A method for controlling water and fertilizer management based on image data processing

Abstract

The present invention relates to a method for controlling water and fertilizer management based on image data processing. Visible light images, multispectral images, and thermal infrared images of crops are collected and feature extraction is performed. The morphological parameters of crop plants, vegetation indices, and the temperature distribution of the vegetation canopy are calculated to create a three-dimensional physiological feature matrix. An image segmentation network model is created. The visible light images, multispectral images, thermal infrared images, and the three-dimensional physiological feature matrix are input into the image segmentation network model to segment plant organs, and organ-level phenotypic parameters are quantified. A water and fertilizer decision-making model based on reinforcement learning is established. With the real-time data of the three-dimensional physiological feature matrix, organ-level phenotypic parameters, and specified soil parameters as the state input and the water and fertilizer ratio scheme as the action space, through a dynamic optimization control strategy, the implementation of the water and fertilizer management scheme is carried out. It solves the technical problems of insufficient integration of single-sensor data, limited phenotypic analysis accuracy, and adaptability defects of the decision-making model existing in traditional water and fertilizer regulation schemes.

Description

Technical Field

[0001] The present invention belongs to the technical field of water and fertilizer management, and particularly relates to a water and fertilizer management control method based on image data processing. Background Art

[0002] In recent years, the technology of agricultural precision management has developed rapidly. As a key link in crop yield increase and quality improvement, water and fertilizer regulation has gradually shifted from an experience-driven mode to a data-driven mode. However, there are still significant technical bottlenecks in the aspects of multi-source data fusion accuracy, dynamic decision-making real-time performance, and organ-level phenotype analysis ability in traditional methods, which are specifically manifested in the following aspects:

[0003] First, insufficient integration of single-sensor data:

[0004] Traditional water and fertilizer regulation schemes mostly rely on single spectral bands (such as visible light or near-infrared) or soil sensor data, and it is difficult to synchronously capture the coupling relationship between crop physiological states and environmental stresses (such as heat stress, nutrient imbalance, soil indicators). Visible light images can only analyze morphological parameters (such as leaf area, plant height), but cannot quantify nitrogen content. Thermal infrared images can detect abnormal canopy temperature, but lack a dynamic association with photosynthetic capacity; soil indicators only reflect the rhizosphere environment and cannot directly map the physiological response of the canopy.

[0005] Second, limited accuracy of phenotype analysis:

[0006] Existing image segmentation algorithms are mostly designed for rough organ-level segmentation, and have insufficient analytical ability for micro-morphological features such as leaf serration and stem curvature, resulting in a prediction error rate of water and fertilizer requirements as high as 15%-20%. In addition, two-dimensional image analysis ignores the three-dimensional spatial distribution (such as the impact of leaf inclination differences on light interception), and it is difficult to accurately evaluate the dynamics of biomass accumulation.

[0007] Third, adaptability defects of decision-making models:

[0008] Mainstream water and fertilizer decision-making models, such as threshold-triggered and expert system-dependent fixed rule bases in the prior art, cannot adapt to the time-varying nature of the field environment and the non-linear characteristics of crop growth. For example: Static fertilization formulas ignore the selective absorption differences of stomatal aperture for different nitrogen forms (nitrate nitrogen / ammonium nitrogen); irrigation strategies based on historical data are difficult to cope with the transpiration wave under extreme weather.

[0009] Therefore, the problems existing in traditional water and fertilizer regulation schemes, namely insufficient integration of single-sensor data, limited accuracy of phenotype analysis, and adaptability defects of decision-making models, are technical problems that need to be solved urgently at present. Summary of the Invention

[0010] The purpose of the present invention is to provide a water and fertilizer management control method based on image data processing, so as to solve the technical problems of insufficient integration of single-sensor data, limited phenotypic analysis accuracy, and adaptability defects of decision-making models existing in traditional water and fertilizer regulation schemes.

[0011] To solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0012] A water and fertilizer management control method based on image data processing includes the following steps:

[0013] S1: Collect visible-light images, multispectral images, and thermal infrared images of crops through an image acquisition module, and respectively extract features from the visible-light images, multispectral images, and thermal infrared images to obtain visible-light image features, multispectral image features, and thermal infrared image features;

[0014] S2: Calculate the plant morphological parameters of the crops through the visible-light image features, calculate the vegetation indices of the crops through the multispectral image features, and obtain the vegetation canopy temperature distribution by providing the thermal infrared image features;

[0015] S3: Create a three-dimensional physiological feature matrix of the crops based on the plant morphological parameters, vegetation indices, and canopy temperature distribution;

[0016] S4: Create an image segmentation network model, input the visible-light images, multispectral images, thermal infrared images, and three-dimensional physiological feature matrix into the image segmentation network model to segment plant organs, and quantify organ-level phenotypic parameters;

[0017] S5: Establish a water and fertilizer decision-making model based on reinforcement learning, use the real-time data of the three-dimensional physiological feature matrix, organ-level phenotypic parameters, and soil specified parameters as state inputs, use the water and fertilizer ratio scheme as the action space, and dynamically optimize the control strategy through the Q-learning algorithm;

[0018] S6: Implement the water and fertilizer management plan based on the optimized control strategy.

[0019] Preferably, the visible-light images of the crops in step S1 are collected through a visible-light image acquisition module, the multispectral images are collected through a multispectral image acquisition module, and the thermal infrared images are collected through a thermal infrared image acquisition module.

