Corn plant nitrogen nondestructive monitoring method and system based on spectrum technology

Abstract

The invention relates to the technical field of plant non-destructive monitoring, in particular to a corn plant nitrogen non-destructive monitoring method and system based on a spectrum technology. The method comprises the following steps: determining a spectrum reliability index according to a three-dimensional point cloud model; screening the point cloud points of the hyperspectral reliability index to determine candidate pixel points; determining an optimal NDRE index according to the change trend of the NDRE index of each candidate pixel point in the multispectral image set; generating a predicted NDRE index of a reference pixel point in a preset influence range of the candidate pixel point through spatial interpolation; calculating a difference value between the original NDRE index and the predicted NDRE index, and taking the difference value as a predicted deviation value; determining the NDRE correction of each pixel point in combination with the growth parameter similarity characteristic and the spectral reliability index difference characteristic between the pixel points; and the corrected NDRE index is obtained by using the NDRE correction, and the plant nitrogen content is predicted in combination with the actually measured nitrogen content, so that the accuracy of the nitrogen monitoring effect is improved.

Description

Technical Field

[0001] This invention relates to the field of non-destructive monitoring technology for plants, specifically to a non-destructive monitoring method and system for nitrogen in maize plants based on spectral technology. Background Technology

[0002] To address the problems of excessive water consumption, inefficient irrigation, excessive nitrogen fertilizer input, and unreasonable application timing and dosage in traditional flood irrigation in corn-growing areas, it is necessary to achieve precise control of water and fertilizer integration through IoT devices and multi-model big data. In this process, how to accurately grasp the nitrogen nutrition status of corn plants in real time is a key link to achieve scientific fertilization, improve nitrogen fertilizer utilization, and reduce environmental pollution.

[0003] Currently, traditional nitrogen monitoring relies on destructive sampling and chemical analysis. However, this method is inefficient and cannot meet the real-time needs of large-scale fields. Existing technologies often use non-destructive spectral monitoring technology based on UAV platforms to invert nitrogen content by analyzing the spectral reflectance characteristics of the maize canopy (such as the NDRE index). However, in practical applications, this technology does not fully consider the impact of factors such as differences in plant spatial structure, leaf shading, and shooting angle on the reliability of spectral data, which can easily cause spectral distortion. This distortion is unrelated to changes in the plant's own biochemical composition but seriously interferes with the mapping relationship between spectral indices such as NDRE and the true nitrogen content, resulting in decreased accuracy and insufficient reliability of monitoring results. Summary of the Invention

[0004] To address the technical problem of decreased accuracy and insufficient reliability of monitoring results due to insufficient consideration of the impact of factors such as differences in plant spatial structure, leaf shading, and shooting angle on the reliability of spectral data, this invention provides a non-destructive monitoring method and system for nitrogen in maize plants based on spectral technology. The specific technical solution adopted is as follows: This invention proposes a non-destructive method for monitoring nitrogen levels in maize plants based on spectral technology. The method includes: Multispectral image sets and 3D point cloud data of the maize plant monitoring area were acquired and fused to generate pixel-level digital area maps, in which each pixel is associated with spatial coordinates, multi-band spectral reflectance, and growth parameters of the maize plant to which it belongs. Based on the spatial geometric differences between each point cloud point and its neighboring point cloud points in the 3D point cloud model, the spectral reliability index is determined; point cloud points with hyperspectral reliability index are screened and mapped to multispectral images to determine candidate pixels. The original NDRE index of each pixel is calculated based on the spectral reflectance in the digital region map; the optimal NDRE index is determined according to the NDRE index variation trend of each candidate pixel in the multispectral image set; and the predicted NDRE index of the reference pixel within the preset influence range of the candidate pixel is generated by spatial interpolation using the nearest neighbor method. The difference between the original NDRE index and the predicted NDRE index of the reference pixel is calculated as the prediction deviation value; the NDRE correction amount of each pixel is determined by combining the similarity characteristics of growth parameters and the differences in spectral reliability indicators among pixels. The corrected NDRE index of each pixel is obtained using the NDRE correction amount, and the prediction model is trained by combining it with the measured nitrogen content to predict the nitrogen content of the plant.

[0005] Furthermore, the process of obtaining the growth parameters of the corn plant includes: Using the Otsu thresholding method, the corn canopy region was segmented from the pixel-level digital region map based on the gray-scale distribution characteristics of the original NDRE index. Within the maize canopy area, individual maize plants are segmented based on 3D point cloud data and / or multispectral images. For each plant, one or more growth parameters, including plant height, canopy width, leaf area index, and stem diameter, are calculated. These growth parameters are then assigned to all pixels of the corresponding plant.

[0006] Furthermore, the process for determining the spectral reliability index includes: For each point in the 3D point cloud model, the neighboring point cloud points within a preset neighborhood are obtained with the point cloud point as the center; and the plant height of the point cloud point and its neighboring point cloud points is counted, and the maximum plant height among multiple neighboring point cloud points is extracted. The local flatness index is determined based on the differences in plant height between point cloud points and their neighboring point cloud points. The ratio of the plant height at the point cloud points to the maximum plant height is calculated as an indicator of leaf exposure. The product of the local flatness index and the leaf exposure index is calculated and used as the spectral reliability index.

[0007] Furthermore, the process for determining the local flatness index includes: Calculate the absolute difference between the height of a point in the point cloud and the height of each of its neighboring point cloud points, and use this as the height difference. The exponential function of the opposite of the height difference, with the natural constant as the base, is used as the first exponent; Calculate the arithmetic mean of all first indices as a local flatness indicator.

