Intelligent mango fertilizing method

Abstract

The invention discloses an intelligent mango fertilizing method. The method comprises the following steps: arranging a drip irrigation pipeline connected with a water and fertilizer mixing device in a drip irrigation area; canopy images including visible light and near-infrared band information of leaves, flower buds, flowers and fruits are periodically collected through a multispectral imaging device; visual features and spectral features are extracted based on an image processing algorithm, the visual features comprise leaf color distribution, form thickness indexes, flower bud density, flower coverage rate and fruit volume ratio, and the spectral features are near-infrared reflectivity ratios of leaf and flower and fruit areas; matching a growth stage standard library according to the characteristic data, judging a flower bud differentiation stage, a flowering stage or a fruit expansion stage, and calling a corresponding stage judgment model to identify a nutrient and moisture deficiency state; and dynamically adjusting the nutrient concentration and dosage of the drip irrigation liquid according to the judgment result. The method is used for accurate water and fertilizer management of mango planting, a fertilization decision is optimized through multi-source data fusion and closed-loop regulation, and the response capability in a complex nutrient demand scene is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent fertilization for agricultural planting, and more specifically, relates to an intelligent mango fertilization method. Background Art

[0002] In mango cultivation, scientific fertilization is a key factor influencing fruit yield and quality. Traditional fertilization methods primarily rely on manual experience or fixed monitoring equipment, which have the following limitations: 1. Insufficient monitoring coverage and data reliability. Traditional methods often use ground-based fixed sensors or handheld devices to collect plant data. Single-point monitoring has a limited range and cannot cover large orchards. For example, fixed multispectral cameras can only image a portion of the tree canopy, requiring frequent relocation of the equipment, resulting in inefficient data collection. Furthermore, natural light intensity fluctuates with weather and time of day, directly affecting the accuracy of canopy spectral reflectance measurements. Unstable lighting conditions can easily distort the spectral characteristics of leaves and flowers and fruits, especially during cloudy or dusk periods, when data deviations in the near-infrared band increase significantly. These issues reduce the reliability of nutrient estimation models based on spectral analysis, making it difficult to support long-term, stable fertilization decisions. 2. Lack of dynamic control capabilities for complex nutrient requirements. The proportions of nitrogen, phosphorus, potassium, and water required by mangoes vary significantly during different growth stages (such as flower bud differentiation, flowering, and fruit expansion). Traditional intelligent fertilization systems typically use single thresholds or static models for control, such as triggering irrigation based solely on soil moisture or leaf color, without considering dynamic changes in growth stages. When multiple nutrients are simultaneously deficient, the system struggles to prioritize, leading to blindly increasing the application of a single nutrient while ignoring other needs. For example, if both phosphorus and water are deficient during the flowering period, directly supplementing with phosphorus fertilizer may reduce fertilizer utilization efficiency due to limited root water absorption capacity. These issues stem from a lack of integrated analysis of multi-dimensional features (such as flower and fruit density and leaf physiological indicators) and a lack of refined design of nutrient requirement models for different periods of time. Low growth stage recognition accuracy leads to inaccurate fertilization timing. Existing technologies often rely on manual observation or single visual features (such as flower count) and lack quantitative indicators and multi-source data validation. For example, determining the flowering period based solely on flower cover can lead to misjudgment due to partial obscuration of flowers by leaves or insufficient imaging resolution. Furthermore, the identification of the fruit expansion period often lags behind actual physiological changes, resulting in untimely potassium fertilizer addition and a negative impact on fruit expansion rate. 4. The logic for coordinated regulation of water and nutrients is imperfect. In scenarios where both water and nutrients are scarce, traditional systems often adopt a parallel replenishment strategy, ignoring 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 barriers. In addition, uneven soil moisture distribution after irrigation may affect the accuracy of secondary nutrient status detection. Solving such problems requires the establishment of a closed-loop control mechanism of "water first, then fertilizer", but existing technologies lack real-time linkage analysis of canopy characteristics, making it difficult to achieve dynamic delay judgment and feedback adjustment.

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

[0004] An object of the present invention is to address at least the above-mentioned drawbacks and to provide at least the advantages which will be described hereinafter.

[0005] The present invention provides an intelligent mango fertilization method, which realizes the accurate identification of nitrogen, phosphorus, potassium and water through multi-dimensional feature fusion analysis, combines priority regulation and feedback supplement mechanism to form a closed-loop fertilization decision, improves the response capability at different periods under complex nutrient deficiency scenarios, and increases the efficiency and yield of mango cultivation.

[0006] The present invention provides an intelligent mango fertilization method, comprising the following steps: S1) Arrange drip irrigation pipes connected to the water-fertilizer mixing device in the drip irrigation area; S2) a multispectral imaging device periodically collects canopy images of mango trees, wherein the canopy images include visible and near-infrared band information of leaves, flower buds, flowers, and fruits; S3) extracting visual and spectral features of the canopy image based on an image processing algorithm, wherein the visual features include leaf color distribution, morphological thickness index, flower bud density, flower coverage, and fruit volume ratio; and the spectral features include the near-infrared reflectance ratio between the leaf region and the flower and fruit region; S4) determining the flower bud differentiation period, flowering period, and fruit expansion period of the mango tree by image comparison based on the visual and spectral features, and then identifying the nitrogen deficiency, phosphorus deficiency, potassium deficiency, and water deficiency status of the mango tree at different periods using a preset judgment model; S5) According to the determination result of the nutrient deficiency state, the water-fertilizer mixing device is controlled to dynamically adjust the concentration of corresponding nutrients in the drip irrigation liquid and the irrigation amount.

[0007] Preferably, the multispectral imaging device is mounted on a drone or a ground mobile platform, and the drone periodically collects canopy images of the mango orchard according to a preset flight path, and the preset lighting conditions are maintained by the drone's own fill light module during image collection.

[0008] The drone's preset flight path is dynamically planned based on the mango orchard's geo-fence map and the tree canopy projection range. This can be executed using the following existing steps: S2.1) Scan the mango orchard terrain using a drone-mounted LiDAR to generate a geo-fence map containing the location of each mango tree and the canopy boundary; S2.2) Based on the geo-fence map, the flight path is divided into several parallel routes. The spacing between adjacent routes is determined by calculating the coverage width of a single shot of the multispectral imaging device and the lateral overlap ratio, wherein the lateral overlap ratio is 60%-70%; S2.3) As the drone flies along the route, it uses its onboard GPS and visual obstacle avoidance module to correct its heading in real time, and hovers to capture images when it reaches the center of the canopy of each mango tree; S2.4) The fill light module includes a ring-shaped LED array and a light intensity sensor. Before shooting, the light intensity sensor detects the ambient light intensity. If it is below a preset threshold, the LED array is activated to provide fill light at a constant color temperature and intensity to ensure uniform lighting on the canopy surface.

[0009] Preferably, the specific method of extracting the visual features and spectral features in step S3 includes: S3.1) Perform semantic segmentation on the canopy image to separate leaves, buds, flowers, and fruit regions; S3.2) Extract green saturation distribution data in the HSV color space from the leaf region, and calculate the standard deviation of the point cloud density in the vein region using 3D point cloud reconstruction technology as an indicator of leaf morphological thickness; S3.3) Count the number of flower buds per unit area in the flower bud area, calculate the percentage of coverage area to total canopy area in the flower area, and calculate the percentage of volume to total canopy volume in the fruit area; S3.4) For the near-infrared band image, calculate the ratio of the average reflectance of the leaf region to the average reflectance of the flower and fruit region as the spectral reflectance ratio (spectral feature data).

[0010] The present invention extracts different target areas (leaves, flowers, and fruits) through semantic segmentation to enhance the comprehensiveness of features; and combines the distribution characteristics of flowers and fruits with leaf indicators to support growth stage determination in multiple dimensions.

[0011] Preferably, the specific process of image comparison and growth stage determination in step S4 includes: S4.1) Based on the leaf morphology thickness index extracted in step S3.2, the flower bud number density, flower cover 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 percentages extracted each time, and the result is matched with a preset growth stage standard library; S4.2) The growth stage standard library includes the flower bud density threshold, flower coverage threshold, fruit volume percentage, fruit volume growth rate threshold and spectral reflectance ratio threshold for each period; if the current flower bud number density exceeds the flower bud density threshold of the differentiation period and the flower coverage percentage is zero, it is determined to be the flower bud differentiation period; if the current flower coverage percentage exceeds the flower coverage threshold of the flowering period and the fruit volume percentage is lower than the expansion period threshold, it is determined to be the flowering period; if the fruit volume percentage growth rate continues to exceed the fruit volume growth rate threshold of the expansion period and the spectral reflectance ratio is lower than the spectral reflectance ratio threshold, it is determined to be the fruit expansion period.

