A method and system for evaluating crop growth status based on low-altitude agricultural remote sensing image data analysis

Abstract

The application provides a crop growth state evaluation method and system for low-altitude agricultural remote sensing image data analysis, comprising: deploying an imaging unit and an inertial measurement attitude unit on a low-altitude collection platform, collecting low-altitude images, recording the shooting time and the imaging pose, extracting the plant surface observation points and the leaf color appearance, and combining the imaging pose to solve the initial spatial coordinates of the observation points; comparing the spatial positions according to the stable spatial layout and the initial spatial coordinates, identifying the observation points deviating from the stable spatial layout and confirming them as a set of false plant points; extracting the canopy top height and the adjacent row exposed ground height from the three-dimensional morphology of the plant, determining the plant height and calculating the canopy fluctuation amplitude, and determining the canopy coverage degree according to the area ratio of the canopy horizontal projection range and the range of the row exposed ground.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for assessing crop growth status through analysis of low-altitude agricultural remote sensing image data. Background Technology

[0002] Low-altitude agricultural remote sensing technology is a core supporting means for modern precision agriculture. It uses drones equipped with imaging devices to acquire farmland images, enabling refined monitoring and assessment of crop growth status. In practical applications, 3D reconstruction technology can accurately retrieve crop height information, providing crucial data for yield prediction and field management. 3D reconstruction relies on repeated observations of the same plant surface across overlapping adjacent images. By obtaining the imaging position of the same observation point in multiple images, a correspondence between different images is established, and then the spatial coordinates of the observation point are calculated to complete the 3D reconstruction. Therefore, the number of overlapping images covering the same plant and the stability of the imaging position of the observation point directly determine the completeness and accuracy of the reconstruction.

[0003] Current mainstream reconstruction methods primarily rely on increasing the forward overlap rate of adjacent images to increase the number of images covering the same plant, ensuring reconstruction integrity. To obtain sufficient overlapping observations, the industry standard practice is to increase the forward overlap rate of adjacent images, from the conventional 60% to 70% or even 80%. Increased overlap means the same plant will appear in more images, theoretically providing more observation angles and denser surface observation points, thus improving reconstruction integrity and accuracy. This static setting is particularly effective in mapping rigid features such as buildings and roads, as these targets remain in a fixed position throughout the shooting time. However, under natural wind forces, crop canopy leaves continuously sway. When the forward overlap rate increases, the shooting time required for the drone to cover the same plant lengthens, and the imaging position of the same observation point shifts in images at different times. With increased forward overlap, the drone needs to continuously capture more images to cover the same plant, and these images are acquired at more dispersed times along the timeline, resulting in observation points that originally belonged to the same plant surface exhibiting different spatial coordinates in multiple images.

[0004] For example, under the action of a crosswind of 5 meters per second, the top of a corn plant may sway laterally by 20 to 30 centimeters. When the overlap rate increases from 60% to 80%, the time span for acquiring images covering the plant increases from 2 seconds to 3.5 seconds. The difference in the position of the plant in the first and last images may exceed the tolerance range of the reconstruction, resulting in breaks or false protrusions in the 3D reconstruction.

[0005] How to effectively identify and eliminate observation points that drift due to wind-induced swaying and false plant points that deviate from the true shape of the plant, while ensuring sufficient overlapping observations of the same plant, and so that the plant height inversion curve remains continuous and complete while avoiding false features, has become a key issue in assessing crop growth status using low-altitude agricultural remote sensing. Summary of the Invention

[0006] This invention provides a method for assessing crop growth status based on low-altitude agricultural remote sensing image data analysis, the method comprising:

[0007] An imaging unit and an inertial attitude measurement unit are deployed on a low-altitude acquisition platform to acquire low-altitude images, record the shooting time and imaging pose, extract observation points on the plant surface and leaf color appearance, and calculate the initial spatial coordinates of the observation points by combining the imaging pose.

[0008] Based on the low-altitude image and the shooting time, the imaging position movement of each observation point between adjacent low-altitude images is detected, and the observation points whose imaging position movement is greater than the preset movement threshold are identified as the unstable plant point set.

[0009] The unstable plant point set is removed from the initial spatial coordinates, and the remaining initial spatial coordinates are used to construct a stable spatial layout of the farmland crop rows and the ground between rows. The plant spacing and row spacing are obtained by counting the distance between adjacent plants and the distance between crop rows.

[0010] Based on the comparison of the stable spatial layout and the initial spatial coordinates, the observation points that deviate from the stable spatial layout are identified and confirmed as a set of false plant points.

[0011] False plant point sets are removed from the initial spatial coordinates. The three-dimensional morphology of the plants is reconstructed by combining the remaining initial spatial coordinates with low-altitude imagery. The horizontal projection range of the canopy and the range of bare ground between rows are extracted from each crop row area.

[0012] The height of the canopy top and the height of the bare ground between rows are extracted from the three-dimensional morphology of the plant to determine the plant height and calculate the canopy undulation amplitude. The canopy coverage is determined by the area ratio of the horizontal projection range of the canopy to the area of ​​the bare ground between rows.

[0013] Check the continuous distribution of the canopy along the plant spacing and row spacing to mark the locations of missing seedlings and broken rows, and determine the growth consistency rate by combining the numerical difference between plant height and leaf color.

[0014] Furthermore, the deployment of an imaging unit and an inertial attitude measurement unit on the low-altitude acquisition platform to acquire low-altitude images and record the shooting time and imaging pose includes:

[0015] The imaging unit uses an area array visible light sensor to continuously acquire low-altitude images of farmland along the flight path. The inertial attitude measurement unit combines a three-axis gyroscope, a three-axis accelerometer, and a global navigation satellite receiver to read the attitude angle and position coordinates of the platform at the moment of shooting. Continuous exposure is triggered according to the preset heading overlap rate. The exposure time is synchronized by the platform's internal clock to the attitude reading of the inertial attitude measurement unit and written into the image metadata field.

[0016] Furthermore, the extraction of plant surface observation points and leaf color appearance, combined with the calculation of the initial spatial coordinates of the observation points based on the imaging pose, includes:

[0017] For the image region where the crop canopy is located in the low-altitude image, the scale-invariant feature transform algorithm is used to extract the leaf tip, leaf vein bifurcation and leaf edge turning point as plant surface observation points. For the same observation point, the correspondence between corresponding points is established between adjacent low-altitude images through descriptor matching. The average pixel intensity of the red, green and blue channels in the neighborhood of the observation point is read as the leaf color appearance. The initial spatial coordinates are obtained by collinear equation intersection solution based on the imaging pose and the focal length, principal point and distortion parameters of the imaging unit.

[0018] Furthermore, the method of detecting the imaging position shift of each observation point between adjacent low-altitude images based on low-altitude imagery and the shooting time includes:

[0019] According to the chronological order of the shooting time, the corresponding relationship of the same point is retrieved between adjacent low-altitude images. The image-side coordinates of the same observation point in the low-altitude images at the previous and subsequent times are combined into temporal image-side coordinate pairs. According to the imaging pose, the image-side coordinates at the previous time are back-projected to the image plane at the subsequent time using the collinearity equation to obtain the image-side coordinates after pose compensation. The Euclidean distance between the image-side coordinates and the actual image-side coordinates at the subsequent time is used as the imaging position movement range. Observation points that exceed the preset movement threshold are classified into the unstable plant point set.

