobreeding Experimental Station (34.28° N, 108.07° E) in Yangling District, Xianyang City, Shaanxi Province, China. Figure 1a presents a panoramic view of the biomass estimation model experimental area; the winter wheat cultivar Xinong 162 was sown as the test crop.

The biomass estimation model experimental area measured 50 m × 90 m, and the entire field was subdivided into 70 plots, as shown in Figure 1a. Five wheat plants per sampling area were measured following the five-point sampling method, and these plants were harvested approximately 2–5 cm above ground level, yielding 70 sample sets. In the laboratory, samples were enzyme-deactivated at 105 °C for 60 min, then oven-dried at 75 °C until constant weight was achieved, and the dry mass of each sample was weighed to determine biomass at respective growth stages. The variable application experimental area is 15 m long and 18 m wide, and each small sub-plot is 5 m long and 3 m wide, with a total of 15 sub-plots.

Figure 2 presents the proposed method flowchart; the proposed methodology starts with the selection of selecting experimental plots and collection of LiDAR point cloud data alongside RGB imagery. Subsequent processing extracts LiDAR feature variables, canopy cover, and visible-light vegetation indices from these datasets, and these parameters serve as inputs for constructing biomass estimation models. For target spraying plots, point cloud and visible-light data are acquired and processed through the biomass model to evaluate plot-level biomass. Based on biomass variations across subplots, corresponding prescription maps are generated. Finally, a variable-rate spraying system is designed, integrating the prescription map, and variable application experiments are conducted to assess the deposition uniformity and chemical-saving efficacy of the implemented system.

2.2. Point Cloud and Visible-Light Data Acquisition

The UAV data acquisition system consists of a payload platform and an airborne LiDAR unit, with its structural configuration illustrated in Figure 3.

The payload platform utilizes a quadrotor UAV model DJI Matrice 300 RTK (SZ DJI Technology Co., Ltd., Shenzhen, China), primarily comprising a flight control system, an inertial measurement unit (IMU) for recording attitude angles, a Real-Time Kinematic (RTK) positioning system providing location information, and a ground control station. The airborne LiDAR system is the Zenmuse L1, which integrates a Livox LiDAR module, high-precision IMU, a 1-inch RGB camera, and a 3-axis stabilized gimbal, while acquiring high-accuracy positioning data from the DJI Matrice 300 RTK. When the UAV uses RTK positioning, its navigation system can achieve a horizontal accuracy of 1 cm and a vertical accuracy of 1.5 cm. The L1 LiDAR has a maximum detection range of 450 m, a point acquisition rate of 240,000 points per second, and supports triple-return echoes. The RGB camera offers a maximum resolution of 5472 × 3648 pixels and supports both JPEG and RAW formats.

LiDAR data and RGB data acquisition occurred on 23 April 2025 and 11 May 2025 between 10:00 and 14:00 h under clear weather conditions in open terrain with negligible wind, and wheat biomass data were collected simultaneously. Flight operations were conducted at 20 m altitude above ground level with a speed of 2 m/s, featuring 70% along-track overlap and 80% cross-track overlap. Pre-flight calibration included boresight alignment between LiDAR and camera using planar targets, and the scanning mode employed triple-return repetitive scanning.

2.3. Data Preprocessing

2.3.1. Point Cloud Data Preprocessing

Raw LiDAR point cloud data cannot be directly applied in practical research and typically requires post-processing. Initially, data from the GNSS receiver mounted on the flight platform and base station are post-processed. The GNSS and IMU data are coupled to derive position and orientation information (POS data) for the LiDAR scanner.

Following POS data computation, precise trajectory calculation is performed using GNSS base station data and POS data. The raw LiDAR data are then processed through DJI Terra software V4.0, encompassing trajectory data matching and 3D point cloud visualization, ultimately exporting point cloud data in LAS file format.

During LiDAR point cloud acquisition and processing, noise interference represents a prevalent challenge. Given that noise primarily originates from non-field objects beyond the wheat plots, and considering the high-density clustered distribution characteristics of wheat plant point clouds, the Cloth Simulation Filter (CSF) technique was ultimately selected for outlier noise removal.

2.3.2. Visible-Light Data Preprocessing

Preprocessing of visible-light data primarily involves image stitching, utilizing Pix4Dmapper v4.5.6 software. This process encompasses four critical stages: feature point extraction and matching, aerial triangulation, dense point cloud generation, and 3D model reconstruction.

