Intelligent mango fertilization method

Abstract

The application discloses an intelligent mango fertilization method, comprising arranging a drip irrigation pipeline connected with a water-fertilizer mixing device in a drip irrigation area; periodically collecting canopy images by a multispectral imaging device, containing visible light and near-infrared band information of leaves, flower buds, flowers and fruits; extracting visual features and spectral features based on an image processing algorithm, the visual features including leaf color distribution, morphological thickness index, flower bud density, flower coverage rate and fruit volume proportion, and the spectral features being near-infrared reflectivity ratios of leaf and flower and fruit areas; matching a growth stage standard library according to the feature data to determine a flower bud differentiation period, a flowering period or a fruit enlargement period, calling corresponding period determination models to identify nutrient and water deficiency states; and dynamically adjusting nutrient concentration and dosage of drip irrigation liquid according to the determination results. The application is used for precise water and fertilizer management of mango planting, optimizes fertilization decisions through multi-source data fusion and closed-loop control, and improves response capability in complex nutrient demand scenarios.

Description

Technical Field

[0001] This invention belongs to the field of intelligent fertilization technology in agricultural planting, and more specifically, this invention relates to an intelligent mango fertilization method. Background Technology

[0002] In mango cultivation, scientific fertilization is a crucial factor affecting fruit yield and quality. Traditional fertilization methods mainly rely on manual experience or fixed monitoring equipment, which has the following limitations: 1. Insufficient monitoring range and data reliability: Traditional methods mostly use ground-based fixed sensors or handheld devices to collect plant data, resulting in limited single-point monitoring range and difficulty in covering large orchards. For example, fixed multispectral cameras can only image a local canopy, requiring frequent relocation of equipment, leading to low data collection efficiency. In addition, natural light intensity fluctuates with weather and time of day, directly affecting the measurement accuracy of canopy spectral reflectance. When light conditions are unstable, the spectral characteristics of leaves and flower / fruit areas are easily distorted, especially during cloudy or twilight periods, when near-infrared data deviations increase significantly. These problems reduce the reliability of nutrient determination models based on spectral analysis, making it difficult to support long-term stable fertilization decisions. 2. Lack of dynamic regulation capability under complex nutrient requirements: The requirements for nitrogen, phosphorus, potassium, and water vary significantly at different growth stages of mangoes (such as flower bud differentiation, flowering, and fruit enlargement). Traditional intelligent fertilization systems typically use single thresholds or static models for regulation, such as triggering irrigation based solely on soil moisture or leaf color, without considering the dynamic changes during growth stages. When multiple nutrients are simultaneously deficient, the system struggles to prioritize, easily leading to the blind application of a single element while ignoring other needs. For example, if phosphorus deficiency and water insufficiency coexist during flowering, direct phosphate fertilizer supplementation may result in reduced fertilizer utilization due to limited root water absorption. These problems stem from a lack of integrated analysis of multi-dimensional features (such as flower and fruit density, leaf physiological indicators), and a lack of refined design of nutrient requirement models for different stages. 3. Low accuracy in growth stage identification leads to fertilization timing deviations. In existing technologies, growth stage determination relies heavily on manual observation or single visual features (such as flower count), lacking quantitative indicators and multi-source data for verification. For example, judging the flowering period solely by flower coverage may result in misjudgment due to some flowers being obscured by leaves or insufficient imaging resolution. Furthermore, the identification of the fruit enlargement period often lags behind actual physiological changes, leading to untimely potassium fertilizer supplementation and affecting the fruit enlargement rate. 4. The logic for the coordinated regulation of water and nutrients is incomplete. In scenarios where both water and nutrients are scarce, traditional systems often employ parallel replenishment strategies, neglecting the impact of water on nutrient absorption efficiency. For example, when soil moisture content is low, directly injecting high-concentration fertilizers may cause root osmotic stress, exacerbating nutrient absorption obstacles. Furthermore, uneven soil moisture distribution after irrigation may affect the accuracy of secondary nutrient status monitoring. Solving these problems requires establishing a closed-loop control mechanism of "water first, fertilizer second," but current technologies lack real-time linkage analysis of canopy characteristics, making it difficult to achieve dynamic delay judgment and feedback adjustment.

[0003] The aforementioned shortcomings limit the application of intelligent fertilization technology in large-scale orchards, and there is an urgent need for a method that can integrate multi-source data, dynamically identify growth stages, and optimize synergistic regulation. Summary of the Invention

[0004] One object of the present invention is to address at least the aforementioned deficiencies and to provide at least the advantages that will be described later.

[0005] This invention provides an intelligent mango fertilization method that achieves accurate identification of nitrogen, phosphorus, potassium, and water through multi-dimensional feature fusion analysis. Combined with priority regulation and feedback replenishment mechanisms, it forms a closed-loop fertilization decision, improving the response capability at different stages under complex nutrient deficiency scenarios, thereby increasing the efficiency and yield of mango cultivation.

[0006] The present invention provides an intelligent mango fertilization method, comprising the following steps:

[0007] S1) Arrange drip irrigation pipelines connecting to the water and fertilizer mixing device in the drip irrigation area;

[0008] S2) A multispectral imaging device periodically acquires canopy images of mango trees, the canopy images containing visible and near-infrared band information of leaves, flower buds, flowers and fruits;

[0009] S3) Based on image processing algorithms, extract visual and spectral features of the canopy image. The visual features include leaf color distribution, morphological thickness index, flower bud density, flower coverage, and fruit volume ratio. The spectral features include the ratio of near-infrared reflectance of the leaf area to that of the flower and fruit area.

[0010] S4) Based on the visual and spectral features, the flower bud differentiation period, flowering period and fruit enlargement period of the mango tree are determined by image comparison, and then the nitrogen deficiency state, phosphorus deficiency state, potassium deficiency state and water deficiency state at different periods are identified by the preset judgment model.

[0011] S5) Based on the determination result of the nutrient deficiency state, control the water-fertilizer mixing device to dynamically adjust the concentration of the corresponding nutrients in the drip irrigation solution and the irrigation amount.

[0012] Preferably, the multispectral imaging device is mounted on a drone or a ground mobile platform. The drone periodically collects images of the mango orchard canopy according to a preset flight path, and the preset lighting conditions are maintained by the drone's built-in supplementary lighting module during image acquisition.

[0013] The drone's preset flight path is dynamically planned based on the electronic fence map of the mango orchard and the canopy projection range, and can be executed using the following existing steps:

[0014] S2.1) The terrain of the mango orchard is scanned by a lidar mounted on a drone to generate an electronic fence map containing the location of each mango tree and the canopy boundary;

[0015] S2.2) Based on the electronic fence map, the flight path is divided into several parallel routes. The distance between adjacent routes is determined by calculating the coverage width of a single shot of the multispectral imaging device and the lateral overlap rate, wherein the lateral overlap rate is 60%-70%.

[0016] S2.3) When the UAV flies along the route, it corrects its course in real time through the onboard GPS and visual obstacle avoidance module, and hovers to take pictures when it reaches the center point of the canopy of each mango tree;

[0017] S2.4) The supplementary lighting module includes a ring LED array and a light intensity sensor. Before shooting, the light intensity sensor detects the ambient light intensity. If it is lower than the preset threshold, the LED array is activated to supplement light with constant color temperature and light intensity to ensure uniform illumination on the canopy surface.

[0018] Preferably, the specific extraction methods for visual and spectral features in step S3 include:

[0019] S3.1) Perform semantic segmentation on the canopy image to separate the leaf, flower bud, flower, and fruit regions;

[0020] S3.2) In the leaf region, extract the green saturation distribution data in the HSV color space, and calculate the standard deviation of the point cloud density in the leaf vein region through three-dimensional point cloud reconstruction technology, which serves as an indicator of leaf morphology and thickness.

[0021] S3.3) Calculate the number density of flower buds per unit area in the flower bud region, calculate the percentage of the coverage area in the flower region relative to the total canopy area, and calculate the percentage of the volume in the fruit region relative to the total canopy volume.

[0022] S3.4) For near-infrared images, calculate the ratio of the average reflectance of the leaf region to the average reflectance of the flower and fruit region, and use it as the spectral reflectance ratio (spectral characteristic data).

[0023] This invention extracts different target regions (leaves, flowers, and fruits) through semantic segmentation, enhancing the comprehensiveness of features; and combines flower and fruit distribution features with leaf indicators to provide multi-dimensional support for determining the growth stage.

[0024] Preferably, the specific process for image comparison and growth stage determination in step S4 includes:

[0025] S4.1) Based on the leaf morphology and thickness index extracted in step S3.2, the flower bud number density, flower coverage percentage, and fruit volume percentage extracted in step S3.3, and the spectral reflectance ratio extracted in step S3.4, the fruit volume growth rate is calculated by accumulating the fruit volume percentage extracted each time, and matched with the preset growth stage standard library.

[0026] S4.2) The growth stage standard library includes thresholds for flower bud density, flower coverage, fruit volume percentage, fruit volume growth rate, and spectral reflectance ratio for each stage. If the current flower bud density exceeds the flower bud density threshold for the differentiation stage and the flower coverage percentage is zero, it is determined to be the flower bud differentiation stage. If the current flower coverage percentage exceeds the flower coverage threshold for the flowering stage and the fruit volume percentage is lower than the threshold for the expansion stage, it is determined to be the flowering stage. If the fruit volume percentage growth rate continuously exceeds the fruit volume growth rate threshold for the expansion stage and the spectral reflectance ratio is lower than the threshold for the spectral reflectance ratio, it is determined to be the fruit expansion stage.

[0027] The specific thresholds are as follows: Flower bud differentiation stage: Flower bud density ≥ 2.5 buds / cm² and flower coverage percentage = 0%;

[0028] Flowering period: Flower coverage percentage ≥ 30% and fruit volume percentage ≤ 15%;

[0029] Fruit enlargement period: The daily average growth rate of fruit volume is ≥3% and the ratio of spectral reflectance is ≤1.0.