[0008] Further, the step of screening point cloud points for hyperspectral reliability indicators and mapping them to multispectral images to determine candidate pixels includes: The monitoring area of ​​corn plants was divided into equal-area grid areas, and the corresponding block areas of each grid area in the three-dimensional point cloud model were obtained. For each segmented region, the point cloud points within the segmented region are sorted from largest to smallest according to the spectral reliability index; a preset number of target point cloud points are then selected according to the sorting order. The target point cloud points are mapped onto the multispectral image to determine candidate pixels.

[0009] Furthermore, the process for determining the optimal NDRE index includes: For each candidate pixel, obtain the NDRE index and the corresponding shooting point position on all corresponding images in the multispectral image set; based on the coordinates of the shooting point position and the spatial coordinates of the candidate pixel, establish an deviation vector; Assuming any shooting point is the optimal shooting position, a target deviation vector is formed; based on the angle between the target deviation vector and each deviation vector, the deviation index of the target deviation vector and each deviation vector is obtained through a preset deviation index calculation function; According to the order of deviation index from smallest to largest, the difference between the NDRE index of the corresponding image and the NDRE index of the image associated with the best shooting position is calculated in turn as the NDRE deviation value, and an image deviation sequence is generated. Based on the changing trend of NDRE deviation values ​​in the image deviation sequence, the conformity index is determined; If the absolute value of the conformity index is not less than the preset threshold, the NDRE index of the image corresponding to the best shooting position will be used as the best NDRE index of the candidate pixel.

[0010] Furthermore, the determination of the conformity index based on the changing trend of NDRE deviation values ​​in the image deviation sequence includes: Calculate the absolute difference between adjacent NDRE deviation values ​​in the image deviation sequence as the first difference; calculate the difference between adjacent NDRE deviation values ​​as the second difference. The ratio of the sum of all second differences to the sum of all first differences is used as the compliance index.

[0011] Furthermore, the NDRE correction determination process includes: For any pixel, the growth parameters of the pixel are constructed into a growth vector, and a target reference pixel is determined within the preset spatial neighborhood of the pixel; and the growth parameters of each target reference pixel are constructed into a reference growth vector respectively. Calculate the cosine similarity between the growth vector and each reference growth vector as a growth state similarity index; calculate the difference in spectral reliability index between the target reference pixel and the pixel as a measurement reliability index. The product of the growth state similarity index and the measurement reliability index is normalized to obtain the weights of the pixels and each target reference pixel. Calculate the weighted average of the prediction deviation values ​​of all target reference pixels as the NDRE correction amount.

[0012] Furthermore, the step of obtaining the corrected NDRE index for each pixel using the NDRE correction amount includes: adding the NDRE correction amount to the original NDRE index of each pixel to obtain the corrected NDRE index.

[0013] A non-destructive monitoring system for nitrogen in maize plants based on spectral technology is provided. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of a non-destructive monitoring method for nitrogen in maize plants based on spectral technology.

[0014] The present invention has the following beneficial effects: By introducing a spectral reliability index and calculating based on the spatial geometric differences in 3D point clouds, spectral noise caused by factors such as leaf occlusion and observation angle can be effectively identified, and highly reliable candidate pixels can be selected, thereby improving the accuracy of NDRE index calculation. When determining the optimal NDRE index, the changing trend of multispectral image set and the deviation of shooting position are considered. The NDRE value under the optimal observation conditions is selected through the conformity index, further reducing the impact of environmental interference. Since environmental structural noise is effectively removed, the constructed model learns a purer "spectral-nitrogen" relationship. Therefore, the trained model has strong robustness (anti-environmental interference ability) and generalization ability (adaptability to different plots or growth stages), and can be more accurately applied to maize nitrogen monitoring in different growth stages, different plots, and even different years, solving the problems of overfitting and poor universality of traditional empirical models. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of a non-destructive monitoring method for nitrogen in maize plants based on spectral technology, provided in one embodiment of the present invention. Figure 2 An example diagram illustrating the process of determining spectral reliability indicators provided in one embodiment of the present invention; Figure 3 This is an example diagram illustrating the process for determining the optimal NDRE index according to an embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a non-destructive monitoring method and system for nitrogen in maize plants based on spectral technology proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The following description, in conjunction with the accompanying drawings, details the specific scheme of the non-destructive monitoring method and system for nitrogen in maize plants based on spectral technology provided by this invention.

[0020] Please see Figure 1 The diagram illustrates a flowchart of a non-destructive monitoring method for nitrogen in maize plants based on spectral technology, according to an embodiment of the present invention. The method includes: S101: Acquire multispectral image sets and 3D point cloud data of the maize plant monitoring area, and generate pixel-level digital region maps through fusion processing, in which each pixel is associated with spatial coordinates, multi-band spectral reflectance, and growth parameters of the maize plant to which it belongs.

[0021] To facilitate subsequent spectral reliability analysis, NDRE index optimization, and nitrogen content prediction, it is necessary to construct a unified, accurate, and information-rich dataset (i.e., a pixel-level digital area map) from the complementary raw data from different sensors. For example, under suitable weather conditions (such as sunny days, windless or light breezes at noon), a multi-sensor system mounted on a drone can be used to synchronously collect data on the corn plant monitoring area.

