A micro-spraying water and fertilizer collaborative precision control method for late-sowing wheat

Abstract

The present application belongs to the field of automatic control technology, and specifically relates to a kind of micro-spraying water and fertilizer collaborative precision control method for late-sowing wheat. It includes the following steps: step 1, arrange dielectric spectrum array probe, ground temperature chain, solar-induced chlorophyll fluorescence narrowband spectrometer and pole-mounted color focal plane polarized camera on the target plot, and gather multi-source seedling and soil moisture condition joint dataset through field gateway time alignment; step 2, send the multi-source seedling and soil moisture condition joint dataset to the vehicle-mounted heterogeneous edge reasoning unit composed of Jetson Orin and Hailo-8L interconnection; step 3, Jetson Orin runs rolling optimization scheduling on the per-area three-branch decision result set according to the preset agronomic priority, and performs closed-loop correction on the nitrogen lower threshold, phosphorus lower threshold, water lower threshold and corresponding window duration according to the per-channel correction rule, and generates water and fertilizer collaborative execution timing of the next rolling time domain after writing back. The present application can achieve the stable yield control objectives of increasing phosphorus and nitrogen, stabilizing seedlings with less water, and moving spring irrigation later.

Description

Technical Field

[0001] This invention belongs to the field of automatic control technology, and particularly relates to the field of automatic control using computer vision technology. Specifically, it is a method for precise control of micro-spraying water and fertilizer synergy for late-sown wheat. Background Technology

[0002] Late-sown wheat refers to wheat varieties in the Huang-Huai-Hai Plain and northern winter wheat regions that are forced to be sown later due to factors such as late clearing of the previous crop, continuous rain, or autumn floods. These wheat varieties generally suffer from inherent disadvantages such as delayed emergence before winter, insufficient accumulated temperature, fewer tillers per plant, and shallow root systems. Upon entering the spring greening stage, they exhibit a higher proportion of weak seedlings and a slower initial growth of the plant population, directly threatening the number of ears and yield. Existing technologies for spring water and fertilizer management of late-sown wheat are developing along three paths.

[0003] The first approach is irrigation decision-making based on ground-based soil moisture monitoring stations. Using frequency-domain or time-domain reflectometry soil moisture sensors in conjunction with a weather station network to trigger irrigation commands based on preset soil moisture thresholds is already quite common in water-saving irrigation in field fields. However, single-frequency reflectometry probes are affected by the aliasing of dielectric responses between unfrozen water and free water during the freeze-thaw transition period. This results in a significant discrepancy between the inverted water content and the actual water available for root use, easily leading to indiscriminate flooding in early spring due to low water content readings, which can actually cause frost damage and yellowing seedlings.

[0004] The second approach is based on variable nitrogen fertilizer application using multispectral UAV or satellite remote sensing. Canopy nitrogen status is retrieved using normalized difference vegetation index (NDVI) and red-edge chlorophyll index (RBCA), generating prescription maps to drive variable spraying operations. This method achieves stable retrieval accuracy during high-coverage growth periods. However, in late-sown wheat during the greening-up stage, when canopy cover is generally less than 30% and surface residue and bare soil are mixed into pixels, vegetation indices severely degrade, and weak seedling zones and bare soil patches are easily misidentified. Furthermore, optical remote sensing of cloud and rain effects has a limited window for obtaining continuous observations in spring.

[0005] The third approach is quantitative replenishment based on integrated water and fertilizer equipment, typically involving Venturi fertilization at a preset ratio combined with drip irrigation or micro-sprinkler main pipelines. While such systems are relatively mature at the execution level, their upstream decision-making is often based on human experience or simple thresholds, lacking a precise identification of spatial unevenness in seedling conditions. Furthermore, they fail to incorporate soil temperature, the greening process, and the state of unfrozen water into triggering constraints. Consequently, the timing and dosage of water and fertilizer replenishment are difficult to precisely match the agronomical requirements of late-sown wheat: "increased phosphorus and nitrogen, less water for stable seedling growth, and delayed spring irrigation."

[0006] In summary, existing technologies have not yet provided a synergistic control method that can accurately identify the spatial distribution of tillers under low canopy coverage, reliably distinguish between moisture content and unfrozen water content during freeze-thaw cycles, and jointly reflect seedling condition, soil moisture, soil temperature, and greening process in the timing of water and fertilizer application. Significant technological gaps remain in the stable yield management of late-sown wheat during the greening period. Summary of the Invention

[0007] The main objective of this invention is to provide a micro-spraying water and fertilizer synergistic precision control method for late-sown wheat. This method can accurately identify the spatial distribution of tillers under low canopy coverage during the greening period, reliably distinguish between water content and unfrozen water status during the freeze-thaw period, and jointly reflect seedling condition, soil moisture, soil temperature, and greening process in the water and fertilizer execution sequence. It overcomes the defects of traditional multispectral vegetation index degradation failure and single-frequency reflectance probe freeze-thaw period inversion bias, and achieves the stable yield control objectives of increasing phosphorus and nitrogen, stabilizing seedlings with less water, and delaying spring irrigation.

[0008] To address the aforementioned technical problems, this invention provides a method for precise control of synergistic water and fertilizer application via micro-spraying for late-sown wheat, comprising the following steps: Step 1: Deploy dielectric spectrum array probes, ground temperature chains, solar-induced chlorophyll fluorescence narrowband spectrometers, and rod-mounted color split-focus plane polarization cameras on the target plot. Output complex reflection coefficient, ground temperature value, photosynthetic potential index value, and polarization channel color images, respectively. These are then time-aligned and aggregated into a multi-source crop and soil moisture joint dataset through a field gateway. Step 2: The multi-source seedling and soil moisture joint dataset is fed into a vehicle-mounted heterogeneous edge inference unit consisting of Jetson Orin and Hailo-8L interconnected. Jetson Orin extracts the Thiessen polygon set and leaf greenness index and suspected phosphorus deficiency index of each Thiessen polygon from the polarization channel color image. Hailo-8L obtains the surface water state value, root water state value and unfrozen water content value from the complex reflectance coefficient. Jetson Orin obtains the greening stage identifier based on the soil temperature value and photosynthetic potential index value, and performs spatial mapping of the dielectric inversion results according to the geometric centroid of the Thiessen polygon, assigning a region class identifier to each Thiessen polygon to form a seedling condition classification mapping. Hailo-8L and Jetson Orin run three-branch decision inference on the seedling condition classification mapping region by region, and outputs a region-by-region three-branch decision result set consisting of three states of nitrogen supplementation decision, three states of phosphorus supplementation decision, and three states of micro-spraying decision. Step 3: Jetson Orin performs rolling optimization scheduling on the three decision result sets of each zone according to the preset agronomic priority, generating a water and fertilizer co-execution sequence including phosphate fertilizer injection window, nitrogen fertilizer injection window and water spraying window. This sequence is then sent to the microfluidic proportional fertilizer mixing module and piezoelectric ceramic micro-sprayer array for zoned spraying via the field gateway. After each window ends, images are re-acquired to obtain the tiller density increment and leaf color recovery value. The nitrogen lower threshold, phosphorus lower threshold, water lower threshold and corresponding window duration are closed-loop corrected according to the channel correction rules. After writing back, the water and fertilizer co-execution sequence for the next rolling time domain is generated.

[0009] Furthermore, in step 1, four sets of dielectric spectrum array probes and four geothermal chains are deployed along two mutually perpendicular sampling baselines. Two sets of dielectric spectrum array probes and two geothermal chains are deployed at equal intervals along each sampling baseline. The planar coordinates of the four sets of dielectric spectrum array probes are associated with the valve-controlled zone coordinate system. A solar-induced chlorophyll fluorescence narrowband spectrometer is deployed in the center of the plot, and a pole-mounted color-focused plane polarization camera is deployed 2m above the plot. The dielectric spectrum array probes have one set of coaxial sensitive electrodes built into each of the three depths of 5cm, 20cm, and 40cm. The frequency is cyclically swept at five frequencies: 1MHz, 10MHz, 100MHz, 500MHz, and 1GHz. The real and imaginary parts of the complex reflection coefficient are output synchronously at each frequency. The geothermal chain has one platinum resistance temperature sensor at each of the four depths of 2cm, 5cm, 10cm, and 20cm, and outputs geothermal values ​​at 10-minute intervals. The solar-induced chlorophyll fluorescence narrowband spectrometer outputs canopy fluorescence radiance at 30-minute intervals with a narrowband channel centered at 760nm and a half-bandwidth of 0.3nm. After fitting and subtracting the baseline from the continuous baseline segments on both sides of the 760nm wavelength, the peak value of fluorescence radiance during the effective illumination period of the day is taken as the photosynthetic potential index value. The pole-mounted color split-focus plane polarization camera is a split-focus plane camera that synchronously outputs color images with four polarization channels (0°, 45°, 90°, and 135°) with red, green, and blue color components in a single frame, and synchronously triggers shooting at 30-minute intervals.

[0010] Furthermore, the Jetson Orin contains a GPU computing array and an ARM Cortex main control processor cluster. The Hailo-8L, as a neural network accelerator, is connected to the Jetson Orin via the PCIe bus. The GPU computing array is responsible for image spatial channel processing and spatial geometry calculations, the ARM Cortex main control processor cluster is responsible for complex number transformation, sliding statistics, threshold rule determination and scheduling control, and the Hailo-8L is responsible for the quantized deployment of neural network tensor inference tasks. All three are executed in parallel. Before acquisition, the pole-mounted color split-focus plane polarization camera undergoes intrinsic parameter calibration, extrinsic parameter calibration and polarization channel gain correction. At least four ground control points are set in the target plot coordinate system. The ARM Cortex main control processor cluster solves the homography transformation matrix between the polarization image pixel coordinates and the plot's two-dimensional coordinates based on the ground control points and stores it in shared memory.

[0011] Furthermore, the GPU computing array extracts the luminance component from the four polarization channel color images to obtain four polarization intensity maps, and then obtains three intermediate channels pixel by pixel: the first intermediate channel is obtained by summing the 0-degree polarization intensity map and the 90-degree polarization intensity map pixel by pixel; the second intermediate channel is obtained by subtracting the 90-degree polarization intensity map from the 0-degree polarization intensity map; and the third intermediate channel is obtained by subtracting the 135-degree polarization intensity map from the 45-degree polarization intensity map. The GPU computing array calculates the sum of squares and takes the square root of the second and third intermediate channels pixel by pixel, and then divides the result of the sum of squares and square root by the first intermediate channel pixel by pixel to obtain a linear polarization degree map. The GPU computing array uses the Otsu adaptive thresholding method to obtain a segmentation threshold for the linear polarization degree map, and marks the pixels with linear polarization degree values ​​greater than the segmentation threshold as canopy pixels to obtain a canopy binary mask.