[0030] The plant physiological mechanism underlying a spectral reflectance ratio threshold of 1.0 is as follows: During fruit enlargement, plants preferentially transport large amounts of photosynthetic products and water to the fruit, leading to decreased photosynthetic activity and reduced chlorophyll content in the leaves. The near-infrared (NIR) reflectance of leaves is closely related to mesophyll cell structure and water content; when leaf physiological activity weakens, its NIR reflectance decreases significantly. During fruit enlargement, rapid cell division, increased water content, and densification of the internal structure enhance the scattering of near-infrared light, resulting in a relative increase in reflectance. The combined effect of decreased leaf reflectance and increased fruit reflectance causes the reflectance ratio (leaf / flower / fruit) to be less than 1.0.

[0031] S4.3) After determining the growth stage, the nutrient determination sub-model for that period is called. The sub-model uses a random forest algorithm to fuse near-infrared reflectance, ambient temperature and humidity and historical fertilization data to output the probability value of nitrogen, phosphorus, potassium or water deficiency.

[0032] This invention identifies each growth stage through quantitative analysis of flower and fruit density and volume growth rate; it matches nutrient requirement models for each stage to avoid fertilization deviations caused by misjudgment.

[0033] Preferably, step S5 also includes steps with the following priority:

[0034] S5.21 When the probability of a single nutrient deficiency is ≥70%, the water and fertilizer mixing device shall be controlled to increase the corresponding nutrient concentration to the preset upper limit and the drip irrigation volume shall be increased to 1.2-1.5 times the standard irrigation volume;

[0035] S5.22 When the probability of a single nutrient deficiency is 30%-70%, adjust the nutrient concentration and irrigation amount linearly.

[0036] S5.23 When the probability of nutrient deficiency is less than 30%, drip irrigation should be performed while maintaining the baseline concentration and irrigation volume;

[0037] S5.24 The specific steps for prioritizing the activation of drip irrigation pipelines for water irrigation are as follows: if both water and nutrient deficiency are detected simultaneously, irrigation is prioritized until the soil moisture content reaches the standard, and then the canopy image is re-acquired for a second determination after a first set time (e.g., 24 hours).

[0038] S5.25 If nutrient deficiency still exists after the second determination, the drip irrigation system will be triggered to supplement the corresponding elements, and the delay time for feedback data collection will be dynamically adjusted according to the following rules:

[0039] a) When the supplementary element is nitrogen, the delay time is 48-72 hours;

[0040] b) When the supplemented element is phosphorus or potassium, the delay period is 5-7 days;

[0041] c) When nitrogen, phosphorus and potassium are simultaneously deficient, the longest delay time corresponding to phosphorus or potassium, 5-7 days, shall be used as the benchmark.

[0042] S5.26 After each adjustment, feedback data is collected according to the dynamic adjustment delay time until the lack probability value is <10%; if the lack probability value is still ≥10% after 3 consecutive adjustments, an abnormal alarm is triggered and automatic control is terminated.

[0043] Preferably, the specific implementation method of "calculating the standard deviation of point cloud density in the leaf vein region as a morphological thickness index through three-dimensional point cloud reconstruction technology" in step S3.2 includes:

[0044] S3.21) The point cloud data of the leaf region is sampled in layers, and the leaf is divided into several sub-regions (square grids with a side length of 5-8cm). The grids containing the center line of the leaf veins are reconstructed with high precision in three dimensions (point cloud density ≥100 points / cm²). The non-vein grids are generated with low-resolution point clouds (point cloud density ≤20 points / cm²) by linear interpolation to reduce the amount of computation.

[0045] S3.2.2) The high-precision leaf vein point cloud and the interpolated non-leaf vein point cloud are locally registered based on the improved ICP (Iterative Closest Point) algorithm with the leaf vein centerline as the reference. The number of iterations (≤10 times) and the search radius (≤1 mm) are limited. The registration data is processed in parallel by half-precision floating-point operations of the UAV's onboard GPU, and the standard deviation of the point cloud density is output.

[0046] Preferably, the specific implementation of calling the period-specific nutrient determination sub-model in step S4.3 includes the following steps:

[0047] S4.31) Based on the growth stage determined in step S4.2, load the corresponding sub-model from the pre-trained model library;

[0048] S4.32) Input the current near-infrared reflectance, ambient temperature and humidity, historical fertilization data, green saturation distribution data, and leaf morphology and thickness indicators into the sub-model; output the probability value of nitrogen, phosphorus, potassium, or water deficiency through the random forest algorithm;

[0049] S4.33) When any probability value is ≥30%, the control logic of step S5 is triggered.

[0050] Preferably, the pre-trained model library is constructed through the following steps:

[0051] S4.311) Collect historical data on mango cultivation, including canopy images, near-infrared reflectance, environmental temperature and humidity, historical fertilization records, green saturation distribution data, and leaf morphology and thickness indicators at different growth stages.

[0052] S4.312) The historical data is annotated, and the annotation content includes the actual state of nitrogen, phosphorus, potassium and water deficiency in each period;

[0053] S4.313) Divide the dataset according to the growth stage and train the flower bud differentiation period sub-model, flowering period sub-model and fruit enlargement period sub-model respectively. Each sub-model adopts the random forest algorithm. The input features are near-infrared reflectance, environmental temperature and humidity, historical fertilizer application, green saturation distribution data and leaf morphology and thickness index. The output is the nutrient deficiency probability value.

[0054] S4.314) Optimize the sub-model parameters through cross-validation until the test set accuracy reaches a preset value (e.g., preset value ≥ 85%).

[0055] Preferably, the specific training and validation methods for the sub-models include:

[0056] Flower bud differentiation period sub-model: In the input features, the weight of green saturation distribution data accounts for 40%, leaf morphology and thickness index accounts for 30%, near-infrared reflectance accounts for 20%, and environmental temperature and humidity accounts for 10%;

[0057] Flowering period sub-model: In the input features, the percentage of flower coverage accounts for 35%, near-infrared reflectance accounts for 30%, green saturation distribution data accounts for 25%, and historical fertilization amount accounts for 10%.

[0058] Fruit enlargement period sub-model: Among the input features, the ratio of spectral reflectance accounts for 50%, the fruit volume growth rate accounts for 30%, the ambient temperature accounts for 15%, and the historical fertilization amount accounts for 5%;

[0059] The random forest algorithm parameters for each sub-model are: 100-150 decision trees, 10-12 layers at maximum depth, and 5 samples per leaf node.

[0060] The present invention has at least the following beneficial effects:

[0061] This invention achieves precise identification of nitrogen, phosphorus, potassium, and water through multi-dimensional feature fusion analysis. Combined with priority control and feedback replenishment mechanisms, it forms a closed-loop fertilization decision-making system, improving responsiveness at different stages in complex nutrient deficiency scenarios. This addresses the problem of traditional intelligent fertilization methods lacking dynamic control logic based on different stages, making them difficult to adapt to scenarios with complex nutrient requirements.

[0062] This invention expands the monitoring range by leveraging the mobility of drones and combines it with active supplemental lighting to ensure consistent image acquisition conditions, thereby improving the reliability of spectral feature analysis. It addresses the limitations of traditional fixed imaging equipment, such as limited coverage and the tendency for spectral data distortion due to fluctuations in natural lighting, making stable monitoring of large-scale orchards difficult.

[0063] This invention dynamically adjusts nitrogen, phosphorus, and potassium concentrations and irrigation volume through a phased sub-model and priority control logic. In the experiment, the fertilizer application rate in the experimental group was reduced by 15%-19%, and the yield per plant increased to 36.5 kg (compared to 29.8 kg in control group 1), verifying the effectiveness of dynamic regulation.

[0064] This invention integrates multispectral reflectance, morphological indicators (such as fruit volume growth rate), and environmental data to construct a cross-validation mechanism, thereby reducing the false positive rate. Experimental results show that the false positive rate for nutrient deficiency was reduced to 8.7% (compared to 23.4% in control group 1).

[0065] This invention employs a closed-loop mechanism of "water first, fertilizer second," with a second assessment of nutrient requirements 24 hours after irrigation. In the experiment, the experimental group saved 25% of water, and due to sufficient water, fertilizer utilization increased by 12%, resulting in a net profit increase of 2051 yuan per mu.

[0066] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of the intelligent mango fertilization method described in this invention; Detailed Implementation

[0068] The present invention will be further described in detail below with reference to embodiments, so that those skilled in the art can implement it based on the description.

[0069] It should be noted that, unless otherwise specified, the experimental methods described in the following embodiments are conventional methods, and the reagents and materials mentioned are commercially available. In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "setting" should be interpreted broadly. For example, they can refer to fixed connection or setting, detachable connection or setting, or integral connection or setting. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The terms "lateral," "longitudinal," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this invention and simplifying the description. They do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0070] Figure 1 A flowchart illustrating an intelligent mango fertilization method is shown, including the following steps:

[0071] S1) Arrange drip irrigation pipelines connecting to the water and fertilizer mixing device in the drip irrigation area;

[0072] S2) A multispectral imaging device periodically acquires canopy images of mango trees, the canopy images containing visible and near-infrared band information of leaves, flower buds, flowers and fruits;

[0073] S3) Based on image processing algorithms, extract visual and spectral features of the canopy image. The visual features include leaf color distribution, morphological thickness index, flower bud density, flower coverage, and fruit volume ratio. The spectral features include the ratio of near-infrared reflectance of the leaf area to that of the flower and fruit area.

[0074] S4) Based on the visual and spectral features, the flower bud differentiation period, flowering period and fruit enlargement period of the mango tree are determined by image comparison, and then the nitrogen deficiency state, phosphorus deficiency state, potassium deficiency state and water deficiency state at different periods are identified by the preset judgment model.