[0022] The sensor system includes at least a multispectral camera: used to acquire reflectivity information of the surface in the monitoring area in multiple specific bands (such as blue light 450nm, green light 550nm, and red light 650nm) to form a multispectral image set; a LiDAR (Light Detection and Ranging) system: used to actively emit laser pulses and receive echoes to directly acquire high-precision three-dimensional spatial coordinates (X, Y, Z) of the surface and corn canopy in the monitoring area to form a three-dimensional point cloud model; and a POS (Positioning and Attitude Determination) system: integrating RTK-GNSS (Real-time Dynamic Differential Global Navigation Satellite System) and IMU (Inertial Measurement Unit) to accurately record the spatial position of the UAV when capturing each frame of image.

[0023] It should be noted that the raw data collected (raw data collected by the sensor system) needs to be preprocessed to eliminate errors. The preprocessing methods are common technical means, which will not be elaborated in this embodiment. For example, radiometric calibration and atmospheric correction are performed on multispectral images to convert the raw DN values ​​(Digital Number) into physically meaningful surface reflectance; noise reduction and filtering are performed on the three-dimensional point cloud data to remove noise points caused by environmental interference (such as dust) or equipment errors.

[0024] To unify the preprocessed raw data into a single coordinate system, as an example, attitude and position data recorded by the POS system are used to perform geometric correction on the raw images acquired by the multispectral camera. This eliminates image distortion caused by changes in UAV attitude, lens distortion, and terrain undulations, generating an orthophoto with accurate geographic reference. Subsequently, multiple orthophotos are stitched together to form a multispectral orthophoto map covering the entire monitoring area. Then, based on the high-precision spatiotemporal labels provided by the POS system, the 3D point cloud data acquired by LiDAR is precisely registered with the multispectral orthophoto to the same geographic coordinate system. Thus, for a pixel in the image, its corresponding position (e.g., coordinates) in 3D space can be found.

[0025] It should be noted that unifying multispectral imagery and 3D point cloud data into the same geographic coordinate system is to determine the pixel-point cloud mapping relationship. For each pixel in the multispectral imagery, one or more point cloud points in the 3D point cloud that are closest to its spatial location are found. For example, a back projection method can be used, that is, using the intrinsic and extrinsic parameters of the multispectral camera, the image pixel coordinates are projected backward into 3D space to form a ray. Then, point cloud points near or intersecting this ray are found in the 3D point cloud as the spatial corresponding point cloud points of the pixel.

[0026] To generate pixel-level digital region maps, as an example, the reflectance values ​​of each pixel in each band are directly extracted from the registered multispectral orthophoto as the core spectral attribute of each pixel; the planar coordinates (X, Y) of each pixel come from the georeferenced image of the orthophoto, and its elevation information (Z) is obtained by interpolation from the registered point cloud data.

[0027] Since the pixel-level digital region map also includes the growth parameters of the corn plant, in this embodiment, the corn canopy region is segmented from the pixel-level digital region map based on the gray-scale distribution characteristics of the original NDRE index using the Otsu thresholding method. Within the corn canopy region, individual corn plants are segmented based on three-dimensional point cloud data and / or multispectral images, and one or more growth parameters, including plant height, canopy width, leaf area index, and stem diameter, are calculated for each plant. The growth parameters are then assigned to all pixels of the corresponding plant.

[0028] The Otsu thresholding method is an unsupervised segmentation algorithm that automatically finds the optimal binarization threshold. Its core logic is to iterate through all possible grayscale thresholds and calculate the inter-class variance of the "foreground" and "background" pixels divided by the first threshold. When the inter-class variance reaches its maximum value, the first threshold is the optimal threshold. At this time, the difference between the foreground (such as the corn canopy area) and the background (such as the soil area) is the most significant, so as to complete the segmentation of the corn canopy area. The Otsu thresholding method is a well-known technique in the art, and the detailed process will not be described in this embodiment.

[0029] For example, based on the calculated original NDRE index, the corn canopy region is segmented from the image using the Otsu thresholding method, and background elements such as soil are removed. Then, point cloud segmentation and / or image segmentation algorithms are used to further identify individual corn plants within the canopy region. For each individual plant, from the 3D point cloud, the plant height (the height difference between the highest point in the point cloud and the ground model), stem diameter (based on the fitting of the stem point cloud), and canopy volume can be extracted. From the 2D spectral image, the canopy width and leaf area index (LAI) can be extracted. Finally, these growth parameters for each plant are assigned to all pixels belonging to that plant. Therefore, in the final pixel-level digital region map, all pixels belonging to the same corn plant will share the same plant-level growth parameters.

[0030] It should be noted that extracting the required growth parameters from 3D point clouds and / or spectral images is a common technique, which will not be elaborated upon in this embodiment.

[0031] S102: Based on the spatial geometric differences between each point cloud point and its neighboring point cloud points in the 3D point cloud model, determine the spectral reliability index; screen the point cloud points with the hyperspectral reliability index and map them to the multispectral image to determine candidate pixel points.

[0032] It should be noted that during the entire growth period of maize, the nitrogen content in the early stage is mainly reflected by the OSAVI index. However, since the maize irrigation area is relatively sparse in the early stage of growth, it is less affected by factors such as differences in plant spatial structure, leaf shading, and shooting angle. Therefore, in order to obtain accurate nitrogen content, this invention only corrects the spectral data of the middle and late stages of maize growth, while the NDRE index is mainly used as the monitoring indicator in the middle and late stages of maize growth.