[0012] Furthermore, the GPU computing array performs a pixel-by-pixel AND operation on the canopy binary mask and the first intermediate channel to obtain a canopy intensity map. Using a preset maximum height threshold, a morphological H-maximum transformation is applied to the canopy intensity map to obtain a set of tiller vertices represented in pixel coordinates. Using this set of tiller vertices as seeds, a gradient-based watershed segmentation algorithm is applied to the canopy intensity map to obtain a set of individual tillers corresponding one-to-one with each tiller vertex. The ARM Cortex main control processor cluster uses a homography transformation matrix to convert the set of tiller vertices from pixel coordinates to plot 2D coordinates. The GPU computing array then applies the Fortune scanline algorithm to the converted set of tiller vertices to obtain a set of Thiessen polygons, where each Thiessen polygon corresponds one-to-one with an individual tiller. The GPU computing array clips each Thiessen polygon according to the plot boundary and the boundary of its respective control zone, and the arithmetic mean of the areas of all clipped Thiessen polygons is obtained. ARM The Cortex main control processor cluster determines the valve control partition to which each Thiessen polygon belongs based on the principle of geometric centroid falling.

[0013] Furthermore, within each Thiessen polygon-covered region of the original color image from the polarization camera, the GPU computing array takes the median of the red, green, and blue color components to form the red median, green median, and blue median, respectively, and takes the median of the canopy light intensity map to form the strong median. The red, green, and blue medians, along with the strong median, form the leaf color quadruple corresponding to the Thiessen polygon. The ARM Cortex main control processor cluster divides the green median in the leaf color quadruple by the red median to obtain the leaf greenness index, divides the blue median in the leaf color quadruple by the red median to obtain the median ratio, and then divides the median ratio by the strong median to complete the intensity normalization, thus obtaining the phosphorus deficiency index. The larger the value of the phosphorus deficiency index, the higher the probability of phosphorus deficiency.

[0014] Furthermore, the Hailo-8L is pre-quantized with a complex dielectric state inversion network. This network is a fully connected neural network with a 30-dimensional dielectric input vector as input and three scalar outputs: water content, unfrozen water content, and conductivity. Before leaving the factory, this fully connected neural network is trained on sample pairs generated by the Cole-Cole relaxation model and quantized into tensor operators executable by the Hailo-8L. The Cortex main control processor cluster converts the 5-frequency complex reflection coefficients uploaded by the dielectric spectrum array probes into measured values ​​of the 5-frequency complex dielectric constant for each probe using a transmission line model with known parameters such as the characteristic impedance and electrode length of the coaxial sensitive electrode. Then, the measured values ​​of the 5-frequency complex dielectric constants of each group of dielectric spectrum array probes at three depths of 5cm, 20cm, and 40cm are split into real and imaginary parts, respectively, and concatenated into a 30-dimensional dielectric input vector according to depth and frequency point order, which is then fed into the complex dielectric state inversion network. The Hailo-8L outputs the water content, unfrozen water content, and conductivity of each group of probes according to depth. The output water content at a depth of 5cm is taken as the surface water content value of the corresponding probe, and the output water content at depths of 20cm and 40cm are fused according to preset depth weights to obtain the root water content value of the corresponding probe. The minimum value of the unfrozen water content at the three depths is taken as the unfrozen water content value of the corresponding probe.

[0015] Furthermore, the ARM Cortex main control processor cluster takes the most recent five 24-hour moving averages of the geothermal temperature for each of the four depth channels to obtain four geothermal moving average sequences. For the photosynthetic potential index, it takes the most recent three-day peak sequences and calculates the difference between adjacent days to obtain the daily increment sequence of photosynthetic potential. The following rules are used to mark the greening-up stage: when the 5cm geothermal moving average sequence is below 3℃, it is considered greening-up preparation; when the 5cm geothermal moving average sequence is above 3℃ for three consecutive 24-hour periods and the daily increment sequence of photosynthetic potential is positive for two consecutive days, it is considered greening-up initiation; when both the 5cm and 10cm geothermal moving average sequences are above 5℃ for three consecutive 24-hour periods... The above conditions are met when the daily increase in photosynthetic potential is positive for two consecutive days, and when the 20cm ground temperature sliding mean is above 5℃ for three consecutive days, the greening process is considered complete. The GPU computing array uses the planar coordinates of four dielectric spectrum array probes as known interpolation nodes and the geometric centroid of each Thiessen polygon as the interpolation point. Using an inverse distance-weighted interpolation method with a distance weighting exponent of 2, the surface water state, root water state, and unfrozen water content values ​​output by the Hailo-8L are interpolated to obtain the corresponding surface water state, root water state, and unfrozen water content of the Thiessen polygon. ARM Cortex main control processor clusters are assigned zone classification labels according to the following rules: areas with an area greater than 1.5 times the average area are weak seedling areas, areas with an area less than 0.5 times the average area are overly dense areas, and areas with an area between 0.5 and 1.5 times the average area are suitable seedling areas. The zone classification label, surface water content, root water content, unfrozen water content, greening stage label, leaf greenness index, suspected phosphorus deficiency index, the valve control zone, and the corresponding Thiessen polygon area are used to form a seedling condition classification mapping.

[0016] Furthermore, the ARM Cortex main control processor cluster outputs three states for nitrogen supplementation, phosphorus supplementation, and micro-spraying decisions for each Thiessen polygon in the seedling condition grading mapping according to three sets of independent dual-threshold rules: Nitrogen supplementation decision uses leaf greenness as evidence value; if it is below the lower nitrogen threshold, nitrogen supplementation is immediately executed; if it is between the lower and upper nitrogen thresholds, nitrogen supplementation is temporarily suspended; if it is above the upper nitrogen threshold, nitrogen supplementation is delayed. Phosphorus supplementation decision uses a suspected phosphorus deficiency index as evidence value; if it is above the upper phosphorus threshold, phosphorus supplementation is immediately executed; if it is between the lower and upper phosphorus thresholds, phosphorus supplementation is temporarily suspended; if it is below the lower phosphorus threshold, phosphorus supplementation is delayed. Micro-spraying decision uses root layer water content as evidence value; if it is below the lower water content threshold, micro-spraying is immediately initiated; if it is above the lower water content threshold, micro-spraying is initiated immediately. When the lower threshold is between the upper threshold and the lower threshold of moisture content, micro-spraying is temporarily suspended for observation; when the upper threshold of moisture content is above the lower threshold of moisture content, micro-spraying is delayed for processing. Additional safety constraints: When the greening stage is marked as greening preparation or the unfrozen water content is lower than the set unfrozen water safety threshold, the corresponding three states of micro-spraying decision for the Thiessen polygon are forcibly set to micro-spraying delayed for processing. Additional zone correction: When the zone is marked as a weak seedling zone, the lower threshold of nitrogen is increased by 1 preset step and the upper threshold of phosphorus is decreased by 1 preset step. When the zone is marked as an overly dense zone, the corresponding three states of nitrogen supplementation decision for the Thiessen polygon are at most set to nitrogen supplementation temporarily suspended for observation. When the zone is marked as a suitable seedling zone, the original dual-threshold judgment result is maintained. The Hailo-8L runs the quantized deployment of the three-branch decision inference operator to perform tensor-level integration of the above rule results and outputs the entire set of three-branch decision results for each zone of the Thiessen polygon.

[0017] This invention provides a precise micro-spraying fertigation method for late-sown wheat, which has the following advantages: It utilizes a rod-mounted color split-focus plane polarization camera to distinguish living leaves from surface stubble and bare soil based on the polarization response differences caused by the waxy cuticle. Even under low cover conditions during the greening stage, it can stably extract tiller vertices and construct Thiessen polygons, achieving precise identification of spatial heterogeneity of seedling conditions and overcoming the defect of traditional multispectral vegetation indices degrading and failing in low-coverage scenarios. Furthermore, it utilizes an array dielectric spectrum probe to simultaneously invert water content, unfrozen water content, and conductivity based on the Cole-Cole relaxation model under multi-frequency sweep conditions, enabling the accurate determination of real water content that can be utilized by the root system during the freeze-thaw transition period. The water used is clearly separated from the frozen water, avoiding the risk of frost damage caused by blind irrigation triggered by low water content values ​​in spring due to the reflection probe based on a single frequency point. The greening stage is determined by combining the photosynthetic potential index value of solar-induced chlorophyll fluorescence with the sliding mean sequence of soil temperature stratification. The greening stage identifier and the safety threshold of unfrozen water are used as safety constraints for micro-spraying, so that the timing of water and fertilizer supplementation strictly matches the physiological activation process of organisms. Based on three-way decision, the three channels of nitrogen supplementation, phosphorus supplementation and micro-spraying are independently determined to be immediately executed, temporarily observed and delayed, and the treatment is postponed. The dual correction of seedling area classification is added to realize the differentiated intervention of actively increasing phosphorus and nitrogen in weak seedling areas and inhibiting tillering in overly dense areas. The vehicle-mounted heterogeneous edge inference unit, consisting of Jetson Orin and Hailo-8L interconnected via PCIe bus, enables parallel and efficient execution of image spatial channel processing, tensor inference, and scheduling control. Combined with a rolling mechanism that re-acquires images after each window ends and performs closed-loop corrections on various thresholds and window durations, the entire method can continuously and adaptively operate in spring, thereby achieving the stable yield control goals of increasing phosphorus and nitrogen supplementation, stabilizing seedlings with less water, and delaying spring irrigation. Attached Figure Description

[0018] Figure 1 A top-view diagram illustrating the deployment of multi-source sensing front-ends in the field for the micro-spraying water and fertilizer synergistic precision control method for late-sown wheat provided in this embodiment of the invention. Figure 2 A schematic diagram of the response curve of linear polarization degree to incident angle during the polarization canopy extraction process provided in this embodiment of the invention; Figure 3 This is a partial schematic diagram of the results of polarization canopy tillering monomer segmentation and Thiessen mosaic provided in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the inversion principle of the complex dielectric spectrum at five sampling frequency points under the Cole-Cole relaxation model provided in an embodiment of the present invention. Detailed Implementation

[0019] The method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0020] The target plot in this embodiment is a late-sown wheat plot measuring 60 meters east-west and 40 meters north-south. A two-dimensional plot coordinate system is established at the southwest corner of the plot, with the east direction as the positive X-axis and the north direction as the positive Y-axis. Two mutually perpendicular sampling baselines are drawn along Y=20 meters and X=30 meters, intersecting at the geometric center of the plot. One set of dielectric spectrum array probes and one geothermal chain are deployed at Y=10 meters and Y=30 meters on the north-south sampling baseline, and one set of dielectric spectrum array probes and one geothermal chain are deployed at X=15 meters and X=45 meters on the east-west sampling baseline, for a total of four sets of dielectric spectrum array probes and four geothermal chains. The cross-shaped layout can take into account the spatial representation of both north-south and east-west gradient directions at four sampling locations, and has a higher probability of capturing patchy weak seedling zones than the linear layout. The planar coordinates of the four dielectric spectrum array probes are recorded as (30, 10), (30, 30), (15, 20), and (45, 20), respectively, in meters. These four coordinates are linked by the field gateway and the valve control zone coordinate system, so that it is possible to directly query which valve control zone each dielectric spectrum array probe falls into.