[0020] Preferably, the specific process of extracting features from the visible-light images, multispectral images, and thermal infrared images in step S1 to obtain visible-light image features, multispectral image features, and thermal infrared image features is as follows:

[0021] S11: standardizing the visible light image, extracting the edge features and skeleton features of plant leaves in the visible light image through an edge detection algorithm, generating a point cloud model by combining the multi-view visible light image, and obtaining leaf inclination angle and canopy opening features, and fruit / stem features;

[0022] S12: Use the correction model to correct the multispectral image, use the improved RANSAC algorithm to align the visible light and multispectral images to ensure the pixel-level spatial consistency, extract features based on the 710nm red edge image and the 840nm near-infrared light image, calculate the vegetation index features, and combine the 660nm red light image to obtain the normalized vegetation index features;

[0023] S13: Convert the grayscale value to absolute temperature through blackbody radiation correction, apply adaptive threshold segmentation to extract vegetation area, obtain temperature thermogram and obtain canopy temperature standard deviation characteristics.

[0024] Preferably, in step S2, the specific process of calculating the plant morphological parameters of the crop by using the visible light image features, calculating the vegetation index of the crop by using the multispectral image features, and obtaining the temperature distribution of the vegetation canopy by providing the thermal infrared image features is as follows:

[0025] S21: Calculate leaf area index, leaf length, leaf width, circumference, shape factor, petiole length, leaf inclination, and canopy opening based on leaf inclination and canopy opening characteristics, plant leaf edge characteristics, and leaf skeleton characteristics; obtain fruit size and color based on fruit / stem characteristics;

[0026] S22: Calculate the vegetation index NDRE based on the vegetation index characteristics. The calculation formula of the vegetation index NDRE is as follows:

[0027] NDRE=( ρ 840- ρ 710) / ( ρ 840+ ρ 710);

[0028] in, ρ 840 is the reflectivity of the 840nm near-infrared band, ρ 710 is the reflectivity of the 710nm red edge band;

[0029] S23: Calculate the Normalized Difference Vegetation Index (NDVI) based on the Normalized Difference Vegetation Index (NDVI) feature. The calculation formula of the Normalized Difference Vegetation Index (NDVI) is as follows:

[0030] NDVI=( ρ 840- ρ 660) / ( ρ 840+ ρ 660);

[0031] Among them, ρ 840 is the reflectance in the 840 nm near-infrared band, ρ 660 is the reflectance in the 660 nm red light band;

[0032] Preferably, in step S4, an image segmentation network model is created, and the visible light image, multi-spectral image, thermal infrared image, and three-dimensional physiological feature matrix are input into the image segmentation network model to segment plant organs, and the specific process of quantifying organ-level phenotypic parameters is as follows:

[0033] S41: Establish the visible light image, multi-spectral image, and thermal infrared image in a unified spatial coordinate system, and perform registration based on the spatial coordinate system;

[0034] S42: Resample the three-dimensional physiological feature matrix to the two-dimensional image resolution through bilinear interpolation to establish a voxel-pixel mapping relationship;

[0035] S43: The image segmentation network model is set with 12 input channels, among which three channels input the visible light image, three channels input the reflectance in the 840 nm near-infrared band, the reflectance in the 710 nm red edge band, and the reflectance in the 660 nm red light band of the multi-spectral image, and the remaining channels input the eigenvalues of the three-dimensional physiological feature matrix;

[0036] S44: The backbone network module of the image segmentation network model adopts a ConvNeXt-Tiny encoder, embeds a dilated spatial pyramid pooling module to expand the receptive field to 128×128 pixels, introduces an efficient channel attention module in the decoder, marks the candidate regions, and outputs the overall plant region;

[0037] S45: The secondary network module of the image segmentation network model performs fine segmentation of plant organs on the candidate regions;

[0038] S46: Quantify the leaf area, leaf serration, and leaf curvature based on the Fourier descriptor analysis of the segmentation mask contour, quantify the stem height, bending degree, and diameter based on the three-dimensional skeletonization process and the thermal infrared temperature gradient mapping, and quantify the fruit volume and surface patches based on the three-dimensional point cloud convex hull algorithm and the multi-spectral abnormal reflection analysis.

[0039] Preferably, the specific process of step S5 is as follows:

[0040] S51: Establish the state space of the water and fertilizer decision-making model, and the state space includes a three-dimensional physiological feature matrix, organ-level phenotypic parameters, and soil specified parameters. The soil specified parameters are collected in real time by a soil moisture sensor and include 8-dimensional data of pH value, organic matter content, nutrient content, heavy metal content, microbial biomass, enzyme activity, humidity, and conductivity, and are denoised by Kalman filtering;

[0041] S52: Create the action space of the water and fertilizer decision-making model, where the action space includes action IDs, water and fertilizer plans, and the corresponding water, nitrogen, phosphorus, and potassium ratios of the water and fertilizer plans. The action IDs include 0, 1, 2; 0, 1, 2 correspond to the baseline plan, high-nitrogen plan, and water-saving plan respectively;

[0042] S53: Create the reward function of the water and fertilizer decision-making model. The formula of the reward function is:

[0043] R = α (△BI - BI target ) + β (WUE - WUE baseline );

[0044] Among them, R is the reward function value; α and β are crop-related weight coefficients, which balance the optimization priorities of biomass growth and resource efficiency. In the initial growth stage: α > β (focusing on biomass accumulation), in the mature stage: β > α (focusing on resource efficiency); △BI is the biomass increment, which reflects the growth rate of plant biomass per unit time, and the calculation method is: dynamically estimated through the three-dimensional reconstruction data of the leaf area index (LAI) and the stem volume; BI target is the target biomass threshold, which is the expected biomass growth value preset according to the crop growth stage and is corrected in real time based on environmental factors (such as accumulated temperature, sunshine hours) through a sliding window algorithm; WUE is the water use efficiency, which is usually defined as the ratio of crop yield to irrigation water volume (kg / m³), and maximizes the biomass output per unit water volume by adjusting the irrigation strategy. WUE baseline is the baseline water use efficiency, which represents the historical average or industry standard level and is used as a reference benchmark for reward calculation, reflecting the improvement amplitude of the current strategy relative to the conventional plan;