2.4. Data Processing

2.4.1. LiDAR Feature Variable Extraction

Feature variables serve as critical data for predicting wheat height, and reconstructed point clouds require sequential processing before feature extraction: ground/non-ground point separation via filtering, Digital Elevation Model (DEM) generation from ground points, and elevation normalization.

Similarly, the CSF method was applied to process the reconstructed LiDAR point clouds, effectively segregating ground and non-ground points.

Based on UAV LiDAR characteristics and regional topography, an irregular triangulation interpolation method generated a 0.5 m resolution DEM using separated ground points. Following filtering, LiDAR points were classified into ground points and vegetation points. Elevation normalization was then performed by calculating the elevation difference between point clouds and the DEM surface, and the point clouds of elevation normalization are illustrated in Figure 4. This terrain-agnostic process produces normalized point clouds that accurately represent topographic relief while preserving multi-echo properties.

Elevation-normalized data enabled the extraction of key LiDAR feature variables, including the maximum, mean, and standard deviation of heights, as well as height percentiles. All feature variables and their specific descriptions are listed in

Table 1. LiDAR feature variables.

Variable	Description
Hmean	Mean height
Hmax	Maximum height
Hsd	Height standard deviation
Hvar	Height variance
Hcv	Coefficient of variation
Hmad	Median absolute deviation
Hskew	Height skewness
H(40th,50th,60th,70th,80th,90th,95th,99th)	Percentile heights
CPR	Canopy relief ratio

.

2.4.2. Canopy Cover Extraction

Canopy cover reflects the occlusion extent of vegetation canopy over the ground surface, formally defined as

$$\mathrm{C}\mathrm{a}\mathrm{n}\mathrm{o}\mathrm{p}\mathrm{y} \mathrm{c}\mathrm{o}\mathrm{v}\mathrm{e}\mathrm{r}={\displaystyle \frac{\mathrm{V}\mathrm{e}\mathrm{r}\mathrm{t}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{l} \mathrm{P}\mathrm{r}\mathrm{o}\mathrm{j}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n} \mathrm{A}\mathrm{r}\mathrm{e}\mathrm{a}}{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{G}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{n}\mathrm{d} \mathrm{A}\mathrm{r}\mathrm{e}\mathrm{a}}}\times 100\%$$

(1)

When derived from LiDAR data, canopy cover extraction leverages active remote sensing principles. Through laser pulse interactions with vegetation and terrain, 3D structural information is acquired, enabling the precise calculation of the vertical projection occlusion ratio. In this study, the LiDAR point clouds obtained from elevation normalization were processed using the Point Cloud Magic V2.0 software to obtain the canopy cover.

2.4.3. Visible-Light Vegetation Index Extraction

Visible-light vegetation indices constitute a class of spectral indicators derived from the red, green, and blue bands of remote sensing imagery. In this study, GIMP image processing v3.2 software was employed to precisely crop 70 experimental plots, and the Python-based processing pipeline was implemented for the batch extraction of mean grayscale values across all plots, leveraging scientific computing libraries including rasterio, numpy, and pandas. The extracted mean grayscale values, denoted as R, G, and B, were utilized to compute 21 visible-band vegetation indices [37]. Given that the biomass estimation model may run on the UAV’s onboard computer, whose computing capacity is limited, the Pearson correlation screening is applied to eliminate indices with weak correlations to biomass in order to reduce the model’s computational load, with those exhibiting strong correlations with wheat biomass selected and redundant variables eliminated. This process enhanced model construction efficiency and prediction accuracy, and ultimately, 19 optimized vegetation indices (

Table 2. Visible-light vegetation index.

Visible-Light Vegetation Index	Formula	Visible-Light Vegetation Index	Formula
r	R/R + G + B	r/b	r/b
g	R/R + G + B	g/b	g/b
r − b	R − b	g − b	g − b
r + b	R + b	(r − b)/(r + b)	(r − b)/(r + b)
(r − g − b)/(r + g)	(r − g − b)/(r + g)	GRVI	(g − r)/(g + r)
EXG	2g − b − r	MGRVI	(g2 − r2)/(g2 + r2)
RGBVI	(g2 − br)/(g2 + br)	NDI	(r − g)/(r + g + 0.01)
EXR	1.4r − g	VARI	(g − r)/(g + r − b)
EXGR	3g − 2.4r − b	GLA	(2G − B − R)/(2G + B + R)
CIVE	0.441r − 0.881g + 0.385b + 18.78745

) were retained for wheat biomass prediction.