[0012] The specific judgment thresholds are: flower bud differentiation period: flower bud density ≥ 2.5 / cm² and flower coverage percentage = 0%; Flowering period: flower coverage percentage ≥ 30% and fruit volume percentage ≤ 15%; Fruit expansion period: the average daily growth rate of fruit volume is ≥3% and the spectral reflectance ratio is ≤1.0.

[0013] The plant physiological mechanism that leads to a spectral reflectance ratio threshold of 1.0 is that during fruit expansion, plants preferentially direct a large amount of photosynthate and water to the fruit, resulting in decreased leaf photosynthetic activity and a reduction in chlorophyll content. Leaf near-infrared (NIR) reflectance is closely related to mesophyll cell structure and water content. When leaf physiological activity weakens, NIR reflectance decreases significantly. During the expansion phase, fruit cells rapidly divide, increasing water content and densifying their internal structure, leading to enhanced near-infrared light scattering and a relative increase in reflectance. The combined effect of decreased leaf reflectance and increased fruit reflectance results in a reflectance ratio (leaf / fruit) between the leaf and flower regions less than 1.0.

[0014] S4.3) After determining the growth stage, a nutrient determination sub-model dedicated to that stage is called. The sub-model integrates near-infrared reflectance, ambient temperature and humidity, and historical fertilization data through a random forest algorithm to output a probability value for nitrogen, phosphorus, potassium, or water deficiency.

[0015] The present invention locks each growth stage through quantitative analysis of flower and fruit density and volume growth rate; matches the nutrient demand model in stages to avoid fertilization deviation caused by misjudgment.

[0016] Preferably, step S5 also includes the priority: S5.21 When the probability value of a single nutrient deficiency is ≥70%, the water-fertilizer mixing device will be controlled to increase the corresponding nutrient concentration to the preset upper limit and the drip irrigation volume will be increased to 1.2-1.5 times the standard irrigation volume; S5.22 When the probability of a single nutrient deficiency is between 30% and 70%, adjust the nutrient concentration and irrigation amount in a linear manner; S5.23 When the probability of nutrient deficiency is less than 30%, drip irrigation is performed while maintaining the basic concentration and irrigation volume; S5.24 Prioritizing the drip irrigation system for water irrigation specifically includes: if both water and nutrient deficiencies are detected, irrigation is prioritized until the soil moisture content reaches the standard, followed by a delay of a first set time (e.g., 24 hours) to recapture canopy images for a second assessment; S5.25 If nutrient deficiency still exists after the second determination, the drip irrigation pipeline is triggered to supplement the corresponding elements, and the delay time for feedback data collection is dynamically adjusted according to the following rules: a) When nitrogen is used as the supplemental element, the delay is 48-72 hours; b) When the supplemental element is phosphorus or potassium, the delay is 5-7 days; c) When nitrogen, phosphorus, and potassium are deficient at the same time, the longest delay of 5-7 days corresponding to phosphorus or potassium shall be used as the benchmark; S5.26 After each adjustment, feedback data is collected according to the dynamically adjusted delay time until the lack probability value is less than 10%; if the lack probability value is still ≥10% after three consecutive adjustments, an abnormal alarm is triggered and automatic control is terminated.

[0017] Preferably, the specific implementation of "calculating the point cloud density standard deviation of the vein area as the morphological thickness indicator by using the three-dimensional point cloud reconstruction technology" in step S3.2 includes: S3.21) Perform stratified sampling of the leaf region point cloud data, dividing the leaf into several sub-regions (square grids with sides of 5-8 cm). Perform high-precision 3D reconstruction (point cloud density ≥ 100 points / cm²) on the grid containing the vein centerline. Use linear interpolation to generate low-resolution point clouds (point cloud density ≤ 20 points / cm²) on the non-vein grid to reduce computational effort. S3.2.2) Locally align the high-precision vein point cloud with the interpolated non-vein point cloud using the improved ICP (Iterative Closest Point) algorithm with the vein centerline as the reference. Limit the number of iterations (≤10) and the search radius (≤1 mm). Process the alignment data in parallel using half-precision floating-point operations on the drone's onboard GPU, and output the point cloud density standard deviation.

[0018] Preferably, the specific implementation of calling the period-specific nutrient determination sub-model in step S4.3 includes the following steps: S4.31) Loading the corresponding sub-model from the pre-trained model library according to the growth stage determined in step S4.2; S4.32) Inputting current near-infrared reflectance, ambient temperature and humidity, historical fertilization data, green saturation distribution data, and leaf morphology and thickness index into the sub-model; outputting a probability value of nitrogen, phosphorus, potassium, or water deficiency using a random forest algorithm; S4.33) When any probability value is ≥30%, the control logic of step S5 is triggered.

[0019] Preferably, the pre-trained model library is constructed by the following steps: S4.311) Collect historical data on mango cultivation, including canopy images at different growth stages, near-infrared reflectance, ambient temperature and humidity, historical fertilization records, green saturation distribution data, and leaf morphology and thickness indicators; S4.312) annotating the historical data, including the actual status of nitrogen, phosphorus, potassium and water deficiencies at each period; S4.313) The dataset was divided by growth stage, and sub-models for flower bud differentiation, flowering, and fruit expansion were trained separately. Each sub-model used a random forest algorithm. The input features were near-infrared reflectance, ambient temperature and humidity, historical fertilizer application, green saturation distribution data, and leaf morphology and thickness indicators. The output was the probability of nutrient deficiency. S4.314) Optimize sub-model parameters through cross-validation until the test set accuracy reaches a preset value (e.g., the preset value ≥ 85%).

[0020] Preferably, the specific training and verification methods of the sub-model include: Flower bud differentiation period sub-model: The weight of the green saturation distribution data in the input features accounts for 40%, the leaf morphology and thickness index accounts for 30%, the near-infrared reflectance accounts for 20%, and the ambient temperature and humidity account for 10%; Flowering period sub-model: Among the input features, flower cover percentage accounts for 35%, near-infrared reflectance accounts for 30%, green saturation distribution data accounts for 25%, and historical fertilizer application accounts for 10%; Fruit expansion period sub-model: Among the input features, the spectral reflectance ratio accounts for 50%, the fruit volume growth rate accounts for 30%, the ambient temperature accounts for 15%, and the historical fertilizer application rate accounts for 5%; The random forest algorithm parameters for each sub-model are: 100-150 decision trees, 10-12 maximum depth, and 5 minimum leaf node samples.

[0021] The present invention has at least the following beneficial effects: This method uses multi-dimensional feature fusion analysis to accurately identify nitrogen, phosphorus, potassium, and water. It combines priority control with a feedback replenishment mechanism to form a closed-loop fertilization decision-making system, improving the ability to respond to different periods of complex nutrient deficiency scenarios. This addresses the problem that traditional intelligent fertilization methods lack dynamic control logic based on different periods and are difficult to adapt to complex nutrient demand scenarios.

[0022] This invention leverages the mobility of drones to expand monitoring range, combines active fill lighting to ensure consistent image acquisition conditions, and improves the reliability of spectral feature analysis. This addresses the limitations of traditional fixed imaging equipment, which suffer from limited coverage and spectral data distortion caused by natural light fluctuations, making stable monitoring of large orchards difficult.

[0023] This invention dynamically adjusts nitrogen, phosphorus, and potassium concentrations and irrigation rates through a phased sub-model and priority control logic. In experiments, fertilizer use in the experimental group decreased by 15%-19%, while yield per plant increased to 36.5 kg (compared to 29.8 kg in the control group), demonstrating the effectiveness of dynamic regulation.

[0024] This method integrates multispectral reflectance, morphological indicators (such as fruit volume growth rate), and environmental data to establish a cross-validation mechanism to reduce the false positive rate. Test results showed that the false positive rate for nutrient deficiency was reduced to 8.7% (compared to 23.4% in control group 1).

[0025] This invention utilizes a closed-loop mechanism of "water first, fertilizer second," with a 24-hour delay after irrigation to determine nutrient needs. In trials, the experimental group achieved a 25% water saving. Furthermore, due to adequate water availability, fertilizer utilization increased by 12%, resulting in a 2,051 yuan / mu increase in overall net profit.

[0026] Other advantages, objectives and features of the present invention will be reflected in part from the following description and will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a schematic flow chart of the intelligent mango fertilization method of the present invention; DETAILED DESCRIPTION

[0028] The present invention is further described in detail below with reference to the embodiments so that those skilled in the art can implement the invention with reference to the description.

[0029] It should be noted that the experimental methods described in the following embodiments are conventional methods unless otherwise specified, and the reagents and materials are commercially available unless otherwise specified. In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "set" should be understood in a broad sense, for example, they can be fixedly connected or set, or detachably connected or set, or integrally connected or set. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances. The orientations or positional relationships indicated by the terms "transverse", "longitudinal", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc. are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description. They do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention.