[0045] S54: The reward function drives strategy optimization through the following mechanism. When the actual biomass increment (△BI) exceeds the preset target BI target , α the β term generates a positive reward, and the

[0046] term ensures that the water use efficiency is not lower than the baseline level, avoids resource waste caused by over-irrigation, and the α / β ratio adjustment realizes the dynamic balance between increasing production and saving water, meeting the sustainable development needs of agricultural production.

[0047] θ Preferably, during the image acquisition process, the image acquisition module performs polarized light filtering and polarization angle adjustment. The polarization angle adjustment formula is as follows:= arctan( n 2 / n 1) + k ·Δ φ ;

[0048] Wherein, θ is the polarization angle adjustment amount, n 1, n 2 are the refractive indices of the medium, and Δ φ is the curvature compensation amount of the blade surface, k is the proportionality coefficient, which is used to describe the sensitivity of the polarization angle to the phase difference change.

[0049] Preferably, when the NDRE value is in the range of 0.35 - 0.42, a hierarchical fertilization strategy is triggered:

[0050] First - level fertilization correction: Execute 80% of the reference fertilization amount;

[0051] Second - level fertilization correction: Dynamically adjust by superimposing the detection result of the leaf stomatal aperture;

[0052] Third - level fertilization correction: Perform acid - base balance compensation in combination with the soil pH value.

[0053] Preferably, the hierarchical fertilization strategy is specifically as follows:

[0054] The calculation formula of the reference fertilization amount is:

[0055] Reference fertilization amount = (Fertilizer requirement for target yield - Soil fertilizer supply) / Fertilizer utilization rate;

[0056] First - level fertilization correction: Execute 80% of the reference fertilization amount;

[0057] Second - level fertilization correction: Dynamically adjust by superimposing the detection result of the leaf stomatal aperture. The leaf stomatal aperture is measured by an infrared thermal imager. When the leaf stomatal aperture is less than 150 mmol, increase the application of nitrate nitrogen fertilizer by 15%. When the leaf stomatal aperture is between 150 - 350 mmol, maintain the current amount with ±0%. When the leaf stomatal aperture is greater than 350 mmol, reduce the application of ammonium nitrogen fertilizer by 10%;

[0058] Third - level fertilization correction: When the soil pH is lower than 6.0, for every 0.5 pH unit lower, increase the application of calcium - magnesium phosphate fertilizer by 3 kg / mu. When the soil pH is higher than 7.5, for every 0.5 pH unit higher, increase the application of sulfur powder by 1.5 kg / mu.

[0059] The beneficial effects of the present invention include:

[0060] The water and fertilizer management control method based on image data processing provided by the present invention collects visible light images, multispectral images, and thermal infrared images of crops and performs feature extraction, calculates the plant morphological parameters, vegetation indices, and vegetation canopy temperature distribution of the crops, creates a three-dimensional physiological feature matrix, creates an image segmentation network model, inputs the visible light images, multispectral images, thermal infrared images, and three-dimensional physiological feature matrix into the image segmentation network model to segment plant organs, quantifies organ-level phenotypic parameters, establishes a water and fertilizer decision-making model based on reinforcement learning, uses the real-time data of the three-dimensional physiological feature matrix, organ-level phenotypic parameters, and soil specified parameters as state inputs, uses the water and fertilizer ratio scheme as the action space, and implements the water and fertilizer management scheme through dynamic optimization control strategies. It solves the technical problems of insufficient integration of single-sensor data, limited phenotypic analysis accuracy, and adaptive defects of the decision-making model existing in traditional water and fertilizer regulation schemes.

[0061] First, calculate the plant morphological parameters of the crop through visible light image features, calculate the vegetation index of the crop through multispectral image features, obtain the vegetation canopy temperature distribution by providing thermal infrared image features, create a three-dimensional physiological feature matrix of the crop, break through the information density limitation of traditional two-dimensional data, and realize the synchronous analysis of multi-scale physiological states of canopy-organ-cell.

[0062] Second, through the Q-learning reinforcement learning algorithm, the composite reward function of biomass increment and water use efficiency is used to drive policy optimization. Compared with the traditional PID control model, it greatly shortens the response time of water and fertilizer scheme adjustment and significantly reduces the resource waste rate.

[0063] Finally, trigger a three-level joint control mechanism through the NDRE threshold. The first-level correction avoids the risk of excessive fertilization; the second-level correction, the stomatal aperture is linked with the adjustment of nitrogen form to improve stress adaptability; the pH compensation of the third-level correction maintains the stability of the rhizosphere microenvironment, forming a full-chain optimization system of "macro control - meso compensation - micro response". Description of the Drawings

[0064] Figure 1 It is a schematic flow chart of the water and fertilizer management control method based on image data processing of the present invention.