2.5. Model Evaluation and Results Analysis

2.5.1. Model Performance Assessment

This study employed the coefficient of determination (R²) and root mean square error (RMSE) as primary evaluation metrics for model performance assessment.

Coefficient of Determination (R²) quantifies the goodness-of-fit in regression models, representing the proportion of variance in the dependent variable explainable by independent variables. Ranging between 0 and 1, it indicates the correlation between measured and predicted values, and the calculation formula is given by (2). Root Mean Square Error (RMSE) measures the average deviation between predicted and true values, and the calculation formula is given by (3).

$${\mathrm{R}}^2=1-{\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}})}^2}{\sum _{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-\overline{{\mathrm{y}}_{\mathrm{i}}})}^2}}$$

(2)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt[]{{\displaystyle \frac{1}{\mathrm{n}}}{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}})}^2}$$

(3)

2.5.2. Comparative Analysis

Three modeling approaches were employed: Stepwise Regression (SR), Partial Least Squares Regression (PLSR), and Random Forest (RF). The biomass estimation model was constructed based on a total dataset of 140 samples, collected at two distinct time points (70 samples per time point), including 100 samples in the training set and 40 samples in the independent test set. Among the 70 samples collected at the first time point, the average biomass per wheat plant was 70.78 g; for the samples collected at the second time point, the average biomass per wheat plant was 92.13 g, which is consistent with the growth trend of wheat. A comparative analysis of biomass estimation models was conducted across these three methods.

The results are shown in

Table 3. Different parameter prediction results for each model.

Model	Features	R2	RMSE (g/m2)
SR	Visible-Light Vegetation Indices	0.618	305.456
	Visible-Light Vegetation Indices, Canopy Cover and LiDAR Feature Variables	0.709	256.544
PLSR	Visible-Light Vegetation Indices	0.626	274.319
	Visible-Light Vegetation Indices, Canopy Cover and LiDAR Feature Variables	0.730	246.351
RF	Visible-Light Vegetation Indices	0.823	210.857
	Visible-Light Vegetation Indices, Canopy Cover and LiDAR Feature Variables	0.834	202.431

.

As shown in Table 3, among the two input parameter combinations, the combination of “visible-light vegetation indices, canopy coverage, and LiDAR feature variables” achieves the highest estimation accuracy. Among the three machine learning methods, the RF algorithm yields the highest accuracy, followed by the PLSR algorithm, while the SR algorithm performs the worst. Therefore, the biomass estimation model constructed in this study adopts the RF algorithm, with input variables including visible-light vegetation indices, canopy coverage, and LiDAR feature variables.

2.5.3. Variable-Rate Spraying Prescription Map Generation

Building upon the biomass estimation model established in the preceding section, the variable-rate spraying experimental area was analyzed. The variable application experimental area is shown in Figure 1c; it consists of 15 plots, and due to differences in organic fertilization measures and nitrogen fertilizer application levels, there are significant differences in the growth of wheat in the 15 plots.

By employing the biomass estimation model that integrates visible-light vegetation indices, canopy cover, and LiDAR feature variables, a high-resolution biomass spatial distribution map was generated, as shown in Figure 5.

In Figure 5, the 15 subplots were classified into three tiers. Subplots with darker coloration exhibited higher biomass levels. Specifically, light green zones were defined as sparse areas, and assigned low-level spray volume. Medium green zones are denoted moderate areas, assigned medium-level spray volume. Dark green zones represented dense areas, assigned high-level spray volume. The wheat biomass spatial distribution map was thereby converted into a variable-rate spraying prescription map, where color gradations directly corresponded to chemical application differentials.

3. Variable-Rate Spraying System Design

The variable-rate spraying system consists of two core components: the prescription map interpretation module and the PWM control module. The prescription map interpretation module receives GNSS positioning data and extracts the corresponding application rate from the prescription map based on the geographic location. The PWM control module then receives the target application rate from the interpretation module and achieves real-time regulation of the pump flow through closed-loop feedback from flow sensors. Guided by the variable-rate spraying prescription map, this system implements LiDAR-assisted variable-rate spraying.

3.1. Hardware Components

This study employed the DJI MG-1P series agricultural plant protection UAV as the carrier platform for the variable-rate spraying system, but it lacked built-in variable-rate spraying and prescription map processing capabilities. The original spraying kit of the plant protection UAV was removed and replaced with the self-designed variable-rate spraying system developed in this paper; as shown in Figure 6, the system has a spraying width of 2 m and a maximum spray volume of 10 L.