[0030] Figure 1 A flow chart of an intelligent mango fertilization method is shown, comprising the following steps: S1) Arrange drip irrigation pipes connected to the water-fertilizer mixing device in the drip irrigation area; S2) a multispectral imaging device periodically collects canopy images of mango trees, wherein the canopy images include visible and near-infrared band information of leaves, flower buds, flowers, and fruits; S3) extracting visual and spectral features of the canopy image based on an image processing algorithm, wherein the visual features include leaf color distribution, morphological thickness index, flower bud density, flower coverage, and fruit volume ratio; and the spectral features include the near-infrared reflectance ratio between the leaf region and the flower and fruit region; S4) determining the flower bud differentiation period, flowering period, and fruit expansion period of the mango tree by image comparison based on the visual and spectral features, and then identifying the nitrogen deficiency, phosphorus deficiency, potassium deficiency, and water deficiency status of the mango tree at different periods using a preset judgment model; S5) According to the determination result of the nutrient or water deficiency state, the water-fertilizer mixing device is controlled to dynamically adjust the concentration of corresponding nutrients in the drip irrigation liquid and the irrigation amount.

[0031] This method uses multi-dimensional feature fusion analysis to accurately identify nitrogen, phosphorus, potassium, and water. It combines priority control with a feedback replenishment mechanism to form a closed-loop fertilization decision-making system, improving the ability to respond to different periods of complex nutrient deficiency scenarios. This addresses the problem that traditional intelligent fertilization methods lack dynamic control logic based on different periods and are difficult to adapt to complex nutrient demand scenarios.

[0032] According to one embodiment of the present invention, the annular water and fertilizer drip irrigation area is 0.9-1.2 times the projection radius of the tree crown, the aperture of the drip irrigation pipe is 1-2 mm, and the drip irrigation flow rate can be adjusted in the range of 0.5-5 liters / minute. The drip irrigation pipe can be made of polyethylene, and the water and 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 rating of 0.3-0.6 MPa. The annular drip irrigation area is set around the tree trunk, and the drip irrigation pipe is buried at a depth of 10-15 cm below the surface and evenly distributed along the drip irrigation area. The water and fertilizer mixing device controls the drip irrigation liquid flow through a solenoid valve, the flow sensor monitors the irrigation amount in real time, and the fertilizer concentration detection module detects the nitrogen, phosphorus and potassium concentrations in the mixed liquid through a conductivity sensor.

[0033] The drone's flight altitude is 3-5 meters, with a single shot covering a width of 2-3 meters. The fill light module's illumination intensity can be adjusted between 800 and 1200 lux. Multispectral imaging equipment can utilize a five-band camera (red, green, blue, near-infrared, and thermal infrared). The fill light module can be configured with a ring-shaped LED array (color temperature 5500K). Alternatively, see the LED fill light module described in "Research on the Application of Automatic Fill Lighting Systems for UAVs in Overhead Transmission Line Inspections" by Yang Junwu et al. The multispectral camera is mounted on the drone's gimbal, and the fill light module is fixed around the camera lens. The drone flies along a pre-set route and hovers in the center of the tree canopy. A light intensity sensor monitors ambient light intensity. If it falls below 800 lux, the LED array is activated to increase the light intensity to 1200 lux to ensure uniform illumination during image acquisition.

[0034] The thresholds for flower bud density are ≥2.5 per cm², flower coverage is ≥30%, daily fruit volume growth is ≥3%, and the spectral reflectance ratio is ≤1.0. Image processing algorithms can implement semantic segmentation based on the OpenCV library, and 3D point cloud reconstruction can utilize a ToF (Time of Flight) depth camera. Point cloud data processing can be powered by an embedded GPU (such as the Jetson Xavier NX). Images captured by the drone are transmitted to a ground server, where semantic segmentation algorithms are used to separate leaves, flower buds, flowers, and fruits. Green saturation distribution data for leaf regions is extracted from the HSV color space, and the standard deviation of vein point cloud density is calculated using stratified sampling (high-precision area point cloud density ≥100 points / cm²). The near-infrared reflectance ratio is calculated by averaging the reflectance of leaves and flowers and fruits. If the ratio is ≤1.0, combined with the fruit volume growth rate, the plant is considered to be in the expansion stage. The soil moisture content threshold is 60%-70% of field capacity, the nutrient deficiency probability trigger threshold is ≥30%, and the irrigation delay is 24 hours. The soil moisture sensor can be a capacitive sensor, and the water-fertilizer mixing unit can be integrated with a PID controller. The sensor probe can be made of corrosion-resistant ceramic. The soil moisture sensor is buried 20 cm below the drip irrigation area to monitor soil moisture in real time. If both water and nutrients are deficient, the water-fertilizer mixing unit prioritizes irrigation until the soil moisture content reaches 65%. After 24 hours, images are collected again to determine nutrient status. If the probability of nutrient deficiency is ≥30%, the drip irrigation solution concentration is adjusted linearly (for example, at a probability of 70%, the concentration is increased to 1.5 times the upper limit).

[0035] This method combines circular drip irrigation with drone-assisted multispectral monitoring to accurately identify mango growth stages and dynamically control nutrient requirements. Layered image analysis and threshold determination reduce the impact of natural light fluctuations on data, improving the reliability of fertilization decisions. The control logic that prioritizes water replenishment optimizes nutrient absorption efficiency, avoids resource waste, and adapts to scenarios with complex nutrient deficiencies.

[0036] If a ground mobile platform is used for mobile collection, build a ground mobile platform with a height of 3-5 meters and let the equipment move on the platform. The other execution steps are the same as above.

[0037] According to another embodiment of the present invention, the drone's flight altitude can be set to 3-5 meters, the distance 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 ratio of 65%), and the lateral overlap ratio can be set to 60%-70%. The lidar scanning accuracy can be configured to ±2 centimeters. The multispectral imaging device can use a five-band camera (red, green, blue, near-infrared, and thermal infrared), the lidar can use a 16-line scanning module, and the visual obstacle avoidance module can use a binocular camera. The drone's fuselage can be constructed of carbon fiber composite materials, and the gimbal bracket can be constructed of aluminum alloy. The multispectral camera is mounted on the gimbal at the bottom of the drone, and the lidar is fixed to the front of the drone. The drone automatically generates parallel flight paths based on the geo-fence map, with the distance between flight paths calculated based on the single-shot coverage width and the overlap ratio. 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 of the tree canopy, the drone hovers for 1-2 seconds to complete image acquisition. The fill light module's light intensity threshold 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 fill light module can include a ring-shaped LED light cluster (15-20 cm in diameter), the light intensity sensor can be a digital light meter, and the control circuit can integrate a PWM dimming chip. The LED lampshade can be made of polycarbonate, and the heat sink can be made of aluminum alloy. The ring-shaped LED array is installed 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 the fill light process, the light intensity sensor continuously monitors and provides feedback to ensure that the light uniformity error is less than 5%.

[0038] This solution uses pre-set flight paths and dynamic route planning to capture full-coverage images of the orchard, reducing the need for manual intervention. The fill light module actively adjusts light intensity, suppressing the impact of natural light fluctuations on spectral data and improving image consistency across time. The collaborative operation of the drone and LiDAR ensures that the flight path precisely adapts to the orchard terrain, avoiding missed or duplicate images.

[0039] 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 of ≥90%. The training dataset can include 5,000 annotated mango canopy images. The semantic segmentation algorithm can be implemented based on a deep learning framework (such as a U-Net network), and the operating platform can utilize an embedded GPU computing module. The training dataset can be stored on a solid-state drive, and the annotation tool can be based on open-source image annotation software. Canopy images captured by the drone are transmitted to a ground-based processing unit, where pixel-level classification is performed using a U-Net network. Leaf regions are initially extracted using a green channel threshold (e.g., an H value range of 100-140 in HSV space). Flower buds and flower regions are separated using morphological filtering. Fruit regions are identified based on color clustering (e.g., a red channel percentage of ≥60%). The segmentation results are output as binary masks for subsequent feature extraction. Green saturation distribution data can be quantized into a histogram ranging from 0 to 255. The grid length for calculating the standard deviation of the 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 a modified ICP (Iterative Closest Point) method. The point cloud data processing unit can be equipped with an embedded GPU (such as the Jetson series), and the heat dissipation module can use a copper thermal pad. Point cloud data of the leaf area is collected using a depth camera, and after gridding, high-precision reconstruction is performed on the vein centerline area. Linear interpolation is used to generate a low-resolution point cloud (density ≤ 20 points / cm²) in non-vein areas. During registration, the vein centerline is used as the reference, the ICP algorithm's search radius is limited to ≤1 mm, and the number of iterations is ≤10. The final output is the point cloud density standard deviation, which serves as a morphological thickness indicator.