[0075] S5) Based on the determination of the nutrient or water deficiency state, control the water and fertilizer mixing device to dynamically adjust the concentration of the corresponding nutrients in the drip irrigation solution and the irrigation amount.

[0076] This invention achieves precise identification of nitrogen, phosphorus, potassium, and water through multi-dimensional feature fusion analysis. Combined with priority control and feedback replenishment mechanisms, it forms a closed-loop fertilization decision-making system, improving responsiveness at different stages in complex nutrient deficiency scenarios. This addresses the problem of traditional intelligent fertilization methods lacking dynamic control logic based on different stages, making them difficult to adapt to scenarios with complex nutrient requirements.

[0077] According to one embodiment of the present invention, the annular drip irrigation zone is 0.9-1.2 times the projected radius of the tree canopy, the drip irrigation pipe orifice diameter is 1-2 mm, and the drip irrigation flow rate is adjustable from 0.5-5 liters / minute. The drip irrigation pipe can be made of polyethylene, and the water-fertilizer mixing device can integrate a solenoid valve, a flow sensor, and a fertilizer concentration detection module. The drip irrigation pipe can be made of UV-resistant polyethylene hose with a pressure resistance rating of 0.3-0.6 MPa. The annular drip irrigation zone is set around the tree trunk, and the drip irrigation pipe is buried 10-15 cm below the ground surface and evenly distributed along the drip irrigation zone. The water-fertilizer mixing device controls the drip irrigation flow rate through a solenoid valve, the flow sensor monitors the irrigation volume in real time, and the fertilizer concentration detection module detects the nitrogen, phosphorus, and potassium concentrations in the mixture through a conductivity sensor.

[0078] The drone flies at an altitude of 3-5 meters, with a single shot covering a width of 2-3 meters. The illumination intensity of the supplementary lighting module can be adjusted from 800-1200 lux. The multispectral imaging equipment can employ a five-band (red, green, blue, near-infrared, thermal infrared) camera, and the supplementary lighting module can be configured with a ring-shaped LED array (color temperature 5500K). Alternatively, refer to the LED supplementary lighting module in the research paper "Application Research of Automatic Supplementary Lighting System for UAVs in Overhead Transmission Line Inspection" by Yang Junwu et al. The multispectral camera is mounted on the bottom gimbal of the drone, and the supplementary lighting module is fixed around the camera lens. The drone flies along a preset route to the center of the tree canopy and hovers. The light intensity sensor detects the ambient light intensity; if it is below 800 lux, the LED array is activated to supplement the light up to 1200 lux, ensuring uniform illumination during image acquisition.

[0079] The thresholds for flower bud density are ≥2.5 buds / cm², flower coverage is ≥30%, daily fruit volume growth rate is ≥3%, and spectral reflectance ratio is ≤1.0. Image processing algorithms can be implemented using the OpenCV library for semantic segmentation, and 3D point cloud reconstruction can be performed using a ToF (Time-of-Flight) depth camera. Point cloud data processing can rely on embedded GPUs (such as Jetson XavierNX). Images collected by the UAV are transmitted to a ground server, where leaf, flower bud, flower, and fruit regions are separated using semantic segmentation algorithms. Green saturation distribution data for leaf regions is extracted from the HSV color space, and the standard deviation of leaf vein point cloud density is calculated through stratified sampling (high-precision region point cloud density ≥100 points / cm²). The near-infrared reflectance ratio is calculated using the average reflectance of leaf and flower / fruit regions; if the ratio is ≤1.0, it is determined to be in the fruit enlargement stage based on the fruit volume growth rate. The soil moisture content threshold is 60%-70% of field capacity, the nutrient deficiency probability trigger threshold is ≥30%, and the irrigation delay duration is 24 hours. The soil moisture sensor can be a capacitive sensor, and the water-fertilizer mixing device can integrate a PID controller. The sensor probe can be made of ceramic material, which is corrosion-resistant. The soil moisture sensor is buried 20 cm below the drip irrigation area to monitor soil moisture content in real time. If both water and nutrients are lacking, the water-fertilizer mixing device will prioritize irrigation until the soil moisture content reaches 65%, and then re-acquire images after 24 hours to determine the nutrient status. If the probability of nutrient deficiency is ≥30%, the drip irrigation solution concentration will be adjusted linearly (e.g., the concentration will be increased to 1.5 times the upper limit when the probability is 70%).

[0080] This method combines ring drip irrigation with UAV multispectral monitoring to achieve precise identification of mango growth stages and dynamic regulation of nutrient requirements. Layered image analysis and threshold determination are employed to reduce the impact of natural light fluctuations on data and improve the reliability of fertilization decisions. The priority water replenishment control logic optimizes nutrient absorption efficiency, avoids resource waste, and adapts to the needs of scenarios with complex nutrient deficiencies.

[0081] If a ground mobile platform is used for mobile data collection, the height of the ground mobile platform should be 3-5 meters, and the equipment should be moved on the platform. Other execution steps are the same as described above.

[0082] According to another embodiment of the present invention, the UAV's flight altitude can be set to 3-5 meters, and the spacing between adjacent flight paths can be calculated to be 2-3 meters (based on a single-shot coverage width of 2 meters and a lateral overlap rate of 65%), with the lateral overlap rate set to 60%-70%. The LiDAR scanning accuracy can be configured to ±2 cm. The multispectral imaging device can employ a five-band (red, green, blue, near-infrared, thermal infrared) camera, the LiDAR can employ a 16-line scanning module, and the visual obstacle avoidance module can employ a binocular camera. The UAV fuselage can be made of carbon fiber composite material, and the gimbal bracket can be made of aluminum alloy. The multispectral camera is mounted on the bottom gimbal of the UAV, and the LiDAR is fixed to the front of the UAV. The UAV automatically generates parallel flight paths based on the electronic fence map, with the flight path spacing calculated based on the single-shot coverage width and overlap rate. During flight, the GPS module provides real-time positioning, and the visual obstacle avoidance module detects obstacles ahead and adjusts its course. Upon reaching the center point of the tree canopy, the UAV hovers for 1-2 seconds to complete image acquisition. The illumination intensity threshold of the supplementary lighting module can be set to 800 lux, the constant color temperature can be adjusted to 5500K, and the light intensity compensation range can be set to 800-1200 lux. The LED array power can be configured to 10-15 watts. The supplementary lighting module can include a ring-shaped LED light group (15-20 cm in diameter), a digital illuminance meter as the light intensity sensor, and a PWM dimming chip integrated into the control circuit. The LED lampshade can be made of polycarbonate, and the heat sink can be made of aluminum alloy. The ring-shaped LED array is mounted around the multispectral camera lens, with the light intensity sensor located outside the LED array. Before shooting, the light intensity sensor detects the ambient light intensity; if it is below 800 lux, the PWM chip adjusts the LED current to increase the light intensity to 1200 lux. During supplementary lighting, the light intensity sensor continuously monitors and feeds back data to ensure that the illumination uniformity error is less than 5%.

[0083] This solution achieves full-coverage image acquisition of the orchard through preset flight paths and dynamic route planning, reducing the need for manual intervention. The supplemental lighting module actively adjusts light intensity to suppress the impact of natural light fluctuations on spectral data, improving image consistency across different time periods. The collaborative operation of the drone and LiDAR ensures that the flight path accurately adapts to the orchard terrain, avoiding missed or repeated images.

[0084] According to another embodiment of the present invention, the input image resolution of the semantic segmentation algorithm can be set to 1920×1080 pixels, and the segmentation accuracy threshold can be set to a pixel classification accuracy ≥90%. The training dataset can contain 5000 labeled mango canopy images. The semantic segmentation algorithm can be implemented based on a deep learning framework (such as the U-Net network), and the running platform can be an embedded GPU computing module. The storage medium for the training dataset can be a solid-state drive, and the annotation tool can be based on open-source image annotation software. The canopy images collected by the UAV are transmitted to the ground processing unit, and the images are classified at the pixel level using the U-Net network. Leaf regions are initially extracted using a green channel threshold (e.g., H value range of 100-140 in HSV space), flower buds and flower regions are separated using morphological filtering, and fruit regions are identified based on color clustering (e.g., red channel proportion ≥60%). The segmentation results are output in the form of a binary mask for subsequent feature extraction. Green saturation distribution data can be quantized into a histogram of 0-255. The grid side length for calculating the standard deviation of leaf vein point cloud density can be set to 5-8 cm, and the point cloud density in high-precision areas can be configured to ≥100 points / cm². A ToF depth camera can be used for 3D point cloud reconstruction, and the point cloud registration algorithm can be based on an improved ICP (Iterative Closest Point) method. The point cloud data processing unit can be equipped with an embedded GPU (such as the Jetson series), and a copper heat sink can be used for the heat dissipation module. Point cloud data of the leaf region is acquired using a depth camera, and after meshing, high-precision reconstruction is performed on the leaf vein centerline region. Non-vein regions use linear interpolation to generate low-resolution point clouds (density ≤20 points / cm²). During registration, the leaf vein centerline is used as the reference, limiting the search radius of the ICP algorithm to ≤1 mm and the number of iterations to ≤10. The final output point cloud density standard deviation is used as a morphological thickness indicator.

[0085] Flower bud density can be statistically analyzed as ≥2.5 per square centimeter, flower coverage percentage can be calculated as ≥30%, and fruit volume percentage can be quantified as ≤15% (flowering period) or ≥15% (enlargement period). Flower bud and flower region statistics can be based on the OpenCV contour detection algorithm, and fruit volume calculation can be combined with stereo vision depth maps. A binocular camera can be used for stereo vision, and a ceramic checkerboard calibration plate can be used. The number density of segmented flower bud regions is statistically analyzed using connected component analysis, and the coverage of flower regions is calculated using the ratio of contour area to total canopy area. Fruit volume is reconstructed based on the fusion of depth maps and RGB images, and the percentage of canopy volume occupied is calculated using a voxel mesh method.