[0033] It is important to understand that during the mid-to-late stages of corn growth, the plant canopy is already highly closed, and the leaves are heavily layered. Leaves near the bottom and edges of the canopy are often affected by neighboring leaves, stems, and the soil background, making them prone to multiple scattering, shadow superposition, and soil reflection mixing. These factors cause a significant shift in the overall baseline of the spectral curve, obscuring absorption characteristics related to chemical components such as nitrogen. However, leaves located at the top of the canopy and directly exposed to sunlight have a relatively flat surface and simple geometric structure, making them less affected by shading and mixed scattering. Their spectral reflectance is closer to the ideal Lambertian representation, so the baseline shift is significantly reduced, and they can more accurately reflect the nitrogen and chlorophyll content characteristics inside the leaves.

[0034] The process of determining spectral reliability indicators is as follows: Figure 2 As shown, it includes: S102-1: For each point in the 3D point cloud model, obtain the neighboring point cloud points within a preset neighborhood range, with the point cloud point as the center; and count the plant height of the point cloud point and its neighboring point cloud points, and extract the maximum plant height among multiple neighboring point cloud points.

[0035] It is important to understand that when drones scan corn irrigation areas, the uneven leaves curl up, reducing the area of ​​the leaves that the drone can collect. Therefore, most of the point cloud data scanned by the drone is located on the flat leaf surface. The cloud points in the highly concentrated area obtained by the drone scan are often from relatively flat leaf surfaces. Such leaf surfaces have a high degree of smoothness, and there is relatively less multiple scattering of light on them. They are less affected by interference from adjacent structures. Therefore, their spectral curves are less affected by multiplicative noise and baseline drift, and can more accurately reflect the chemical composition characteristics of the leaves themselves.

[0036] It should be noted that the specific value of the preset neighborhood range is determined according to the actual situation, and this embodiment does not make a specific limitation. For example, it is necessary to combine the spatial scale of the corn plant (such as the canopy diameter is usually 0.5-1m) and set the neighborhood as a spherical neighborhood or a cubic neighborhood (spherical neighborhood is commonly used). In order to ensure that the range of 1-3 leaves around the point cloud point is covered, it is necessary to avoid the range being too small to cause insufficient samples, and also to avoid the range being too large to introduce the point clouds of other plants. Therefore, the neighborhood radius is usually taken as 0.1-0.3m.

[0037] It should be noted that in the 3D point cloud data, the plant height of the point cloud is equal to the Z-coordinate (vertical height) of the point cloud and the ground Z-coordinate reference value of the monitoring area. The ground Z-coordinate reference value can be obtained by fitting the ground point cloud (for example, by taking the minimum Z-coordinate of all point clouds or the average Z-coordinate of the ground area).

[0038] S102-2: Determine the local flatness index based on the plant height difference characteristics between point cloud points and their neighboring point cloud points.

[0039] The local flatness index reflects the degree of leaf flatness. Among them, the smaller the difference in plant height between a point cloud and its neighboring point cloud, the flatter the surrounding leaf surface is. This indicates that the leaves are arranged regularly without obvious protrusions / indentations, and the less structural interference is encountered during spectral acquisition, the higher the reliability.

[0040] In this embodiment, the absolute difference between the height of a point cloud point and the height of each neighboring point cloud point is calculated as the height difference; an exponential function of the opposite of the height difference is calculated with the natural constant as the base, as the first exponent; and the arithmetic mean of all the first exponents is calculated as the local flatness index.

[0041] Since a smaller height difference indicates a very small difference in plant height between the point cloud point and all neighboring points, it is highly likely that the top of the canopy is a flat area with uniform local surface height, minimal interference from structural undulations in spectral acquisition, and excellent flatness. Therefore, the larger the first index, the larger the local flatness index. Thus, the local flatness index can be expressed by the following formula: in, represents the local flatness index; C represents the total number of point cloud points in the neighborhood of a point cloud point. This represents the height difference between a point in the point cloud and its i-th neighboring point in the point cloud. This represents an exponential function with the natural constant as its base.

[0042] It should be noted that when the neighborhood range is preset, the selection of an appropriate number of neighborhood point cloud points is already taken into account. Therefore, the total number of neighborhood point cloud points cannot be zero.

[0043] S102-3: Calculate the ratio of the plant height of the point cloud to the maximum plant height, as an indicator of leaf exposure.

[0044] It is important to understand that the larger the plant height of the cloud point, the closer it will be to the maximum plant height. This reflects that the closer the plant height of the cloud point is to the local highest canopy height (such as the top leaves of the canopy), the less area of ​​the cloud point is blocked by other leaves, and the more direct sunlight it receives. At this time, the spectral reflectance can truly reflect the chlorophyll status of the leaves, and the exposure is high.

[0045] It should be noted that in actual planting scenarios, corn plants cannot be without growth height; that is, the maximum plant height cannot be zero.

[0046] S102-4: Calculate the product of the local flatness index and the leaf exposure index as a spectral reliability index.

[0047] It should be noted that, based on the spectral reliability index, it is possible to determine which part of the corn plant the corresponding point cloud point is located in. For example, if the spectral reliability index of a certain point cloud point is greater than 0.8, it is very likely to correspond to the flat leaf area at the top of the canopy. In this case, the spectral data representing the location of the point cloud point is more reliable.

[0048] In this embodiment, the corn plant monitoring area is divided into equal-area grid areas, and the corresponding block areas in the three-dimensional point cloud model are obtained for each grid area; for each block area, the point cloud points in the block area are sorted from largest to smallest according to the spectral reliability index; a preset number of target point cloud points are selected according to the sorting order; the target point cloud points are mapped to the multispectral image to determine candidate pixel points.