[0040] Flower bud density can be calculated as ≥2.5 per square centimeter, flower coverage as ≥30%, and fruit volume as ≤15% (during the flowering phase) or ≥15% (during the expansion phase). Flower bud and flower area statistics can be calculated using the OpenCV contour detection algorithm, and fruit volume calculation can be combined with a stereo depth map. A binocular camera can be used as the stereo camera, and a ceramic checkerboard calibration plate can be used. The segmented flower bud area is analyzed for density using connected domain analysis, and flower coverage is calculated by the ratio of the contour area to the total canopy area. Fruit volume is reconstructed by fusing the depth map with the RGB image, and the percentage of the canopy volume is calculated using a voxel grid method.

[0041] The reflectance ratio threshold for the near-infrared band 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 can be averaged. The near-infrared band of the multispectral imaging device can be selected from the 900-950nm wavelength range, and the reflectance calibration plate can be made of polytetrafluoroethylene. The calibration plate holder can be made of aluminum alloy, and the reflectance data processing unit can integrate an FPGA acceleration module. After radiometric correction of the near-infrared band image, the average reflectance of the leaf region (based on the 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 expansion period.

[0042] This solution achieves precise quantitative analysis of leaf, flower, and fruit regions through semantic segmentation and multi-dimensional feature extraction. Combining point cloud density with spectral reflectance ratios enhances cross-validation for growth stage determination. Hierarchical processing strategies (such as high-precision vein reconstruction and non-vein interpolation) balance computational efficiency and accuracy requirements, adapting to real-time data processing in orchards. The integration of flower and fruit statistical features with spectral data provides a reliable basis for determining complex nutrient deficiencies.

[0043] According to another embodiment of the present invention, the flower bud density threshold during the flower bud differentiation stage can be set at ≥2.5 buds / cm², the flower coverage threshold during the flowering stage can be set at ≥30%, the volume growth rate threshold during the fruit expansion stage can be set at ≥3% daily average fruit volume growth rate, and the spectral reflectance ratio threshold can be set at ≤1.0. The growth stage standard library can be stored on a cloud server, the feature matching algorithm can be implemented using a Python script, 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 housing of the edge computing device can be made of aluminum alloy. Feature data collected by the drone (e.g., flower bud density of 2.8 buds / cm², 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%, the stage is determined to be flower bud differentiation; if the flower coverage is ≥30% and the fruit volume is ≤15%, the stage is determined to be flowering; if the daily average fruit volume growth rate is ≥3% and the spectral reflectance ratio is ≤1.0, the stage is determined to be expansion. The determination results trigger the corresponding sub-model loading instruction.

[0044] The input feature weights for the random forest sub-model can be configured to: 40% for green saturation during the bud differentiation period, 30% for leaf thickness, 35% for flower cover during the flowering period, 30% for near-infrared reflectance, and 50% for spectral reflectance ratio during the expansion period. The trigger threshold for nutrient deficiency probability is ≥30%. The sub-model can be deployed on edge computing devices. The random forest algorithm can be implemented using the Scikit-learn library. Digital probes can be used for the temperature and humidity sensors. ABS plastic can be used for the sensor housing, and shielded twisted-pair cables can be used for the data transmission. After the flowering period is determined, the flowering period sub-model is loaded, inputting the current near-infrared reflectance of 0.68, an ambient temperature of 28°C, and historical phosphorus application data (e.g., 0.5 kg / plant in the past week). The sub-model uses a random forest algorithm (120 decision trees, with a maximum depth of 10 layers) to calculate a probability of phosphorus deficiency of 45%. If the probability is ≥30%, water and fertilizer control instructions are triggered.

[0045] This solution achieves precise correlation between growth stage and nutrient status through quantitative thresholding and staged sub-model invocation. Local storage of standard libraries reduces cloud-based reliance and improves real-time response. The random forest algorithm integrates multi-source data to enhance robustness in complex deficiency scenarios. Staged weighting avoids misjudgment of single features and adapts to physiological differences at different stages.

[0046] According to another embodiment of the present invention, the soil moisture threshold can be set at 60%-70% of field capacity, the irrigation delay can be set to 24 hours, and the irrigation rate can be adjusted to 1.2-1.5 times the standard rate. The soil moisture sensor can be a capacitive sensor, the control module of the water-fertilizer mixing device can be integrated with a PID controller, and the irrigation solenoid valve can be a normally closed, two-way valve. The sensor probe can be made of corrosion-resistant ceramic; 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 in real time. If both water deficiency (e.g., water content ≤ 55%) and nutrient deficiency (e.g., nitrogen deficiency probability ≥ 30%) are detected, the PID controller prioritizes opening the irrigation solenoid valve to increase the soil moisture content to 65%. After irrigation is completed, the system delays for 24 hours and controls the drone to recapture canopy images for a secondary assessment.

[0047] 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, and the drip irrigation rate adjustment range can be set to 1.2-1.5 times the standard rate. The feedback data collection interval can be configured as follows: when supplementing nitrogen, the delay is 48-72 hours; when supplementing phosphorus or potassium, the delay is 5-7 days; when supplementing nitrogen, phosphorus, and potassium, the longest delay of 5-7 days for phosphorus or potassium is used as the reference. The probability calculation module can be integrated into the edge computing device. Feedback data can be stored using an SD card, and 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 the second determination, if the nitrogen deficiency probability is 70%, the conductivity sensor detects the current drip irrigation fluid concentration (for example, 0.8 g / L), and the PID controller increases the nitrogen concentration to the upper limit of 1.2 g / L and increases the drip irrigation rate to 1.3 times the standard rate. After the adjustment, the system collects canopy images and soil data again after a delay of 48 hours until the probability of nitrogen deficiency drops below 10%. If the probability is 40%, the concentration is adjusted linearly to 0.96g / L, and the drip irrigation rate is adjusted to 1.1 times the standard amount.

[0048] This solution optimizes the order of water and nutrient replenishment through priority control and feedback mechanisms, preventing low fertilizer utilization during drought conditions. Secondary assessment reduces the risk of misoperation, dynamically adjusts to varying levels of deficiency, and improves fertilization accuracy. Delayed data collection and continuous feedback ensure traceability of control effects, adapting to the dynamic needs of the orchard environment.

[0049] According to another embodiment of the present invention, the probability thresholds for single nutrient deficiency are set at 70% and 30%, with the upper limit for concentration adjustment set at 1.5 times the standard value (for example, a base nitrogen concentration of 1.0 g / L with an upper limit of 1.5 g / L). The irrigation rate adjustment range is 1.2-1.5 times the standard rate (for example, a standard rate of 5 L / plant becomes 6-7.5 L / plant after adjustment). The linear proportionality formula is: Adjustment value = Base value + (Current probability value - 30%) × (Upper limit - 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 use a four-electrode probe (range 0-5 mS / cm). The drip irrigation solution concentration can be adjusted using a stepper motor-driven peristaltic pump (flow accuracy ±2%). The peristaltic pump hose can be made of corrosion-resistant silicone, the sensor electrodes can be platinum-coated, and the data transmission cable can use a shielded twisted pair cable (waterproof rating IP67). Conductivity sensors are installed at the outlet of the drip irrigation pipe to monitor nitrogen, phosphorus, and potassium concentrations in real time (e.g., nitrogen concentration is 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 x 1.5). If the probability is 50%, the concentration is adjusted to 1.25 g / L using a linear formula, and the irrigation volume is adjusted to 6.25 L / plant.

[0050] The delay after nitrogen supplementation should be 48-72 hours (preferably 60 hours), and the delay after phosphorus / potassium supplementation should be 5-7 days (preferably 6 days). The threshold for consecutive adjustment failures is three, and the threshold for achieving the deficiency probability is <10%. Feedback data collection can rely on a drone multispectral imaging system (with a configurable recollection cycle). Data storage can use an industrial-grade SD card (capacity ≥ 64GB). The timer module can be based on a hardware RTC chip (with an accuracy of ±1 minute / month). The SD card protective case can be made of ABS plastic (dust and water resistant), and the timer circuit board can be made of FR-4 substrate (high temperature resistant). After adjustment is completed, the system starts a countdown (for example, a 60-hour nitrogen delay). At the end of the countdown, the drone is controlled to reacquire canopy images. If the second assessment shows a nitrogen deficiency probability of 8%, adjustment is terminated. If the probability reaches 12%, the concentration is adjusted again to 1.0 g / L and a new 48-hour countdown is initiated. If the probability remains ≥10% after three consecutive adjustments (for example, 12%, 15%, or 13%), an audible and visual alarm (such as a buzzer and flashing LED) is triggered, automatic adjustment is suspended, and manual intervention is required.