[0086] The near-infrared reflectance ratio threshold can be set to ≤1.0, the reflectance calculation window for the leaf region can be set to 10×10 pixels, and the reflectance of the flower and fruit regions is taken as the average value. The near-infrared band of the multispectral imaging device can be selected in the 900-950nm wavelength range, and the reflectance calibration plate can be made of polytetrafluoroethylene (PTFE). The calibration plate bracket can be made of aluminum alloy profile, and the reflectance data processing unit can integrate an FPGA acceleration module. After radiometric correction, the average reflectance of the leaf region (based on a semantic segmentation mask) and the flower and fruit region is extracted. When calculating the ratio, if the ratio of the average reflectance of the leaf region (e.g., 0.65) to the average reflectance of the flower and fruit region (e.g., 0.72) is 0.90 (≤1.0), it is determined to meet the spectral characteristics of the fruit enlargement stage.

[0087] This solution achieves precise quantitative analysis of leaf and flower / fruit regions through semantic segmentation and multi-dimensional feature extraction. Combining point cloud density with spectral reflectance ratios enhances the cross-validation capability for growth stage determination. Layered processing strategies (such as high-precision leaf vein reconstruction and non-vein interpolation) balance computational efficiency and accuracy requirements, adapting to real-time data processing scenarios in orchards. The fusion of statistical features of flowers and fruits with spectral data provides a reliable basis for determining compound nutrient deficiencies.

[0088] According to another embodiment of the present invention, the flower bud density threshold during the flower bud differentiation period can be set to ≥2.5 buds / cm², the flower coverage threshold during the flowering period can be set to ≥30%, the fruit volume growth rate threshold during the fruit enlargement period can be set to an average daily fruit volume growth rate of ≥3%, and the spectral reflectance ratio threshold can be set to ≤1.0. The growth stage standard library can be stored on a cloud server, the feature matching algorithm can be implemented based on Python scripts, and the threshold determination module can be integrated into an edge computing device. The server storage medium can be an SSD hard drive, and the heat dissipation casing of the edge computing device can be made of aluminum alloy. Feature data collected by the drone (such as flower bud density of 2.8 buds / cm² and flower coverage of 0%) is transmitted to the edge computing device and compared with the locally stored standard library. If the flower bud density is ≥2.5 buds / cm² and the flower coverage is 0%, it is determined to be the flower bud differentiation period; if the flower coverage is ≥30% and the fruit volume is ≤15%, it is determined to be the flowering period; if the average daily fruit volume growth rate is ≥3% and the spectral reflectance ratio is ≤1.0, it is determined to be the enlargement period. The determination result triggers the corresponding sub-model loading instruction.

[0089] The input feature weights for the random forest sub-model can be configured as follows: 40% for green saturation during flower bud differentiation, 30% for leaf thickness; 35% for flower coverage during flowering, 30% for near-infrared reflectance; and 50% for spectral reflectance ratio during fruit enlargement. The nutrient deficiency probability trigger threshold is ≥30%. The sub-model can be deployed on edge computing devices. The random forest algorithm can be implemented using the Scikit-learn library, and digital probes can be used for temperature and humidity sensors. The sensor probe housing can be made of ABS plastic, and shielded twisted-pair cables can be used for data transmission. After determining that the flowering period has begun, the flowering period sub-model is loaded, and the current near-infrared reflectance (0.68), ambient temperature (28℃), and historical phosphorus application data (e.g., 0.5 kg / plant of phosphorus applied in the most recent week) are input. The sub-model calculates a phosphorus deficiency probability of 45% using the random forest algorithm (120 decision trees, maximum depth 10 layers). If the probability value is ≥30%, a water and fertilizer regulation command is triggered.

[0090] This solution achieves a precise correlation between growth stages and nutrient status through quantified thresholds and phased sub-model calls. Local storage of the standard library reduces cloud dependency and improves real-time response speed. The random forest algorithm integrates multi-source data, enhancing robustness in complex deficiency scenarios. Phased weight configuration avoids misjudgment based on a single feature and adapts to the differences in physiological characteristics at different stages.

[0091] According to another embodiment of the present invention, the soil moisture content threshold can be set to 60%-70% of field capacity, the irrigation delay time can be set to 24 hours, and the irrigation amount can be adjusted to 1.2-1.5 times the standard amount. The soil moisture sensor can be a capacitive sensor, the control module of the water-fertilizer mixing device can integrate a PID controller, and the irrigation solenoid valve can be a normally closed two-position two-way valve. The sensor probe can be made of ceramic material for corrosion resistance; the solenoid valve body can be made of brass, and the sealing ring can be made of fluororubber. The soil moisture sensor is buried 20 cm below the drip irrigation area to monitor soil moisture content in real time. If both water deficiency (e.g., moisture content ≤ 55%) and nutrient deficiency (e.g., nitrogen deficiency probability ≥ 30%) are detected simultaneously, the PID controller prioritizes opening the irrigation solenoid valve to increase the soil moisture content to 65%. After irrigation, the system delays for 24 hours and controls the drone to re-acquire canopy images for secondary judgment.

[0092] The nutrient deficiency probability trigger threshold can be set to ≥30%, the concentration adjustment upper limit can be set to 1.5 times the standard value, the drip irrigation volume adjustment range can be set to 1.2-1.5 times the standard volume, and the feedback data acquisition interval can be configured as follows: when the supplemented element is nitrogen, the delay time is 48-72 hours; when the supplemented element is phosphorus or potassium, the delay time is 5-7 days; when nitrogen, phosphorus, and potassium are all deficient, the longest delay time corresponding to phosphorus or potassium, 5-7 days, is used as the benchmark. The probability value calculation module can be integrated into the edge computing device, the feedback data storage can use an SD card, and the drip irrigation fluid concentration detection can be based on a conductivity sensor. The conductivity sensor electrode can be made of platinum, and the data storage medium can be an industrial-grade TF card. After secondary judgment, if the nitrogen deficiency probability value is 70%, the conductivity sensor detects the current drip irrigation fluid concentration (e.g., 0.8 g / L), the PID controller increases the nitrogen concentration to the upper limit of 1.2 g / L, and at the same time increases the drip irrigation volume to 1.3 times the standard volume. After adjustment, the system will reacquire canopy images and soil data after a 48-hour delay until the nitrogen deficiency probability drops below 10%. If the probability is 40%, the concentration will be adjusted to 0.96 g / L linearly, and the drip irrigation rate will be adjusted to 1.1 times the standard rate.

[0093] This solution optimizes the order of water and nutrient replenishment through priority control and feedback mechanisms, avoiding low fertilizer utilization under drought conditions. Secondary assessment reduces the risk of misoperation, dynamically adjusting to different levels of deficiency and improving fertilization accuracy. Delayed data collection and continuous feedback ensure traceability of the control effects, adapting to the dynamic needs of the orchard environment.

[0094] According to another embodiment of the present invention, the single nutrient deficiency probability threshold is set to 70% and 30%, the upper limit of concentration adjustment is 1.5 times the standard value (e.g., nitrogen base concentration 1.0 g / L, upper limit 1.5 g / L), and the irrigation amount adjustment range is 1.2-1.5 times the standard amount (e.g., standard amount 5 L / plant, adjusted to 6-7.5 L / plant). The linear proportional formula is: Adjustment value = Base value + (Current probability value - 30%) × (Upper limit value - Base value) / 40%. The probability calculation module can be integrated into an edge computing device (such as a Raspberry Pi 4B), the conductivity sensor can be a four-electrode probe (range 0-5 mS / cm), and the drip irrigation solution concentration adjustment can be based on a stepper motor driven peristaltic pump (flow accuracy ±2%). The peristaltic pump hose can be made of silicone (corrosion resistant), the sensor electrodes can be platinum coated, and the data transmission line can be shielded twisted pair cable (waterproof rating IP67). A conductivity sensor is installed at the drip irrigation line outlet to monitor nitrogen, phosphorus, and potassium concentrations in real time (e.g., nitrogen concentration 0.8 g / L). If the probability of nitrogen deficiency is 75%, the edge computing device sends a command to the peristaltic pump to increase the nitrogen concentration to 1.5 g / L and the irrigation volume to 7.5 L / plant (the standard volume is 5 L × 1.5 times). When the probability is 50%, the concentration is adjusted to 1.25 g / L according to a linear formula, and the irrigation volume is adjusted to 6.25 L / plant.

[0095] The delay time after nitrogen supplementation is 48-72 hours (preferably 60 hours), and the delay time for phosphorus / potassium supplementation is 5-7 days (preferably 6 days). The threshold for consecutive adjustment failures is 3 times, and the threshold for the deficiency probability to reach the target is <10%. Feedback data acquisition can rely on a UAV multispectral imaging system (reacquisition cycle is configurable). Data storage can use an industrial-grade SD card (capacity ≥64GB), and the timer module can be based on a hardware RTC chip (error ±1 minute / month). The SD card protective shell can be made of ABS plastic (dustproof and waterproof), and the timer circuit board can be made of FR-4 substrate (high temperature resistant). After adjustment, the system starts a countdown (e.g., nitrogen delayed by 60 hours). After the countdown ends, the UAV is controlled to reacquire canopy images. If the second determination shows a nitrogen deficiency probability of 8%, the adjustment is terminated; if the probability is 12%, the concentration is adjusted again to 1.0 g / L and a new 48-hour countdown is started. If the probability is still ≥10% after 3 consecutive adjustments (e.g., 12%, 15%, 13%), an audible and visual alarm is triggered (e.g., buzzer + LED flashing), and automatic adjustment is paused, requiring manual intervention for inspection.

[0096] This solution uses probability grading and linear proportional adjustment to achieve precise matching between nutrient supplementation and deficiency levels, avoiding salinization or resource waste caused by over-fertilization. A dynamic delay mechanism adapts to the physiological absorption cycles of different nutrients (e.g., nitrogen is fast-acting, phosphorus and potassium are slow-acting), ensuring the validity of feedback data. Continuous failure alarms prevent the system from falling into an ineffective loop, improving the safety and reliability of orchard management.