[0021] Each set of dielectric spectrum array probes is inserted along a steel coaxial rod perpendicularly into the soil, with one set of coaxial sensitive electrodes embedded at depths of 5cm, 20cm, and 40cm. The inner conductor of each coaxial sensitive electrode is typically 3cm long, with an outer diameter of 2.0mm and an inner diameter of 6.4mm. The dielectric material is low-loss polytetrafluoroethylene (PTFE). Each set of coaxial sensitive electrodes is cyclically swept across five frequencies: 1MHz, 10MHz, 100MHz, 500MHz, and 1GHz. At each frequency, the real and imaginary parts of the complex reflection coefficient are output synchronously. Using these five discrete frequency points spanning three orders of magnitude allows for effective sampling of the inflection points of the Cole-Cole relaxation curve and significantly reduces the amount of data collected. A five-frequency sweep cycle is completed every 3 minutes.

[0022] Each geothermal chain is a slender stainless steel tube buried in the soil, with one Class A platinum resistance temperature sensor embedded at four depths: 2cm, 5cm, 10cm, and 20cm. The 2cm depth is used to capture the rapid response of nighttime radiative cooling and daytime radiative warming; the 5cm depth corresponds to the surface freeze-thaw transition; the 10cm depth is close to the location of the tillering node of wheat before winter; and the 20cm depth corresponds to the root activation threshold. Each platinum resistance temperature sensor reports the geothermal value at 10-minute intervals, with an accuracy better than ±0.1℃.

[0023] A solar-induced chlorophyll fluorescence narrowband spectrometer was deployed at the geometric center of the plot, outputting canopy fluorescence radiance at 30-minute intervals through a narrowband channel centered at 760 nm with a half-width of 0.3 nm. The 760 nm wavelength was chosen to utilize the Fraunhofer dark line formed by the atmospheric oxygen absorption band to enhance the resolvability of the fluorescence signal. The photosynthetic potential index was calculated as follows: two continuous baseline segments were taken on either side of this wavelength, from 758 nm to 759 nm and from 761 nm to 762 nm, respectively. Linear interpolation was performed to obtain the reflected baseline radiance at 760 nm. The reflected baseline radiance was subtracted from the measured radiance in the 760 nm narrowband channel to obtain the instantaneous fluorescence radiance at the current moment. The maximum value of all instantaneous fluorescence radiances within the effective illumination period of the day was taken as the photosynthetic potential index for that day. This method is robust to downward irradiance fluctuations caused by cloud transients.

[0024] A pole-mounted color split-focus plane polarization camera is suspended 2 meters above the site, with its optical axis pointing vertically downwards. Its field of view covers a canopy area of ​​approximately 4 meters x 3 meters at the geometric center of the site. The camera integrates a micro-polarization array with four polarization directions (0°, 45°, 90°, and 135°) at the pixel level. Each polarization direction outputs red, green, and blue color components using a Bayer arrangement, simultaneously acquiring color images from all four polarization channels in a single exposure. Shooting is triggered synchronously at 30-minute intervals. The split-focus plane polarization camera ensures strict synchronization of the four polarization channels, preventing motion blur caused by wind-blown canopy movement from contaminating the accuracy of Stokes parameter calculations.

[0025] After the pole-mounted color split-plane polarization camera is initially installed, the following calibration process is performed: Intrinsic parameter calibration of the camera is performed using a calibration board to obtain the focal length, principal point coordinates, and distortion coefficients; the camera's pose relative to the ground coordinate system is measured using a total station or differential real-time dynamic positioning receiver to obtain extrinsic parameter calibration results; uniform light field images are acquired for each of the four polarization channels using an integrating sphere uniform light source, and the correction factor for the gain of each of the four polarization channels is calculated pixel-by-pixel to ensure consistent response of the four polarization channels under a uniform light field. At least four ground control points are set up within the target ground coordinate system, typically 15cm diameter red and black alternating targets are placed at the four corners and the geometric center of the ground. The coordinates of all targets are determined using a differential real-time dynamic positioning receiver. Before each image capture, the ARM Cortex main control processor cluster automatically identifies four or more ground control points in the image, obtains the pixel coordinates of the ground control points in the image, combines the coordinates of the ground control points in the plot coordinate system, and solves the homography transformation matrix between the polarization image pixel coordinates and the plot two-dimensional coordinates using least squares. Homography transformation matrix It is a 3×3 matrix used to store homogeneous pixel coordinates. Mapped to homogeneous block coordinates The obtained homography transformation matrix The shared memory stored in Jetson Orin is repeatedly accessed in subsequent steps. This transformation can absorb the geometric deviations caused by camera tilt, lens distortion residuals, and slight pole shaking in one go, so that the area measurement of subsequent tiller vertices and Thiessen polygons is performed in meters and will not drift due to changes in camera attitude.

[0026] The outputs from the dielectric spectrum array probe, the soil temperature chain, the solar-induced chlorophyll fluorescence narrowband spectrometer, and the pole-mounted color-focused plane polarization camera are converged via a field gateway. The field gateway uses industrial Ethernet switching supporting Precision Time Protocol (PTP) to timestamp the outputs of the four sensors at sub-millisecond timing accuracy. These timestamps are then aligned and assembled in 30-minute windows to obtain a multi-source crop and soil moisture joint dataset. Since soil temperature and dielectric sampling frequencies are higher than those of polarization imagery and fluorescence, the gateway uses the nearest available soil temperature and dielectric data points for assembly, ensuring that each frame of the multi-source crop and soil moisture joint dataset contains both complete image and fluorescence data, as well as synchronized soil temperature and dielectric data.

[0027] The multi-source crop and soil moisture joint dataset was then fed into the vehicle-mounted heterogeneous edge inference unit. The vehicle-mounted heterogeneous edge inference unit was installed in the equipment compartment of the field inspection vehicle, using a Jetson Orin AGX 64GB module as the motherboard. The GPU computing array within the Jetson Orin handled image spatial channel processing and spatial geometry calculations, while the ARM Cortex main control processor cluster handled complex number transformations, sliding statistics, threshold rule determination, and scheduling control. The Hailo-8L, connected to the Jetson Orin via a PCIe Gen3 x4 bus in M.2 2280 form factor, served as a neural network accelerator for quantized neural network tensor inference tasks. The GPU computing array, ARM Cortex main control processor cluster, and Hailo-8L executed in parallel, with ample computing power margin within a 30-minute sampling interval.

[0028] For the four polarization channels of color images output by the pole-mounted color split-plane polarization camera, the GPU computing array first extracts the luminance component of each polarization channel color image to obtain four polarization intensity maps, denoted as follows: , , , The luminance component is obtained by weighting according to ITU-R BT.601. Then, three intermediate channels are calculated pixel by pixel: the first intermediate channel is obtained by summing the 0-degree polarization intensity map and the 90-degree polarization intensity map pixel by pixel, denoted as... The total light intensity of the pixel is represented by ; the second intermediate channel is calculated by subtracting the 90-degree polarization intensity map from the 0-degree polarization intensity map, denoted as . The third intermediate channel is calculated by subtracting the 135-degree polarization intensity map from the 45-degree polarization intensity map, denoted as... . , , The first three components of the Stokes vector can be used as intermediate values ​​for subsequent polarization analysis. Then, the linear polarization degree map is calculated pixel-by-pixel, and the linear polarization degree value at each pixel position is defined as: ;in This is the linear polarization degree value at this pixel, ranging from 0 to 1. A larger value indicates a more definite polarization direction of the received radiation at that pixel. Live late-sown wheat leaves are covered by a waxy cuticle layer several micrometers thick. This cuticle layer produces significant specular reflection of obliquely incident light, causing the reflected light to be linearly polarized parallel to the incident plane. Therefore, live leaves appear as high-value pixels on the linear polarization map. Surface stubble and bare soil, due to their high surface roughness and dominant scattering components, exhibit depolarized reflected light, appearing as low-value pixels on the linear polarization map. The GPU computing array uses the Otsu adaptive thresholding method to obtain the segmentation threshold from the linear polarization map. This method automatically selects the segmentation threshold with the goal of maximizing inter-class variance and is adaptive to changes in illumination caused by early spring clouds. Pixels with linear polarization values ​​greater than the segmentation threshold in the linear polarization map are marked as canopy pixels, and the remaining pixels are non-canopy pixels, thus obtaining a canopy binary mask. In one alternative implementation, in cloudy or twilight low-light scenarios, a local thresholding method based on an adaptive window can be used instead. The linear polarization map is divided into 16×16 pixel sub-blocks, and the Otsu adaptive thresholding method is run independently before spatial smoothing.

[0029] The polarization response characteristics of the polarization sensing model used in this embodiment on two typical types of ground features are as follows: Figure 2 As shown. Figure 2 The horizontal axis represents the incident angle of light, in degrees, ranging from 0 to 90 degrees, with scales of 0, 15, 30, 45, 60, 75, and 90. The vertical axis represents the linear polarization degree at the pixel, dimensionless, ranging from 0 to 0.85. Figure 2 Two curves, distinguished by different symbols, are plotted: the first curve represents the linear polarization degree of the surface of late-sown wheat leaves as a function of the incident angle, and the second curve represents the linear polarization degree of the surface stubble and bare soil as a function of the incident angle. The two curves are plotted side-by-side in the same coordinate system to visually compare the differences in polarization response between the two types of land cover. (Using...) Figure 2The comparison relationship shown in this embodiment is based on the difference in polarization degree as an effective criterion for separating the pixels of late-sown wheat leaves from the background surface stubble-bare soil pixels. In particular, the significant difference in polarization degree between the two curves near the Brewster angle is used to achieve highly robust extraction of live wheat pixels, avoiding the problem that the discrimination based solely on the difference in hue or brightness of the visible light RGB channel is easily affected by changes in illumination and fluctuations in the reflectivity of stubble-bare soil.

[0030] The definition of linear polarization follows the standard expression of the Stokes parameter, i.e. ,in This represents the linear polarization degree value at that pixel location. The total light intensity at that pixel location is equal to the pixel-by-pixel sum of the 0-degree polarization intensity map and the 90-degree polarization intensity map. This represents the result of subtracting the 90-degree polarization intensity map from the 0-degree polarization intensity map. This represents the result of subtracting the 135-degree polarization intensity map from the 45-degree polarization intensity map. The closer it is to 1, the more certain the polarization direction of the radiation received by that pixel is; The closer it is to 0, the more likely the reflected light is to be depolarized.

[0031] The linear polarization degree curves of live late-sown wheat leaves exhibit a distinct bell-shaped distribution, with the peak value appearing at an incident angle of approximately 55 degrees and averaging around 0.72. Within the range of incident angles less than 30 degrees or greater than 80 degrees, [the curves show a clear polarization distribution]. It drops below 0.2. The root cause of this bell-shaped distribution is that the surface of the living leaf is covered by a waxy cuticle layer several micrometers thick. This cuticle layer produces specular reflection of obliquely incident light, and the electric field vector of the reflected light tends to dominate in the direction perpendicular to the incident plane. Therefore, the linear polarization degree is significantly enhanced near the Brewster angle. On the graph, an arrow points from the peak position near the incident angle of 55 degrees to the side of the curve, and is labeled "specular reflection of waxy cuticle layer" to explain the physical origin of this peak.