[0051] This solution uses probabilistic grading and linear proportional adjustment to precisely match nutrient replenishment to the degree of deficiency, avoiding salinization and resource waste caused by over-fertilization. A dynamic delay mechanism adapts to the physiological absorption cycles of different nutrients (e.g., nitrogen takes effect quickly, while phosphorus and potassium take effect slowly), ensuring the validity of feedback data. A continuous failure alarm prevents the system from falling into ineffective cycles, improving the safety and reliability of orchard management.

[0052] According to another embodiment of the present invention, the sub-region grid side length can be set to 5-8 cm. The point cloud density of the vein centerline grid can be configured to ≥100 points / cm², while the point cloud density of non-vein areas can be limited to ≤20 points / cm². 3D point cloud layered sampling can be implemented using a ToF depth camera, and the linear interpolation algorithm can be integrated into an embedded GPU computing module. The depth camera's protective housing can be made of aluminum alloy, and the interpolation algorithm's storage medium can be LPDDR4 memory. The ToF depth camera is mounted on a drone gimbal to collect raw point cloud data from the leaf area. The ground processing unit divides the leaf into a 6-cm square grid, identifies the grids containing the vein centerlines (e.g., grid numbers G01-G05), and performs high-precision reconstruction (point cloud density 120 points / cm²) on these grids. Linear interpolation is used to generate low-resolution point clouds (density 15 points / cm²) for non-vein grids (e.g., grid numbers G06-G20), reducing the GPU computing load.

[0053] The number of ICP algorithm iterations 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, with parallel computing enabled by embedded GPUs (such as the Jetson series). Registration results can be stored on NVMe solid-state drives. Copper fins can be used as GPU heat sinks, and SRAM chips can be used for the registration data cache. The high-precision vein point cloud and the interpolated non-vein point cloud are transmitted to the GPU, where improved ICP registration is performed based on the vein centerline. The algorithm limits the search radius to 0.8 mm, and after 8 iterations, the error threshold of 0.04 mm is reached. The density standard deviation (e.g., 0.12 points / cm²) of the vein region is calculated from the registered point cloud data and output as a morphological thickness indicator to the judgment module.

[0054] This solution leverages layered sampling and differentiated reconstruction strategies to balance point cloud data accuracy and computational resource consumption, improving the efficiency of large-scale leaf processing. An improved ICP algorithm, combined with hardware acceleration, ensures registration accuracy meets the requirements for morphological metric calculations. Separating vein and non-vein areas reduces redundant data interference and enhances the reliability of thickness metrics.

[0055] According to another embodiment of the present invention, the training dataset can include 2,000 sets of historical data, the test set accuracy threshold can be set to ≥85%, and the cross-validation fold can be set to 5. 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 edge computing device's SSD, data annotation can be based on open-source annotation tools, and model training can rely on Python's Scikit-learn library. The SSD storage medium can be 3D NAND flash memory, and the training server's heat sink can be made of aluminum alloy. The flower bud differentiation period sub-model is stored in the edge device's model partition M1. When the flower bud differentiation period is determined, the system loads the model from M1. Input data includes the current green saturation distribution histogram (e.g., a peak in the 120-140 range), leaf morphological thickness indicators (e.g., a point cloud density standard deviation of 0.15), and a near-infrared reflectivity of 0.72.

[0056] The input feature weights can be set to 40% for green saturation and 30% for leaf thickness in the flower bud differentiation submodel, 35% for flower cover in the flowering submodel, and 50% for spectral reflectance ratio in the plant growth submodel. The trigger threshold for nutrient deficiency probability can be set to ≥30%. The random forest algorithm can run on embedded GPUs (such as the Jetson series). The temperature and humidity sensors can use digital probes with an I2C interface, and 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 submodel, the inputs are a current flower cover of 32%, a near-infrared reflectance of 0.68, an ambient temperature of 28°C, and a historical phosphorus application rate of 0.5 kg / plant. The random forest algorithm (with 120 decision trees and a maximum depth of 10 layers) calculates a probability of phosphorus deficiency of 45%. If the probability is ≥30%, an adjustment instruction is sent to the water and fertilizer mixing device.

[0057] This solution improves the targeted assessment of nutrient deficiencies at different growth stages by loading sub-models in stages and configuring differentiated feature weights. Local deployment of pre-trained model libraries reduces cloud-based reliance and enhances real-time responsiveness. The random forest algorithm integrates multi-source data, reducing interference from single environmental variables and adapting to the dynamic changes in orchards.

[0058] According to another embodiment of the present invention, the historical dataset can include at least 3,000 data sets, with an annotation accuracy requirement of ≥95%. A single data set can include canopy images (resolution 1920×1080 pixels), near-infrared reflectance (range 0.3-0.8), and ambient temperature and humidity (temperature 15-35°C, humidity 30-80%). Canopy images can be collected using a multispectral camera, ambient temperature and humidity data can be collected using a digital temperature and humidity sensor, and historical fertilization records can be stored in an electronic ledger system. Enterprise-grade SSD hard drives can be used for data storage, and ABS plastic can be used for the sensor housing. The multispectral camera is mounted on a drone gimbal, and the temperature and humidity sensor is fixed to the orchard weather station bracket. Data collection is conducted weekly for two years. During annotation, each data set is manually labeled with its true status based on leaf yellowing (nitrogen deficiency), flower bud drop rate (phosphorus deficiency), and fruit shrinkage rate (potassium deficiency). The annotation results are stored in database table T1.

[0059] The flower bud differentiation dataset can contain 1000 data sets, 800 data sets for the flowering stage, and 1200 data sets for the swell stage. The random forest algorithm parameters can be configured to have 120 decision trees, a maximum depth of 12 layers, and a minimum leaf node sample size of 5. Model training can be performed using 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 performed using the Pandas library. Copper heat pipes can be used as the server heatsink, and DDR4 memory can be used. The flower bud differentiation dataset is loaded into the training server's memory. Input features include a green saturation distribution histogram (HSV space H value range 100-140), leaf morphological thickness index (point cloud density standard deviation 0.1-0.3), and near-infrared reflectance 0.5-0.7. The random forest model outputs a nitrogen deficiency probability value, which is saved as file M1 after training.

[0060] The cross-validation fold can be set to 50%, the test set accuracy threshold can be set to ≥85%, and the number of parameter optimization iterations can be limited to 100. Cross-validation can be implemented using the K-fold algorithm, parameter tuning can be performed using a grid search tool, and validation results can be stored in local log files. An NVMe solid-state drive can be used as the log file storage medium, and a 16-thread CPU can be allocated for grid search computing resources. After training the flowering period sub-model, a 50-fold cross-validation is performed, splitting the dataset into a training set (640 groups) and a validation set (160 groups). A grid search is performed over the number of decision trees (100-150) and the maximum depth (10-12 layers). The parameter combination with the highest validation accuracy (e.g., 87.3%) is selected and saved as the final model M2.

[0061] This solution ensures that model training aligns with actual growth needs through phased data collection and annotation. Cross-validation optimizes parameter combinations to enhance model generalization. The integration of historical data and multi-source features enhances robustness in different nutrient deficiency scenarios. Independent training of sub-models in different phases prevents cross-interference between feature weights at different times, adapting to the dynamic management needs of orchards.

[0062] According to another embodiment of the present invention, the weight of the green saturation distribution data in the flower bud differentiation period sub-model can be set to 40%, the weight of the leaf morphology thickness index to 30%, the weight of the near-infrared reflectivity to 20%, and the weight of the ambient temperature and humidity to 10%. In the flowering period sub-model, the weight of the flower cover percentage is 35%, the weight of the near-infrared reflectivity is 30%, the weight of the green saturation distribution data is 25%, and the weight of the historical fertilizer application amount is 10%. In the fruit expansion period sub-model, the weight of the spectral reflectance ratio is 50%, the weight of the fruit volume growth rate is 30%, the weight of the ambient temperature is 15%, and the weight of the historical fertilizer application amount is 5%. The weight distribution 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 the flower bud differentiation period sub-model is loaded, 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 weighted and summed up (e.g., green saturation score × 0.4 + leaf thickness score × 0.3) as the input vector of the random forest algorithm.

[0063] The number of decision trees can be set to 100-150, the maximum depth can be configured to 10-12 layers, and the minimum number of leaf node samples can be set to 5. The Gini impurity criterion can be used for feature splitting during training, 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 be performed on an embedded GPU (such as the Jetson series). The JSON configuration file can be stored on an eMMC chip, and a copper thermal pad can be used as the GPU heat sink. When training the flowering submodel, 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, and the minimum number of leaf node samples is limited to 5. After training, the model is saved as file M2 and deployed to the model partition of the edge computing device. √ represents the square root symbol.