[0065] Figure 2 It is a schematic flow chart of the image segmentation network model of the present invention for segmenting plant organs and quantifying organ-level phenotypic parameters. Detailed Embodiments

[0066] The following combines the attached Figures 1 - 2 Further detailed description of the present invention will be given:

[0067] Example 1

[0068] See the attachedFigure 1 As shown in the figure, a water and fertilizer management control method based on image data processing includes the following steps:

[0069] S1: Collect visible light images, multispectral images, and thermal infrared images of corn crops through an image acquisition module, and perform feature extraction on the visible light images, multispectral images, and thermal infrared images respectively to obtain visible light image features, multispectral image features, and thermal infrared image features. The visible light images, multispectral images, and thermal infrared images are collected by a drone equipped with visible light, multi-light including 710nm, 840nm, and 660nm bands, and a thermal infrared camera. The images are collected along a preset flight path with a 70% overlap rate to ensure the integrity of canopy coverage. The visible light images are collected through the 400 - 700nm band, with a resolution greater than 20 million pixels, to capture the details of plant morphology. The multispectral images cover the 710nm red edge, 840nm near-infrared, and 660nm red light bands to quantify chlorophyll fluorescence and photosynthetic activity. The thermal infrared images based on the 8 - 14μm wavelength range have a spatial resolution less than or equal to 0.1°C to detect abnormal transpiration areas in the canopy.

[0070] S2: Calculate the plant morphological parameters of the corn crop through the visible light image features, calculate the vegetation index of the corn crop through the multispectral image features, and obtain the vegetation canopy temperature distribution by providing the thermal infrared image features;

[0071] S3: Create a three-dimensional physiological feature matrix of the corn crop based on the plant morphological parameters, vegetation index, and canopy temperature distribution;

[0072] S4: Create an image segmentation network model, input the visible light images, multispectral images, thermal infrared images, and three-dimensional physiological feature matrix into the image segmentation network model to segment the plant organs, and quantify the organ-level phenotypic parameters;

[0073] S5: Establish a water and fertilizer decision-making model based on reinforcement learning, use the real-time data of the three-dimensional physiological feature matrix, organ-level phenotypic parameters, and soil specified parameters as the state input, use the water and fertilizer ratio scheme as the action space, and dynamically optimize the control strategy through the Q-learning algorithm;

[0074] S6: Implement the water and fertilizer management plan based on the optimized control strategy.

[0075] In this embodiment, the visible light images of the corn crop in step S1 are collected through a visible light image acquisition module, the multispectral images are collected through a multispectral image acquisition module, and the thermal infrared images are collected through a thermal infrared image acquisition module.

[0076] In step S1, the specific process of extracting features from the visible light image, multispectral image, and thermal infrared image to obtain visible light image features, multispectral image features, and thermal infrared image features is as follows:

[0077] S11: Standardize the visible light image, extract the plant leaf edge features and leaf skeleton features in the visible light image through an edge detection algorithm, generate a three-dimensional point cloud model by combining multi-view visible light images, and obtain leaf inclination angle and canopy openness features, fruit / stem features. The canopy openness feature measures the canopy porosity.

[0078] In the process of extracting the plant leaf edge features and leaf skeleton features in the visible light image through the edge detection algorithm, use a σ 5×5 Gaussian kernel with = 1.5 to smooth the image and suppress noise interference. Calculate the gradient magnitude in the specified direction, retain the local maximum pixels in the gradient direction, thin the edge to a single-pixel width, set high and low thresholds, connect strong edges and eliminate weak responses. First, initialize the clustering space, select the center point in the image grid, adjust it to the position with the minimum gradient in the 3×3 neighborhood, then combine color similarity and spatial proximity, and iteratively update the clustering within the specified search window. Finally, perform forced connectivity processing to merge fragmented superpixels.

[0079] In the process of generating the three-dimensional point cloud model, match the superpixel feature points through the structure from motion algorithm to generate a dense point cloud, calculate the leaf inclination angle by fitting the angle between the normal vector of the point cloud and the vertical axis, calculate the canopy porosity based on the voxelized model by statistically analyzing the light penetration rate, and canopy porosity = penetrated voxels / total canopy voxels × 100%.

[0080] S12: Use a calibration model to correct the multispectral image, eliminate the geometric distortion caused by the viewing angle difference, ensure pixel-level spatial consistency, use an improved RANSAC algorithm to align the visible light and multispectral images to ensure pixel-level spatial consistency, extract features from the 710nm red edge image and 840nm near-infrared light image, calculate the vegetation index features, and combine the 660nm red light image to obtain the normalized vegetation index features.

[0081] Extract scale-invariant feature points from the 710nm red edge band image, 840nm near-infrared light image, or 660nm red light image in the visible light and multispectral images, accelerate the nearest neighbor matching through KD-Tree, eliminate obvious outliers, dynamically adjust the inlier threshold, and fit the affine transformation matrix for iterative optimization. Terminate the iteration when the root mean square error is less than 0.5 pixels. Resample the multispectral image to the visible light image coordinate system, maintain the integrity of the spectral information, and verify the registration accuracy through the maximization of mutual information (MI) to ensure the consistency of NDVI calculation.

[0082] S13: Convert the gray value to the absolute temperature through blackbody radiation correction, apply adaptive threshold segmentation to extract the vegetation area, obtain the temperature thermal map, and obtain the standard deviation feature of the canopy temperature.