[0032] The linear polarization curves corresponding to surface residue and bare soil are significantly lower than those of live leaves, with an overall peak value of only about 0.16. Furthermore, the peak value is relatively flat, remaining within the range of 0.05 to 0.16 over a wide incident angle range of 0 to 70 degrees. The root cause of this low polarization distribution lies in the relatively high geometric roughness of the surface of the residue and bare soil, the widespread and random distribution of the normals of the micro-area elements, and the depolarization of the reflected light after multiple scatterings. The Stokes parameter... and They cancel each other out, making Significantly reduced. On the map, arrows point from the surface stubble curve to the side, labeled "diffuse reflection depolarization" to illustrate the physical source of this low value.

[0033] Between the two curves, a horizontal dashed line is used to plot the value levels of the Otsu adaptive threshold, typically as follows: Nearby, and labeled "Otsu Adaptive Threshold". This threshold is automatically obtained by the Otsu adaptive thresholding method with the goal of maximizing the inter-class variance, without the need for manual preset; pixels in the linear polarization degree map that are greater than this threshold are marked as canopy pixels, and the remaining pixels are marked as non-canopy pixels, thus obtaining a canopy binary mask.

[0034] Next, tillering monomer segmentation and Thiessen tessellation are performed on the canopy binary mask. The GPU computing array connects the canopy binary mask with the first intermediate channel. Perform a pixel-by-pixel AND operation to obtain a canopy intensity map with non-zero pixel values ​​only in the canopy region; then apply a preset maximum height threshold. Apply morphological H-maxima transform to the canopy light intensity map to transform values ​​with amplitudes lower than [the specified value]. Local maxima are smoothed out, and pixel locations with significant local maxima features are retained as candidate tillering vertices. A preset maximum height threshold is used. A typical value for H is 15% of the median value in the canopy light intensity map. This value has been proven effective in distinguishing the main stem from the tips of adjacent tillers. The H-maximum transformation suppresses spurious extrema generated by specular reflection in the middle of the leaf, thus preventing multiple tiller vertices from splitting from a single leaf. Each tiller vertex in the resulting set of tiller vertices is first represented by pixel coordinates. Using the set of tiller vertices as seeds, a gradient-based watershed segmentation algorithm is applied to the canopy light intensity map to obtain a set of individual tillers corresponding one-to-one with each tiller vertex. Each individual tiller corresponds to the area occupied by one tiller in the canopy projection.

[0035] The ARM Cortex main control processor cluster then invoked the homography transformation matrix stored in shared memory. The pixel coordinates of each tiller vertex in the tiller vertex set. Convert to two-dimensional coordinates of the land parcel The transformation relationship is as follows: ;in These are pixel coordinates. These are the homogeneous, normalized 2D coordinates of the plot. The position of each tiller vertex in the transformed set of tiller vertices is represented in meters, with the southwest corner of the plot as the origin. The GPU computing array uses the Fortune scanline algorithm to construct a set of Thiessen polygons in the 2D plane of the plot using the transformed set of tiller vertices; the time complexity of the Fortune scanline algorithm is O(log n). ,in This represents the number of tillering vertices in the set of tillering vertices, per frame on a typical plot. In the range of thousands to tens of thousands, the Fortune scanline algorithm can be completed in milliseconds on a Jetson Orin GPU computing array. Each Thiessen polygon in the set corresponds one-to-one with a tiller. Geometrically, each Thiessen polygon represents the set of all points in the plot plane that are closer to that tiller vertex than to any other tiller vertex. Therefore, the area of ​​a Thiessen polygon can be used as the equivalent space share occupied by that tiller within the plot. Since the Fortune scanline algorithm generates open Thiessen polygons extending to infinity at the plot edges, the GPU computing array clips all Thiessen polygons according to the plot boundaries and their respective control zone boundaries. After clipping, all Thiessen polygons become bounded polygons. The arithmetic mean of the areas of all clipped Thiessen polygons is then obtained. The ARM Cortex main control processor cluster then determines the control zone to which each Thiessen polygon belongs based on the principle of geometric centroid falling within the zone.

[0036] For each Thiessen polygon, the GPU computing array extracts the median of the red, green, and blue color components within the original color image region covered by the polarization camera, forming the red median, green median, and blue median respectively. The median of the canopy light intensity map is then used to form the strong median. Using the median provides robustness against outliers. The red, green, and blue medians, along with the strong median, form the leaf color quadruple for the corresponding Thiessen polygon. The ARM Cortex main control processor cluster then calculates the leaf greenness index. ;in Indicates the greenness index of leaves. It is the middle green position. It is red in the middle position. (Leaf greenness index) The larger the value, the darker the leaf color within the area covered by the Tyson polygon, indicating a more abundant nitrogen supply. Next, calculate the suspected phosphorus deficiency index, first calculating the median ratio. ,in It is the blue median, then the median ratio Divide by strong median After completing intensity normalization, the suspected phosphorus deficiency index was obtained: ;in This indicates a possible phosphorus deficiency index. Early spring wheat leaves with phosphorus deficiency will appear purplish-red due to anthocyanin accumulation, and the leaves will show relatively reduced blue-green wavelength reflectance. However, the design of a possible phosphorus deficiency index needs to isolate the influence of overall canopy light intensity, as the blue / red ratio will also shift in shaded areas or under twilight. A strong median is introduced. Intensity normalization is performed to ensure that the suspected phosphorus deficiency index only reflects changes in spectral shape and is insensitive to absolute radiance. The higher the value, the greater the likelihood of phosphorus deficiency.

[0037] The results of the mosaic acquisition of the pole-mounted color split-focus plane polarization camera on the plot used in this embodiment are as follows: Figure 3 As shown, the sample plot is a typical sample plot located at any location on the entire plot, and can represent the treatment method of other sample plots on the entire plot. Figure 3 The horizontal axis represents the land parcel. The direction and vertical axis represent the land parcel. The direction and unit are both meters, and the range is 0 to 4 meters. The horizontal axis is marked with 0, 1, 2, 3, and 4 meters respectively, and the vertical axis is marked with 0, 1, 2, 3, and 4 meters respectively. Figure 3 One is marked in a grid format. The mosaic results of a typical sample plot show that each grid cell corresponds to the field of view coverage of a single image taken by a pole-mounted color split-plane polarization camera at that location. There is a pre-defined overlap band between adjacent grid cells to ensure seamless mosaicking. (Using...) Figure 3 The mosaic method shown in this embodiment is used to uniformly stitch together the continuous push-broom data from one end to the other of the plot by the pole-mounted color split-focus plane polarization camera according to geometric coordinates, thereby obtaining a full-coverage polarization-light intensity joint image of the entire plot, so that the subsequent extraction of live late-sown wheat pixels based on polarization response differences has spatial consistency at the plot scale.

[0038] The image contains approximately 70 solid black dots, each representing a tiller vertex. The vertex coordinates are obtained by extracting local maxima pixel locations from the canopy intensity map obtained from the polarized canopy, performing a morphological H-maxima transform, and then converting the results into two-dimensional plot coordinates using a homography transformation matrix. The homography transformation satisfies... ,in Here are the pixel coordinates of the tiller apex in the polarization image. The transformed two-dimensional coordinates of the land parcel This is a pre-calibrated 3×3 homography transformation matrix.

[0039] Using the set of tiller vertices as seeds, a set of Thiessen polygons is constructed in the two-dimensional plane of the plot using the Fortune scanline algorithm. Each polygon enclosed by a thin black line in the diagram represents a Thiessen polygon, and each Thiessen polygon corresponds one-to-one with a tiller vertex. Geometrically, each Thiessen polygon represents the set of all points in the plot plane that are closer to that tiller vertex than to any other tiller vertex. Therefore, the area of ​​a Thiessen polygon can be used as the equivalent spatial share occupied by that tiller in the plot, directly reflecting the tiller density near that location.

[0040] Each Thiessen polygon is assigned a region classification label based on its seedling condition classification. The area mean is obtained by first taking the arithmetic mean of the areas of all cropped Thiessen polygons. Then assign values ​​according to the following rules: Area greater than The area was designated as a weak seedling area, with a size smaller than [missing information]. The area is classified as an overly dense area, with a size of [missing information]. and The areas between these points are designated as suitable seedling zones. Multiple large Thiessen polygons near the upper right corner of the quadrat correspond to weak seedling zones, while multiple small, dense polygons near the lower left corner correspond to overly dense zones. Most polygons in the main body of the quadrat correspond to suitable seedling zones. The distribution of these three types of zones directly reflects the invention's effective ability to identify spatial heterogeneity of seedling conditions in a plot.

[0041] For the complex reflection coefficient of the dielectric spectrum array probe, the Hailo-8L pre-quantizes and deploys a complex dielectric state inversion network. This network is a fully connected neural network with a 30-dimensional dielectric input vector and three scalar outputs: water content, unfrozen water content, and conductivity. A typical structure includes an input layer of 30 neurons, three hidden layers of 64 neurons each, a ReLU activation function, and an output layer of 3 neurons. This fully connected neural network is trained on sample pairs generated by a Cole-Cole relaxation model before leaving the factory. The sample pairs are generated as follows: first, the Cole-Cole relaxation model... ;in It is an angular frequency of The complex permittivity at that point, It is the static dielectric constant. It is the high-frequency limiting dielectric constant. It is the relaxation time. It is the Cole-Cole relaxation index. It is an imaginary unit. Latin hypercube sampling was performed within the range of water content from 1% to 45%, unfrozen water content from 0% to 45%, and conductivity from 0.01 to 3 Siemens per meter. Each sampling point corresponds to one set. , , , Based on published empirical relationships of soil dielectric properties, moisture content, unfrozen water content, and conductivity were correlated with the aforementioned four Cole-Cole parameters. Then, values ​​were taken at five frequency points: 1 MHz, 10 MHz, 100 MHz, 500 MHz, and 1 GHz. The real and imaginary parts of the vector are concatenated to form a 30-dimensional dielectric input vector sample, which is then paired with the corresponding water content, unfrozen water content, and conductivity to form sample pairs. Approximately 500,000 sample pairs are generated. After training to obtain a fully connected neural network, the neural network is then quantized into INT8 tensor operators using the Hailo Dataflow Compiler and deployed in Hailo-8L.