[0049] It should be noted that the grid side length can be set by combining the plant spacing or row spacing of corn (for example, the plant spacing of conventional corn is 0.2-0.5m and the row spacing is 0.6-0.8m) and the point cloud density (for example, the point cloud density of lidar is 50-200 points / ㎡). This ensures that each grid contains at least 1-3 complete corn plants. For example, if the corn row spacing in the monitoring area is 0.7m and the plant spacing is 0.3m, the grid can be set as a "0.7m×0.3m" rectangle (consistent with the direction of the planting rows). The area of ​​a single grid is about 0.25-0.21㎡, which is just enough to cover 1-2 corn plants.

[0050] To accurately map to the segmented regions, as an example, we traverse all point cloud points in the 3D point cloud model and extract the planar coordinates (X, Y) of each point; we determine whether the (X, Y) of each point falls within the planar coordinate range of a certain grid; and we group all point cloud points that fall within the range of that grid to form the point cloud segmented region corresponding to that grid.

[0051] It should be noted that the specific value of the preset quantity is determined according to actual needs, and this embodiment does not make a specific limitation. For example, the preset quantity is determined according to the number of point cloud points in the segmented area. Usually, the preset quantity is ensured to account for 5%-10% of the total number of point cloud points in the segmented area. Assuming that the total number of point cloud points in the segmented area is large, such as 200, then the preset quantity is in the range of 5-10.

[0052] Based on the point cloud point-image pixel mapping relationship established in the aforementioned steps (implemented through camera intrinsic parameters, extrinsic parameters, and back projection algorithms), the three-dimensional coordinates of each point cloud point can be calculated to obtain its pixel coordinates on the multispectral image through projection.

[0053] S103: Calculate the original NDRE index of each pixel based on the spectral reflectance in the digital region map; determine the optimal NDRE index based on the NDRE index variation trend of each candidate pixel in the multispectral image set; generate the predicted NDRE index of reference pixels within the preset influence range of the candidate pixels through nearest neighbor spatial interpolation.

[0054] It is important to understand that because drones scan corn irrigation areas continuously, multiple spectral images exist for any given location. The spectral curves of the same location differ in these different images. When the leaves are unobstructed and unaffected by neighboring leaves or soil, if the drone camera's observation direction is nearly perpendicular to the leaf surface, the camera receives more reflected light energy, resulting in a higher overall amplitude of the spectral curve. Simultaneously, due to a shorter scattering path and less mixing interference, the shape of the spectral curve is more stable, and absorption valleys are easier to identify. Therefore, as the observation angle (camera shooting angle) gradually deviates from the normal direction of the leaf surface (i.e., becomes "more oblique"), the effective signal received by the sensor weakens, while more scattered light (noise) from the background or other leaves is received. This leads to a decrease in the overall spectral reflectance and distortion of the spectral shape. This indicates that the NDRE index systematically weakens as the observation angle deviates from the normal, reflecting the true optical and physical properties of the leaves.

[0055] It should be noted that the specific method for calculating the NDRE index based on spectral reflectance is a well-known technique in the art, and will not be described in detail in this embodiment.

[0056] The process of determining the optimal NDRE index is as follows: Figure 3 As shown, it includes: S103-1: For each candidate pixel, obtain the NDRE index and the corresponding shooting point position on all corresponding images in the multispectral image set; establish an deviation vector based on the coordinates of the shooting point position and the spatial coordinates of the candidate pixel.

[0057] It's important to understand that, with the help of LiDAR and RTK-GNSS, we not only have the pixel coordinates of the image, but also the precise three-dimensional geographic coordinates of each point cloud point and each drone shooting location. With these precise three-dimensional coordinates, a clear spatial vector, or offset vector, can be constructed from a point on the plant canopy to the optical center of the drone camera lens. Then, the angle between the two offset vectors is calculated, realizing the conversion from physical data to digital data. This step transforms the light propagation path and angular relationship in the physical world into a calculable spatial vector and angle in the digital world through precise spatial coordinates, enabling quantitative analysis and reproduction of the impact of observation geometry on spectral data within a computer.

[0058] The shooting point location refers to the spatial coordinates of the camera when the corresponding image is captured, for example, represented by (x, y, z).

[0059] The deviation vector is a vector pointing from the shooting point to the candidate pixel. It precisely describes the shooting angle, and the direction of the deviation vector represents the direction from which the camera took the image (e.g., directly above, left front, right rear).

[0060] S103-2: Assuming any shooting point is the optimal shooting position, a target deviation vector is formed; based on the angle between the target deviation vector and each deviation vector, the deviation index of the target deviation vector and each deviation vector is obtained through a preset deviation index calculation function.

[0061] Since candidate pixels correspond to multiple spectral images, a strategy of making assumptions one by one can be adopted to find the best shooting angle. For example, if the shooting position of the first image in the image set is assumed to be the "best shooting position", then the deviation vector from this "best position" to the candidate pixel is the target deviation vector. Then, the angle between the target deviation vector and the deviation vectors of all other images is obtained.

[0062] Since the size of the included angle represents the degree of difference between "other shooting angles" and the assumed "optimal angle," the smaller the included angle, the closer the shooting angles are and the less obvious the deviation. Therefore, the deviation index can be represented by the following preset deviation index calculation function: Where F represents the deviation index; This represents the angle between the target deviation vector and the deviation vector.

[0063] It should be noted that in real-world scenarios, since the directions of all deviation vectors are basically pointing towards the canopy region, the angles between the target deviation vector and each other deviation vector are... It is most likely an acute angle or a right angle, therefore, The value range is [0, 1].