[0064] This solution uses differentiated weight allocation in different stages to strengthen the discriminative contribution of key features at each growth stage and reduce the interference of secondary features. Standardizing random forest parameters improves model training efficiency and ensures algorithm consistency across different orchard environments. The coordinated configuration of weights and parameters adapts to the physiological characteristics of mangoes and enhances the robustness of judgment in scenarios with complex nutrient deficiencies.

[0065] According to an embodiment 1 of the intelligent mango fertilization method of the present invention, the following steps are included: S1) setting up a circular water and fertilizer drip irrigation area centered on the tree trunk, wherein a drip irrigation pipe connected to a water and fertilizer mixing device is arranged in the drip irrigation area; S2) Periodically capturing canopy images of mango trees using a multispectral imaging device carried by a drone. The canopy images include visible and near-infrared information of leaves, flower buds, flowers, and fruits. Preset lighting conditions are maintained during image acquisition using the drone's built-in fill light module. The drone's preset flight path is dynamically planned based on the mango orchard's geo-fence map and the tree canopy projection range. This can be executed using the following existing steps: S2.1) Scan the mango orchard terrain using a drone-mounted LiDAR to generate a geo-fence map containing the location of each mango tree and the canopy boundary; S2.2) Based on the geo-fence map, the flight path is divided into several parallel routes. The spacing between adjacent routes is determined by calculating the coverage width of a single shot of the multispectral imaging device and the lateral overlap ratio, wherein the lateral overlap ratio is 60%-70%; S2.3) As the drone flies along the route, it uses its onboard GPS and visual obstacle avoidance module to correct its heading in real time, and hovers to capture images when it reaches the center of the canopy of each mango tree; S2.4) The fill light module includes an LED array and a light intensity sensor. Before shooting, the light intensity sensor detects the ambient light intensity. If it is below a preset threshold, the LED array is activated to provide fill light at a constant color temperature and intensity to ensure uniform lighting on the canopy surface.

[0066] S3) Extracting visual and spectral features of the canopy image based on an image processing algorithm. 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 between the leaf region and the flower and fruit region. Specific extraction methods include: S3.1) Perform semantic segmentation on the canopy image to separate leaves, buds, flowers, and fruit regions; S3.2) Extract green saturation distribution data in the HSV color space from the leaf region, and calculate the standard deviation of the point cloud density in the vein region using 3D point cloud reconstruction technology as an indicator of leaf morphological thickness; S3.3) Count the number of flower buds per unit area in the flower bud area, calculate the percentage of coverage area to total canopy area in the flower area, and calculate the percentage of volume to total canopy volume in the fruit area; S3.4) For the near-infrared band image, calculate the ratio of the average reflectance of the leaf region to the average reflectance of the flower and fruit region as the spectral reflectance ratio (spectral feature data).

[0067] S4) Based on the visual and spectral features, the flower bud differentiation, flowering, and fruit expansion stages of the mango tree are determined by image comparison, and the nitrogen, phosphorus, potassium, and water deficiency states of the mango tree at different stages are respectively identified using a preset determination model. The specific process includes: S4.1) Based on the leaf morphology thickness index extracted in step S3.2, the flower bud number density, flower cover 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 percentages extracted each time, and the result is matched with a preset growth stage standard library; S4.2) The growth stage standard library includes the flower bud density threshold, flower coverage threshold, fruit volume percentage, fruit volume growth rate threshold and spectral reflectance ratio threshold for each period; if the current flower bud number density exceeds the flower bud density threshold of the differentiation period and the flower coverage percentage is zero, it is determined to be the flower bud differentiation period; if the current flower coverage percentage exceeds the flower coverage threshold of the flowering period and the fruit volume percentage is lower than the expansion period threshold, it is determined to be the flowering period; if the fruit volume percentage growth rate continues to exceed the fruit volume growth rate threshold of the expansion period and the spectral reflectance ratio is lower than the spectral reflectance ratio threshold, it is determined to be the fruit expansion period.

[0068] The specific judgment thresholds are: flower bud differentiation period: flower bud density ≥ 2.5 / cm² and flower coverage percentage = 0%; Flowering period: flower coverage percentage ≥ 30% and fruit volume percentage ≤ 15%; Fruit expansion period: average daily growth rate of fruit volume ≥3% And the spectral reflectance ratio is ≤1.0.

[0069] S4.3) After determining the growth stage, a nutrient determination sub-model specific to that stage is invoked. This sub-model uses a random forest algorithm to integrate near-infrared reflectance, ambient temperature and humidity, and historical fertilization data to output a probability value for nitrogen, phosphorus, potassium, or water deficiency. The specific implementation of invoking the nutrient determination sub-model specific to that stage includes the following steps: S4.31) Loading the corresponding sub-model from the pre-trained model library according to the growth stage determined in step S4.2; wherein the pre-trained model library is constructed by the following steps: S4.311) Collect historical data on mango cultivation, including canopy images at different growth stages, near-infrared reflectance, ambient temperature and humidity, historical fertilization records, green saturation distribution data, and leaf morphology and thickness indicators; S4.312) annotating the historical data, including the actual status of nitrogen, phosphorus, potassium and water deficiencies at each period; S4.313) The dataset was divided by growth stage, and sub-models for flower bud differentiation, flowering, and fruit expansion were trained separately. Each sub-model used a random forest algorithm. The input features were near-infrared reflectance, ambient temperature and humidity, historical fertilizer application, green saturation distribution data, and leaf morphology and thickness indicators. The output was the probability of nutrient deficiency. S4.314) Optimize sub-model parameters through cross-validation until the test set accuracy reaches a preset value (e.g., the preset value ≥ 85%).

[0070] The specific training and verification methods of the sub-model include: Flower bud differentiation period sub-model: The weight of the green saturation distribution data in the input features accounts for 40%, the leaf morphology and thickness index accounts for 30%, the near-infrared reflectance accounts for 20%, and the ambient temperature and humidity account for 10%; Flowering period sub-model: Among the input features, flower cover percentage accounts for 35%, near-infrared reflectance accounts for 30%, green saturation distribution data accounts for 25%, and historical fertilizer application accounts for 10%; Fruit expansion period sub-model: Among the input features, the spectral reflectance ratio accounts for 50%, the fruit volume growth rate accounts for 30%, the ambient temperature accounts for 15%, and the historical fertilizer application rate accounts for 5%; The random forest algorithm parameters for each sub-model are: 100-150 decision trees, 10-12 maximum depth, and 5 minimum leaf node samples.

[0071] S4.32) Inputting current near-infrared reflectance, ambient temperature and humidity, historical fertilization data, green saturation distribution data, and leaf morphology and thickness index into the sub-model; outputting a probability value of nitrogen, phosphorus, potassium, or water deficiency using a random forest algorithm; S4.33) When any probability value is ≥30%, the control logic of step S5 is triggered.

[0072] S5) controlling the water-fertilizer mixing device to dynamically adjust the concentration of corresponding nutrients in the drip irrigation liquid and the irrigation amount based on the determination result of the nutrient deficiency state; specifically comprising: S5.21 When the probability value of a single nutrient deficiency is ≥70%, the water-fertilizer mixing device will be controlled to increase the corresponding nutrient concentration to the preset upper limit and the drip irrigation volume will be increased to 1.2-1.5 times the standard irrigation volume; S5.22 When the probability of a single nutrient deficiency is between 30% and 70%, adjust the nutrient concentration and irrigation amount in a linear manner; S5.23 When the probability of nutrient deficiency is less than 30%, drip irrigation is performed while maintaining the basic concentration and irrigation volume; S5.24 Prioritizing the drip irrigation system for water irrigation specifically means that if both water and nutrient deficiencies are detected, irrigation is prioritized until the soil moisture content reaches the standard, followed by a delay of 24 hours to recapture canopy images for a second assessment. S5.25 If nutrient deficiency still exists after the second determination, the drip irrigation pipeline is triggered to supplement the corresponding elements, and the delay time for feedback data collection is dynamically adjusted according to the following rules: a) When nitrogen is used as the supplemental element, the delay is 48-72 hours; b) When the supplemental element is phosphorus or potassium, the delay is 5-7 days; c) When nitrogen, phosphorus, and potassium are deficient at the same time, the longest delay of 5-7 days corresponding to phosphorus or potassium shall be used as the benchmark; S5.26 After each adjustment, feedback data is collected according to the dynamically adjusted delay time until the lack probability value is less than 10%; if the lack probability value is still ≥10% after three consecutive adjustments, an abnormal alarm is triggered and automatic control is terminated.