[0097] According to another embodiment of the present invention, the side length of the sub-region grid can be set to 5-8 cm, the point cloud density of the vein centerline grid can be configured to ≥100 points / cm², and the point cloud density of the non-vein region can be limited to ≤20 points / cm². Three-dimensional point cloud layer sampling can be implemented based on a ToF depth camera, and the linear interpolation algorithm can be integrated into an embedded GPU computing module. The protective shell of the depth camera can be made of aluminum alloy, and the storage medium for the interpolation algorithm can be LPDDR4 memory. The ToF depth camera is mounted on the UAV gimbal to collect raw point cloud data of the leaf region. The ground processing unit divides the leaf into square grids with a side length of 6 cm, identifies the grids where the vein centerline is located (e.g., grid numbers G01-G05), and performs high-precision reconstruction on them (point cloud density 120 points / cm²). Non-vein grids (e.g., grid numbers G06-G20) are generated into low-resolution point clouds (density 15 points / cm²) using linear interpolation to reduce the GPU computational load.

[0098] The number of iterations of the ICP algorithm can be limited to ≤10, the search radius can be set to ≤1 mm, and the registration error threshold can be set to ±0.05 mm. The improved ICP algorithm can be implemented based on the CUDA acceleration library, parallel computing can rely on embedded GPUs (such as Jetson series), and NVMe solid-state drives can be used for storing the registration results. Copper fins can be used for the GPU heatsink, and SRAM chips can be configured for the registration data cache. The high-precision leaf vein point cloud and the interpolated non-leaf vein point cloud are transmitted to the GPU, and the improved ICP registration is performed with the leaf vein centerline as the reference. The algorithm limits the search radius to 0.8 mm, and reaches the error threshold of 0.04 mm after 8 iterations. The density standard deviation of the leaf vein region (e.g., 0.12 points / cm²) is calculated from the registered point cloud data and output as a morphological thickness index to the judgment module.

[0099] This solution employs a hierarchical sampling and differentiated reconstruction strategy to balance point cloud data accuracy with computational resource consumption, thereby improving the efficiency of large-scale leaf processing. An improved ICP algorithm, combined with hardware acceleration, ensures that registration accuracy meets the requirements for morphological index calculation. Separation of vein and non-vein regions reduces redundant data interference and enhances the reliability of thickness indicators.

[0100] According to another embodiment of the present invention, the training dataset can contain 2000 sets of historical data, the test set accuracy threshold can be set to ≥85%, and the cross-validation folds can be set to 5 folds. The number of decision trees in the sub-model can be configured to 100-150, and the maximum depth can be set to 10-12 layers. The pre-trained model library can be stored in the SSD of the edge computing device, the data annotation can be based on open-source annotation tools, and the model training can rely on the Python Scikit-learn library. The SSD storage medium can be 3D NAND flash memory, and the heat sink of the training server can be made of aluminum alloy. The flower bud differentiation period sub-model is stored in the model partition M1 of the edge device. When the flower bud differentiation period is determined, the system loads the model from M1. The input data includes the current green saturation distribution histogram (e.g., the peak value is in the 120-140 range), leaf morphology and thickness indicators (e.g., point cloud density standard deviation 0.15), and near-infrared reflectance 0.72.

[0101] The input feature weights can be set as follows: 40% for green saturation and 30% for leaf thickness in the flower bud differentiation sub-model; 35% for flower coverage in the flowering sub-model; and 50% for spectral reflectance ratio in the swelling sub-model. The nutrient deficiency probability trigger threshold can be set to ≥30%. The random forest algorithm can run on embedded GPUs (such as Jetson series), and the temperature and humidity sensor can be a digital probe with an I2C interface. Historical fertilization data can be stored in an SQLite database. The sensor housing can be made of ABS plastic, and the database storage medium can be an eMMC chip. After loading the flowering sub-model, the current flower coverage is 32%, near-infrared reflectance is 0.68, ambient temperature is 28℃, and historical phosphorus application rate is 0.5 kg / plant. The random forest algorithm (120 decision trees, maximum depth 10 layers) calculates a phosphorus deficiency probability of 45%. If the probability value is ≥30%, an adjustment command is sent to the water and fertilizer mixing device.

[0102] This solution improves the accuracy of nutrient deficiency assessment at different growth stages by using phased sub-model loading and differentiated feature weight configuration. Local deployment of the pre-trained model library reduces cloud dependency and enhances real-time response capabilities. The random forest algorithm integrates multi-source data, reducing interference from single environmental variables and adapting to dynamic orchard scenarios.

[0103] According to another embodiment of the present invention, the historical dataset can contain at least 3000 sets of data, with a labeling accuracy rate of ≥95%. A single set of data may include canopy images (resolution 1920×1080 pixels), near-infrared reflectance (range 0.3-0.8), and ambient temperature and humidity (temperature 15-35℃, humidity 30-80%). Canopy images can be acquired using a multispectral camera, and ambient temperature and humidity data can be obtained using a digital temperature and humidity sensor. Historical fertilization records can be stored in an electronic ledger system. The data storage medium can be an enterprise-grade SSD, and the sensor housing can be made of ABS plastic. The multispectral camera is mounted on a drone gimbal, and the temperature and humidity sensor is fixed to an orchard weather station bracket. Data collection is conducted weekly for two years. During labeling, manual annotation of each set of data is performed based on leaf yellowing (nitrogen deficiency), flower bud abscission rate (phosphorus deficiency), and fruit wrinkling rate (potassium deficiency). The annotation results are stored in database table T1.

[0104] The flower bud differentiation dataset can contain 1000 data sets, the flowering period dataset 800 sets, and the fruit enlargement period dataset 1200 sets. The random forest algorithm parameters can be configured with 120 decision trees, a maximum depth of 12 layers, and a minimum number of leaf node samples of 5. Model training can be based on the Python Scikit-learn library. The training server can be equipped with a multi-core CPU (e.g., 16 cores) and 64GB of RAM. Dataset partitioning can be implemented using the Pandas library. Copper heat pipes can be used for server cooling, and DDR4 RAM is recommended. The flower bud differentiation dataset is loaded into the training server's memory. Input features include a histogram of green saturation distribution (HSV space H value range 100-140), leaf morphology and thickness indices (point cloud density standard deviation 0.1-0.3), and near-infrared reflectance 0.5-0.7. The random forest model outputs nitrogen deficiency probability values, and the completed training is saved as file M1.

[0105] The cross-validation fold count can be set to 5 folds, the test set accuracy threshold can be set to ≥85%, and the parameter optimization iteration count can be limited to 100. Cross-validation can be implemented based on the K-fold algorithm, parameter tuning can rely on a grid search tool, and validation results can be stored in a local log file. The log file storage medium can be an NVMe solid-state drive, and the grid search computing resources can be allocated to a 16-thread CPU. After the flowering stage sub-model is trained, 5-fold cross-validation is used to divide the dataset into a training set (640 sets) and a validation set (160 sets). The grid search traverses the number of decision trees (100-150) and the maximum depth (10-12 layers), selecting the parameter combination with the highest validation accuracy (e.g., 87.3%), and saving it as the final model M2.

[0106] This solution employs phased data collection and annotation to ensure model training aligns with actual growth needs. Cross-validation optimizes parameter combinations, enhancing model generalization ability. The fusion of historical data and multi-source features strengthens the robustness in identifying different nutrient deficiency scenarios. Independent training of phased sub-models avoids interference between feature weights at different stages, adapting to the dynamic management requirements of orchards.

[0107] According to another embodiment of the present invention, in the flower bud differentiation stage sub-model, the weight of green saturation distribution data can be set to 40%, leaf morphology and thickness index to 30%, near-infrared reflectance to 20%, and environmental temperature and humidity to 10%. In the flowering stage sub-model, the weight of flower coverage percentage is 35%, near-infrared reflectance to 30%, green saturation distribution data to 25%, and historical fertilization amount to 10%. In the fruit enlargement stage sub-model, the weight of spectral reflectance ratio is 50%, fruit volume growth rate to 30%, environmental temperature to 15%, and historical fertilization amount to 5%. The weight allocation parameters can be stored in the model configuration file. The model training platform can implement weight loading based on Python scripts, and feature data normalization can rely on the NumPy library. The configuration file storage medium can be a YAML format text file, and the data normalization processing unit can be integrated into the embedded GPU memory. When loading the flower bud differentiation stage sub-model, the green saturation weight of 40% and the leaf thickness weight of 30% are read from the configuration file W1. After the input data is normalized, it is summed by weight (e.g., green saturation score × 0.4 + leaf thickness score × 0.3) and used as the input vector for the random forest algorithm.

[0108] The number of decision trees can be set to 100-150, the maximum depth to 10-12 layers, and the minimum number of leaf node samples to 5. The feature splitting criterion during training can be Gini impurity, and the proportion of random feature subsets can be set to √n (where n is the total number of input features). Random forest training can be implemented using the Scikit-learn library, parameter configuration can be stored in a JSON file, and model inference can rely on an embedded GPU (such as the Jetson series). The JSON configuration file can be stored on an eMMC chip, and a copper heatsink can be used for GPU cooling. During the flowering stage sub-model training, 120 decision trees with a maximum depth of 10 layers are loaded from configuration file P1. During training, each decision tree randomly selects √7 ≈ 2 features (out of a total of 7 input features) for node splitting, with the minimum number of leaf node samples limited to 5. After training, the model is saved as file M2 and deployed to the model partition of an edge computing device. Here, √ represents the square root symbol.

[0109] This scheme strengthens the discriminative contribution of key features at each growth stage and reduces the interference of secondary features through phased differentiated weight allocation. Standardization of random forest parameters improves model training efficiency and ensures algorithm consistency across different orchard environments. Coordinated configuration of weights and parameters adapts to changes in mango physiological characteristics, enhancing the robustness of judgment in scenarios of compound nutrient deficiency.