[0064] S103-3: Calculate the difference between the NDRE index of the corresponding image and the NDRE index of the image associated with the best shooting position in ascending order of deviation index, and use it as the NDRE deviation value to generate an image deviation sequence.

[0065] It is understandable that the image at the very front has the closest shooting angle to our assumed "optimal shooting angle," while the image at the very back has the greatest difference in shooting angle.

[0066] Image deviation sequence describes how the NDRE index changes as the shooting angle gradually changes from the assumed "optimal shooting angle". Generally speaking, for stable and reliable spectral data, the trend of the NDRE index should be smooth.

[0067] S103-4: Determine the compliance index based on the changing trend of NDRE deviation values ​​in the image deviation sequence.

[0068] In this embodiment, the absolute difference between adjacent NDRE deviation values ​​in the image deviation sequence is calculated as the first difference; the difference between adjacent NDRE deviation values ​​is calculated as the second difference; and the ratio of the sum of all second differences to the sum of all first differences is calculated as the compliance index.

[0069] The conformity index is used to measure the smoothness or stability of a deviation sequence.

[0070] Based on the actual situation analysis, the relationship between the sum of all second differences and the sum of all first differences always satisfies the following: the absolute value of the sum of all second differences never exceeds the sum of all first differences. Therefore, the range of the conformity index is [-1, 1]. Thus, the larger the absolute value of the ratio, the closer the sum of all second differences is to the sum of all first differences, and the more uniform the trend of the image deviation sequence (monotonically increasing or decreasing). The more stable the NDRE deviation value changes with the shooting angle, the higher the conformity.

[0071] It should be noted that since the image deviation sequence is "NDRE deviation value under different shooting angles", and the change of shooting angle will cause changes in the incident angle of light on corn leaves and slight changes in the shading relationship, it is understandable that even if the NDRE deviation values ​​are very close (e.g. [2.01, 2.02, 2.01]), there will be a first difference due to the slight difference. Therefore, the sum of all the first differences cannot be zero.

[0072] S103-5: If the absolute value of the conformity index is not less than the preset threshold, the NDRE index of the image corresponding to the best shooting position is taken as the best NDRE index of the candidate pixel.

[0073] It should be noted that the specific value of the preset threshold is determined according to the actual situation, and this embodiment does not impose a specific limitation. For example, if the preset threshold is 0.8, and the absolute value of the compliance index is not less than 0.8, it is obvious that the trend is stable enough regardless of whether it is increasing or decreasing, then the selection of the best shooting point is reliable.

[0074] The reference pixel is the target for optimization of the candidate pixel. Since the candidate pixel has high spectral reliability, the best NDRE index of the candidate pixel can be used as a "benchmark" to correct the original NDRE index of the reference pixel and improve the accuracy of the reference pixel.

[0075] It should be noted that the specific value of the preset influence range is determined according to actual needs, and this embodiment does not impose a specific limitation. For example, a 3×3 neighborhood.

[0076] For example, the reference pixel with the smallest Euclidean distance is found by the nearest neighbor method, and the best NDRE index of the candidate pixel is assigned to the reference pixel as the predicted NDRE index of the reference pixel.

[0077] S104: Calculate the difference between the original NDRE index and the predicted NDRE index of the reference pixel as the prediction deviation value; combine the similarity characteristics of growth parameters and the differences in spectral reliability indicators among pixels to determine the NDRE correction amount for each pixel.

[0078] It's important to understand that in the spatial distribution of crop nitrogen, soil and plant nutrient status typically exhibit spatial autocorrelation, meaning that nitrogen content in adjacent areas is often similar and doesn't fluctuate drastically within small ranges. Therefore, in a two-dimensional spectral image, the nitrogen content at a representative location can be used as a predicted value for pixels within its neighborhood, thus inferring changes in nitrogen content in maize plants in actual agricultural scenarios.

[0079] It should be noted that a reference pixel has two NDRE values: one is the original NDRE index (calculated directly from the spectral reflectance of the reference pixel's location); the other is the predicted NDRE index (assigned by spatial interpolation from the best NDRE index of the nearest candidate pixel).

[0080] It is important to understand that the predicted NDRE index represents a relatively reliable theoretical expectation based on the principle of spatial autocorrelation. Therefore, a positive prediction deviation value indicates that the original measurement value of the reference pixel is higher than the expected level of its region, and there may be an overestimation; a negative prediction deviation value indicates that the original measurement value of the reference pixel is lower than expected, and there may be an underestimation due to occlusion, shadows, etc.

[0081] In this embodiment, for any pixel, the growth parameters of the pixel are constructed into a growth vector, and a target reference pixel is determined within a preset spatial neighborhood of the pixel; the growth parameters of each target reference pixel are constructed into a reference growth vector; the cosine similarity between the growth vector and each reference growth vector is calculated as a growth state similarity index; the difference between the spectral reliability index of the target reference pixel and the pixel is calculated as a measurement reliability index; the product of the growth state similarity index and the measurement reliability index is normalized to obtain the weights of the pixel and each target reference pixel; the weighted average of the prediction deviation values ​​of all target reference pixels is calculated as the NDRE correction amount.

[0082] For example, parameters that are strongly correlated with the nitrogen level of maize and can reflect the growth status of the plant need to be selected. If three parameters are selected, namely "plant height, crown width, and leaf density", the growth vector of a certain pixel can be represented as: growth vector V=[plant height H, crown width W, leaf density D]=[2.1m, 0.6m, 120 points / ㎡].

[0083] The target reference pixel is the pixel within the preset spatial neighborhood of the current pixel that has a highly reliable NDRE index prediction, that is, the pixel assigned the best NDRE index.