[0020] Furthermore, the step of removing unstable plant point sets from the initial spatial coordinates and constructing a stable spatial layout of crop rows and inter-row ground using the remaining initial spatial coordinates includes:

[0021] The remaining initial spatial coordinates are divided into stable parts of the plant canopy and stable parts of the ground according to the height value. Based on the planar projection distribution of the stable parts of the plant canopy and stable parts of the ground on the horizontal plane, the alternating rows of farmland crops and the ground between the rows are identified. The planar distance between the center projection points of the stable parts of the canopy of adjacent plants in the same crop row is calculated as the plant spacing along the extension direction of the crop row. The vertical distance between the center lines of the extension of two adjacent crop rows is used as the row spacing.

[0022] Furthermore, the step of comparing the spatial positions based on the stable spatial layout and the initial spatial coordinates to identify observation points that deviate from the stable spatial layout and confirm them as a set of false plant points includes:

[0023] Based on the distribution range of the extended center line of the crop row and the ground between the rows in the stable spatial layout, the permissible distribution area of ​​the plants is delineated on each crop row along the crop row direction with the plant spacing as the interval, and the permissible distribution area between adjacent crop rows is delineated with the row spacing as the interval, and the two constitute the permissible distribution map of the location.

[0024] Compare the horizontal projection coordinates of each observation point in the initial spatial coordinates with the positional allowable distribution map, and mark the observation points that exceed the plant's allowable distribution area and the row's allowable distribution area as horizontal deviation observation points.

[0025] The height difference between the horizontal deviation observation point and the average height of the stable ground part is read. Horizontal deviation observation points whose height difference is higher than the maximum height value of the stable part of the plant canopy are identified as false plant points and included in the set of false plant points.

[0026] Furthermore, the step of deleting the set of false plant points from the initial spatial coordinates and reconstructing the three-dimensional morphology of the plant using the remaining initial spatial coordinates in conjunction with low-altitude imagery includes:

[0027] The observation point numbers of the false plant point set are traversed, and the corresponding three-dimensional coordinate records are deleted from the initial spatial coordinates to obtain the effective initial spatial coordinate set. The observation points in the effective initial spatial coordinate set are matched with dense corresponding points using multi-view stereo matching. The dense corresponding points are combined with the imaging pose and substituted into the collinear equation for forward intersection to obtain the dense plant surface point cloud. The dense plant surface point cloud and the effective initial spatial coordinate set are merged to form the three-dimensional shape of the plant.

[0028] Furthermore, the step of extracting the canopy top height and the height of bare ground between rows from the three-dimensional morphology of the plant, determining the plant height and calculating the canopy undulation amplitude, and determining the canopy coverage by the area ratio of the horizontal projection range of the canopy to the area of ​​bare ground between rows includes:

[0029] The dense plant surface point cloud with a height higher than the preset canopy height threshold within the same crop row area in the three-dimensional morphology of the plant is classified into the canopy top point set and the top height value is read. The average height of the point cloud within the bare ground area between adjacent rows is taken as the average ground height. The difference between the top height value and the average ground height is taken as the plant height. The difference between the maximum and minimum top height values ​​within the canopy top point set is taken as the canopy undulation amplitude. The sum of the closed area of ​​the horizontal projection range of the canopy and the closed area of ​​the bare ground area between adjacent rows is taken as the reference area. The ratio of the closed area of ​​the horizontal projection range of the canopy to the reference area is taken as the canopy coverage.

[0030] Furthermore, the step of checking the continuous distribution of the canopy along the plant spacing and row spacing to mark the locations of missing seedlings and gaps in the rows includes:

[0031] Plant sampling points are set sequentially along the crop rows in the farmland at the plant spacing. Each plant sampling point is checked to see if it is covered by the horizontal projection range of the canopy. If two or more consecutive adjacent plant sampling points are not covered by the horizontal projection range of the canopy, the row segment where the line connecting the consecutive uncovered plant sampling points is located is marked as the location of the missing seedlings and broken rows.

[0032] Furthermore, the determination of growth uniformity rate by combining the numerical difference between plant height and leaf color appearance includes:

[0033] The average pixel intensity of the three channels corresponding to each plant is read as the leaf color sample value, and the plant height of the plant is read as the plant height sample value. The absolute value of the difference between the plant height sample value of each plant and the median of the plant height sample values ​​of the whole field is calculated. The absolute value of the difference between the leaf color sample value of each plant and the median of the leaf color sample values ​​of the whole field is calculated. The two are weighted and summed to obtain the consistency difference of the plant. The proportion of the number of plants with a consistency difference lower than the preset consistency threshold to the total number of plants in the whole field is determined as the growth consistency rate. Plant height, canopy undulation amplitude, canopy coverage, missing seedling and broken row position and growth consistency rate are summarized and archived according to crop row area number.

[0034] This invention provides a crop growth status assessment system for low-altitude agricultural remote sensing image data analysis, comprising an acquisition module, an observation point processing module, a statistical module, an identification module, an extraction module, and an output module.

[0035] The acquisition module is used to deploy imaging units and inertial attitude measurement units on the low-altitude acquisition platform to acquire low-altitude images, record the shooting time and imaging pose, extract observation points on the plant surface and leaf color appearance, and calculate the initial spatial coordinates of the observation points in combination with the imaging pose.

[0036] The observation point processing module is used to detect the imaging position movement of each observation point between adjacent low-altitude images based on low-altitude images and shooting time, and to identify the observation points whose imaging position movement is greater than a preset movement threshold as a set of unstable plant points.

[0037] The statistics module is used to remove the set of unstable plant points from the initial spatial coordinates, construct a stable spatial layout of the farmland crop rows and the ground between rows with the remaining initial spatial coordinates, and calculate the plant spacing and row spacing by counting the distance between adjacent plants and the distance between crop rows.

[0038] The identification module is used to compare the spatial position with the initial spatial coordinates based on the stable spatial layout, identify observation points that deviate from the stable spatial layout and confirm them as a set of false plant points;

[0039] The extraction module is used to remove the set of false plant points from the initial spatial coordinates, and to reconstruct the three-dimensional morphology of the plants by combining the remaining initial spatial coordinates with low-altitude imagery. It also extracts the horizontal projection range of the canopy and the range of bare ground between rows within each crop row area.

[0040] The output module is used to extract the height of the top of the canopy and the height of the bare ground between rows from the three-dimensional morphology of the plant, determine the plant height and calculate the canopy undulation amplitude, and determine the canopy coverage by the area ratio of the horizontal projection range of the canopy to the area of ​​the bare ground between rows; check the continuous distribution of the canopy along the plant spacing and row spacing to mark the location of missing seedlings and broken rows, and determine the growth consistency rate by combining the numerical difference between the plant height and the leaf color appearance.

[0041] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0042] This invention discloses a method for assessing crop growth status based on low-altitude agricultural remote sensing image data analysis. It acquires image and pose data by deploying imaging units and inertial attitude measurement units on a low-altitude acquisition platform, calculates the initial spatial coordinates of observation points, and extracts leaf color appearance. Addressing the dynamic interference and false point cloud problems in low-altitude remote sensing data, this invention innovatively introduces a dual filtering mechanism of temporal motion analysis and spatial layout constraints. First, it identifies and removes unstable plant points caused by factors such as wind by detecting the amplitude of imaging position movement of observation points between adjacent images. Second, it uses the remaining stable points to construct a spatial layout model of the crop rows and the ground between rows to statistically analyze plant and row spacing. Finally, it identifies and deletes false plant points that deviate from the layout through spatial position comparison. Reliable spatial coordinates after double filtering are used to reconstruct the three-dimensional morphology of plants. The height of the top of the canopy and the height of the bare ground are extracted to calculate the plant height and the canopy undulation. The coverage is determined based on the horizontal projection range of the canopy and the range of bare ground between rows. The continuity of the canopy is checked along the plant and row spacing, the location of missing seedlings and broken rows is marked, and the growth consistency rate is determined by combining the difference between plant height and leaf color values. This achieves a high-precision assessment of crop growth status. Attached Figure Description

[0043] Figure 1 This is a flowchart of the crop growth status assessment method based on low-altitude agricultural remote sensing image data analysis according to the present invention.