[0042] The relationship between the complex permittivity of the soil samples collected in this embodiment and frequency is as follows: Figure 4 As shown. Figure 4The system is divided into two independent coordinate systems from top to bottom: the upper coordinate system shows the change of the real part of the complex permittivity with frequency, and the lower coordinate system shows the change of the imaginary part of the complex permittivity with frequency. The horizontal axis of both coordinate systems is frequency, in Hertz, and uses a logarithmic scale. Hertz Hertz ranges from 1 MHz to 1 GHz; the vertical axis of the upper coordinate system represents the real part of the complex permittivity, ranging from 0 to 70, while the vertical axis of the lower coordinate system represents the imaginary part of the complex permittivity, ranging from 0 to 16. Multiple spectrum curves of the complex permittivity corresponding to different soil moisture contents are plotted within each coordinate system, with different symbols used to distinguish the curves. From lower to higher moisture contents, the real part curve of the complex permittivity shows an overall increase from approximate to approximately tens in the low-frequency range, gradually decreasing after exceeding 100 MHz; the imaginary part curve of the complex permittivity shows a monotonically increasing spectrum in the low-frequency range, reaching a peak and then decreasing. (Using...) Figure 4 The comparison relationship shown in this embodiment transforms the soil moisture content inversion problem into a complex permittivity spectrum fitting problem, thus avoiding the problem that single-frequency measurements are easily affected by interface polarization interference in the low-frequency band and by scattering loss in the high-frequency band, which can lead to misjudgment of soil moisture content.

[0043] Four curves are plotted within each coordinate system, corresponding to the complex dielectric spectra under four typical soil moisture conditions: 15%, 25%, 35%, and 45%. Each curve is generated using a Cole-Cole relaxation model, specifically expressed as follows: ,in Represents angular frequency The complex permittivity at that point, This represents the static dielectric constant, i.e., the low-frequency limit. Represents the high-frequency limiting dielectric constant. Indicates the relaxation time. This represents the Cole-Cole relaxation index. Represents the imaginary unit. , Indicates frequency.

[0044] As can be seen from the left coordinate system, the real part of the complex permittivity... In the low frequency band The near-static dielectric constant is close to that of Hertz. In the high frequency band The dielectric constant near Hertz is close to the high-frequency limiting dielectric constant. There is a monotonically decreasing transition region in the middle, with the center frequency of the transition region being approximately... to The center frequency is on the order of Hertz, and it corresponds to the characteristic frequency of the Cole-Cole relaxation. From the 15% moisture content curve to the 45% moisture content curve, the real part shifts upwards overall. The low-frequency limit of the 15% curve is about 14, and the low-frequency limit of the 45% curve is about 62. It can be seen that the real part of the soil complex dielectric constant is extremely sensitive to moisture content. This is the physical basis for inverting moisture content based on dielectric spectrum.

[0045] As can be seen from the coordinate system on the right, the imaginary part of the complex permittivity... It exhibits a typical unimodal distribution, with the peak position appearing near the aforementioned relaxation characteristic frequency. The peak amplitude increases significantly with increasing water content, with the peak value of the curve at 15% water content being approximately 2 and the peak value of the curve at 45% water content being approximately 14.

[0046] The positions of the five sampling frequency points are indicated by vertical dashed lines and markers scattered on each curve in the two coordinate systems, from left to right: 1MHz, 10MHz, 100MHz, 500MHz, and 1GHz. The frequency point labels are located below the horizontal axis.

[0047] During real-time inference, the ARM Cortex main control processor cluster converts the 5-frequency complex reflection coefficients uploaded in real-time by the dielectric spectrum array probes into measured values ​​of the 5-frequency complex dielectric constant for each probe, using a transmission line model with known parameters such as the characteristic impedance and electrode length of the coaxial sensitive electrode. For open-ended coaxial sensitive electrodes, the complex reflection coefficients... It satisfies an approximate relationship with the complex permittivity. ;in It is angular frequency Complex reflection coefficient at that location, This is the characteristic impedance of the coaxial sensitive electrode, typically 50 ohms. This is the geometrically distributed capacitance at the beginning, measured by calibration under air-filled conditions using the inner and outer conductor radii and electrode length of the coaxial sensitive electrode. Typical values ​​are on the order of 0.05 picofarads. (It is known that...) and , Then, by inversely solving the above relationship, the measured value of the complex permittivity can be obtained. This process is performed independently at three depths: 5cm, 20cm, and 40cm for each group of dielectric spectrum array probes. Then, the measured values ​​of the complex permittivity at five frequencies for each group of dielectric spectrum array probes at the three depths are decomposed into real and imaginary parts, and concatenated into a 30-dimensional dielectric input vector in the order of "depth first, then frequency, then real and imaginary parts." The Hailo-8L independently performs three complex dielectric state inversion network inferences for each group of dielectric spectrum array probes at different depths, obtaining the water content, unfrozen water content, and conductivity at depths of 5cm, 20cm, and 40cm. The output water content at 5cm depth is used as the surface water state value for the corresponding dielectric spectrum array probe. The output water content at 20cm and 40cm depths is fused according to preset depth weights to obtain the root water state value for the corresponding dielectric spectrum array probe. The typical preset depth weights are 0.4 and 0.6, i.e., 0.4 for 20cm depth and 0.6 for 40cm depth, making the root water state value more sensitive to the water environment of deep roots. The minimum unfrozen water content at the three depths is taken as the unfrozen water content value for the corresponding dielectric spectrum array probe. Taking the minimum value is the most conservative criterion from a safety perspective.

[0048] Based on the outputs of the geothermal chain and solar-induced chlorophyll fluorescence narrowband spectrometer, the ARM Cortex main control processor cluster performs the greening-return stage determination. First, the moving averages of the past five days are taken for each of the four geothermal depth channels to obtain four geothermal moving average sequences. Using a five-day window smooths out the instantaneous fluctuations caused by single-day cold waves, ensuring that the greening-return determination is based on a stable trend temperature. Next, the peak sequence of the photosynthetic potential index over the past three days is taken, and then the difference between adjacent days is taken from this peak sequence to obtain the daily increment sequence of photosynthetic potential. Two consecutive positive daily increments of photosynthetic potential indicate a continuous enhancement of canopy photosynthetic activity, serving as the physiological basis for the initiation of greening-return. The greening-up stage is indicated according to the following rules: When the 5cm soil temperature moving average sequence is below 3℃, it is considered a preparatory stage for greening-up, with root activity almost stagnant; when the 5cm soil temperature moving average sequence is above 3℃ for three consecutive days and nights, and the daily increase in photosynthetic potential is positive for two consecutive days, it is considered the start of greening-up, with surface roots beginning to activate; when both the 5cm and 10cm soil temperature moving average sequences are above 5℃ for three consecutive days and nights, and the daily increase in photosynthetic potential is positive for two consecutive days, it is considered greening-up in progress, with tillering nodes entering an active elongation stage; when the 20cm soil temperature moving average sequence is above 5℃ for three consecutive days and nights, it is considered greening-up complete, with the entire root system activated. The greening-up stage indicators are updated twice daily, at 00:00 and 12:00.

[0049] The GPU computes the array and then performs spatial mapping on the dielectric inversion results. Using the planar coordinates of the four sets of dielectric spectrum array probes as known interpolation nodes, and the geometric centroid coordinates of each Thiessen polygon as the interpolation point, an inverse distance-weighted interpolation method with a distance weighting exponent of 2 is called. For each Thiessen polygon, the inverse distance-weighted interpolation formula is as follows: ;in The point to be interpolated The interpolation result at the location, It is the first The dielectric spectrum array probe corresponds to the surface water content value, root layer water content value, or unfrozen water content value. The point to be interpolated With the The Euclidean distances between the planar coordinates of the dielectric spectrum array probes. A distance weighting exponent of 2 is a commonly used empirical setting for inverse distance-weighted interpolation in two-dimensional space, balancing smoothness and locality, ensuring that the influence of nearest neighbor probes dominates in the sense of the inverse square of the distance. When the distance to a point to be interpolated... When the depth is less than 0.5 meters, the inversion result of the corresponding probe is directly used as the interpolation result of the point to be interpolated to avoid numerical overflow caused by the denominator approaching zero. After interpolating the surface water state value, root water state value, and unfrozen water content value respectively, the surface water state, root water state, and unfrozen water content corresponding to the Thiessen polygon are obtained.

[0050] The ARM Cortex main control processor cluster then assigns a zone classification to each Thiessen polygon according to the following rules: when the area is greater than 1.5 times the average area, the space allocated to the tillering unit corresponding to the Thiessen polygon within the plot is significantly higher than the average, indicating that the nearby tillering density is relatively sparse, and it is classified as a weak seedling area; when the area is less than 0.5 times the average area, the space for the tillering unit corresponding to the Thiessen polygon is cramped, indicating that the nearby tillering density is relatively dense, and it is classified as an overly dense area; when the area is between 0.5 and 1.5 times the average area, it is classified as a suitable seedling area. These two ratios of 1.5 and 0.5 times, calibrated through field trials, can identify approximately 15% to 25% of weak seedling areas and 10% to 20% of overly dense areas. The seedling condition classification mapping is formed by combining the zone identifier, surface water content, root water content, unfrozen water content, greening stage identifier, leaf color greenness index, suspected phosphorus deficiency index, the corresponding valve control zone, and the area of ​​the corresponding Thiessen polygon. The internal structure of the seedling condition classification mapping is a data table with the Thiessen polygon number as the key.

[0051] Finally, a three-way decision inference is performed based on the seedling condition classification mapping. The ARM Cortex main control processor cluster outputs three states for nitrogen supplementation, phosphorus supplementation, and micro-spraying for each Thiessen polygon according to three sets of independent dual-threshold rules. Nitrogen supplementation decision is based on leaf greenness index. As evidence value, when When the nitrogen level is below the lower threshold, nitrogen supplementation will be performed immediately. When the nitrogen level is between the lower and upper thresholds, nitrogen supplementation should be initiated and observation temporarily suspended. When nitrogen levels exceed the upper threshold, nitrogen supplementation is postponed. Typical initial values ​​for the lower and upper nitrogen thresholds are 1.10 and 1.30, respectively; specific values ​​can be determined based on the variety and region. Phosphorus supplementation decisions are based on suspected phosphorus deficit indicators. As evidence value, When the phosphorus content exceeds the upper threshold, phosphorus supplementation is initiated immediately. When the phosphorus content is between the lower and upper thresholds, phosphorus supplementation should be initiated and observation temporarily suspended. When the phosphorus content is below the lower threshold, phosphorus supplementation is postponed. Note that a higher value for the suspected phosphorus deficiency index indicates a higher probability of phosphorus deficiency, which is opposite to the direction of the leaf greenness index. Therefore, the "immediate execution" option in the dual threshold rule also corresponds to the opposite direction. Micro-spraying decisions are based on root zone water content as evidence. When root zone water content is below the lower threshold, micro-spraying is initiated immediately; when root zone water content is between the lower and upper thresholds, micro-spraying is temporarily suspended; and when root zone water content is above the upper threshold, micro-spraying is postponed.