[0110] An embodiment 1 of the intelligent mango fertilization method according to the present invention includes the following steps:

[0111] S1) A ring-shaped water and fertilizer drip irrigation zone is set up with the tree trunk as the center, and the drip irrigation zone is arranged with drip irrigation pipelines connected to the water and fertilizer mixing device;

[0112] S2) Periodically collect canopy images of mango trees using a multispectral imaging device mounted on a drone. The canopy images include visible light and near-infrared band information of leaves, flower buds, flowers and fruits. During image acquisition, the drone's built-in supplementary lighting module maintains preset lighting conditions.

[0113] The drone's preset flight path is dynamically planned based on the electronic fence map of the mango orchard and the canopy projection range, and can be executed using the following existing steps:

[0114] S2.1) The terrain of the mango orchard is scanned by a lidar mounted on a drone to generate an electronic fence map containing the location of each mango tree and the canopy boundary;

[0115] S2.2) Based on the electronic fence map, the flight path is divided into several parallel routes. The distance between adjacent routes is determined by calculating the coverage width of a single shot of the multispectral imaging device and the lateral overlap rate, wherein the lateral overlap rate is 60%-70%.

[0116] S2.3) When the UAV flies along the route, it corrects its course in real time through the onboard GPS and visual obstacle avoidance module, and hovers to take pictures when it reaches the center point of the canopy of each mango tree;

[0117] S2.4) The supplementary lighting module includes an LED array and a light intensity sensor. Before shooting, the light intensity sensor detects the ambient light intensity. If it is lower than a preset threshold, the LED array is activated to supplement light with constant color temperature and light intensity to ensure uniform illumination on the canopy surface.

[0118] S3) Extract visual and spectral features of the canopy image based on image processing algorithms. The visual features include leaf color distribution, morphological thickness indicators, flower bud density, flower coverage, and fruit volume ratio. The spectral features include the near-infrared reflectance ratio of the leaf area to the flower and fruit area. Specific extraction methods include:

[0119] S3.1) Perform semantic segmentation on the canopy image to separate the leaf, flower bud, flower, and fruit regions;

[0120] S3.2) In the leaf region, extract the green saturation distribution data in the HSV color space, and calculate the standard deviation of the point cloud density in the leaf vein region through three-dimensional point cloud reconstruction technology, which serves as an indicator of leaf morphology and thickness.

[0121] S3.3) Calculate the number density of flower buds per unit area in the flower bud region, calculate the percentage of the coverage area in the flower region relative to the total canopy area, and calculate the percentage of the volume in the fruit region relative to the total canopy volume.

[0122] S3.4) For near-infrared images, calculate the ratio of the average reflectance of the leaf region to the average reflectance of the flower and fruit region, and use it as the spectral reflectance ratio (spectral characteristic data).

[0123] S4) Based on the visual and spectral features, the mango tree's flower bud differentiation period, flowering period, and fruit enlargement period are determined through image comparison. Then, a preset judgment model is used to identify the nitrogen deficiency, phosphorus deficiency, potassium deficiency, and water deficiency states at different stages. The specific process includes:

[0124] S4.1) Based on the leaf morphology and thickness index extracted in step S3.2, the flower bud number density, flower coverage percentage, and fruit volume percentage extracted in step S3.3, and the spectral reflectance ratio extracted in step S3.4, the fruit volume growth rate is calculated by accumulating the fruit volume percentage extracted each time, and matched with the preset growth stage standard library.

[0125] S4.2) The growth stage standard library includes thresholds for flower bud density, flower coverage, fruit volume percentage, fruit volume growth rate, and spectral reflectance ratio for each stage. If the current flower bud density exceeds the flower bud density threshold for the differentiation stage and the flower coverage percentage is zero, it is determined to be the flower bud differentiation stage. If the current flower coverage percentage exceeds the flower coverage threshold for the flowering stage and the fruit volume percentage is lower than the threshold for the expansion stage, it is determined to be the flowering stage. If the fruit volume percentage growth rate continuously exceeds the fruit volume growth rate threshold for the expansion stage and the spectral reflectance ratio is lower than the threshold for the spectral reflectance ratio, it is determined to be the fruit expansion stage.

[0126] The specific thresholds are as follows: Flower bud differentiation stage: Flower bud density ≥ 2.5 buds / cm² and flower coverage percentage = 0%;

[0127] Flowering period: Flower coverage percentage ≥ 30% and fruit volume percentage ≤ 15%;

[0128] Fruit enlargement period: Daily average fruit volume growth rate ≥ 3%

[0129] And the ratio of spectral reflectance is ≤1.0.

[0130] S4.3) After determining the growth stage, a nutrient determination sub-model specific to that stage is invoked. This sub-model integrates near-infrared reflectance, ambient temperature and humidity, and historical fertilization data using a random forest algorithm to output the probability value of nitrogen, phosphorus, potassium, or water deficiency. The specific implementation of invoking the nutrient determination sub-model specific to that stage includes the following steps:

[0131] S4.31) Based on the growth stage determined in step S4.2, load the corresponding sub-model from the pre-trained model library; wherein, the pre-trained model library is constructed through the following steps:

[0132] S4.311) Collect historical data on mango cultivation, including canopy images, near-infrared reflectance, environmental temperature and humidity, historical fertilization records, green saturation distribution data, and leaf morphology and thickness indicators at different growth stages.

[0133] S4.312) The historical data is annotated, and the annotation content includes the actual state of nitrogen, phosphorus, potassium and water deficiency in each period;

[0134] S4.313) Divide the dataset according to the growth stage and train the flower bud differentiation period sub-model, flowering period sub-model and fruit enlargement period sub-model respectively. Each sub-model adopts the random forest algorithm. The input features are near-infrared reflectance, environmental temperature and humidity, historical fertilizer application, green saturation distribution data and leaf morphology and thickness index. The output is the nutrient deficiency probability value.

[0135] S4.314) Optimize the sub-model parameters through cross-validation until the test set accuracy reaches a preset value (e.g., preset value ≥ 85%).

[0136] The specific training and validation methods for the sub-models include:

[0137] Flower bud differentiation period sub-model: In the input features, the weight of green saturation distribution data accounts for 40%, leaf morphology and thickness index accounts for 30%, near-infrared reflectance accounts for 20%, and environmental temperature and humidity accounts for 10%;

[0138] Flowering period sub-model: In the input features, the percentage of flower coverage accounts for 35%, near-infrared reflectance accounts for 30%, green saturation distribution data accounts for 25%, and historical fertilization amount accounts for 10%.

[0139] Fruit enlargement period sub-model: Among the input features, the ratio of spectral reflectance accounts for 50%, the fruit volume growth rate accounts for 30%, the ambient temperature accounts for 15%, and the historical fertilization amount accounts for 5%;

[0140] The random forest algorithm parameters for each sub-model are: 100-150 decision trees, 10-12 layers at maximum depth, and 5 samples per leaf node.

[0141] S4.32) Input the current near-infrared reflectance, ambient temperature and humidity, historical fertilization data, green saturation distribution data, and leaf morphology and thickness indicators into the sub-model; output the probability value of nitrogen, phosphorus, potassium, or water deficiency through the random forest algorithm;

[0142] S4.33) When any probability value is ≥30%, the control logic of step S5 is triggered.

[0143] S5) Based on the determination of the nutrient deficiency state, control the water-fertilizer mixing device to dynamically adjust the concentration of the corresponding nutrients in the drip irrigation solution and the irrigation amount; specifically including:

[0144] S5.21 When the probability of a single nutrient deficiency is ≥70%, the water and fertilizer mixing device shall be controlled to increase the corresponding nutrient concentration to the preset upper limit and the drip irrigation volume shall be increased to 1.2-1.5 times the standard irrigation volume;

[0145] S5.22 When the probability of a single nutrient deficiency is 30%-70%, adjust the nutrient concentration and irrigation amount linearly.

[0146] S5.23 When the probability of nutrient deficiency is less than 30%, drip irrigation should be performed while maintaining the baseline concentration and irrigation volume;

[0147] S5.24 The specific steps for prioritizing the activation of drip irrigation pipelines for water irrigation are as follows: if both water and nutrient deficiency are detected simultaneously, irrigation is prioritized until the soil moisture content reaches the standard, and then the canopy image is re-acquired for a second determination after a first set time of 24 hours.

[0148] S5.25 If nutrient deficiency still exists after the second determination, the drip irrigation system will be triggered to supplement the corresponding elements, and the delay time for feedback data collection will be dynamically adjusted according to the following rules:

[0149] a) When the supplementary element is nitrogen, the delay time is 48-72 hours;

[0150] b) When the supplemented element is phosphorus or potassium, the delay period is 5-7 days;

[0151] c) When nitrogen, phosphorus and potassium are simultaneously deficient, the longest delay time corresponding to phosphorus or potassium, 5-7 days, shall be used as the benchmark.

[0152] S5.26 After each adjustment, feedback data is collected according to the dynamic adjustment delay time until the lack probability value is <10%; if the lack probability value is still ≥10% after 3 consecutive adjustments, an abnormal alarm is triggered and automatic control is terminated.