[0044] Figure 2 This is a schematic diagram of the crop growth status assessment method based on low-altitude agricultural remote sensing image data analysis according to the present invention.

[0045] Figure 3 This is another schematic diagram of the crop growth status assessment method based on low-altitude agricultural remote sensing image data analysis according to the present invention. Detailed Implementation

[0046] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0047] Example 1

[0048] like Figures 1-3 The crop growth status assessment method based on low-altitude agricultural remote sensing image data analysis in this embodiment may specifically include:

[0049] S101. Deploy an imaging unit and an inertial attitude measurement unit on a low-altitude acquisition platform to acquire low-altitude images, record the shooting time and imaging pose, extract observation points on the plant surface and leaf color appearance, and calculate the initial spatial coordinates of the observation points in combination with the imaging pose.

[0050] An imaging unit and an inertial attitude measurement unit are deployed on the airframe of the low-altitude acquisition platform. The imaging unit uses an area array visible light sensor to continuously acquire low-altitude images of farmland along the flight path. The inertial attitude measurement unit uses a combination of a three-axis gyroscope, a three-axis accelerometer, and a global navigation satellite receiver to read the platform's attitude angle and position coordinates at the moment of shooting. The imaging unit is triggered to continuously expose according to a preset heading overlap rate. The exposure time is synchronized to the attitude reading of the inertial attitude measurement unit by the platform's internal clock and written into the image metadata field to obtain the corresponding record of the low-altitude image, the shooting time, and the imaging attitude.

[0051] For the image region where the crop canopy is located in the low-altitude image, the scale-invariant feature transform algorithm is used to extract the leaf tip, leaf vein bifurcation and leaf edge turning point as plant surface observation points. For the same observation point, the corresponding relationship of the same point is established between adjacent images through descriptor matching. The average pixel intensity of the red, green and blue channels is read in the neighborhood window of the observation point as the leaf color appearance record. If the descriptor distance is lower than the preset matching threshold, the corresponding relationship of the same point is retained; otherwise, it is discarded.

[0052] Based on the attitude angle and position coordinates in the imaging pose, and the focal length, principal point, and distortion parameters of the imaging unit, a collinearity equation is used to construct the projection constraint between the image-side coordinates and the object-side coordinates of the observation point. The image-side coordinates of the corresponding points in multiple low-altitude images are substituted into the collinearity equation for intersection calculation. The three-dimensional position of each observation point on the plant surface in the object-side space is solved as the initial spatial coordinates. The initial spatial coordinates are bound and archived one by one with the shooting time and the leaf color appearance according to the observation point index.

[0053] The starting point for the 3D reconstruction process of low-altitude agricultural remote sensing is to acquire original images and pose data with spatiotemporal consistency. The low-altitude acquisition platform is usually a multi-rotor UAV or a fixed-wing UAV, with a gimbal or shock-absorbing bracket installed under the airframe to support the imaging unit and the inertial attitude measurement unit. The relative positions of the two units are precisely calibrated before flight and kept rigidly connected.

[0054] In one possible implementation, the imaging unit employs an area array visible light sensor with a resolution of at least 2 megapixels. The lens focal length is pre-selected based on the operating altitude and ground resolution requirements. The ground resolution R is calculated as R = H × a / f, where H is the operating altitude, a is the sensor pixel size, and f is the lens focal length. For example, when operating at an altitude of 70 meters and a pixel size of 0.0045 mm, a 20 mm fixed-focus lens is used to obtain a ground sampling distance of 1.575 cm. The sensor continuously acquires the low-altitude images along a preset flight path at fixed exposure intervals, maintaining preset forward overlap and lateral overlap rates between adjacent images. The forward overlap rate is set to 75% to 85%, and the lateral overlap rate is set to 65% to 75%. The inertial attitude measurement unit is composed of a three-axis gyroscope, a three-axis accelerometer, and a global navigation satellite receiver. A three-axis gyroscope is used to read the platform's attitude angles around the roll, pitch, and yaw axes at the moment of shooting. A three-axis accelerometer assists in compensating for attitude drift. A global navigation satellite receiver uses real-time dynamic differential mode to obtain the platform's position coordinates in the geocentric-ground-fixed coordinate system. To meet the requirements of subsequent collinearity equation solving and spatial coordinate archiving, the obtained geocentric-ground-fixed coordinates need to be converted to the local northeast-sky coordinate system or the national standard projected coordinate system, thereby achieving a unified spatial reference for different data sources.

[0055] It should be noted that time synchronization is a prerequisite for ensuring a strict correspondence between pose and image. Each exposure trigger signal is issued by the platform's internal clock, and the pose reading of the inertial attitude measurement unit is synchronized at the same trigger moment, writing the attitude angle, position coordinates, and exposure time into the image's metadata field. The shooting time accuracy is controlled at the millisecond level for subsequent time-series tracking of observation points with the same name.

[0056] Specifically, for the image region containing the crop canopy in the low-altitude image, a scale-invariant feature transform algorithm is used to extract observation points on the plant surface. This algorithm first constructs a multi-scale Gaussian pyramid on the image, then searches for local extrema as candidate feature points in the difference images of adjacent scales. Subsequently, low-contrast points and edge response points are removed, retaining stable local extrema. In crop canopy scenes, stable extrema typically fall at leaf tips, vein bifurcation points, and leaf edge transitions. These locations have distinct texture features and stable geometric shapes, making them suitable as observation points on the plant surface.

[0057] For example, hundreds of such observation points can be extracted from the canopy region of a single corn plant in a single image, covering the top and side leaves of the canopy.

[0058] In one embodiment, a neighborhood window is drawn centered on the pixel coordinates of each observation point on the plant surface. The window size W is dynamically determined based on the image ground resolution G and the average physical width L of the target crop leaves, calculated using the formula W=L / G. For example, when the average physical width L of a corn leaf is 0.08 meters and the ground resolution G corresponding to the UAV's flight altitude is 0.005 meters per pixel, the calculated window size W is 16 pixels, i.e., a 16x16 pixel neighborhood window is drawn. This dynamic size ensures accurate coverage of the single leaf area under different flight altitudes and crop types. The average pixel intensity of the red, green, and blue channels within the window is read as the leaf color appearance record for that observation point. The leaf color appearance reflects the differences in crop growth and also serves as an auxiliary criterion in the corresponding point matching process, used to distinguish adjacent leaves with similar shapes but different colors. The correspondence between the same observation point and adjacent images is established through descriptor matching. The scale-invariant feature transformation algorithm generates a 128-dimensional oriented gradient histogram descriptor for each feature point, and the similarity between descriptors of adjacent images is measured by Euclidean distance. If the distance between two descriptors is less than the preset matching threshold of 0.6, the two feature points are determined to be corresponding points and retained; if the distance is greater than the preset matching threshold of 0.6, the corresponding relationship is removed to avoid incorrect matching being passed on to subsequent calculation stages.