[0052] Two additional sets of enhancement constraints are added to the above. The first set is a safety constraint on micro-spraying decisions based on the greening stage indicator and the unfrozen water content: when the greening stage indicator is greening preparation or the unfrozen water content is lower than the set unfrozen water safety threshold, the three states of the micro-spraying decision for this Thiessen polygon are forcibly set to micro-spraying delayed processing. The unfrozen water safety threshold is typically set to 60% of the water content at the corresponding dielectric spectrum array probe position, meaning that as long as the unfrozen water ratio is lower than this value, it is considered that there are enough ice crystals occupying the pores. At this time, irrigation will damage the root system due to capillary blockage and freeze-thaw shear. The second group modifies the nitrogen and phosphorus supplementation decisions based on the zone classification: When the zone classification is a weak seedling area, the nitrogen lower threshold is increased by one preset step and the phosphorus upper threshold is decreased by one preset step, actively intervening in the spring management of the weak seedling area by increasing phosphorus and nitrogen; when the zone classification is an overly dense area, the three states of nitrogen supplementation decision for this Thiessen polygon are set to at most "nitrogen supplementation temporarily suspended" to avoid further promoting tillering and causing ineffective tillers in the overly dense area; when the zone classification is a suitable seedling area, the original dual-threshold judgment results are maintained. The typical value of the preset step is 5% of the dynamic range of each threshold, for example, the preset step of the nitrogen lower threshold is 0.01, and the preset step of the moisture lower threshold is 1% of the volumetric moisture content.

[0053] The three decision-making operators described above form a lightweight decision fusion network. It takes the evidence value matrix, threshold matrix, and constraint matrix of all Thiessen polygons as input and the three-state result vector of all Thiessen polygons as output. This network achieves parallel execution of rule determination and dual correction at the tensor level, significantly accelerating the process compared to polygon-by-polygon iterative determination on a CPU. The overall decision-making time for plots with 5000 to 10000 Thiessen polygons is controlled within tens of milliseconds, leaving computational margin for downstream rolling optimization scheduling. The output consists of a three-state decision set per region, composed of the three states of nitrogen supplementation, phosphorus supplementation, and micro-spraying decisions for all Thiessen polygons. This set is stored in the output cache of the onboard heterogeneous edge inference unit for subsequent reading by the rolling optimization scheduling module.

[0054] In one alternative implementation, the aforementioned vehicle-mounted heterogeneous edge inference unit can be extended to a multi-machine collaborative form. Specifically, only the Jetson Orin is placed on the field inspection vehicle for image processing and spatial geometry calculations, while the Hailo-8L is deployed at the field gateway. Data exchange between the two is achieved via RTSP or gRPC, suitable for applications requiring miniaturized inspection vehicles. Furthermore, if the Hailo-8L is not available on-site, the complex dielectric state inversion network and the three decision inference operators can be re-deployed on the Jetson Orin's GPU computing array, using TensorRT for FP16 quantization. This results in a slight loss of accuracy but lowers the implementation threshold.

[0055] In the vehicle-mounted heterogeneous edge inference unit, the Jetson Orin ARM Cortex main control processor cluster uses the next 7 days as the rolling time domain. The 7-day timescale was chosen because it takes 3 to 5 days for late-sown wheat to stabilize from spring nitrogen and phosphorus supplementation to the emergence of leaf color and tillering responses. The 7-day timescale covers the entire process from execution to response and allows for adjustment margins. The rolling process is repeated every 24 hours, meaning that at 00:05 each day, the latest three-way decision result set from the previous time is used as input to recalculate the water and fertilizer synergistic execution sequence for the next 7 days, ensuring that the agronomic response of the most recent day is promptly reflected in the arrangements for the following 6 days.

[0056] The ARM Cortex main control processor cluster first performs an execution set merge on the three decision result sets for each zone. For each Thiessen polygon within the plot, its three states of phosphorus supplementation, nitrogen supplementation, and micro-spraying decisions are checked: Thiessen polygons that are in the immediate execution state of phosphorus supplementation are assigned to the phosphate fertilizer execution set according to their respective valve control zones; Thiessen polygons that are in the immediate execution state of nitrogen supplementation are assigned to the nitrogen fertilizer execution set according to their respective valve control zones; Thiessen polygons that are in the immediate initiation state of micro-spraying are assigned to the micro-spraying execution set according to their respective valve control zones. Thiessen polygons in the state of nitrogen supplementation postponement observation, phosphorus supplementation postponement observation, and micro-spraying postponement observation are not included in any execution set in this rolling time domain; Thiessen polygons in the state of nitrogen supplementation delayed processing, phosphorus supplementation delayed processing, and micro-spraying delayed processing are also not included in any execution set in this rolling time domain, but the delayed processing state will be recorded, and if the evidence value is still in the corresponding trigger interval in the next rolling time domain, it will be processed again. This transformation from the three states of "immediate execution, delayed observation, and postponed processing" to the two states of "execution / non-execution" follows the approach of determining the "acceptance / rejection" boundary in the three-way decision-making process, avoiding the decision drift caused by forcibly dividing in the intermediate state.

[0057] The phosphate fertilizer execution set, nitrogen fertilizer execution set, and micro-spraying execution set use the valve-controlled zone number as the key and the list of Thiessen polygon serial numbers included in the execution set under that valve-controlled zone as the value. The plot is divided into 16 valve-controlled zones in a 4×4 rectangular grid, with each valve-controlled zone having an area of ​​approximately 150 square meters.

[0058] After the set is executed and completed, the execution order is arranged in the rolling time domain according to the preset agronomic priority. The preset agronomic priority is phosphorus supplementation first, nitrogen supplementation second, and water spraying last. The basis for this priority order is: phosphorus is a difficult-to-move element that requires a long time for root zone infiltration and absorption, so supplementing phosphorus first can continue to be effective in subsequent operations; nitrogen absorption requires a certain soil water activity in the root zone, so arranging it after phosphorus supplementation can enhance nitrogen solubility through the wetting effect of fertilizer solution; placing water spraying last can avoid fertilizer solution being diluted and rinsing off residual fertilizer solution from the leaves.

[0059] The specific arrangement process is as follows: First, allocate one phosphate fertilizer injection window to each valve-controlled zone in the phosphate fertilizer execution set; then allocate one nitrogen fertilizer injection window to each valve-controlled zone in the nitrogen fertilizer execution set, with the start time of this nitrogen fertilizer injection window after the end time of any existing phosphate fertilizer injection window in that valve-controlled zone; if a valve-controlled zone is not included in the phosphate fertilizer execution set, its nitrogen fertilizer injection window is placed in the earliest available operating window within the rolling time domain; then allocate one clean water replenishment spray window to each valve-controlled zone in the micro-spraying execution set, with the start time of this clean water replenishment spray window after the end time of any existing nitrogen fertilizer injection window in that valve-controlled zone; if a valve-controlled zone is not included in the nitrogen fertilizer execution set, its clean water replenishment spray window is placed in the earliest available clean water operating window within the rolling time domain. The duration of each window segment is determined by the following formula: ;in, Indicates the duration of the window, in minutes; This indicates the number of Thiessen polygons that fall within this valve-controlled partition and are included in the corresponding execution set of this window; This indicates the preset unit execution time, which is the basic spraying time required for a single Thiessen polygon. Typical values ​​are: 0.8 minutes for phosphate fertilizer injection, 1.0 minute for nitrogen fertilizer injection, and 0.6 minutes for water replenishment. This setting links the window duration to the number of seedlings to be treated, avoiding over-irrigation in areas with few seedlings or insufficient supply in areas with dense seedlings.

[0060] To avoid operational conflicts between different valve-controlled zones, the ARM Cortex main control processor cluster discretizes the rolling time domain into time slots with a granularity of 5 minutes. A greedy slotting algorithm is used to sequentially find the earliest available time slot start for each valve-controlled zone in the execution set. Before inserting a window segment, the time slot occupancy table of the valve-controlled zone is scanned, skipping time slots already occupied by other windows within the same valve-controlled zone. Simultaneously, the total occupancy tables of global water pumps and fertilizer pumps are scanned. If more than three valve-controlled zones are executing simultaneously in the same time slot, the execution is postponed to the start of the next available time slot. This constraint originates from the upper limit of the main pump flow rate and avoids pipeline pressure drop. In an alternative implementation, the greedy slotting algorithm can be replaced with mixed-integer linear programming, aiming to minimize the total execution completion time. Joint optimization is performed on the start points of all windows, resulting in a more compact execution sequence in large irrigation areas. However, the cost is that the solution time increases from milliseconds to seconds, requiring a trade-off based on the plot size.

[0061] After the arrangement is completed, a tuple of start and end times and valve control zone numbers for the phosphate fertilizer injection window, nitrogen fertilizer injection window, and clean water replenishment window is obtained. These are then packaged together to form a water-fertilizer coordinated execution sequence. The data format of the water-fertilizer coordinated execution sequence is a sequence of execution items arranged in ascending order of start time. Each execution item includes a window type label, start time, end time, valve control zone number, and the number of Thiessen polygons participating in that window.

[0062] The field gateway distributes the water and fertilizer co-processing sequence to the microfluidic proportional fertilizer mixing module and the piezoelectric ceramic micro-sprayer array one by one. The microfluidic proportional fertilizer mixing module has a 3-in-1-out structure. The three inlets are connected to the potassium dihydrogen phosphate mother liquor tank, the urea mother liquor tank, and the dilution water tank, respectively. Each inlet is equipped with a peristaltic pump and a pressure sensor. The outlet connects to the fertilizer mixing chamber and then to the piezoelectric ceramic micro-sprayer array via the main pipeline. The preset mother liquor ratio is set according to the window type: under the phosphate fertilizer injection window, the volumetric flow rate ratio of potassium dihydrogen phosphate mother liquor to dilution water is 1:9, that is, when the mother liquor mass concentration is 50 g / L, the working concentration of potassium dihydrogen phosphate at the outlet is 5 g / L; under the nitrogen fertilizer injection window, the volumetric flow rate ratio of urea mother liquor to dilution water is 1:14, and the working concentration of urea at the outlet is about 6 g / L; under the clean water supplementation window, the dilution water flow rate is 100%, and the peristaltic pumps of the two mother liquor inlets are stopped. The rotational speed of the three peristaltic pumps is dynamically adjusted by the ARM Cortex main control processor cluster according to the aforementioned volumetric flow rate ratio, ensuring that the outlet concentration strictly follows the preset mother liquor ratio while maintaining a constant total flow rate in the main pipeline. The microfluidic chip incorporates a serpentine mixing channel for thorough mixing of the components.

[0063] The piezoelectric ceramic micro-sprinkler array consists of several piezoelectric ceramic micro-sprinklers fixedly distributed according to valve-controlled zones, approximately 2 meters above the field, below the sprinkler branch pipes. Each valve-controlled zone carries 4 to 8 piezoelectric ceramic micro-sprinklers. Each piezoelectric ceramic micro-sprinkler is driven by a ring-shaped piezoelectric ceramic actuator to produce a 100-micrometer diameter nozzle. Under pulse excitation at a working voltage of 40 to 60 volts and a frequency of 20 to 50 kHz, it achieves digital water droplet spraying with a controllable spray particle size of 30 to 80 micrometers. The atomized particle size of the piezoelectric ceramic micro-sprinkler can be adjusted according to the working frequency, and the spray volume per unit time can be directly controlled by the PWM duty cycle. With no moving mechanical parts, it is suitable for precise control during the greening period when water is scarce and seedlings are stable. Each valve-controlled zone inlet is equipped with a two-position, two-way solenoid valve, which receives opening and closing commands from the field gateway; the solenoid valve opens when the window starts and closes when the window ends.