The variable-rate spraying system designed in this study primarily consists of the following components: the controller (STM32F1), a LoRa module (ATK-LORA-01), a GNSS module (GT-U12), an electronic speed controller (Hobbywing X-ROTOR 40A), a brushless water pump, and a vortex flow sensor. The structural diagram of the variable-rate spraying system is shown in Figure 7.

The GNSS module comprises an HD8040X chip from HuaDa BeiDou and an active antenna. It transmits signals to the controller via a USART serial interface at a baud rate of 115,200 and a frequency of 5 Hz. The GT-U12 module provides positional accuracy within 1 m. Upon receiving the data, the controller parses the GNGGA string to acquire information such as latitude, longitude, and altitude. The controller then directly outputs PWM signals via its advanced timer to control the electronic speed controller, which drives the water pump; the system board controls the water pump’s flow rate by reading the pulse signal output from the vortex flow meter. To facilitate the real-time monitoring of the UAV’s operational status, its operational information is transmitted to the host computer via the LoRa module.

Figure 6 depicts the plant protection UAV integrated with the variable-rate spraying system; the electronic speed controller operates at 50 Hz. Regulation of the pump’s flow rate and on/off state is achieved by modulating the duty cycle of the input signal to the ESC; upon system initialization, the ESC configures the pump’s maximum and minimum flow rate thresholds and engages the arming sequence. However, zero-flow conditions persist at this stage. When the UAV enters a geofenced application zone predefined in the prescription map, the controller retrieves the corresponding prescription data and generates a PWM waveform to command the target flow rate. Conversely, upon departure from the designated operational boundary, the controller terminates pump operation.

The vortex flow meter (Moqiao Sensing MJ-A68-1) employs a turbine impeller rotated by fluid flow, which intersects magnetic flux lines to generate a frequency-modulated pulse output. Real-time flow rate quantification is derived from this pulse signal via frequency-to-flow conversion algorithms, and this measurement is fed back to the control system to dynamically adjust the PWM duty cycle, ensuring conformity between the actual flow rate and prescribed setpoints. The flow meter is installed in series between the solution tank and the pump inlet, and the discharge flow from each brushless DC pump is distributed through a flow bifurcation unit to two pressure-swirl nozzles, yielding a total system configuration of two pumps and four nozzles. Each pump delivers a maximum flow capacity of 3 L·min⁻¹.

3.2. Variable-Rate Spray Control Model

To achieve precise control of nozzle discharge rates, laboratory experiments were conducted to establish the correlation between the flow rates of four pressure-swirl nozzles and the duty cycle applied to the pump. A linear calibration model relating flow rate to duty cycle was derived, with the corresponding calibration curve presented in Figure 8. This model provides the governing equation for PWM waveform generation.

Concurrently, under constant flow conditions, the frequency-modulated pulse output from the vortex flow meter was recorded, and linear regression analysis was performed to establish the functional relationship between pulse frequency and actual flow rate. By implementing the inverse solution of the aforementioned calibration equations, target flow rates were converted into command duty cycles, thereby enabling closed-loop control of the pump’s output.

3.3. Software Design

During the execution of spraying operations based on prescription maps, the GNSS module is responsible for providing real-time positioning data of the UAV. When the system detects that the UAV has entered or left the prescription area demarcated by the geofence, the core controller will promptly acquire GNSS coordinates and perform comparative calculations with the coordinate information in the prescription map. If the UAV is within the scope of the prescription plot, the core controller will read the prescription information to control the brushless water pump to operate at the rated flow rate. Meanwhile, by receiving real-time data from the flowmeter sensor, it will use a PID algorithm to carry out closed-loop regulation of the water pump flow, ensuring that the spraying volume precisely matches the prescription requirements. The target application rate obtained from the prescription map analysis is set as the setpoint, while the control signal sent to the pump serves as the output. Meanwhile, the current actual application rate measured in real time by the flow meter is fed back to the controller, and together with the target rate, they form the input parameters of the controller. When the UAV flies out of the prescription plot, the spraying system will automatically trigger the shutdown mechanism to stop the operation. The UAV’s spraying operation and flight paths are demonstrated in Figure 9. In field trials, the aircraft flight speed was set to 1 m/s, and the spray nozzle working height was set to 1.5 m.

3.4. Field Experiment

To investigate spray coverage and deposition characteristics across differential application rates in heterogeneous field zones, a spray experiment was conducted on 21 May 2025 during the wheat’s grain-filling stage.