[0084] It should be noted that the specific value of the preset spatial neighborhood is determined according to the actual situation. This embodiment does not make a specific limitation. For example, it needs to be combined with the spatial scale of corn plants (such as plant spacing of 0.2-0.5m and row spacing of 0.6-0.8m). It is usually set as a 5×5 or 7×7 pixel neighborhood centered on the current pixel.

[0085] It is important to understand that, since the growth state similarity index is used to measure the directional consistency between the growth vector and each reference growth vector, in actual agricultural scenarios, since all growth parameters are positive (plant height and crown width cannot be negative), the actual value range of the growth state similarity index is [0, 1]. Therefore, the greater the cosine similarity, the more similar the growth states of the two plants are, and the type and degree of interference experienced by the spectral measurement values ​​of the pixel and the target reference pixel may also be similar, and the higher the reference value of the target reference pixel.

[0086] It is important to understand that in actual agricultural scenarios, the spectral reliability index of the target reference pixel is not less than the spectral reliability index of the pixel. Therefore, the measurement reliability index must be greater than zero. Thus, if the index is smaller, the reliability of the spectral acquisition environment of the pixel and the target reference pixel is more similar, reflecting that the "reliability advantage" of the target reference pixel is not obvious, and the correction weight needs to be reduced.

[0087] As an example, the prediction bias value of each target reference pixel is multiplied by the weight to obtain the contribution value, and then the arithmetic mean of all weighted values ​​is calculated to obtain the NDRE correction amount of the current pixel.

[0088] For example, suppose the target reference pixels for a certain pixel are A and B, where target reference pixel A has a prediction bias of -0.15, a weight of 0.65, and a contribution of -0.15 × 0.65 ≈ -0.0975; and target reference pixel B has a prediction bias of -0.12, a weight of 0.35, and a contribution of -0.12 × 0.35 ≈ -0.042. Then, the NDRE correction for this pixel is -0.0975 + (-0.042) ≈ -0.1395 (approximately -0.14).

[0089] In this embodiment, the NDRE correction amount is added to the original NDRE index of each pixel to obtain the corrected NDRE index.

[0090] It's important to understand that the sign of the NDRE correction already includes the direction of the deviation of the original NDRE index. If the NDRE correction is positive, it means that the original NDRE index is lower than the true value (for example, the spectral reflectance of the lower layer pixels of the corn canopy is underestimated due to shading by the upper leaves). In this case, the original NDRE index needs to be increased. If the NDRE correction is negative, it means that the original NDRE index is higher than the true value (for example, the spectral reflectance of the top pixels of the corn canopy is overestimated due to direct strong light). In this case, the original NDRE index needs to be decreased.

[0091] S105: The corrected NDRE index of each pixel is obtained by using the NDRE correction amount, and the prediction model is trained by combining the measured nitrogen content to predict the nitrogen content of the plant.

[0092] It is important to understand that, based on the corrected NDRE index, a brand-new calibrated NDRE index map with spatial consistency correction can be generated. The values ​​on this calibrated NDRE index map have eliminated interference such as baseline drift caused by the complexity of the canopy structure to the greatest extent, and can more realistically reflect the nitrogen nutrition status inside the plant.

[0093] To train the prediction model, a training dataset can be constructed. Instead of using a single NDRE value for each pixel (or the plant region represented by a pixel), a multi-dimensional feature vector can be constructed as the model's input to improve model accuracy. This multi-dimensional feature vector includes at least the corrected NDRE index (the most important feature), other indices calculated from multi-band spectral reflectance (such as the NDVI index), and plant growth parameters. Furthermore, each feature vector needs a corresponding "answer," namely, the measured nitrogen content (obtained through traditional destructive sampling methods to collect leaf or plant samples from the corresponding area and perform chemical analysis in a laboratory). Each feature vector is then paired with a measured nitrogen content value to form a complete training dataset.

[0094] It should be noted that the model training method is a well-known technique in the art, and will not be described in detail in this embodiment. For example, the model can use various machine learning or deep learning algorithms, and then train the prediction model based on the training dataset.

[0095] It's important to understand that a trained prediction model can predict the nitrogen content for each pixel. These predictions are then used to generate an intuitive spatial distribution map of nitrogen, where different colors can indicate nitrogen abundance or deficiency. This map enables various downstream applications. For example, agricultural managers can generate variable fertilizer prescription maps based on the nitrogen spatial distribution map, guiding fertilizer applicators to apply more fertilizer in nutrient-deficient areas and less or no fertilizer in nitrogen-sufficient areas.

[0096] A non-destructive monitoring system for nitrogen in maize plants based on spectral technology is disclosed. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of a non-destructive monitoring method for nitrogen in maize plants based on spectral technology.

[0097] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0098] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A method for non-destructive monitoring of nitrogen status in corn plants based on spectroscopy, characterized in that, The method comprises: Obtaining a multi-spectral image set and three-dimensional point cloud data of a corn plant monitoring area, and generating a pixel-level digital region map through fusion processing, wherein each pixel point is associated with a spatial coordinate, a multi-band spectral reflectance, and a growth parameter of the corn plant to which it belongs; According to the spatial geometric difference characteristics between each point cloud point and its neighborhood point cloud points in the three-dimensional point cloud model, a spectral reliability index is determined; point cloud points with high spectral reliability indexes are screened and mapped to the multi-spectral image to determine candidate pixel points; Based on the spectral reflectance in the digital region map, the original NDRE index of each pixel point is calculated; according to the NDRE index variation trend of each candidate pixel point in the multi-spectral image set, the optimal NDRE index is determined; through nearest neighbor method spatial interpolation, the predicted NDRE index of the reference pixel points within the preset influence range of the candidate pixel points is generated; The difference between the original NDRE index and the predicted NDRE index of the reference pixel points is calculated as the prediction bias value; combined with the growth parameter similarity characteristics and the spectral reliability index difference characteristics between the pixel points, the NDRE correction amount of each pixel point is determined; The corrected NDRE index of each pixel point is obtained by using the NDRE correction amount, and a prediction model is trained combined with the measured nitrogen content to predict the nitrogen content of the plant.