[0083] Establish a gray-temperature linear relationship using a blackbody reference source with a known temperature. The specific formula is: T = a × DN + b , where T is the output absolute temperature value, in °C or K, obtained by reverse derivation through the blackbody radiation model; DN is the original gray value of the sensor, directly output by the thermal infrared camera without temperature conversion; a is the temperature conversion coefficient, a linear scale factor that determines DN the change in temperature when b increases by 1 unit, and is calibrated through two blackbody reference sources with known temperatures;

[0084] Example 2

[0085] Based on Example 1, in step S2, calculate the plant morphological parameters of the corn crop through the visible light image features, calculate the vegetation index of the corn crop through the multi-spectral image features. The specific process of obtaining the vegetation canopy temperature distribution by providing the thermal infrared image features is as follows:

[0086] S21: Calculate the leaf area index, leaf length, leaf width, perimeter, shape factor, petiole length, leaf inclination angle, and canopy opening based on the leaf inclination angle and canopy opening features, plant leaf edge features, and leaf skeleton features. Obtain the fruit size and color based on the fruit / stem characteristics;

[0087] S22: Calculate the vegetation index NDRE based on the vegetation index features. The calculation formula of the vegetation index NDRE is as follows:

[0088] NDRE = ( ρ 840 - ρ 710) / ( ρ 840 + ρ 710);

[0089] where ρ 840 is the reflectance of the 840 nm near-infrared band, ρ 710 is the reflectance of the 710 nm red-edge band;

[0090] S23: Calculate the normalized difference vegetation index NDVI based on the normalized difference vegetation index features. The calculation formula of the normalized difference vegetation index NDVI is as follows:

[0091] NDVI = ( ρ 840 - ρ 660) / ( ρ 840 + ρ 660);

[0092] Among them, ρ 840 is the reflectance of the 840 nm near - infrared band, ρ 660 is the reflectance of the 660 nm red - light band.

[0093] In this embodiment, by fitting the included angle between the leaf normal vector and the vertical axis through a three - dimensional point - cloud model, the leaf inclination angle is calculated by using the least - squares plane fitting. The super - pixel feature points are matched through the structure - from - motion algorithm to generate a dense point cloud, and the included angle between the point - cloud normal vector and the vertical axis is fitted to calculate the leaf inclination angle. Based on the voxelized model, the light penetration rate is statistically calculated, and the canopy porosity is calculated. Canopy porosity = penetrated voxels / total canopy voxels × 100%. The leaf contour is detected, the leaf perimeter is calculated by chain - code encoding, the shape factor is obtained by the convex - hull algorithm, the leaf - vein skeleton is extracted by medial - axis transformation, the petiole length is determined by branch - point detection. Combining the leaf - inclination angle distribution and the canopy porosity, through the porosity inversion model calculation, LAI = - ln(porosity / extinction coefficient).

[0094] The fruit color is extracted by quantifying the fruit hue (H channel) and saturation (S channel) in the HSV space, and the fruit volume is calculated by the three - dimensional point - cloud convex - hull algorithm. Align the 710 nm red - edge band and the 840 nm near - infrared band to the same coordinate system, eliminate the interference of cloud shadows and soil background, calculate the mean value of effective pixels. Based on NDRE = ( ρ 840 - ρ 710) / ( ρ 840 + ρ 710) to obtain the vegetation index NDRE. The 660 nm band needs to be compensated for atmospheric scattering. Based on NDVI = ( ρ 840 - ρ 660) / ( ρ 840 + ρ 660) to obtain the normalized vegetation index NDVI.

[0095] Example 3

[0096] Based on Example 1 or Example 2, refer to Figure 2 , in step S4, an image - segmentation network model is created. The process of inputting the visible - light image, multi - spectral image, thermal - infrared image, and three - dimensional physiological - feature matrix into the image - segmentation network model to segment plant organs and quantify organ - level phenotypic parameters is as follows:

[0097] S41: Establish the visible light image, multispectral image, and thermal infrared image in a unified spatial coordinate system, and perform registration based on the spatial coordinate system;

[0098] S42: Resample the three-dimensional physiological feature matrix to the two-dimensional image resolution through bilinear interpolation, establish a voxel-pixel mapping relationship, and retain the spatial topological relationship;

[0099] S43: The image segmentation network model is set with 12 input channels, where three channels input the visible light image, three channels input the reflectance of the 840nm near-infrared band, the 710nm red edge band, and the 660nm red light band of the multispectral image, and the remaining channels input the eigenvalues of the three-dimensional physiological feature matrix;

[0100] S44: The backbone network module of the image segmentation network model uses a ConvNeXt-Tiny encoder, embeds a dilated spatial pyramid pooling module to expand the receptive field to 128×128 pixels, introduces an efficient channel attention module in the decoder to optimize the feature weights, marks the candidate regions, and outputs the overall plant region;

[0101] S45: The secondary network module of the image segmentation network model performs fine segmentation of the plant organs on the candidate regions, and realizes pixel-level segmentation of leaves, stems, and fruits through a lightweight U-Net++ structure;

[0102] S46: Quantify the leaf area, leaf serration, and leaf curvature based on the Fourier descriptor analysis of the segmentation mask contour, quantify the stem height, curvature, and diameter based on three-dimensional skeletonization processing and thermal infrared temperature gradient mapping, and quantify the fruit volume and surface patches based on the three-dimensional point cloud convex hull algorithm and multispectral abnormal reflection analysis.

[0103] Analyze the contour based on the Fourier descriptor, calculate the area based on pixel statistics, calculate the serration based on the curvature variance, calculate the curvature based on the Fourier coefficients, extract the center line through three-dimensional skeletonization, combine the thermal infrared temperature gradient mapping for the stem diameter and curvature, calculate the volume using the convex hull algorithm, and identify the surface patches through multispectral abnormal reflection. The above process realizes accurate phenotypic analysis from the canopy to the organ level through multi-modal data fusion and hierarchical segmentation.