The experiment included two spraying modes. First, a constant-rate spraying experiment was conducted in which the plant protection UAV maintained a constant spraying rate of 1.3 L/min. After an interval, a variable-rate spraying experiment was carried out; based on the prescription map information, the 15 plots were divided into three pesticide application types—sparse, moderate, and dense. The plots of each type were sprayed at 0.5 L/min, 0.9 L/min, and 1.3 L/min, respectively.

Water-Sensitive Paper (WSP) sensors were affixed to adaxial surfaces of upper, middle, and lower canopy leaves, as shown in Figure 10, and WSP deployment is also illustrated in the figure. Overall, 45 WSP replicates per treatment were deployed, yielding a total of 90 specimens for analysis. WSPs were retrieved immediately post-application and sealed in barrier envelopes to prevent droplet coalescence.

Deposition patterns were digitized using an HP Laser MFP 1135w scanner at 600 dpi resolution for quantitative droplet analysis.

4. Results

4.1. Spray Deposition

Figure 11 shows the comparative analysis of spray deposition under constant-rate application (CRA) and variable-rate application (VRA) regimes across three vegetation density zones. It can be seen from the figure that the CRA treatment consistently exhibited the highest deposition (maximum in the sparse zone: 1.32 μL·cm⁻²), which was attributed to the maximum application rate and reduced canopy interception. VRA achieved uniform deposition across regions, with the minimum deposition (0.272 μL·cm⁻²) occurring on dense canopies due to leaf arrangement-induced shading. Except for the dense zone, deposition under CRA exceeded that under VRA in all regions.

Figure 12 illustrates the spatial distribution of spray deposition represented by color map across canopy strata under CRA and VRA. The color intensity in the color map indicates the magnitude of deposition. Overall, the chromaticity of the CRA consistently remained higher than that of the VRA. Additionally, the maximum average deposition was observed in the upper part of the constant spraying zone within the sparse plot, while the minimum average deposition was found in the lower part of the variable spraying zone within the dense experimental area. The difference between the maximum and minimum deposition was 1.048 μL/cm², which is primarily attributed to variations in spraying flow rates and vegetation density, leading to different degrees of leaf shading.

Compared with VRA, CRA leads to obvious color differences and significant changes in deposition, and this indicates that when the plot density is uneven, the use of CRA will result in uneven pesticide application and is highly likely to cause over-spraying. In contrast, when VRA is carried out using prescription map information, the color will change slowly, and wheat spraying will be concentrated in the middle and upper parts.

4.2. Spray Coverage

As depicted in Figure 13, mean spray coverage is primarily governed by application modality and canopy density; the maximum coverage occurred in sparse zones (CRA: 21.60%), while minimum values were recorded in dense zones (CRA: 10.80%; VRA: 11.50%). In addition, the average coverage of the two spraying systems is very close in the dense plots, while in the medium and sparse plots, the coverage of the constant spraying system is much higher than that of the variable system. This change trend caused by the uneven density of the plots is consistent with the droplet deposition trend shown in Figure 12. In the three types of plots with different densities, the variable spraying system can maintain the consistency of coverage well.

Given that elevated spray coverage does not necessarily correlate with enhanced agronomic efficacy, a judicious reduction in the application volume is feasible while preserving bioactive performance. This paradigm arises because increased application rates may augment foliar coverage, yet exhibit diminishing returns in biological efficacy beyond critical thresholds.

The efficacy of spray distribution is intrinsically linked to droplet deposition density on target surfaces. Figure 14 delineates the correlation between mean droplet deposition density (spots·cm⁻²) and coverage for both application systems. Field validation via WSPs established that coverage exceeding 30% (vertical partition line in Figure 14) constitutes over-application, incurring significant wastage. Consequently, elevated coverage does not inherently enhance application efficiency. To achieve optimal control of pests and diseases, thresholds of 30 and 70 spots·cm⁻² (the horizontal partition lines in Figure 14, respectively) were used to assess the effective droplet deposition density of pesticides [38].

Analysis of the CRA scatter plot in Figure 14 reveals no instances of over-application (coverage > 30%) across sparse, moderate, and dense zones. Notably, the deposition density of VRA mostly falls within the range of 30–70 droplets·cm⁻². In contrast, some deposition points of CRA exceed the 70 droplets·cm⁻² threshold, and the overall performance is inferior to VRA; this is because CRA cannot adjust the spray volume according to the biomass, and in sparse canopy areas, it is prone to excessive overlapping of spray deposition, resulting in high droplet deposition density in some zones, which may cause damage to plant leaves. On the contrary, the VRA can dynamically adjust the flow rate according to the biomass, so its droplet deposition density distribution is relatively more uniform, and there are few cases where it exceeds the threshold of 70 spots·cm⁻², which ensures pesticide efficacy while saving the amount of pesticides used.