2. The method for non-destructive monitoring of nitrogen status in maize plants based on spectroscopy according to claim 1, characterized in that, The process of obtaining the growth parameters of the corn plant comprises: Through Otsu threshold method, the corn canopy area is segmented from the pixel-level digital region map based on the gray distribution characteristics of the original NDRE index; In the corn canopy area, based on the three-dimensional point cloud data and / or the multi-spectral image, the individual corn plants are segmented, and one or more growth parameters including plant height, crown width, leaf area index and stem diameter are counted for each plant, and then the growth parameters are assigned to all pixel points of the corresponding plant.

3. The method for non-destructive monitoring of nitrogen status in maize plants based on spectroscopy according to claim 2, characterized in that, The process of determining the spectral reliability index comprises: For each point cloud point in the three-dimensional point cloud model, the neighborhood point cloud points within a preset neighborhood range are obtained with the point cloud point as the center, and the plant heights of the point cloud point and its neighborhood point cloud points are counted, and the maximum plant height among the neighborhood point cloud points is extracted; According to the plant height difference characteristics between the point cloud point and each of its neighborhood point cloud points, a local flatness index is determined; The ratio of the plant height of the point cloud point to the maximum plant height is calculated as a leaf exposure index; The product of the local flatness index and the leaf exposure index is calculated as the spectral reliability index.

4. The method according to claim 3, wherein, The process of determining the local flatness index comprises: The absolute difference of the plant heights of the point cloud point and each of its neighborhood point cloud points is calculated as a height difference value; The exponential function of the inverse of the height difference value is calculated with a natural constant as the base as a first index; The arithmetic mean of all first indexes is calculated as the local flatness index.

5. The method for non-destructive monitoring of nitrogen status in maize plants based on spectroscopy according to claim 1, characterized in that, The process of screening the point cloud points with high spectral reliability indexes and mapping them to the multi-spectral image to determine the candidate pixel points comprises: Divide the corn plant monitoring area into equal-area grid regions, and obtain the corresponding block regions in the three-dimensional point cloud model for each grid region; For each block region, sort the point cloud points in the block region according to the spectral reliability index from large to small; select a preset number of target point cloud points according to the sorting order; Map the target point cloud points to the multi-spectral image to determine the candidate pixel points.

6. The method for non-destructive monitoring of nitrogen status in maize plants based on spectroscopy according to claim 1, characterized in that, The optimal NDRE index determination process comprises: For each candidate pixel point, the NDRE index and the corresponding shooting point position on all corresponding images in the multi-spectral image set are obtained; and a deviation vector is established based on the coordinates of the shooting point position and the spatial coordinates of the candidate pixel point; Assuming that any shooting point position is the optimal shooting position, a target deviation vector is formed; and a deviation index of the target deviation vector and each deviation vector is obtained through a preset deviation index calculation function based on the included angle between the target deviation vector and each deviation vector; According to the arrangement order from small to large of the deviation index, the difference between the NDRE index of the corresponding image and the NDRE index of the image associated with the optimal shooting position is calculated as the NDRE deviation value, and an image deviation sequence is generated; The conformity index is determined based on the change trend of the NDRE deviation value in the image deviation sequence; If the absolute value of the conformity index is not less than a preset threshold, the NDRE index of the image corresponding to the optimal shooting position is taken as the optimal NDRE index of the candidate pixel point.

7. The method according to claim 6, wherein, The conformity index is determined based on the change trend of the NDRE deviation value in the image deviation sequence, comprising: The absolute difference between adjacent NDRE deviation values in the image deviation sequence is calculated as a first difference value; and the difference between adjacent NDRE deviation values is calculated as a second difference value; The ratio of the sum of all second difference values to the sum of all first difference values is calculated as the conformity index.

8. The method for non-destructive monitoring of nitrogen status in maize plants based on spectroscopy according to claim 1, characterized in that, The NDRE correction amount determination process comprises: For any pixel point, the growth parameter of the pixel point is constructed as a growth vector, and target reference pixel points are determined in a preset spatial neighborhood of the pixel point; and the growth parameter of each target reference pixel point is constructed as a reference growth vector; The cosine similarity of the growth vector and each reference growth vector is calculated as a growth state similarity index; and the difference between the spectral reliability index of the target reference pixel point and the pixel point is calculated as a measurement reliability index; The product of the growth state similarity index and the measurement reliability index is normalized to obtain the weight of the pixel point and each target reference pixel point; The weighted average of the predicted deviation values of all target reference pixel points is calculated as the NDRE correction amount.

9. The method according to claim 8, wherein, The modified NDRE index of each pixel point is obtained by adding the NDRE correction amount to the original NDRE index of each pixel point.

10. A non-destructive monitoring system for nitrogen status of corn plants based on spectroscopy technology, characterized in that, The system comprises a memory, a processor, and a computer program stored in the memory and running on the processor, and the processor implements the steps of the method according to any one of claims 1-9 when executing the computer program.

Images