[0104] Example 4

[0105] Based on Example 1 or Example 2 or Example 3, the specific process of step S5 is as follows:

[0106] S51: Establish the state space of the water-fertilizer decision-making model, where the state space includes a three-dimensional physiological feature matrix, organ-level phenotypic parameters, and specified soil parameters. The specified soil parameters are collected in real time by a soil moisture sensor and include eight-dimensional data of pH value, organic matter content, nutrient content, heavy metal content, microbial biomass, enzyme activity, humidity, and conductivity, and are denoised by Kalman filtering.

[0107] S52: Create the action space of the water-fertilizer decision-making model, where the action space includes an action ID, a water-fertilizer plan, and the corresponding water and nitrogen, phosphorus, and potassium ratios of the water-fertilizer plan. The action ID includes 0, 1, 2; 0, 1, 2 correspond to the baseline plan, the high-nitrogen plan, and the water-saving plan, respectively.

[0108] S53: Create the reward function of the water-fertilizer decision-making model. The formula of the reward function is:

[0109] R = α (△BI - BI target ) + β (WUE - WUE baseline ) ;

[0110] Where R is the reward function value; α and β are the weight coefficients related to corn crops, which balance the optimization priorities of biomass growth and resource efficiency. In the initial growth stage: α > β (focusing on biomass accumulation), in the mature stage: β > α (focusing on resource efficiency); △BI is the biomass increment, which reflects the growth rate of plant biomass per unit time, and the calculation method is: dynamically estimated through the three-dimensional reconstruction data of the leaf area index (LAI) and the stem volume; BI target is the target biomass threshold, which is the expected biomass growth value preset according to the growth stage of corn crops and is corrected in real time through a sliding window algorithm based on environmental factors (such as accumulated temperature, sunshine hours); WUE is the water use efficiency, which is usually defined as the ratio of the corn crop yield to the irrigation water volume (kg / m³), and maximizes the biomass output per unit water volume by adjusting the irrigation strategy. WUE baseline is the baseline water use efficiency, which represents the historical average or industry standard level and is used as a reference benchmark for reward calculation, reflecting the improvement of the current strategy compared with the conventional plan.

[0111] S54: The reward function drives policy optimization through the following mechanism. When the actual biomass increment (△BI) exceeds the preset target BI target , α the βThis item ensures that the water use efficiency is not lower than the baseline level, avoids resource waste caused by over-irrigation, and realizes the dynamic trade-off between yield increase and water conservation through the α / β ratio adjustment, meeting the sustainable development needs of agricultural production.

[0112] During the image acquisition process, the image acquisition module performs polarized light filtering and polarization angle adjustment. The polarization angle adjustment formula is as follows: θ =arctan( n 2 / n 1) + k ·Δ φ ;

[0113] Where θ is the polarization angle adjustment amount, n 1, n 2 are the refractive indices of the media, and Δ φ is the blade surface curvature compensation amount, k is the proportionality coefficient, which is used to describe the sensitivity of the polarization angle to the phase difference change.

[0114] During the polarization angle adjustment process, it is necessary to first calculate the basic polarization angle. According to the incident medium refractive index n 1 and the transmitted medium refractive index n 2, calculate the basic polarization angle θ 0, θ 0=arctan( n 2 / n 1).

[0115] Then, obtain the blade surface curvature radius through laser scanning or structured light three-dimensional reconstruction. The compensation amount is defined by the following formula: Δ φ = d / R , d is the empirical coefficient, which is related to the optical properties of the material, R is the blade surface curvature radius. If the surface curvature changes in real time, such as blade movement, it is necessary to feedback the R value through a photoelectric sensor and update the compensation amount.

[0116] The comprehensive polarization angle adjustment formula is θ =arctan( n 2 / n 1) + k ·Δ φ . Rotate the analyzer to the target angle θ through a stepper motor, and the positioning accuracy is ±0.1°. The polarizer is fixed at the light source end to generate linearly polarized light, and the analyzer is dynamically adjusted to the θ direction to suppress interference in a specific polarization direction. For different wavelength bands, such as visible light and near-infrared, it is necessary to synchronously adjust the polarizer material, such as quartz or polymer polarizer, to match the transmittance characteristics. This process realizes the adaptive polarized light control in complex surface scenarios through refractive index matching and dynamic curvature compensation.

[0117] A water and fertilizer management control method based on image data processing according to the present invention further includes the following process. When the NDRE value is in the range of 0.35 - 0.42, a hierarchical fertilization strategy is triggered:

[0118] Primary fertilization correction: Execute 80% of the benchmark fertilization amount;

[0119] Secondary fertilization correction: Dynamically adjust by superimposing the detection result of leaf stomatal aperture;

[0120] Tertiary fertilization correction: Compensate for acid-base balance in combination with soil pH value.

[0121] The hierarchical fertilization strategy is specifically as follows:

[0122] The calculation formula for the benchmark fertilization amount is:

[0123] Benchmark fertilization amount = (fertilizer requirement for target yield - soil fertilizer supply) / fertilizer utilization rate;

[0124] Primary fertilization correction: Execute 80% of the benchmark fertilization amount;

[0125] Secondary fertilization correction: Dynamically adjust by superimposing the detection result of leaf stomatal aperture. The leaf stomatal aperture is measured by an infrared thermal imager. When the leaf stomatal aperture is less than 150 mmol, apply 15% more nitrate nitrogen fertilizer. When the leaf stomatal aperture is between 150 - 350 mmol, maintain the current amount with ±0%. When the leaf stomatal aperture is greater than 350 mmol, reduce the application of ammonium nitrogen fertilizer by 10%;

[0126] Tertiary fertilization correction: When the soil pH is lower than 6.0, for every 0.5 pH unit lower, apply 3 kg / mu more calcium magnesium phosphate fertilizer. When the soil pH is higher than 7.5, for every 0.5 pH unit higher, apply 1.5 kg / mu more sulfur powder.