[0064] During spraying, the ARM Cortex main control processor cluster monitors pipeline pressure, flow rate, and solenoid valve status in real time. If the actual flow rate does not reach 80% of the preset flow rate within 30 seconds after the start of the window, it is determined to be a main pipeline blockage or pump failure, and is automatically skipped and prioritized for rearrangement in the next rolling time domain; if the pipeline pressure drops below the threshold in the middle of the window, all ongoing windows are globally paused and a maintenance process is triggered to avoid closed-loop correction misjudging execution failures as insufficient agronomic response.

[0065] Seventy-two hours after each window ends, a closed-loop correction is initiated. The 72-hour interval is chosen because changes in leaf greenness after nitrogen supplementation require 48 to 96 hours to stabilize at spring temperatures around 10°C; 72 hours is a suitable interval while also taking into account the lag in tillering response. At the closed-loop moment, the pole-mounted color split-focus plane polarization camera is re-triggered for one 4-polarization channel color capture. Following the polarization canopy extraction, tiller individual segmentation, Thiessen mosaic, and leaf color quaternion extraction process described in step 2, the executed tiller vertex set and leaf greenness index are re-obtained.

[0066] For each valve-controlled partition with execution, the ARM Cortex mainframe processor cluster calculates two evaluation metrics: tiller density increment and leaf color recovery value. The tiller density increment is obtained as follows: ;in, This indicates the increase in tiller density in the valve-controlled zone, expressed in units of plants per square meter; This indicates the number of tillering vertices contained in the set of tillering vertices after execution within the valve-controlled partition; This indicates the number of tillering vertices contained in the set of tillering vertices before execution within the valve-controlled partition; This indicates the area of ​​the valve-controlled zone, in square meters. A positive value indicates an increase in new tillers in the valve-controlled zone within 72 hours. Tiller density difference is used instead of Thiessen polygon area mean difference as the evaluation metric because new tillers further divide the original tillering space, causing the area mean to decrease, which is inversely related to the increase in density; the sign direction after using tiller density is consistent with agronomic intuition.

[0067] The leaf color recovery value is obtained by the following formula: ;in, This represents the leaf color recovery value of the valve-controlled zone, and is dimensionless. This represents the arithmetic mean of the leaf greenness index after the execution of all Thiessen polygons within the valve-controlled zone; This represents the arithmetic mean of the leaf greenness index of all Thiessen polygons within the valve-controlled zone before execution. A positive value indicates that the leaves have turned green again and the nitrogen response is good; A value of zero or negative indicates insufficient nitrogen response or disturbance from other stresses. Taking the arithmetic mean can more sensitively reflect the overall trend of greenness recovery.

[0068] The ARM Cortex main control processor cluster independently corrects the three evidence channels according to the channel correction rules.

[0069] Regarding the nitrogen pathway, when the leaf color recovers... Below the preset recovery threshold If the nitrogen supplementation effect is deemed insufficient, a two-way escalation is implemented: the nitrogen lower threshold is increased by one preset step, ensuring that more Thiessen polygons are included in the immediate nitrogen supplementation process in the next rolling time domain; simultaneously, the nitrogen fertilizer injection window duration in the next rolling time domain is extended by one preset step. The preset execution time for nitrogen fertilizer has been changed from 1.0 minute to 1.0 plus the preset step size. Preset recovery threshold. The typical value is 0.03, meaning that if the green-red ratio does not rise to 0.03 within 72 hours, the nitrogen response is considered insufficient; the preset step size for the nitrogen lower threshold is 0.01, and the preset step size for the nitrogen fertilizer injection window duration is 0.2 minutes.

[0070] For phosphorus channels, when tillering density increases Below the preset incremental threshold In this case, two scenarios need to be distinguished before processing. First, if the valve-controlled zone received phosphorus supplementation in the previous rolling time domain, but the tillering response is still insufficient, it indicates that the existing phosphorus lower threshold is too strict or the phosphorus application amount is too low. Therefore, the phosphorus lower threshold is increased by one preset step, and the phosphate fertilizer injection window duration in the next rolling time domain is extended by one preset step. Second, if the valve-controlled zone was not included in the phosphate fertilizer execution set in the previous rolling time domain, but the tillering density increment is still low, it indicates that the valve-controlled zone may have a hidden phosphorus deficiency that has not been captured by the existing suspected phosphorus deficiency indicators. In this case, the valve-controlled zone is directly forced into the phosphate fertilizer execution set in the next rolling time domain, bypassing the dual threshold rule. (Preset incremental threshold) The typical value is 2 plants per square meter. The preset step size for the phosphorus lower threshold is 5% of the dynamic range of the leaf greenness index, which is about 0.01. The preset step size for the duration of the phosphate fertilizer injection window is 0.15 minutes. This dual-path correction method of "first check if it has been done, then decide whether to increase or supplement" avoids the permanent misjudgment caused by blindly adjusting the threshold for unexecuted valve control zones.

[0071] For the moisture channel, if the root layer moisture content is still below the lower moisture threshold after re-interpolation at the closed-loop time, the lower moisture threshold is increased by 1 preset step, and the duration of the clean water re-spraying window in the next rolling time domain is extended by 1 preset step. The preset step for the lower moisture threshold is 1% of the volumetric moisture content, and the preset step for the clean water re-spraying window duration is 0.1 minutes.

[0072] To prevent monotonic drift of the aforementioned thresholds and window durations under continuous rolling time domain corrections, the preset unit execution duration of all three thresholds and three window durations is constrained by their respective preset upper and lower limits. The preset upper limit for the nitrogen threshold is 1.20, and the preset lower limit is 1.00; the preset upper limit for the phosphorus threshold is 0.50, and the preset lower limit is 0.30; the preset upper limit for the moisture threshold is 28% volumetric moisture content, and the preset lower limit is 18% volumetric moisture content; the preset unit execution duration for the nitrogen fertilizer injection window is 2.5 minutes, and the lower limit is 0.5 minutes; the preset upper limit for the phosphorus fertilizer injection window is 2.0 minutes, and the lower limit is 0.4 minutes; the preset upper limit for the clean water replenishment spray window is 1.5 minutes, and the lower limit is 0.3 minutes. After each correction, the ARM Cortex main control processor cluster first verifies whether the corrected value falls between the corresponding preset upper and lower limits. If it exceeds the limit, the corresponding upper or lower limit value is taken as the final correction value in the direction of the out-of-bounds deviation. This out-of-bounds pruning mechanism prevents the threshold from diverging to an unreasonable range.

[0073] The ARM Cortex main control processor cluster writes back the corrected nitrogen, phosphorus, and water thresholds and the corresponding preset unit execution time to the seedling condition classification mapping, triggering Hailo-8L to re-run the three-branch decision reasoning at the beginning of the next rolling time domain, and then re-executes the execution set union, agronomic priority arrangement, and greedy slot to obtain the water and fertilizer synergistic execution sequence of the next rolling time domain, so that the whole method continues to run continuously from the spring greening to the tillering stage.

[0074] In one alternative implementation, the tiller density increment and leaf color recovery value can be further combined into a comprehensive response index. ,in Indicates the comprehensive response index, Indicates the tillering response weight. , , , These represent the tillering density increment, preset increment threshold, leaf color recovery value, and preset recovery threshold, respectively. The typical value is 0.4. When the comprehensive response index is less than 1.0, the weight is uniformly increased according to the above-mentioned channel correction rules. The advantage is that single scalar decision-making is more convenient for cross-plot comparison; the disadvantage is that the asynchrony between tillering and leaf color responses may be masked by the homogenization process.

[0075] In another alternative implementation, if the signal-to-noise ratio of the polarization image at the closed-loop time is insufficient to extract the set of tiller apex, the closed-loop correction can be postponed to the most recent time period that meets the shooting quality, and the tiller density increment can be linearly normalized according to the time interval; if a qualified image is still not obtained after two consecutive postponements, the water channel correction is performed using the root layer water content change trend as a proxy evaluation index.

[0076] In another alternative implementation, when the plot is a gentle slope with an average slope greater than 3 degrees, a slope aspect-aware sorting rule can be introduced, with the slope top valve control zone prioritized in the first half of the clear water replenishment window and the slope toe zone in the second half, to reduce nutrient loss.

[0077] The specific embodiments of the present invention have been described above. Those skilled in the art may omit, substitute, or modify the details of the above methods without departing from the principles and essence of the present invention, and all such modifications shall fall within the scope of the present invention. The scope of the present invention is defined by the appended claims.

Claims

1. A method for precise control of fertigation and water spraying for late-sown wheat, characterized in that, Includes the following steps: Step 1: Deploy dielectric spectrum array probes, ground temperature chains, solar-induced chlorophyll fluorescence narrowband spectrometers, and rod-mounted color split-focus plane polarization cameras on the target plot. Output complex reflection coefficient, ground temperature value, photosynthetic potential index value, and polarization channel color images, respectively. These are then time-aligned and aggregated into a multi-source crop and soil moisture joint dataset through a field gateway. Step 2: The multi-source seedling and soil moisture joint dataset is fed into a vehicle-mounted heterogeneous edge inference unit consisting of Jetson Orin and Hailo-8L interconnected. Jetson Orin extracts the Thiessen polygon set and leaf greenness index and suspected phosphorus deficiency index of each Thiessen polygon from the polarization channel color image. Hailo-8L obtains the surface water state value, root water state value and unfrozen water content value from the complex reflectance coefficient. Jetson Orin obtains the greening stage identifier based on the soil temperature value and photosynthetic potential index value, and performs spatial mapping of the dielectric inversion results according to the geometric centroid of the Thiessen polygon, assigning a region class identifier to each Thiessen polygon to form a seedling condition classification mapping. Hailo-8L and Jetson Orin run three-branch decision inference on the seedling condition classification mapping region by region, and outputs a region-by-region three-branch decision result set consisting of three states of nitrogen supplementation decision, three states of phosphorus supplementation decision, and three states of micro-spraying decision. Step 3: Jetson Orin performs rolling optimization scheduling on the three decision result sets of each zone according to the preset agronomic priority, generating a water and fertilizer co-execution sequence including phosphate fertilizer injection window, nitrogen fertilizer injection window and water spraying window. This sequence is then sent to the microfluidic proportional fertilizer mixing module and piezoelectric ceramic micro-sprayer array for zoned spraying via the field gateway. After each window ends, images are re-acquired to obtain the tiller density increment and leaf color recovery value. The nitrogen lower threshold, phosphorus lower threshold, water lower threshold and corresponding window duration are closed-loop corrected according to the channel correction rules. After writing back, the water and fertilizer co-execution sequence for the next rolling time domain is generated.