[0153] The supplemental lighting module employs an LED array (such as the Osram OSLON Square series), with an adjustable light intensity range of 800-1200 lux. Dynamic adjustment via PWM dimming technology (duty cycle 10%-100%) ensures uniform light intensity on the canopy surface under varying ambient lighting conditions (standard deviation ≤5%). It is equipped with a diffuser lens (such as a polycarbonate diffuser) to eliminate light spots, with a coverage area adapted to the drone's flight altitude (3-5 meters) to ensure uniform overall illumination of the canopy. An integrated digital illuminance sensor (such as the ams AS7341L multispectral light sensor) monitors ambient light intensity (visible and near-infrared bands) in real time and feeds the data back to the drone's flight control system. Basic mode: When the ambient light intensity is below 800 lux, the supplemental lighting module is activated, providing supplemental lighting at a preset intensity (e.g., 1000 lux). Staged mode: The spectral ratio is dynamically adjusted according to the mango's growth stage: Flower bud differentiation stage: Increase the proportion of blue light (blue:red = 3:1) to inhibit excessive vegetative growth and promote flower bud formation; Flowering stage: Balance red and blue light (1:1) to maintain photosynthetic efficiency; Fruit enlargement stage: Increase the proportion of red light (red:blue = 3:1) to promote fruit sugar accumulation. If the multispectral imaging equipment is collecting data, the supplemental lighting module prioritizes maintaining a constant light intensity; during non-collection periods, it automatically reduces power to save energy. The circuit design uses a totem-pole switch control circuit (referencing "Application Research of Automatic Supplemental Lighting System in Overhead Transmission Line Inspection by UAVs"), controlling the LED supplemental lighting's on / off state via high and low level signals, with a response delay ≤10ms.

[0154] In the implementation example of Jin Huang mango cultivation, a circular drip irrigation zone with a canopy projection radius of 0.9-1.2 times was set up around the tree trunk in one of the Jin Huang mango planting areas of a standardized mango planting base. The drip irrigation pipes had a diameter of 1.2 mm, a burial depth of 10 cm, and covered all 330 fruit trees in the orchard. The water and fertilizer mixing device used was the Netafim NutriFlex 3.0, with a basic irrigation volume of 6 liters per tree and basic nitrogen, phosphorus, and potassium concentrations of 1.2 g / L, 0.9 g / L, and 1.5 g / L, respectively. The drip irrigation pipes were connected to solenoid valves, and the flow rate and concentration were dynamically adjusted based on the judgment results. The drone used was a DJI Motrice 300 RTK, equipped with a MicaSense RedEdge-MX multispectral camera (wavelengths: blue 475nm, green 560nm, red 668nm, red edge 717nm, near-infrared 840nm) and a Velodyne VLP-16Lite lidar. The obstacle avoidance module APAS 5.0 plans the flight path in real time, with a lateral overlap rate of 65% and a flight altitude of 3.5 meters. On cloudy days (ambient light of 600 lux), the LED array is activated to supplement the light to 1000 lux (color temperature of 5500K) to ensure clear canopy images during the flower bud differentiation period.

[0155] Flower bud differentiation period (March): Images collected by drone showed a flower bud density of 3.2 buds / cm² (threshold ≥ 2.5 buds / cm²) and a flower coverage of 0%, which was determined to be the flower bud differentiation period.

[0156] The near-infrared reflectance of the leaves is 0.65, and the reflectance of the flower and fruit area is 0.72, with a ratio of 0.90 (threshold ≤ 1.0). Combined with the daily average growth rate of fruit volume of 4% (threshold 3%), it is determined that the fruit has entered the expansion stage.

[0157] Nitrogen deficiency during flower bud differentiation: The sub-model outputs a nitrogen deficiency probability of 55%, and the nitrogen concentration is increased from 1.2 g / L to 1.6 g / L (preset upper limit) in a linear proportion, and the irrigation amount is increased to 7.2 liters / plant (1.2 times the standard amount).

[0158] Potassium deficiency during the fruit enlargement stage: The probability of potassium deficiency was determined to be 70% in the second assessment. The potassium concentration was directly adjusted to 2.25 g / L (upper limit 1.5 times), and the irrigation amount was increased to 9 liters / plant (1.5 times).

[0159] Field comparison trial

[0160] Location: A standardized mango planting base

[0161] Tree age: 5-year-old Golden Mango, grafted seedling, tree height 3.2-3.8 meters, crown width 2.5-3.0 meters;

[0162] Planting density: 4 meters between plants × 5 meters between rows, 33 plants per mu, experimental area 30 mu (990 plants);

[0163] Soil conditions: Red soil, pH 5.8-6.2, organic matter content 1.8%, basic fertility: available nitrogen 85mg / kg, available phosphorus 12mg / kg, available potassium 105mg / kg;

[0164] Experimental group: 330 plants, using the intelligent fertilization system of this invention (Example 1: Dynamic growth stage determination + phased regulation).

[0165] Control group 1: 330 plants, fertilized using traditional manual experience (fixed cycle: nitrogen fertilizer before flowering, phosphorus and potassium fertilizer after flowering, and manual inspection every week), specifically: once a week, fertilization time is 6:00-8:00 am (avoiding high temperature period);

[0166] Flower bud differentiation period (March): 0.5 kg / plant of urea (46% nitrogen content) and 0.3 kg / plant of superphosphate (12% phosphorus content);

[0167] Flowering period (April): Potassium dihydrogen phosphate (52% phosphorus, 34% potassium) application rate 0.4 kg / plant;

[0168] During the fruit enlargement period (May-June): Apply 0.6 kg of potassium sulfate (50% potassium content) per plant;

[0169] Adjustment criteria: manual observation of leaf color (degree of yellowing) and flower bud abscission rate, combined with soil electrical conductivity testing using a soil rapid tester (model: Hanna HI9835), with an error tolerance of ±10%.

[0170] Control group 2: 330 plants, static model using traditional intelligent fertilization (equipment: Netafim NutriFlex™ Basic fertigation system. Nitrogen, phosphorus, and potassium were added in a fixed ratio (N:P2O5:K2O=2:1:3), with preset concentrations of 1.2 g / L nitrogen, 0.6 g / L phosphorus, and 1.8 g / L potassium. Irrigation volume was fixed at 6 L / plant, controlled by a mechanical flow valve, ignoring changes in growth stages. It relied on a soil moisture sensor (buried 20 cm deep), triggering irrigation when soil moisture content fell below 60% of field capacity. It lacked canopy image acquisition or multispectral analysis capabilities; fertilization decisions were entirely based on preset parameters. It lacked stage-specific adjustments: using the same fertilization ratio during flower bud differentiation, flowering, and fruit enlargement stages, failing to adapt to varying nutrient requirements at different stages. It also lacked a feedback mechanism: soil or plant conditions were not re-monitored after fertilization, lacking dynamic optimization capabilities.

[0171] The criteria for determining the growth stage are as follows: three main branches are randomly selected from each plant, and two agronomists independently record the flower bud density, flower coverage, and fruit volume growth rate. If the results are inconsistent, they are reviewed by a third-party expert.

[0172] False positive rate calculation: If the difference between the system's judgment result and the manual review result exceeds the threshold (e.g., flower bud density error ≥ 0.5 buds / cm²), it is recorded as a false positive.

[0173] Criteria for determining nutrient deficiency: Leaf samples (the third mature leaf) are collected every half month and sent to the laboratory for testing (Kjeldahl method, molybdenum antimony colorimetric method, flame photometry).

[0174] False positive rate calculation: If the system outputs a probability value ≥ 30% but the laboratory test shows no deficiency, or if the probability value is < 30% but the laboratory test shows no deficiency, it is considered a false positive.

[0175] Yield per plant: Weigh all fruits on the plant at harvest time, and remove cracked or diseased fruits.

[0176] Fruit quality: 100 fruits were randomly selected to measure the weight of a single fruit, soluble solids (sugar content), and flesh thickness.

[0177] Direct costs: fertilizer, water resources, labor, and equipment depreciation (drones and water and fertilizer systems are amortized over 5 years).

[0178] Indirect costs: losses from fertilizer application due to misjudgment, and losses from reduced yield.

[0179] Statistical experimental results (data from 2023):

[0180] The misclassification rates for each growth stage are as follows:

[0181] Experimental group: 6.2% (Possible reasons for the main error: misjudgment due to overlap between the early stage of fruit enlargement and the late stage of flowering).

[0182] Control group 1: 17.8% (Possible cause of the main error: missed flowers and fruits inside the canopy during manual observation).

[0183] Control group 2: 12.5% ​​(Possible cause of major error: Static model failed to identify abrupt changes in fruit volume growth rate).

[0184] The misjudgment rate of nutrient deficiency in each group is as follows:

[0185] Experimental group: 8.7% (concentrated on spectral data distortion on cloudy and rainy days);

[0186] Control group 1: 23.4% (misjudgment due to lag in soil testing);

[0187] Control group 2: 15.6% (fixed threshold cannot adapt to temperature fluctuations).

[0188] The average yield per plant in each group is as follows:

[0189] Experimental group: 36.5 kg (standard deviation ± 2.1);

[0190] Control group 1: 29.8 kg (standard deviation ± 3.5);

[0191] Control group 2: 32.4 kg (standard deviation ± 2.8).

[0192] The fruit quality statistics for each group are as follows:

[0193] Experimental group: Single fruit weight 495 grams, sugar content 14.8%, flesh thickness 18.2 mm;

[0194] Control group 1: Single fruit weight 405 grams, sugar content 13.3%, flesh thickness 15.6 mm;

[0195] Control group 2: Single fruit weight 435g, sugar content 14.0%, flesh thickness 16.9mm.

[0196] The cost-benefit comparison statistics for each group are as follows:

[0197] Experimental group: Nitrogen fertilizer reduced by 19%, phosphorus fertilizer reduced by 16%, potassium fertilizer reduced by 13%, water saved by 25%;

[0198] Control group 1: Fertilizer was applied in excess at a fixed cycle, with nitrogen, phosphorus and potassium dosages being 1.4 times that of the experimental group;

[0199] Control group 2: Due to misjudgment, fertilizer was frequently applied, and the amount of nitrogen, phosphorus and potassium was 1.2 times that of the experimental group.

[0200] The overall net profit statistics are as follows:

[0201] Golden Mango Market Price: Wholesale price at the place of origin: 8 yuan / kg (general price, with premium for sugar content grade).

[0202] Fertilizer costs: Nitrogen fertilizer 4.5 yuan / kg, Phosphate fertilizer 5.2 yuan / kg, Potassium fertilizer 5.8 yuan / kg.