[0073] The fill light module utilizes an LED array (such as the Osram OSLON Square series) with an adjustable light intensity range of 800-1200 lux. This light is dynamically adjusted using PWM dimming technology (10%-100% duty cycle) to ensure uniform light intensity across the canopy surface (standard deviation ≤5%) under varying ambient lighting conditions. A scattering lens (such as a polycarbonate diffuser) eliminates light spots, and its coverage range is adapted to the drone's flight altitude (3-5 meters), ensuring uniform lighting across the canopy. An integrated digital light sensor (such as the ams AS7341L multispectral light sensor) monitors ambient light intensity (visible and near-infrared bands) in real time and feeds this data back to the drone's flight control system. Basic mode: When the ambient light intensity falls below 800 lux, the fill light module activates and provides supplemental light at a preset intensity (e.g., 1000 lux). Staged mode: Dynamically adjusts the spectral ratio based on the mango's growth stage: During flower bud differentiation, the blue light ratio is increased (blue:red = 3:1) to inhibit excessive growth and promote flower bud formation; during flowering, the red and blue light ratios are balanced (1:1) to maintain photosynthetic efficiency; during fruit expansion, the red light ratio is increased (red:blue = 3:1) to promote sugar accumulation. When the multispectral imaging device is collecting data, the fill light module prioritizes maintaining a constant light intensity; during non-collection periods, the power is automatically reduced to save energy. The circuit design utilizes a totem pole switch control circuit (referenced in the "Research on the Application of Automatic Fill Lighting Systems for UAVs in Overhead Transmission Line Inspection"), controlling the LED fill light on and off via high and low level signals, with a response delay of ≤10ms.

[0074] In a Jinhuang mango planting implementation example, a circular drip irrigation zone with a radius of 0.9-1.2 times the projected canopy radius was installed in one of the Jinhuang mango planting areas within a standardized mango planting base, centered on the tree trunk. The drip irrigation pipe had a pore diameter of 1.2 mm and was buried at a depth of 10 cm, covering all 330 trees in the orchard. A Netafim NutriFlex 3.0 water and fertilizer mixing device was used, with a base irrigation volume of 6 liters per tree and base concentrations of 1.2 g / L for nitrogen, 0.9 g / L for phosphorus, and 1.5 g / L for potassium, respectively. Solenoid valves were connected to the drip irrigation pipes to dynamically adjust flow and concentration based on the results of the irrigation. The drone used was a DJI Matrice 300 RTK equipped with a MicaSense RedEdge-MX multispectral camera (bands: blue 475 nm, green 560 nm, red 668 nm, red edge 717 nm, near-infrared 840 nm) and a Velodyne VLP-16Lite lidar. The obstacle avoidance module, APAS 5.0, plans the flight path in real time, achieving a 65% lateral overlap rate and a flight altitude of 3.5 meters. On cloudy days (ambient light of 600 lux), the LED array is activated and supplemented with 1000 lux (color temperature of 5500K) to ensure clear canopy images during the flower bud differentiation period.

[0075] Flower bud differentiation period (March): UAV-collected images showed a flower bud density of 3.2 buds / cm² (threshold ≥ 2.5 buds / cm²) and a flower coverage rate of 0%, indicating the period of flower bud differentiation.

[0076] The near-infrared reflectivity of the leaves is 0.65, the reflectivity of the flower and fruit area is 0.72, the ratio is 0.90 (threshold ≤ 1.0), combined with the average daily growth rate of the fruit volume of 4% (threshold 3%), it is determined that it has entered the expansion period.

[0077] Nitrogen deficiency during flower bud differentiation: The sub-model outputs a nitrogen deficiency probability of 55%, and the nitrogen concentration is linearly increased from 1.2 g / L to 1.6 g / L (preset upper limit), and the irrigation amount is increased to 7.2 liters per plant (1.2 times the standard amount).

[0078] Potassium deficiency during the expansion period: The secondary judgment shows that the probability of potassium deficiency is 70%, and the potassium concentration is directly adjusted to 2.25g / L (1.5 times the upper limit), and the irrigation amount is increased to 9 liters / plant (1.5 times).

[0079] Field comparative trials Location: A standardized mango planting base Age: 5-year-old Golden Miscanthus sinensis, grafted seedling, 3.2-3.8 meters tall, 2.5-3.0 meters wide; Planting density: plant spacing 4 meters × row spacing 5 meters, 33 plants per mu, the experimental area is 30 mu (990 plants); Soil conditions: red soil, pH 5.8-6.2, organic matter content 1.8%, basic fertility: alkaline nitrogen 85mg / kg, available phosphorus 12mg / kg, available potassium 105mg / kg; Experimental group: 330 plants, using the intelligent fertilization system of the present invention (dynamic growth stage determination + stage-by-stage regulation in Example 1); Control group 1: 330 plants, traditional manual fertilization (fixed cycle: nitrogen fertilizer before flowering, phosphorus and potassium fertilizer after flowering, manual inspection every week), specifically: fixed once a week, fertilization time is 6:00-8:00 in the morning (avoid high temperature period); Flower bud differentiation period (March): urea (nitrogen content 46%) 0.5kg / plant, superphosphate (phosphorus content 12%) 0.3kg / plant; Flowering period (April): potassium dihydrogen phosphate (52% phosphorus, 34% potassium) dosage 0.4 kg / plant; Swelling period (May-June): potassium sulfate (potassium content 50%), dosage 0.6 kg / plant; Adjustment basis: Manual observation of leaf color (degree of yellowing) and flower bud abscission rate, combined with soil conductivity detection using a soil rapid tester (model: Hanna HI9835), with an error tolerance of ±10%.

[0080] Control group 2: 330 plants, static model traditional intelligent fertilization (equipped with Netafim NutriFlex™ Basic integrated water and fertilizer dispenser). Nitrogen, phosphorus, and potassium were dosed at 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 was fixed at 6 L / plant, controlled by a mechanical flow valve, regardless of growth stage. A soil moisture sensor (buried at a depth of 20 cm) triggered irrigation when soil moisture fell below 60% of field capacity. Fertilization decisions were based entirely on preset parameters, without canopy image acquisition or multispectral analysis. Fertilization decisions were made based entirely on preset parameters. No staged adjustments were made: the same fertilization rate was used during bud differentiation, flowering, and expansion, failing to adapt to varying nutrient requirements. There was no feedback mechanism: soil or plant status was not retested after fertilization, lacking dynamic optimization capabilities. The basis for determining the growth stage is: 3 main branches are randomly selected for each plant, and 2 agronomists independently record the flower bud density, flower coverage rate, and fruit volume growth rate. If the results are inconsistent, they will be reviewed by a third-party expert.

[0081] Calculation of misjudgment rate: If the difference between the system judgment result and the manual review result exceeds the threshold (such as the flower bud density error ≥ 0.5 / cm²), it will be recorded as a misjudgment.

[0082] Basis for determining nutrient deficiency: Collect leaf samples (the third mature leaf) every half month and send them to the laboratory for testing (Kjeldahl nitrogen determination method, molybdenum antimony colorimetry, flame photometry).

[0083] Calculation of false positive rate: If the system output probability value is ≥30% but the laboratory test shows no deficiency, or the probability value is <30% but the laboratory test shows deficiency, it will be recorded as a false positive.

[0084] Statistical analysis of single-plant yield: During the harvest period, the fruits of the entire plant are weighed and cracked and diseased fruits are removed.

[0085] Fruit quality: 100 fruits were randomly selected to measure single fruit weight, soluble solids (sugar content), and flesh thickness.

[0086] Direct costs: fertilizer, water resources, labor, and equipment depreciation (drones and water and fertilizer systems are amortized over 5 years).

[0087] Indirect costs: losses due to fertilizer supplementation and yield reduction caused by misjudgment.

[0088] Statistical test results (2023 data): The statistics of misjudgment rate of growth stage in each group are as follows: Experimental group: 6.2% (possible cause of the main error: misjudgment of the early expansion period and the late flowering period); Control group 1: 17.8% (possible cause of the main error: manual observation of flowers and fruits inside the canopy); Control group 2: 12.5% ​​(Possible cause of the major error: the static model did not identify the sudden change in fruit volume growth rate).

[0089] The statistics of nutrient deficiency misjudgment rate in each group are as follows: Experimental group: 8.7% (concentrated on spectral data distortion on rainy days); Control group 1: 23.4% (misjudgment due to soil detection lag); Control group 2: 15.6% (fixed threshold cannot adapt to temperature fluctuations).

[0090] The average yield of each group is as follows: experimental group: 36.5 kg (SD ± 2.1); Control group 1: 29.8 kg (SD ± 3.5); Control group 2: 32.4 kg (SD ± 2.8).

[0091] The statistics of fruit quality of each group are as follows: Experimental group: single fruit weight 495g, sugar content 14.8%, flesh thickness 18.2mm; Control group 1: single fruit weight 405 g, sugar content 13.3%, flesh thickness 15.6 mm; Control group 2: single fruit weight 435 grams, sugar content 14.0%, flesh thickness 16.9 mm.