[0059] Specifically, for the same observation point on the plant surface, the correspondence between its corresponding points in multiple low-altitude images provides multiple sets of image-space coordinates. Using these image-space coordinates as observations and the three-dimensional coordinates of the observation point in the object space as unknowns, the multiple sets of image-space coordinates, along with the imaging pose of the corresponding images, are substituted into the collinearity equation. The collinearity equation is linearized using Taylor expansion, constructing an error equation V=AX-L containing corrections for the unknowns, where V is the image-space coordinate residual vector, A is the partial derivative design matrix, X is the object-space three-dimensional coordinate correction vector, and L is a constant term vector. The observation weight matrix P is set according to the feature point extraction accuracy; for example, the weight of each observation is set to 1.0. An indirect adjustment model is used to construct the normal equation and solve for X. The object-space three-dimensional coordinates are updated iteratively until the modulus of X is less than 0.001 meters, thereby obtaining the unique three-dimensional position of the plant surface observation point in the object space, i.e., the initial spatial coordinates. The initial spatial coordinates, the shooting time corresponding to the observation point, and the leaf color appearance are bound one-to-one according to the observation point index to form an observation point attribute table, which is then archived and stored. Each record in the observation point attribute table includes the observation point number, three-dimensional coordinates, shooting time sequence, and three-channel color mean, serving as the original data basis for subsequent determination of observation point stability and construction of crop row spatial layout.

[0060] It is understandable that, through the joint deployment of the imaging unit and the inertial attitude measurement unit and the time synchronization mechanism, strict registration of low-altitude images and pose data on a millisecond-level time scale is achieved, providing a traceable spatiotemporal reference for the extraction of observation points and coordinate calculation of non-rigid targets such as crop canopies.

[0061] S102. Based on the low-altitude image and the shooting time, detect the imaging position movement of the observation point corresponding to each initial spatial coordinate between adjacent low-altitude images, and determine the observation point whose imaging position movement is greater than the preset movement threshold as the unstable plant point set.

[0062] S1021. For the plant surface observation point corresponding to the initial spatial coordinates, search for the corresponding relationship of the same point between adjacent low-altitude images according to the order of the shooting time, and combine the image-side coordinates of the same plant surface observation point in the low-altitude image at the previous time and the low-altitude image at the next time into a temporal image-side coordinate pair.

[0063] S1022. Simultaneously read the imaging poses bound to each of the two low-altitude images to obtain the association record between the temporal image coordinate pairs and the imaging poses.

[0064] S1023. Based on the imaging pose in the associated record, the image-side coordinates of the low-altitude image at the previous moment are projected in reverse to the image plane of the low-altitude image at the next moment using the collinearity equation to obtain the image-side coordinates of the plant surface observation point after pose compensation. The Euclidean distance between the image-side coordinates after pose compensation and the actual image-side coordinates on the low-altitude image at the next moment is calculated as the imaging position movement amplitude to obtain the imaging position movement amplitude record for each plant surface observation point.

[0065] S1024. Regarding the recording of the imaging position movement amplitude, if the imaging position movement amplitude of a certain observation point on the plant surface is greater than a preset movement threshold, then the observation point on the plant surface is marked as an unstable plant point and included in the unstable plant point set, thus obtaining the unstable plant point set.

[0066] Crop canopies sway continuously under natural wind forces. The imaging position of the same observation point on the surface of a plant in two adjacent low-altitude images is affected not only by changes in platform pose but also by additional image-side displacement caused by leaf swaying. The observation point on the previous image is reprojected using the relative pose parameters between adjacent images to compensate for the pose. The residual displacement D between the compensated coordinates and the actual observed coordinates in the current image is calculated. When the residual displacement D is greater than a preset movement threshold T, the observation point is determined to be a drifting observation point due to swaying, thus forming a set of unstable plant points.

[0067] In one possible implementation, for each plant surface observation point corresponding to the initial spatial coordinates, a corresponding relationship of corresponding points is retrieved between adjacent low-altitude images according to the chronological order of the shooting time. This corresponding relationship is obtained by extracting and matching feature descriptors from adjacent low-altitude images. The image-side coordinates of the same plant surface observation point in the low-altitude image at the previous time are designated as the first coordinate point, and the image-side coordinates of the same plant surface observation point in the low-altitude image at the subsequent time are designated as the second coordinate point. The first coordinate point and the second coordinate point are then combined to form the temporal image-side coordinate pair.

[0068] Specifically, firstly, the flight logs recorded by the flight control system are analyzed to extract the numbers of each observation point and their corresponding imaging poses, thereby establishing an observation point attribute table. Next, the image-side coordinates extracted from low-altitude images of the same ground target at two different times are paired to generate temporal image-side coordinate pairs. Subsequently, the imaging poses bound to each of the two low-altitude images are read from the observation point attribute table. These imaging poses include three attitude angles and three position coordinates at the moment of capture. Finally, the temporal image-side coordinate pairs and the two sets of imaging poses are organized into an associated record according to the observation point numbers.

[0069] Preferably, the Euclidean distance between the pose-compensated image-side coordinates and the actual image-side coordinates on the low-altitude image at the next moment is calculated. Let the x-coordinate and y-coordinate of the pose-compensated image-side coordinates be u1 and v1, respectively, and the x-coordinate and y-coordinate of the actual image-side coordinates on the low-altitude image at the next moment be u2 and v2, respectively. The formula for calculating the Euclidean distance d is d = ((u1 - u2)). 2 +(v1-v2) 2 ) 0.5 The Euclidean distance value represents the imaging position movement amplitude. This amplitude reflects the residual image-side displacement after subtracting the platform motion component, directly corresponding to the leaf's oscillation response between two exposures. The imaging position movement amplitude of each observation point on the plant surface is summarized according to the observation point number to obtain the imaging position movement amplitude record.

[0070] In one embodiment, the preset movement threshold T is determined by the ground resolution GSD and the wind force level w of the operating environment, where GSD is the ground distance per unit pixel and w is the wind force level value. The basic threshold T0 is set to 3 to 5 pixels. When the wind force level is less than or equal to level 3, the T0 value is used; when the wind force level is level 4 to 5, the upper limit of T0 is increased by 1.2 to 1.5 times; and when the wind force level is greater than level 5, the upper limit of T0 is increased by 1.5 to 2 times. The average object depth value Zc is the sum of the UAV's flight altitude H and the average ground altitude Hg. round The difference, i.e., Zc=HH ground H is read from the position coordinates of the inertial attitude measurement unit. groundThe prior value of the ground altitude of the operational area is obtained from a publicly available digital altitude model (DEM) or from pre-flight ground control point measurements.

[0071] For example, when the operating altitude is 70 meters and the ground sampling distance (GSD) is 0.02 meters per pixel, the threshold for lateral sway amplitude of 0.06 to 0.1 meters corresponding to 3 to 5 pixels is taken when the wind force is below level 3. When the wind force is level 4 to 5, the upper limit of the threshold is expanded to 0.12 meters corresponding to 6 pixels. For the recorded imaging position movement amplitude, each observation point on the plant surface is traversed. If the imaging position movement amplitude of a certain observation point is greater than the preset movement threshold, the observation point is marked as an unstable plant point and included in the unstable plant point set, thereby realizing the centralized identification of observation points that have image-space drift due to canopy leaf sway.

[0072] S103. Remove the set of unstable plant points from the initial spatial coordinates, and assign the remaining initial spatial coordinates to stable locations to construct a stable spatial layout of farmland crop rows and the ground between rows. Calculate the spacing between adjacent plants and the spacing between adjacent crop rows along the crop row direction to obtain the plant spacing and row spacing.

[0073] S1031. For the initial spatial coordinates, traverse the observation point number of each unstable plant point in the set of unstable plant points, delete the three-dimensional coordinate record bound to the corresponding observation point number from the initial spatial coordinates, and use the remaining three-dimensional coordinate record as the stable initial spatial coordinates. Divide the stable initial spatial coordinates into stable parts of the plant canopy and stable parts of the ground according to the height value.