[0127] In summary, the water and fertilizer management control method based on image data processing provided by the present invention collects visible light images, multispectral images, and thermal infrared images of corn crops and performs feature extraction, calculates the plant morphological parameters, vegetation indices, and vegetation canopy temperature distribution of corn crops, creates a three-dimensional physiological feature matrix, creates an image segmentation network model, inputs the visible light images, multispectral images, thermal infrared images, and three-dimensional physiological feature matrix into the image segmentation network model to segment plant organs, quantifies organ-level phenotypic parameters, and establishes a water and fertilizer decision-making model based on reinforcement learning. Using the real-time data of the three-dimensional physiological feature matrix, organ-level phenotypic parameters, and specified soil parameters as state inputs and the water and fertilizer ratio scheme as the action space, it implements a water and fertilizer management plan through dynamic optimization control strategies. This solves the technical problems of insufficient integration of single-sensor data, limited phenotypic analysis accuracy, and adaptive defects of the decision-making model existing in traditional water and fertilizer regulation schemes.

[0128] The present invention calculates the plant morphological parameters of corn crops through visible light image features, calculates the vegetation indices of corn crops through multispectral image features, obtains the vegetation canopy temperature distribution by providing thermal infrared image features, creates a three-dimensional physiological feature matrix of corn crops, breaks through the information density limitation of traditional two-dimensional data, and realizes the synchronous analysis of multi-scale physiological states of canopy-organ-cell. Through the Q-learning reinforcement learning algorithm, the policy optimization is driven by a composite reward function of biomass increment and water use efficiency. Compared with the traditional PID control model, it greatly shortens the response time of water and fertilizer scheme adjustment and significantly reduces the resource waste rate at the same time. Through the NDRE threshold triggering a three-level joint control mechanism, the first-level correction avoids the risk of excessive fertilization; the second-level correction, the stomatal aperture is linked to the adjustment of nitrogen form to improve stress adaptability; the pH compensation of the third-level correction maintains the stability of the rhizosphere microenvironment, forming a full-chain optimization system of "macro-control-meso-compensation-micro-response".

Claims

1. A method for controlling water and fertilizer management based on image data processing, characterized in that, Including the following steps: S1: Collect the visible light image, multispectral image, and thermal infrared image of the crop through the image acquisition module, extract features from the visible light image, multispectral image, and thermal infrared image respectively, and obtain the visible light image features, multispectral image features, and thermal infrared image features; S2: Calculate the plant morphological parameters of the crop through the visible light image features, calculate the vegetation index through the multispectral image features, and obtain the vegetation canopy temperature distribution by providing the thermal infrared image features; S3: Create a three-dimensional physiological feature matrix of the crop based on the plant morphological parameters, vegetation index, and canopy temperature distribution; S4: Create an image segmentation network model, input the visible light image, multispectral image, thermal infrared image, and three-dimensional physiological feature matrix into the image segmentation network model to segment the plant organs, and quantify the organ-level phenotypic parameters; S41: Establish the visible light image, multispectral image, and thermal infrared image in a unified spatial coordinate system, and perform registration based on the spatial coordinate system; S42: Resample the three-dimensional physiological feature matrix to the two-dimensional image resolution through bilinear interpolation, and establish a voxel-pixel mapping relationship; S43: The image segmentation network model is set with 12 input channels, where three channels input the visible light image, and three channels input the reflectance of the 840nm near-infrared band, 710nm red-edge band, and 660nm red-light band of the multispectral image, and the remaining channels input the eigenvalues of the three-dimensional physiological feature matrix; S44: The backbone network module of the image segmentation network model adopts a specified encoder, embeds a dilated spatial pyramid pooling module to expand the receptive field to the specified pixels, introduces an efficient channel attention module in the decoder to mark the candidate regions, and outputs the overall plant region; S45: The secondary network module of the image segmentation network model performs plant organ-level segmentation on the candidate regions; S46: Quantify the leaf area, leaf serration, and leaf curvature based on the Fourier descriptor analysis of the segmentation mask contour, quantify the stem height, curvature, and diameter based on the three-dimensional skeletonization process and thermal infrared temperature gradient mapping, and quantify the fruit volume and surface patches based on the three-dimensional point cloud convex hull algorithm and multispectral abnormal reflection analysis; S5: Establish a water and fertilizer decision-making model based on reinforcement learning, use the real-time data of the three-dimensional physiological feature matrix, organ-level phenotypic parameters, and specified soil parameters as the state input, use the water and fertilizer ratio scheme as the action space, and dynamically optimize the control strategy through the Q-learning algorithm; S6: Implement the water and fertilizer management plan based on the optimized control strategy.

2. The water and fertilizer management control method based on image data processing according to claim 1, characterized in that, In step S1, the visible light image of the crop is collected through the visible light image acquisition module, the multispectral image is collected through the multispectral image acquisition module, and the thermal infrared image is collected through the thermal infrared image acquisition module.

3. A method for controlling water and fertilizer management based on image data processing according to claim 2, characterized in that The specific process of extracting features from the visible light image, multispectral image, and thermal infrared image in step S1 to obtain the visible light image features, multispectral image features, and thermal infrared image features is as follows: S11: Normalize the visible light image, extract the edge features and leaf skeleton features of plant leaves in the visible light image through an edge detection algorithm, generate a point cloud model by combining multi-view visible light images, and obtain the leaf inclination angle, canopy openness features, and fruit / stem features; S12: Use a calibration model to correct the multi-spectral image, use an improved RANSAC algorithm to align the visible light and multi-spectral images, extract features from the 710nm red edge image and 840nm near-infrared light image, calculate the vegetation index features, and combine the 660nm red light image to obtain the normalized vegetation index features; S13: Convert the gray value to absolute temperature through blackbody radiation correction, apply adaptive threshold segmentation to extract the vegetation area, obtain the temperature heat map and the standard deviation feature of the canopy temperature.