2. The method as described in claim 1, characterized in that, In step 1, four sets of dielectric spectrum array probes and four geothermal chains are deployed along two mutually perpendicular sampling baselines. Two sets of dielectric spectrum array probes and two geothermal chains are equidistantly deployed along each sampling baseline. The planar coordinates of the four sets of dielectric spectrum array probes are associated with the valve-controlled zone coordinate system. A solar-induced chlorophyll fluorescence narrowband spectrometer is deployed in the center of the plot, and a pole-mounted color-focused plane polarization camera is deployed 2m above the plot. The dielectric spectrum array probes have one set of coaxial sensitive electrodes embedded at depths of 5cm, 20cm, and 40cm, and are cyclically swept at five frequencies: 1MHz, 10MHz, 100MHz, 500MHz, and 1GHz. The real and imaginary parts of the complex reflection coefficient are output synchronously at each frequency. The geothermal... The chain has one platinum resistance temperature sensor at each of the four depths of 2cm, 5cm, 10cm, and 20cm, and outputs the ground temperature value at 10-minute intervals. The solar-induced chlorophyll fluorescence narrowband spectrometer outputs the canopy fluorescence radiance at 30-minute intervals with a narrowband channel centered at a wavelength of 760nm and a half-bandwidth of 0.3nm. After fitting and subtracting the baseline from the continuous baseline segments on both sides of the 760nm wavelength, the peak value of the fluorescence radiance during the effective illumination period of the day is taken as the photosynthetic potential index value. The pole-mounted color split-focus plane polarization camera is a split-focus plane camera that synchronously outputs color images with four polarization channels of 0 degrees, 45 degrees, 90 degrees, and 135 degrees with red, green, and blue color components in a single frame, and synchronously triggers shooting at 30-minute intervals.

3. The method as described in claim 1, characterized in that, The Jetson Orin contains a GPU computing array and an ARM Cortex main control processor cluster. The Hailo-8L, as a neural network accelerator, is connected to the Jetson Orin via a PCIe bus. The GPU computing array is responsible for image spatial channel processing and spatial geometry calculations, while the ARM Cortex main control processor cluster is responsible for complex number transformation, sliding statistics, threshold rule determination, and scheduling control. The Hailo-8L is responsible for the quantized neural network tensor inference task, and the three are executed in parallel. Before acquisition, the pole-mounted color split-focus plane polarization camera undergoes intrinsic parameter calibration, extrinsic parameter calibration, and polarization channel gain correction. At least four ground control points are set in the target plot coordinate system. The ARM Cortex main control processor cluster solves the homography transformation matrix between the pixel coordinates of the polarization image and the two-dimensional coordinates of the plot based on the ground control points and stores it in shared memory.

4. The method as described in claim 3, characterized in that, The GPU computing array extracts the luminance component from the four polarization channels of the color image to obtain four polarization intensity maps. Then, it obtains three intermediate channels pixel by pixel: the first intermediate channel is obtained by summing the 0-degree polarization intensity map and the 90-degree polarization intensity map pixel by pixel; the second intermediate channel is obtained by subtracting the 90-degree polarization intensity map from the 0-degree polarization intensity map; and the third intermediate channel is obtained by subtracting the 135-degree polarization intensity map from the 45-degree polarization intensity map. The GPU computing array then calculates the sum of squares and takes the square root of the second and third intermediate channels pixel by pixel. The result of the sum of squares and square root is then divided pixel by the first intermediate channel to obtain a linear polarization degree map. The GPU computing array uses the Otsu adaptive thresholding method to obtain a segmentation threshold for the linear polarization degree map. Pixels with linear polarization degree values ​​greater than the segmentation threshold in the linear polarization degree map are marked as canopy pixels, thus obtaining a canopy binary mask.

5. The method as described in claim 4, characterized in that, The GPU computing array performs a pixel-by-pixel AND operation between the canopy binary mask and the first intermediate channel to obtain a canopy intensity map. A morphological H-maximum transformation is then applied to the canopy intensity map using a preset maximum height threshold, resulting in a set of tiller vertices represented in pixel coordinates. Using this set of tiller vertices as seeds, a gradient-based watershed segmentation algorithm is applied to the canopy intensity map to obtain a set of individual tiller vertices corresponding to each tiller vertex. The ARM Cortex main control processor cluster uses a homography transformation matrix to convert the set of tiller vertices from pixel coordinates to two-dimensional plot coordinates. The GPU computing array then applies the Fortune scanline algorithm to the converted set of tiller vertices to obtain a set of Thiessen polygons, where each Thiessen polygon corresponds to one individual tiller. The GPU computing array clips each Thiessen polygon according to the plot boundary and the boundary of its respective control zone, and the arithmetic mean of the areas of all clipped Thiessen polygons is obtained. The ARM Cortex main control processor cluster determines the valve control zone to which each Thiessen polygon belongs based on the principle of geometric centroid falling.

6. The method as described in claim 5, characterized in that, The GPU computing array takes the median of the red, green, and blue color components in each Thiessen polygon-covered polarized camera original color image region to form the red median, green median, and blue median, respectively. It also takes the median of the canopy light intensity map to form the strong median. The red, green, and blue medians, along with the strong median, form the leaf color quadruple corresponding to the Thiessen polygon. The ARM Cortex main control processor cluster divides the green median in the leaf color quadruple by the red median to obtain the leaf greenness index, divides the blue median in the leaf color quadruple by the red median to obtain the median ratio, and then divides the median ratio by the strong median to complete the intensity normalization, thus obtaining the phosphorus deficiency index. The larger the value of the phosphorus deficiency index, the higher the probability of phosphorus deficiency.

7. The method as described in claim 3, characterized in that, Hailo-8L is pre-quantized with a complex dielectric state inversion network. This network is a fully connected neural network with a 30-dimensional dielectric input vector and three scalar outputs: water content, unfrozen water content, and conductivity. Before leaving the factory, this fully connected neural network is trained on sample pairs generated by a Cole-Cole relaxation model and quantized into tensor operators executable by Hailo-8L. The Cortex main control processor cluster converts the 5-frequency complex reflection coefficients uploaded by the dielectric spectrum array probes into measured values ​​of the 5-frequency complex dielectric constant for each probe using a transmission line model with known parameters such as the characteristic impedance and electrode length of the coaxial sensitive electrode. Then, the measured values ​​of the 5-frequency complex dielectric constants of each group of dielectric spectrum array probes at three depths of 5cm, 20cm, and 40cm are split into real and imaginary parts, respectively, and concatenated into a 30-dimensional dielectric input vector according to depth and frequency point order, which is then fed into the complex dielectric state inversion network. The Hailo-8L outputs the water content, unfrozen water content, and conductivity of each group of probes according to depth. The output water content at a depth of 5cm is taken as the surface water content value of the corresponding probe, and the output water content at depths of 20cm and 40cm are fused according to preset depth weights to obtain the root water content value of the corresponding probe. The minimum value of the unfrozen water content at the three depths is taken as the unfrozen water content value of the corresponding probe.

8. The method as described in claim 5, characterized in that, The ARM Cortex main control processor cluster takes the most recent five 24-hour moving averages of the geothermal temperature from each of the four depth channels to obtain four geothermal moving average sequences. For the photosynthetic potential index, it takes the most recent three-day peak sequences and calculates the difference between adjacent days to obtain the daily increment sequence of photosynthetic potential. The following rules are used to mark the greening stage: when the 5cm geothermal moving average sequence is below 3℃, it is considered greening preparation; when the 5cm geothermal moving average sequence is above 3℃ for three consecutive 24-hour periods and the daily increment sequence of photosynthetic potential is positive for two consecutive days, it is considered greening initiation; when both the 5cm and 10cm geothermal moving average sequences are above 5℃ for three consecutive 24-hour periods... The above conditions are met when the daily increase in photosynthetic potential is positive for two consecutive days, and when the 20cm ground temperature sliding mean is above 5℃ for three consecutive days, the greening process is considered complete. The GPU computing array uses the planar coordinates of four dielectric spectrum array probes as known interpolation nodes and the geometric centroid of each Thiessen polygon as the interpolation point. Using an inverse distance-weighted interpolation method with a distance weighting exponent of 2, the surface water state, root water state, and unfrozen water content values ​​output by the Hailo-8L are interpolated to obtain the corresponding surface water state, root water state, and unfrozen water content of the Thiessen polygon. ARM Cortex main control processor clusters are assigned zone classification labels according to the following rules: areas with an area greater than 1.5 times the average area are weak seedling areas, areas with an area less than 0.5 times the average area are overly dense areas, and areas with an area between 0.5 and 1.5 times the average area are suitable seedling areas. The zone classification label, surface water content, root water content, unfrozen water content, greening stage label, leaf greenness index, suspected phosphorus deficiency index, the valve control zone, and the corresponding Thiessen polygon area are used to form a seedling condition classification mapping.

9. The method as described in claim 3, characterized in that, The ARM Cortex main control processor cluster outputs three states for nitrogen supplementation, phosphorus supplementation, and micro-spraying decisions for each Thiessen polygon in the seedling condition grading mapping according to three sets of independent dual-threshold rules: Nitrogen supplementation decision uses leaf greenness as evidence value: if it is below the lower nitrogen threshold, nitrogen supplementation is immediately implemented; if it is between the lower and upper nitrogen thresholds, nitrogen supplementation is temporarily suspended; if it is above the upper nitrogen threshold, nitrogen supplementation is delayed. Phosphorus supplementation decision uses a suspected phosphorus deficiency index as evidence value: if it is above the upper phosphorus threshold, phosphorus supplementation is immediately implemented; if it is between the lower and upper phosphorus thresholds, phosphorus supplementation is temporarily suspended; if it is below the lower phosphorus threshold, phosphorus supplementation is delayed. Micro-spraying decision uses root layer water content as evidence value: if it is below the lower water content threshold, micro-spraying is immediately initiated; if it is above the lower water content threshold, micro-spraying is delayed. When the lower threshold is between the upper threshold and the lower threshold of moisture content, micro-spraying is temporarily suspended for observation; when the upper threshold of moisture content is above the lower threshold of moisture content, micro-spraying is delayed for processing. Additional safety constraints: When the greening stage is marked as greening preparation or the unfrozen water content is lower than the set unfrozen water safety threshold, the corresponding three states of micro-spraying decision for the Thiessen polygon are forcibly set to micro-spraying delayed for processing. Additional zone correction: When the zone is marked as a weak seedling zone, the lower threshold of nitrogen is increased by 1 preset step and the upper threshold of phosphorus is decreased by 1 preset step. When the zone is marked as an overly dense zone, the corresponding three states of nitrogen supplementation decision for the Thiessen polygon are at most set to nitrogen supplementation temporarily suspended for observation. When the zone is marked as a suitable seedling zone, the original dual-threshold judgment result is maintained. The Hailo-8L runs the quantized deployment of the three-branch decision inference operator to perform tensor-level integration of the above rule results and outputs the entire set of three-branch decision results for each zone of the Thiessen polygon.

Images