[0074] S1032. Based on the planar projection distribution of the stable parts of the plant canopy and the stable parts of the ground on the horizontal plane, identify the rows of farmland crops and the ground between the rows, which are arranged alternately on the horizontal plane to form a stable spatial layout.

[0075] S1033. Based on the extension direction of the crop row in the stable spatial layout as the crop row direction, calculate the plane distance between the center projection points of the stable canopy parts of adjacent plants in the same crop row as the plant spacing along the crop row direction, and take the vertical distance between the extension center lines of two adjacent crop rows as the row spacing to obtain the plant spacing and the row spacing.

[0076] The crop rows in the farmland are arranged in regular strips, with bare ground between the rows, forming the most distinctive spatial pattern in the low-altitude remote sensing scene. After removing observation points affected by wind drift, the spatial distribution of the remaining observation points retains the true geometric shape of the crop rows and the ground between them, providing a reliable spatial benchmark for subsequent statistics on plant spacing and row spacing.

[0077] In one possible implementation, an observation point attribute table is first constructed, containing fields such as observation point number, 3D coordinate X / Y / Z values, acquisition timestamp, and associated image number. For the initial spatial coordinates, the observation point number of each unstable plant point in the unstable plant point set is traversed, and the 3D coordinate record bound to that observation point number is retrieved from the observation point attribute table and deleted.

[0078] The remaining three-dimensional coordinate records are aggregated into stable initial spatial coordinates. These stable initial spatial coordinates are stratified according to their height values. First, the height distribution of all stable initial spatial coordinates is statistically analyzed, and the median height H is calculated. median and height standard deviation σ H H median σ represents the median height of all stable coordinates. H The standard deviation represents the height. Then, the stratification threshold H is calculated. threshold =H median +k×σ H H median σ represents the median height of all stable coordinates. H The standard deviation represents the height, and k is the stratification coefficient, ranging from 0.5 to 1.0.

[0079] The stratification coefficient k is determined as follows: In a calibration plot with known crop rows and inter-row ground distribution, the height distribution of stable plant canopy locations and stable ground locations is statistically analyzed. The height of the intersection of these two distributions is used as the optimal stratification threshold. The value of k is then calculated by reverse calculation, and the average of multiple calibration results is taken as the final value of k. In the corn planting scenario, k is preferably set to 0.5. If the height H of a certain stable initial spatial coordinate is greater than H... threshold If H is less than or equal to H, then the coordinate is assigned to a stable part of the plant canopy; threshold Then it is classified as a stable part of the ground.

[0080] For the initial spatial coordinates, the observation point number of each unstable plant point in the unstable plant point set is traversed. The three-dimensional coordinate record bound to that observation point number is retrieved from the observation point attribute table and deleted. The remaining three-dimensional coordinate records are aggregated into stable initial spatial coordinates. The stable initial spatial coordinates are divided into layers according to their height values. After layering, the average height of all coordinates of the stable parts of the ground is used as the reference surface for subsequent processing.

[0081] It should be noted that the construction of a stable spatial layout relies on the projection distribution characteristics of the stable parts of the plant canopy and the stable parts on the ground onto a horizontal plane. All initial stable spatial coordinates are projected onto the horizontal plane, forming a two-dimensional point set. The projection points of the stable parts of the plant canopy exhibit a densely clustered, strip-like distribution on the horizontal plane, with the stripes aligned with the sowing direction; the projection points of the stable parts on the ground form a low-density, continuous strip-like distribution between the stripes.

[0082] Density-based clustering is used to group projection points on the horizontal plane. The clustering process uses the horizontal distance between points as a similarity measure. The preset neighborhood radius is determined based on the crop row spacing, ranging from 0.2 to 0.4 times the row spacing. In corn planting scenarios, the row spacing is typically 60 to 80 centimeters, corresponding to a preset neighborhood radius of 12 to 32 centimeters. Projection points with a horizontal distance less than the preset neighborhood radius are grouped into the same initial cluster. Strip feature recognition is performed on all initial clusters. The aspect ratio L / W and the angle θ between the main direction and the sowing direction are calculated for each initial cluster, where L is the length of the long side of the minimum bounding rectangle of the cluster point set, and W is the length of the short side. When L / W is greater than or equal to 3.0 and θ is less than or equal to 15 degrees, the initial cluster is determined to have strip features. All initial clusters exhibiting strip-like characteristics are projected along the sowing direction onto a cross-section perpendicular to the sowing direction. Initial clusters whose center positions differ by less than 1.5 times the preset neighborhood radius on the cross-section are merged into the same strip-like cluster group. Each strip-like cluster group is identified as a crop row. Initial clusters that do not meet the strip-like characteristic criteria and isolated points that are not clustered are identified as ground areas between rows. The crop rows and the ground areas between rows are arranged alternately on the horizontal plane, together forming a stable spatial layout.

[0083] In one embodiment, the extension direction of the crop rows in the farmland is obtained by fitting a straight line using the least squares method to all projection points within each linear cluster.

[0084] Specifically, linear regression is performed on the set of projection point coordinates {(xi,yi)} within the strip-shaped cluster to solve for the linear equation y=kx+b that minimizes the sum of squared residuals, where k is the slope. The direction vector of the fitted line is the crop row direction. The direction vectors of multiple crop rows in the field are averaged using an angle averaging method to obtain the unified crop row direction for the entire field. Specifically, first, the direction vector of each crop row is converted into an angle θi with the horizontal direction; then, the arithmetic mean θavg of all angles is calculated; finally, θavg is converted back into a direction vector as the unified crop row direction. When the angle of a crop row's direction vector differs from the angles of other crop rows by more than 15 degrees, it is considered an outlier and removed before the angle averaging calculation is performed.

[0085] Preferably, the plant spacing is statistically analyzed along the crop row direction. The first clustering uses the DBSCAN algorithm to cluster projected points into strip-shaped clusters to identify the crop row. The neighborhood radius ε is set to the preset neighborhood radius, i.e., 0.2 to 0.4 times the row spacing, corresponding to 12 to 32 centimeters in a corn planting scenario. The minimum number of points, MinPts, is set to 50 to 100. The value of MinPts is based on the fact that in a crop row area with complete canopy coverage, the number of observation points that can be extracted per unit length (1 meter) is typically 100 to 200. 50% to 100% of this number is taken as MinPts to ensure the continuity of the clustering results. For the stable canopy areas of the plants within the same crop row, a second clustering is performed to identify individual plants. A distance-threshold-based clustering algorithm is used to cluster projected points with a distance of less than 0.15 to 0.25 meters along the crop row direction into a cluster, with each cluster representing one plant. The arithmetic mean of the coordinates of all projection points in each cluster is calculated as the centroid, which is the central projection point of the plant. Along the crop row direction, let the unit vector in this direction be v. The planar distance d between two adjacent central projection points P1 and P2 is calculated as the projected length of the line connecting the two points along the crop row direction; that is, d equals the absolute value of the dot product of the vector from P1 to P2 and v. The planar distances between all adjacent central projection points within the same crop row are counted, sorted from smallest to largest, and the median is taken as the plant spacing.

[0086] Specifically, row spacing is calculated perpendicular to the crop row direction. A local coordinate system is established using the crop row direction angles obtained earlier. All projection points within each crop row are fitted with an extended centerline parallel to the crop row direction along the horizontal axis of this coordinate system. Adjacent extended centerlines are parallel straight lines, and the constant vertical distance between them serves as a sample row spacing. The median of all adjacent row spacing samples across the entire field is taken as the final row spacing value. The plant spacing and row spacing describe the planting intervals of crops within and between rows, respectively, and are used for subsequent crop growth analysis and yield estimation.