Comprehensive analysis demonstrates that the constant spraying system shows higher droplet deposition and coverage throughout the experiment area. The results are shown in

Table 4. Deposition (μL·cm⁻²).

Zone	CRA	VRA
Sparse	1.035	0.489
Moderate	0.608	0.488
Dense	0.480	0.486

and

Table 5. Coverage (%).

Zone	CRA	VRA
Sparse	21.60	10.72
Moderate	13.09	11.57
Dense	10.80	11.05

. This apparent advantage stems from CRA’s inability to adapt application volume to vegetation density, inducing excessive droplet coalescence in sparse zones (deposition: 1.035 μL·cm⁻²).

Consequently, variable-rate spraying can achieve biological efficacy close to that of constant rate spraying in dense areas with lower application rates under conditions of uneven vegetation density, and it outperforms constant-rate spraying in both deposition efficiency and coverage uniformity. The average spraying rate of VRA is 0.9 L/min, while the spraying rate of CRA is constantly 1.3 L/min, since the operation time of both applications is the same; VRA saves 30.1% of the pesticide volume. These experimental results show that increasing the application rate of chemical agents does not necessarily improve the spraying efficiency; on the contrary, it will bring negative economic and environmental impacts.

5. Discussion

This study constructed a wheat biomass estimation model by combining wheat structural parameters obtained via LiDAR with vegetation indices extracted from RGB images, and generated a variable-rate spraying prescription map. Additionally, a variable-rate spraying system was designed which can read the prescription map information and use a PID algorithm to perform closed-loop regulation of the water pump flow, ensuring that the spraying volume precisely matches the prescription requirements. Experimental results show that compared with uniform spraying, variable-rate spraying reduces pesticide dosage by 30.1% while achieving equivalent control effects, effectively avoiding excessive pesticide application. This provides a methodological reference for wheat precision spraying.

Compared with prescription map generation methods based on RGB and multispectral cameras [39,40,41], this system obtains plant structural parameters through LiDAR and adjusts the spraying amount according to biomass, thereby reducing pesticide usage. Compared with LiDAR-based prescription map generation methods, this study differs significantly from existing research: existing studies mostly use crop volume as the basis for generating prescription maps [35,36], while this study selects biomass. Since wheat is a dense crop, biomass is more instructive than volume for generating prescription maps. Furthermore, existing methods only rely on LiDAR to obtain plant structural information, resulting in a single dimension of information that cannot distinguish spectral differences between leaves, stems, and ears. In this study, on the basis of acquiring structural information, vegetation indices extracted from RGB images are integrated. The established model considers the correlation between biomass and spectral characteristics, improving the accuracy and stability of biomass estimation.

Despite these promising results, there are limitations to note: merely comparing variable-rate spraying with uniform spraying is insufficient to fully highlight the advantages of the variable-rate spraying system in this study. Future research should consider comparing it with methods based on spectral technology to identify the system’s shortcomings and further improve it. Meanwhile, this method does not involve research on wheat disease detection. However, wheat diseases have significant responses in specific spectral bands beyond visible light, and multispectral and hyperspectral cameras have great advantages in disease detection. Therefore, future research will consider constructing a biomass estimation model that integrates LiDAR information and multispectral cameras to enhance the system’s responsiveness to diseases.

6. Conclusions

This study proposes a LiDAR-assisted variable-rate spraying system for plant protection UAVs, innovatively employing wheat biomass estimates derived from LiDAR point clouds and RGB imagery to generate prescription maps. A wheat biomass estimation model and a variable-rate pesticide application system were constructed in sequence, and experiments were designed to verify the application effect of the designed system. The experiments results show that compared with the traditional constant spraying method, the variable spraying system proposed in this study can adjust the spraying amount in real time according to the crop biomass, thereby significantly reducing the amount of pesticide used. Looking to the future, it will be necessary to integrate LiDAR, RGB cameras, and edge computing devices into plant protection UAVs, deploy a wheat biomass estimation model on the edge computing devices, and send pesticide application prescription maps to the variable-rate pesticide application system in real time to further improve the real-time performance of plant protection UAV spraying operations.