[0092] The cost-effectiveness comparison statistics of each group are as follows: Experimental group: nitrogen fertilizer reduced by 19%, phosphorus fertilizer reduced by 16%, potassium fertilizer reduced by 13%, and water saved by 25%; Control group 1: excessive fertilization was applied at a fixed cycle, with the amount of nitrogen, phosphorus and potassium being 1.4 times that of the experimental group; Control group 2: Due to misjudgment and frequent fertilization, the amount of nitrogen, phosphorus and potassium used was 1.2 times that of the experimental group.

[0093] The comprehensive net profit statistics are as follows: Market price of Golden Mango: wholesale price at the place of origin is 8 yuan / kg (general fruit price, including sugar content grade premium).

[0094] Fertilizer cost: nitrogen fertilizer 4.5 yuan / kg, phosphorus fertilizer 5.2 yuan / kg, potassium fertilizer 5.8 yuan / kg.

[0095] Water resource cost: irrigation water fee is 0.8 yuan / ton.

[0096] Labor cost: Orchard workers’ daily wage is 150 yuan / person.

[0097] Equipment depreciation: UAVs 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.

[0098] The comprehensive net profit (average per mu) of the experimental group increased by 2051 yuan compared with the control group 1 and increased by 1238 yuan compared with the control group 2; Analysis of the differences between the present invention and the control group revealed increased yield, resource conservation, and reduced labor costs. This demonstrates that the present invention, through its design mechanism combining drone monitoring and dynamic models with closed-loop control, addresses the challenges of low monitoring efficiency, high error rates, and resource waste associated with traditional fertilization methods. The experimental results demonstrate its feasibility in large-scale orchards, making it particularly suitable for high-value-added crops (such as golden mangoes) that require refined nutrient management to improve quality. Furthermore, fully automated control significantly reduces labor reliance in labor-scarce regions, and its water- and fertilizer-saving properties are well-suited for arid or soil-poor environments in resource-constrained areas, making it highly marketable.

[0099] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiment. They can be applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily realized.

Claims

1. Intelligent mango fertilization method, characterized in that, The following steps are involved: S1) Arrange drip irrigation pipes connected to the water-fertilizer mixing device in the drip irrigation area; S2) a multispectral imaging device periodically collects canopy images of mango trees, wherein the canopy images include visible and near-infrared band information of leaves, flower buds, flowers, and fruits; S3) extracting visual and spectral features of the canopy image based on an image processing algorithm, wherein the visual features include leaf color distribution, morphological thickness index, flower bud density, flower coverage, and fruit volume ratio; and the spectral features include the near-infrared reflectance ratio between the leaf region and the flower and fruit region; S4) determining the flower bud differentiation period, flowering period, and fruit expansion period of the mango tree by image comparison based on the visual and spectral features, and then identifying the nitrogen deficiency, phosphorus deficiency, potassium deficiency, and water deficiency status of the mango tree at different periods using a preset judgment model; S5) According to the determination result of the nutrient or water deficiency state, the water-fertilizer mixing device is controlled to dynamically adjust the concentration of corresponding nutrients in the drip irrigation liquid and the irrigation amount.

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 canopy images of the mango orchard according to a preset flight path, and the preset lighting conditions are maintained through the drone's built-in fill light module during image collection.

3. The method according to claim 1, characterized in that The specific extraction method of the visual features and spectral features in step S3 includes: S3.1) Perform semantic segmentation on the canopy image to separate leaves, buds, flowers, and fruit regions; S3.2) Extract green saturation distribution data in the HSV color space from the leaf region, and calculate the standard deviation of the point cloud density in the vein region using 3D point cloud reconstruction technology as an indicator of leaf morphological thickness; S3.3) Count the number of flower buds per unit area in the flower bud area, calculate the percentage of coverage area to total canopy area in the flower area, and calculate the percentage of volume to total canopy volume in the fruit area; S3.4) For the near-infrared band image, calculate the ratio of the average reflectance of the leaf area to the average reflectance of the flower and fruit area as the spectral reflectance ratio.

4. The method according to claim 1, wherein The specific process of image comparison and growth stage determination in step S4 includes: S4.1) Calculating the fruit volume growth rate based on the leaf thickness index, flower bud number density, flower cover percentage, fruit volume percentage, and spectral reflectance ratio by accumulating the fruit volume percentages extracted each time, and matching the calculated fruit volume growth rate with a pre-set growth stage standard library; S4.2) The growth stage standard library includes flower bud density thresholds, flower coverage thresholds, fruit volume percentages, fruit volume growth rate thresholds, and spectral reflectance ratio thresholds for each period; if the current flower bud density exceeds the flower bud density threshold for the differentiation period and the flower coverage percentage is zero, the stage is determined to be the flower bud differentiation period; if the current flower coverage percentage exceeds the flower coverage threshold for the flowering period and the fruit volume percentage is lower than the expansion period threshold, the stage is determined to be the flowering period; if the fruit volume percentage growth rate continues to exceed the fruit volume growth rate threshold for the expansion period and the spectral reflectance ratio is lower than the spectral reflectance ratio threshold, the stage is determined to be the fruit expansion period; S4.3) After determining the growth stage, a nutrient determination sub-model dedicated to that stage is called. The sub-model integrates near-infrared reflectance, ambient temperature and humidity, and historical fertilization data through a random forest algorithm to output a probability value for nitrogen, phosphorus, potassium, or water deficiency.

5. The method according to claim 1, wherein The S5 step also includes the following priorities: S5.1) When both nutrient and water deficiency are detected, the drip irrigation system is activated first for watering, and image data is collected again after a set delay. S5.2) If nutrient deficiency still exists after recollection, trigger the drip irrigation line to supplement the corresponding elements.

6. The method according to claim 5, characterized in that The specific implementation method of executing the trigger drip irrigation pipeline to supplement the corresponding elements includes: S5.21 When the probability value of a single nutrient deficiency is ≥70%, the water-fertilizer mixing device will be controlled to increase the corresponding nutrient concentration to the preset upper limit and the drip irrigation volume will be increased to 1.2-1.5 times the standard irrigation volume; S5.22 When the probability of single nutrient deficiency is 30%-70%, adjust the nutrient concentration and irrigation amount in a linear proportion; S5.23 When the probability of nutrient deficiency is less than 30%, drip irrigation is performed while maintaining the basic concentration and irrigation volume; S5.24 Prioritizing the drip irrigation system for water irrigation specifically includes: if both water and nutrient deficiencies are detected, irrigation is prioritized until the soil moisture content reaches the standard, followed by a delay of the first set time to recapture canopy images for a second determination; S5.25 If nutrient deficiency still exists after the second determination, the drip irrigation pipeline is triggered to supplement the corresponding elements, and the delay time for feedback data collection is dynamically adjusted according to the following rules: a) When nitrogen is used as the supplemental element, the delay is 48-72 hours; b) When the supplemental element is phosphorus or potassium, the delay is 5-7 days; c) When nitrogen, phosphorus, and potassium are deficient at the same time, the longest delay of 5-7 days corresponding to phosphorus or potassium shall be used as the benchmark; S5.26 After each adjustment, feedback data is collected according to the dynamically adjusted delay time until the lack probability value is less than 10%; if the lack probability value is still ≥10% after three consecutive adjustments, an abnormal alarm is triggered and automatic control is terminated.

7. The method according to claim 5, characterized in that The specific implementation method of "calculating the point cloud density standard deviation of the vein area as the morphological thickness indicator by using the three-dimensional point cloud reconstruction technology" in step S3.2 includes: S3.21) Perform hierarchical sampling on the point cloud data of the leaf area, dividing the leaf into several sub-areas. Perform high-precision 3D reconstruction on the mesh containing the vein centerline, and use linear interpolation to generate a low-resolution point cloud for the non-vein mesh. S3.2.2) Locally align the high-precision vein point cloud with the interpolated non-vein point cloud using the improved ICP algorithm with the vein centerline as the reference. Limit the number of iterations and search radius. Process the alignment data in parallel using half-precision floating-point operations on the UAV's onboard GPU, and output the point cloud density standard deviation.

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

9. The method according to claim 8, characterized in that The pre-trained model library is constructed by the following steps: S4.311) Collect historical data on mango cultivation, including canopy images at different growth stages, near-infrared reflectance, ambient temperature and humidity, historical fertilization records, green saturation distribution data, and leaf morphology and thickness indicators; S4.312) annotating the historical data, including the actual status of nitrogen, phosphorus, potassium and water deficiencies at each period; S4.313) The dataset was divided by growth stage, and sub-models for flower bud differentiation, flowering, and fruit expansion were trained separately. Each sub-model used a random forest algorithm. The input features were near-infrared reflectance, ambient temperature and humidity, historical fertilizer application, green saturation distribution data, and leaf morphology and thickness indicators. The output was the probability of nutrient deficiency. S4.314) Optimize sub-model parameters through cross-validation until the test set accuracy reaches the preset value.

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

Images