[0087] Understandably, by removing unstable plant points whose imaging position shifts beyond a preset movement threshold T due to wind swaying, and by layering the remaining coordinates into stable canopy and ground-level regions based on a height difference threshold, and then constructing a stable spatial layout through horizontal plane clustering of each layer, a precise restoration of the crop row geometry is achieved. The plant spacing and row spacing accurately reflect the planting patterns of the field. This method, combined with layering the point cloud into stable canopy and ground-level regions based on a height difference threshold, and then constructing a stable spatial layout through horizontal plane clustering of each layer, achieves a precise restoration of the crop row geometry. The plant spacing and row spacing accurately reflect the planting patterns of the field.

[0088] S104. Based on the stable spatial layout and the initial spatial coordinates, compare the spatial positions to identify observation points that deviate from the stable spatial layout and confirm them as a set of false plant points.

[0089] Based on the distribution range of the extended center line of the crop rows and the ground between the rows in the stable spatial layout, the permissible distribution area of ​​the plants is delineated on each crop row along the direction of the crop row at intervals of the plant spacing, and the permissible distribution area between adjacent crop rows is delineated at intervals of the row spacing. The permissible distribution area of ​​the plants and the permissible distribution area between the rows together constitute a permissible distribution map.

[0090] For each observation point in the initial spatial coordinates, its horizontal projected coordinates are compared with the spatial position of the positional allowable distribution map. If the horizontal projected coordinates of the observation point fall within the plant allowable distribution area or the inter-row allowable distribution area, the observation point is retained. If the horizontal projected coordinates of the observation point exceed both the plant allowable distribution area and the inter-row allowable distribution area, the observation point is marked as a horizontal deviation observation point.

[0091] For the horizontal deviation observation point, the height difference between its height value and the average height of the stable ground part is read. If the height difference is higher than the maximum height value of the stable part of the plant canopy, the horizontal deviation observation point is identified as a false plant point that is suspended outside the stable spatial layout and is included in the false plant point set, thus obtaining the false plant point set.

[0092] After motion feature filtering based on time-series images, the image-side drift of observation points caused by wind-induced swaying has been initially eliminated. However, a few observation points still fall outside the true plant morphology due to random errors caused by multi-view image intersection calculation. The multi-view image intersection calculation refers to calculating three-dimensional coordinates by intersecting spatial rays from images from different perspectives. Its random errors mainly come from fluctuations in image matching accuracy and camera calibration residuals, causing some calculated points to deviate from the geometric pattern of crop rows and form suspended false points.

[0093] The identification of false plant sites requires a stable spatial layout as a reference and is completed through spatial location comparison. Specifically, after motion feature filtering based on time-series images, the image drift of observation points caused by wind-induced swaying has been initially eliminated, but there are still a few observation points that fall outside the true plant morphology due to random errors caused by multi-view image intersection calculation.

[0094] The multi-view image intersection solution refers to calculating three-dimensional coordinates by intersecting spatial rays from images at different viewpoints. Its random errors mainly originate from fluctuations in image matching accuracy and camera calibration residuals, causing some solution points to deviate from the geometric pattern of the crop rows, forming suspended false points. Identification of false plant points requires using a stable spatial layout as a reference, and is completed through spatial position comparison and height verification using a location tolerance distribution map.

[0095] In one possible implementation, the location tolerance distribution map is constructed using the extended center lines of the crop rows in the stable spatial layout as a framework. Along the direction of the crop rows, on the extended center lines of each crop row, plant center projection points are marked at intervals equal to the plant spacing. A permissible plant distribution area is delineated around each plant center projection point with a preset plant spacing tolerance as the radius. The preset plant spacing tolerance is the maximum projection radius *r* of the stable part of the plant canopy on the horizontal plane in the stable spatial layout. When the ratio of the projection radius of the plant canopy perpendicular to the crop row direction to that along the crop row direction is between 0.8 and 1.2, the permissible plant distribution area is circular with a radius of *r*.

[0096] When the ratio is less than 0.8 or greater than 1.2, the permissible distribution area of ​​the plants is elliptical, with the major and minor axes corresponding to the maximum projection radii in two directions, respectively. The width of the permissible distribution area of ​​the plants in the direction perpendicular to the crop row is defined as 2r. The horizontal strip between the extended center lines of two adjacent crop rows is defined as the permissible distribution area between rows, with the middle strip remaining after subtracting the width of 4r on both sides from the row spacing. The permissible distribution area between rows covers a reasonable distribution range of bare ground.

[0097] The permissible plant distribution areas are distributed sequentially along the crop row direction at plant spacing intervals. The permissible distribution areas between rows are located between the permissible plant distribution areas of adjacent crop rows. The two alternately cover the entire field, together forming a permissible positional distribution map. The preset plant spacing tolerance is equal to the maximum projection radius of the stable part of the plant canopy on the horizontal plane in the stable spatial layout; the preset row spacing tolerance is equal to 0.5 times the width of the horizontal strip area, which is determined based on the row spacing; the plant spacing tolerance and row spacing tolerance are independent of each other and are used to determine whether the positional deviation of the plant in the in-row direction and the inter-row direction is within the allowable range, respectively.

[0098] For each observation point in the initial spatial coordinates, its projected coordinates on the horizontal plane are read, and the projected coordinates are compared point by point with the permissible distribution map. In practical applications, both the permissible distribution area of ​​the plant and the permissible distribution area between rows are represented by polygon boundaries, and the comparison method uses geometric determination that the point is inside the polygon.

[0099] If the projected coordinates of an observation point fall within the permissible distribution area of ​​the plants, it is marked as a normal plant point and retained for subsequent 3D reconstruction; if it falls within the permissible distribution area between rows, it is marked as a normal ground point and retained in the same way; if the projected coordinates of the observation point fall neither within the permissible distribution area of ​​the plants nor within the permissible distribution area between rows, it is determined that the observation point deviates from the stable spatial layout, is classified into the set of false plant points, and is removed in subsequent processing. Through the above comparison, the classification and labeling of all observation points are completed, and the observation points marked as false plant points are the target point set to be identified in step S104.

[0100] Preferably, if the horizontal projected coordinates of an observation point exceed the boundary between the permissible distribution area of ​​the plant and the permissible distribution area between rows, that is, if no permissible area on the horizontal plane covers the projected coordinates, then the observation point is marked as a horizontal deviation observation point and included in the candidate list of horizontal deviation observation points.

[0101] In one embodiment, each observation point in the initial spatial coordinates is automatically assigned a unique number according to the acquisition sequence. For each horizontal deviation observation point in the candidate list, its height value is read and subtracted from the average height of the stable ground position to obtain the height difference Δh. If Δh is higher than the maximum height value of the stable plant canopy position, the horizontal deviation observation point is located above the top of the real canopy in the vertical direction and is considered a suspended point. Its observation point number and three-dimensional coordinates are recorded in the set of false plant points. If Δh is between the maximum height values ​​of the stable ground position and the stable plant canopy position, the horizontal deviation observation point is classified as a small obstacle or weed in the field, retained in the initial spatial coordinates but marked as a non-false plant point, and not included in the set of false plant points. After the set of false plant points is completed, the centralized marking of suspended observation points deviating from the stable spatial layout is achieved. By constraining the horizontal projection coordinates with the location tolerance distribution map and verifying the height values, dual identification of false plant points caused by random errors in intersection calculation was achieved. This process eliminated interference data that deviated from the true plant morphology for subsequent 3D morphological reconstruction.

[0102] S105. Delete the set of false plant points from the initial spatial coordinates, and reconstruct the three-dimensional morphology of the plants by combining the remaining initial spatial coordinates with low-altitude images. Extract the horizontal projection range of the canopy and the range of bare ground between rows in each crop row area.