4. A water and fertilizer management control method based on image data processing according to claim 3, characterized in that, In step S2, the plant morphological parameters of the crop are calculated through the visible light image features, the vegetation index of the crop is calculated through the multi-spectral image features, and the specific process of obtaining the vegetation canopy temperature distribution by providing the thermal infrared image features is as follows: S21: Calculate the leaf area index, leaf length, leaf width, perimeter, shape factor, petiole length, leaf inclination angle, and canopy openness based on the leaf inclination angle and canopy openness features, plant leaf edge features, and leaf skeleton features, and obtain the fruit size and color based on the fruit / stem features; S22: Calculate the vegetation index NDRE based on the vegetation index features; S23: Calculate the normalized vegetation index NDVI based on the normalized vegetation index features.

5. A method for controlling water and fertilizer management based on image data processing according to claim 1, characterized in that, The specific process of step S5 is as follows: S51: Establish the state space of the water and fertilizer decision model. The state space includes a three-dimensional physiological feature matrix, organ-level phenotypic parameters, and soil specified parameters. The soil specified parameters are collected in real time by a soil moisture sensor. The soil specified parameters include 8-dimensional data of pH value, organic matter content, nutrient content, heavy metal content, microbial biomass, enzyme activity, humidity, and conductivity, and are denoised by Kalman filtering; S52: Create the action space of the water and fertilizer decision model. The action space includes action IDs, water and fertilizer schemes, and the corresponding water and nitrogen, phosphorus, and potassium ratios of the water and fertilizer schemes. The action IDs include 0, 1, 2; 0, 1, 2 correspond to the baseline scheme, high-nitrogen scheme, and water-saving scheme respectively; S53: Create the reward function of the water and fertilizer decision model. The reward function formula is: R= α ( △ BI - BI target ) + β (WUE - WUE baseline ); Among them, R is the reward function value; α and β are crop-related weight coefficients that balance the optimization priorities of biomass growth and resource efficiency; △BI is the biomass increment, which reflects the growth rate of plant biomass per unit time and is calculated as follows: dynamically estimated through the three-dimensional reconstruction data of leaf area index and stem volume; BI target is the target biomass threshold, which is the expected biomass growth value preset according to the crop growth stage and is corrected in real time based on environmental factors through a sliding window algorithm; WUE is the water use efficiency, defined as the ratio of crop yield to irrigation water volume, and maximizes the biomass output per unit water volume by adjusting the irrigation strategy. WUE baseline is the baseline water use efficiency, representing the historical average or industry standard level, serving as the reference benchmark for reward calculation, and reflecting the improvement amplitude of the current strategy relative to the conventional plan; S54: The reward function drives policy optimization through the following mechanism. When the actual biomass increment △BI exceeds the preset target BI target at this time, α a positive reward is generated for the item, β the item ensures that the water use efficiency is not lower than the baseline level, avoiding resource waste caused by over-irrigation. The α / β ratio adjustment realizes the dynamic balance between yield increase and water conservation, meeting the sustainable development requirements of agricultural production.

6. The water and fertilizer management control method based on image data processing according to claim 1, characterized in that, During the image acquisition process, the image acquisition module performs polarized light filtering and polarization angle adjustment. The polarization angle adjustment formula is as follows: θ = arctan( n 2 / n 1) + k ·Δ φ ; Among them, θ is the polarization angle adjustment amount, n 1, n 2 is the refractive index of the medium, Δ φ is the curvature compensation amount of the blade surface, k is the proportionality coefficient, which is used to describe the sensitivity of the polarization angle to the phase difference change.

7. A method for controlling water and fertilizer management based on image data processing according to claim 4, characterized in that When the NDRE value is in the range of 0.35 - 0.42, trigger the hierarchical fertilization strategy: First-level fertilization correction: Execute 80% of the baseline fertilization amount; Second-level fertilization correction: Dynamically adjust by superimposing the detection results of leaf stomatal aperture; Third-level fertilization correction: Perform acid-base balance compensation by combining the soil pH value.

8. A method for controlling water and fertilizer management based on image data processing according to claim 7, characterized in that, The specific hierarchical fertilization strategy is as follows: The calculation formula for the baseline fertilization amount is: Baseline fertilization amount = (fertilizer requirement for target yield - soil fertilizer supply) / fertilizer utilization rate; First-level fertilization correction: Execute 80% of the baseline fertilization amount; Secondary fertilization correction: The detection results of leaf stomatal aperture are superimposed for dynamic adjustment. The leaf stomatal aperture is measured by an infrared thermal imager. When the leaf stomatal aperture is less than 150 mmol, 15% of nitrate nitrogen fertilizer is additionally applied. When the leaf stomatal aperture is between 150 - 350 mmol, ±0% maintains the current amount. When the leaf stomatal aperture is greater than 350 mmol, 10% of ammonium nitrogen fertilizer is reduced; Tertiary fertilization correction: When the soil pH is lower than 6.0, for every 0.5 pH unit decrease, 3 kg / mu of calcium magnesium phosphate fertilizer is additionally applied. When the soil pH is higher than 7.5, for every 0.5 pH unit increase, 1.5 kg / mu of sulfur powder is additionally applied.

Images