[0203] Water resource costs: Irrigation water fee 0.8 yuan / ton.

[0204] Labor costs: Orchard workers are paid 150 yuan per day.

[0205] Equipment depreciation: Drones and water and fertilizer systems are depreciated on a straight-line basis over 5 years, with an annual depreciation cost of 12,000 yuan / 30 mu = 400 yuan / mu.

[0206] The overall net profit (average per mu) of the experimental group increased by 2,051 yuan compared with control group 1 and by 1,238 yuan compared with control group 2.

[0207] Analysis shows that the differences between this invention and the control group stem from increased yield, resource conservation, and reduced labor costs. This demonstrates that the invention, through its closed-loop control mechanism combining drone monitoring and dynamic models, solves the problems of low monitoring efficiency, high misjudgment rates, and resource waste inherent in traditional fertilization methods. The experimental results demonstrate its feasibility in large-scale orchards, particularly suitable for high-value crops (such as the Golden Mango), where refined nutrient management is needed to improve quality. Furthermore, in areas with labor shortages, fully automated control significantly reduces reliance on manual labor. In resource-constrained areas, its water- and fertilizer-saving characteristics are well-suited for arid or infertile soil environments, making it highly valuable for widespread application.

[0208] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Further modifications can be readily implemented by those skilled in the art.

Claims

1. An intelligent mango fertilization method, characterized in that, Includes the following steps: S1) Arrange drip irrigation pipelines connecting to the water and fertilizer mixing device in the drip irrigation area; S2) A multispectral imaging device periodically acquires canopy images of mango trees, the canopy images containing visible and near-infrared band information of leaves, flower buds, flowers and fruits; S3) Based on image processing algorithms, extract visual and spectral features of the canopy image. The visual features include leaf color distribution, morphological thickness index, flower bud density, flower coverage, and fruit volume ratio. The spectral features include the near-infrared reflectance ratio of the leaf area to the flower and fruit area. S4) Based on the visual and spectral features, the flower bud differentiation period, flowering period and fruit enlargement period of the mango tree are determined by image comparison, and then the nitrogen deficiency state, phosphorus deficiency state, potassium deficiency state and water deficiency state at different periods are identified by the preset judgment model. S5) Based on the determination of nutrient or water deficiency status, the water-fertilizer mixing device dynamically adjusts the concentration of the corresponding nutrient in the drip irrigation solution and the irrigation amount; wherein, step S5 further includes the following steps: S5.1) When a state of deficiency of both nutrients and water is detected, the drip irrigation pipeline is activated first to irrigate water, and the image data is re-acquired after a set delay after irrigation. S5.2) If the nutrient deficiency persists after re-collection, trigger the drip irrigation system to replenish the corresponding elements; The specific implementation methods for triggering the replenishment of corresponding elements in the drip irrigation pipeline include: S5.21 When the probability of a single nutrient deficiency is ≥70%, the water and fertilizer mixing device shall be controlled to increase the corresponding nutrient concentration to the preset upper limit and the drip irrigation volume shall be increased to 1.2-1.5 times the standard irrigation volume; S5.22 When the probability of a single nutrient deficiency is 30%-70%, adjust the nutrient concentration and irrigation amount linearly. S5.23 When the probability of nutrient deficiency is less than 30%, drip irrigation should be performed while maintaining the baseline concentration and irrigation volume; S5.24 The specific steps for prioritizing the activation of drip irrigation pipelines for water irrigation are as follows: if both water and nutrient deficiency are detected simultaneously, irrigation is prioritized until the soil moisture content reaches the standard, and then the canopy image is re-acquired for a second determination after a first set time delay. S5.25 If nutrient deficiency still exists after the second determination, the drip irrigation system will be triggered to supplement the corresponding elements, and the delay time for feedback data collection will be dynamically adjusted according to the following rules: a) When the supplementary element is nitrogen, the delay time is 48-72 hours; b) When the supplemented element is phosphorus or potassium, the delay time is 5-7 days; c) When nitrogen, phosphorus and potassium are simultaneously deficient, the longest delay time corresponding to phosphorus or potassium, 5-7 days, shall be used as the benchmark. S5.26 After each adjustment, feedback data is collected according to the dynamic adjustment delay time until the lack probability value is <10%; if the lack probability value is still ≥10% after 3 consecutive adjustments, an abnormal alarm is triggered and automatic control is terminated.

2. The method according to claim 1, characterized in that, The multispectral imaging device is mounted on a drone or a ground mobile platform. The drone periodically collects images of the mango orchard canopy according to a preset flight path, and the preset lighting conditions are maintained by the drone's built-in supplementary lighting module during image acquisition.

3. The method according to claim 1, characterized in that, The specific methods for extracting visual and spectral features in step S3 include: S3.1) Perform semantic segmentation on the canopy image to separate the leaf, flower bud, flower, and fruit regions; S3.2) In the leaf region, extract the green saturation distribution data in the HSV color space, and calculate the standard deviation of the point cloud density in the leaf vein region through three-dimensional point cloud reconstruction technology, which serves as an indicator of leaf morphology and thickness. S3.3) Calculate the number density of flower buds per unit area in the flower bud region, calculate the percentage of the coverage area in the flower region relative to the total canopy area, and calculate the percentage of the volume in the fruit region relative to the total canopy volume. S3.4) For near-infrared images, calculate the ratio of the average reflectance of the leaf region to the average reflectance of the flower and fruit region, and use it as the spectral reflectance ratio.

4. The method according to claim 1, characterized in that, The specific process for image comparison and growth stage determination in step S4 includes: S4.1) Based on the leaf morphology and thickness index, flower bud number density, flower coverage percentage, fruit volume percentage, and spectral reflectance ratio, and by accumulating the fruit volume percentage extracted each time, calculate the fruit volume growth rate and match it with the preset growth stage standard library. S4.2) The growth stage standard library includes thresholds for flower bud density, flower coverage, fruit volume percentage, fruit volume growth rate, and spectral reflectance ratio for each stage. If the current flower bud density exceeds the flower bud density threshold for the differentiation stage and the flower coverage percentage is zero, it is determined to be the flower bud differentiation stage. If the current flower coverage percentage exceeds the flower coverage threshold for the flowering stage and the fruit volume percentage is lower than the threshold for the enlargement stage, it is determined to be the flowering stage. If the fruit volume percentage growth rate continuously exceeds the fruit volume growth rate threshold for the enlargement stage and the spectral reflectance ratio is lower than the threshold for the spectral reflectance ratio, it is determined to be the fruit enlargement stage. S4.3) After determining the growth stage, the nutrient determination sub-model for that period is called. The sub-model uses a random forest algorithm to fuse near-infrared reflectance, ambient temperature and humidity and historical fertilization data to output the probability value of nitrogen, phosphorus, potassium or water deficiency.

5. The method according to claim 1, characterized in that, The specific implementation of step S3.2, "calculating the standard deviation of point cloud density in the leaf vein region as a morphological thickness index using three-dimensional point cloud reconstruction technology," includes: S3.21) The point cloud data of the leaf region is sampled in layers, the leaf is divided into several sub-regions, and the grid containing the center line of the leaf vein is reconstructed with high precision in three dimensions. The non-vein grid is generated into a low-resolution point cloud using linear interpolation. S3.2.2) The high-precision leaf vein point cloud and the interpolated non-leaf vein point cloud are locally registered based on the improved ICP algorithm with the leaf vein centerline as the reference. The number of iterations and the search radius are limited. The registration data is processed in parallel by the half-precision floating-point operation of the UAV's onboard GPU, and the standard deviation of the point cloud density is output.

6. The method according to claim 4, characterized in that, The specific implementation of calling the period-specific nutrient determination sub-model in step S4.3 includes the following steps: S4.31) Based on the growth stage determined in step S4.2, load the corresponding sub-model from the pre-trained model library; S4.32) Input the current near-infrared reflectance, ambient temperature and humidity, historical fertilization data, green saturation distribution data, and leaf morphology and thickness indicators into the sub-model; output the probability value of nitrogen, phosphorus, potassium, or water deficiency through the random forest algorithm; S4.33) When any probability value is ≥30%, the control logic of step S5 is triggered.

7. The method according to claim 6, characterized in that, The pre-trained model library is constructed through the following steps: S4.311) Collect historical data on mango cultivation, including canopy images, near-infrared reflectance, environmental temperature and humidity, historical fertilization records, green saturation distribution data, and leaf morphology and thickness indicators at different growth stages. S4.312) The historical data is annotated, and the annotation content includes the actual state of nitrogen, phosphorus, potassium and water deficiency in each period; S4.313) Divide the dataset according to the growth stage and train the flower bud differentiation period sub-model, flowering period sub-model and fruit enlargement period sub-model respectively. Each sub-model adopts the random forest algorithm. The input features are near-infrared reflectance, environmental temperature and humidity, historical fertilizer application, green saturation distribution data and leaf morphology and thickness index. The output is the nutrient deficiency probability value. S4.314) Optimize the sub-model parameters through cross-validation until the test set accuracy reaches the preset value.

8. The method according to claim 6, characterized in that, The specific training and validation methods for the sub-models include: Flower bud differentiation period sub-model: In the input features, the weight of green saturation distribution data accounts for 40%, leaf morphology and thickness index accounts for 30%, near-infrared reflectance accounts for 20%, and environmental temperature and humidity accounts for 10%; Flowering period sub-model: In the input features, the percentage of flower coverage accounts for 35%, near-infrared reflectance accounts for 30%, green saturation distribution data accounts for 25%, and historical fertilization amount accounts for 10%. Fruit enlargement period sub-model: Among the input features, the ratio of spectral reflectance accounts for 50%, the fruit volume growth rate accounts for 30%, the ambient temperature accounts for 15%, and the historical fertilization amount accounts for 5%; The random forest algorithm parameters for each sub-model are: 100-150 decision trees, 10-12 layers at maximum depth, and 5 samples per leaf node.

Images