[0103] S1051. For the initial spatial coordinates, traverse the observation point number of each false plant point in the set of false plant points, delete the three-dimensional coordinate record bound to the observation point number from the initial spatial coordinates, and form a valid initial spatial coordinate set with the remaining three-dimensional coordinate records. Each observation point in the valid initial spatial coordinate set retains its corresponding shooting time and leaf color appearance record.

[0104] S1052. Based on the image-side coordinates of each observation point in the effective initial spatial coordinate set corresponding to the low-altitude image, a multi-view stereo matching method is used to perform dense matching of corresponding points in the neighborhood pixel region of the observation point between adjacent low-altitude image pairs. The dense corresponding points are combined with the imaging pose and substituted into the collinearity equation to perform forward intersection to obtain a dense plant surface point cloud. The dense plant surface point cloud and the effective initial spatial coordinate set are merged to form the three-dimensional morphology of the plant.

[0105] S1053. Based on the distribution range of the extended center line of the crop row in the stable spatial layout and the ground between the rows, the dense plant surface point cloud in the three-dimensional morphology of the plant is assigned to the corresponding crop row area according to the horizontal projection coordinates. The outer polygon of the horizontal projection coordinates of all the dense plant surface point clouds in the same crop row area is taken as the horizontal projection range of the canopy. The horizontal surface area between adjacent crop row areas that is not covered by the horizontal projection range of the canopy is taken as the bare ground area between the rows, thus obtaining the horizontal projection range of the canopy and the bare ground area between the rows.

[0106] The accuracy of 3D reconstruction from low-altitude imagery directly determines the degree of restoration of crop canopy geometry. Initial observation points are extracted from multiple viewpoints using image feature matching algorithms. These feature points are automatically distributed across all visible surfaces with distinct textures in the scene, including the top and side structures of the crop canopy, as well as exposed ground between plants. The effective observation points, after removing false plant points, cover the true locations of the crop canopy and the ground. Expanding sparse observation points into a dense point cloud through multi-view stereo matching is a crucial step in obtaining the 3D morphology of the plants.

[0107] In one possible implementation, the observation point attribute table includes an observation point number field and a 3D coordinate field, where the observation point number is a unique identifier, and the 3D coordinate field stores the X, Y, and Z coordinate values ​​corresponding to the observation point. For the initial spatial coordinates, the observation point number of each dummy plant point in the dummy plant point set is read one by one. The 3D coordinate record bound to that observation point number is retrieved from the observation point attribute table, and the retrieved 3D coordinate record is deleted from the initial spatial coordinates. The remaining 3D coordinate records are aggregated to form a valid initial spatial coordinate set. Then, the SIFT feature matching algorithm is used to establish a correspondence between corresponding points among multiple low-altitude images. For successfully matched feature points, their pixel coordinates on each participating image are recorded, forming an image-side coordinate index table. This index table uses the observation point identifier as the primary key, associating the row and column coordinates of the point in different images and the corresponding image number. After bundle adjustment calculations, a valid initial spatial coordinate set is obtained, where each observation point retains the shooting time, leaf color appearance record, and the aforementioned established image-side coordinate index.

[0108] The observation points in this set are sparsely distributed, falling only at distinctive locations such as leaf tips and vein bifurcation points, and cannot directly describe the complete geometry of the canopy surface. The core of multi-view stereo matching lies in finding pixel-level corresponding points between adjacent low-altitude image pairs using the known imaging poses. Using the image-side coordinates of each observation point in the effective initial spatial coordinate set as a seed point, a neighborhood of 5 to 11 pixels in radius around the seed point is expanded to form the pixel region to be matched. For each pixel in the pixel region to be matched, each adjacent low-altitude image is processed sequentially. Based on the polar geometric constraints between the imaging poses, the corresponding pixel with the most similar grayscale is searched within a 2 to 5 pixel distance range from the corresponding epipolar line in the adjacent low-altitude image. Grayscale similarity is measured using the Normalized Cross-Correlation Coefficient (NCC), which is calculated using the following formula: Where I1(i,j) and I2(i,j) are the pixel grayscale values ​​at coordinates (i,j) within the neighborhood window of the pixel to be matched and the neighborhood window of the candidate corresponding pixel, respectively. Ī1 and Ī2 are the mean grayscale values ​​of all pixels within the two neighborhood windows, and the summation is performed on all pixels within the neighborhood windows. The NCC value ranges from -1 to 1; the closer the value is to 1, the more similar the grayscale distributions of the two neighborhood windows are. If the normalized cross-correlation coefficient between two pixels is higher than 0.7, they are confirmed as a pair of dense corresponding points; otherwise, the match is discarded. After completing the above matching process for all adjacent low-altitude images, a set of dense corresponding points of the pixel to be matched on multiple images is obtained.

[0109] For example, in a corn canopy scenario, initial feature points with a texture gradient greater than 15 are first extracted from the image using a feature point detection algorithm as seed points. These seed points are typically located at leaf edges or areas with significant texture changes. Then, a preset matching similarity threshold is set to 0.75 to 0.85. This threshold is calculated using the normalized cross-correlation coefficient (NCC), where an NCC value close to 1 indicates high similarity. This threshold effectively covers areas with clear leaf textures while excluding shadow areas with similarity below 0.75, avoiding mismatches caused by leaf shadows. After pixel-by-pixel scanning, the neighboring pixels within an 8-neighborhood or 16-neighborhood of each seed point that meet the similarity threshold are expanded into a large number of dense corresponding points. A drone equipped with a camera flies at a low altitude (e.g., 70 meters) above the canopy to acquire multiple consecutive low-altitude images. Simultaneously, the imaging pose of each image is recorded in real time using an airborne differential global navigation satellite system and an inertial measurement unit, including the camera's three-dimensional coordinates and attitude angles in object space.

[0110] Preferably, the image-space coordinates of the dense corresponding points on multiple low-altitude images, along with the imaging pose of the corresponding images, are substituted into the collinearity equation, and the three-dimensional position of each pair of dense corresponding points in the object space is calculated by forward intersection. Forward intersection is solved using least squares adjustment to obtain a dense plant surface point cloud covering the entire canopy surface. The point spacing of the dense plant surface point cloud typically reaches 1 to 50 millimeters, which can characterize the leaf edge contours, leaf undulation textures, and canopy top morphology.

[0111] In one embodiment, after feature point extraction and matching of multi-view images using a scale-invariant feature transform algorithm, the three-dimensional object-space coordinates of sparse feature points are obtained through triangulation calculation. This set of coordinates constitutes the effective initial spatial coordinate set, which forms the skeleton structure of the plant. The dense plant surface point cloud and the effective initial spatial coordinate set are merged according to object-space coordinates to jointly form the three-dimensional morphology of the plant. The three-dimensional morphology of the plant includes both the sparse observation skeleton extracted by the scale-invariant feature transform algorithm and the dense surface point cloud from multi-view stereo matching; the two types of points are fused together in the same object-space coordinate system. First, a stable spatial layout of crop rows in the farmland is obtained through point cloud data fitting. Specifically, the collected farmland surface point cloud is horizontally projected, and density clustering analysis is performed on the projected points along the crop planting direction to identify the positional distribution of each crop row. The point cloud coordinates within each crop row are fitted using the least squares method to obtain the extended centerline of the crop row. The extended centerline is the central axis representing the direction of the crop row. After obtaining the extended centerline, the crop row area is assigned. Using each extended center line as an axis of symmetry, extend to both sides along the direction perpendicular to the crop row until reaching the perpendicular bisector of the adjacent extended center line, forming strip-shaped crop row areas. Assign each dense plant surface point cloud in the three-dimensional morphology to the strip-shaped crop row area according to its horizontal projection coordinates.

[0112] Specifically, the construction of the horizontal projection range of the canopy is based on the set of horizontal projection coordinates of the dense plant surface point cloud within each row of crops. For the set of horizontal projection coordinates within the same row of crops, the outer polygon algorithm is used to calculate the most compact two-dimensional polygon contour that can enclose all points. When the plant canopy within the crop row is continuous and has no obvious gaps, a convex hull generation method is used. Based on the Graham scan method, the outermost points are sequentially connected after being sorted by angle to form an outwardly convex closed contour. When there are gaps or edge depressions in the plant canopy within the crop row, a concave hull generation method is used. Based on the convex hull, the contour is contracted inward according to the neighborhood radius to capture the concave parts of the canopy edge, forming a closed contour that fits the real edge of the canopy. The neighborhood radius r is taken as 0.15 to 0.25 times the average row spacing d of the crops, i.e., r = 0.15d to 0.25d, preferably 0.2d. The horizontal coverage area of ​​the outer polygon is the horizontal projection range of the canopy.

[0113] For example, the horizontal projection range of the canopy during the corn tasseling stage is typically a strip-shaped closed area extending along the crop row direction, with the edges exhibiting serrated undulations as the plant grows laterally.

[0114] Preferably, the horizontal area between two adjacent crop rows that is not covered by any of the canopy horizontal projection ranges is marked as the inter-row bare ground area. The boundary of the inter-row bare ground area is jointly defined by the outer contours of the adjacent two canopy horizontal projection ranges and the boundary of the crop row area, covering the horizontal area where the bare soil is located.

[0115] It is understandable that by expanding the dense point cloud through multi-view stereo matching, solving the three-dimensional position by forward intersection of collinear equations, and constructing the canopy outline by enclosing polygons, the complete restoration of the three-dimensional morphology of the plant and the horizontal distribution of the crop rows is achieved. The horizontal projection range of the canopy and the range of the bare ground between the rows depict the lateral coverage pattern of the crop canopy in the field.

[0116] S106. Extract the height of the top of the canopy and the height of the bare ground between adjacent rows from the three-dimensional morphology of the plant to determine the plant height and calculate the canopy undulation amplitude. Determine the canopy coverage by the area ratio of the horizontal projection range of the canopy to the area of ​​the bare ground between rows.

[0117] Based on the height values ​​of the dense plant surface point cloud in the three-dimensional morphology of the plant, point clouds with height values ​​higher than a preset canopy height threshold within the same crop row area are grouped into the canopy top point set. The top height values ​​are read from the canopy top point set. The average ground height value is obtained by averaging the height values ​​of corresponding points in the dense plant surface point cloud within the bare ground area between adjacent rows. The plant height is obtained by the difference between the top height value and the average ground height value. The canopy undulation amplitude is obtained by the difference between the maximum and minimum values ​​of all top height values ​​within the canopy top point set. The closed area of ​​the horizontal projection range of the canopy within the same crop row area and the closed area of ​​the bare ground area between adjacent rows are summed as a reference area. The ratio of the closed area of ​​the horizontal projection range of the canopy to the reference area is determined as the canopy coverage. Thus, the plant height, the canopy undulation amplitude, and the canopy coverage are obtained.

[0118] The height difference between the top of the canopy and the bare ground between adjacent rows directly reflects the vertical growth status of the plant. A greater height difference indicates more vigorous plant growth. For maize, a plant height below 1.5 meters at the tasseling stage is considered weak growth, 1.5 to 2.2 meters is considered normal growth, and greater than 2.2 meters is considered vigorous growth. This criterion is determined based on the field statistical distribution of maize plant height at the tasseling stage. Specifically, multiple groups of maize plant heights at the tasseling stage are measured in a designated field, and the mean μ and standard deviation σ are calculated. μ-2σ is taken as the upper limit of weak growth (approximately 1.5 meters), and μ+1σ is taken as the lower limit of vigorous growth (approximately 2.2 meters). Values ​​equal to 1.5 meters at the endpoint of the interval are classified as normal growth, and values ​​equal to 2.2 meters are classified as vigorous growth.

[0119] In one possible implementation, for the point cloud on the surface of the dense plant in the three-dimensional morphology of the plant, the height value of each point is read, and points within the same crop row area whose height value is higher than a preset canopy height threshold are grouped into the canopy top point set. The preset canopy height threshold is calculated by adjusting the average height of the stable ground location by 40 to 60 centimeters, based on the crop growth stage.

[0120] Specifically, the height is increased by 40 cm during the seedling stage, 50 cm during the growth stage, and 60 cm during the maturity stage. The vertical growth height of the plant can be quantitatively assessed by statistically analyzing the difference between the average height of the canopy apex point set and the average ground height. The apex height values ​​are read from the canopy apex point set. Ground point clouds are extracted from the bare ground area between adjacent crop rows, and the arithmetic mean of the heights of the ground points in this area is calculated to obtain the average ground height Hground. The difference between the apex height Hapex and the average ground height is the plant height, calculated as: plant height = Hapex - Hground. Crop row areas are numbered sequentially from left to right according to the spatial location of the planting rows in the field, with row 1 designated as number 1, row 2 as number 2, and so on. Plant height measured during the heading stage is typically between 1.8 and 2.5 meters, and the measurement results are archived and stored according to the corresponding row area number. The difference between the maximum and minimum values ​​of all apex height values ​​within the canopy apex point set yields the canopy undulation amplitude, which reflects the degree of height fluctuation of the canopy top surface within the same crop row area. For the same crop row area, the projection outline of the canopy on the horizontal plane is identified by the horizontal projection boundary extraction algorithm, and its closed area is calculated and denoted as Scanopy. At the same time, the projection outline of the bare ground between adjacent rows on one side of the row area on the horizontal plane is extracted and its closed area is calculated and denoted as Sbetween. The sum of Scanopy and Sbetween is used as the reference area Sreference. The ratio C = Scanopy / Sreference is calculated and determined as the canopy coverage. The canopy coverage ranges from 0 to 1. This parameter, by introducing adjacent bare ground as a normalization benchmark, jointly characterizes the two-dimensional state of crop vertical growth and lateral shading.

[0121] S107. Check the continuous distribution of the canopy along the plant spacing and row spacing, mark the locations of missing seedlings and broken rows, and determine the growth consistency rate by combining the numerical difference between plant height and leaf color appearance. This yields the plant height, canopy undulation amplitude, canopy coverage, missing seedling and broken row locations, and growth consistency rate.

[0122] Based on the extended centerline of the crop rows in the stable spatial layout, sampling points are sequentially set along the crop row direction at intervals of the plant spacing. The location of each sampling point is checked to see if it is covered by the horizontal projection range of the canopy. If two or more consecutive adjacent sampling points are not covered by the horizontal projection range of the canopy, the row segment connecting the consecutive uncovered sampling points is marked as a missing seedling / broken row location. Based on the leaf color appearance record corresponding to each plant in the observation point attribute table, the average three-channel pixel intensity of that plant is read as the leaf color sample value, and the plant height of that plant is read as the plant height sample value. Calculate the absolute value of the difference between the sample height of each plant and the median of all sample heights in the entire field. Calculate the absolute value of the difference between the sample leaf color of each plant and the median of all sample leaf colors in the entire field. Summate these two values ​​with weights to obtain the consistency difference for that plant. The proportion of plants with a consistency difference lower than a preset consistency threshold is determined as the growth consistency rate. Based on the plant height, canopy undulation amplitude, canopy coverage, location of missing seedlings / row breaks, and growth consistency rate, the data is summarized and archived according to crop row area numbers to obtain an evaluation index record of the crop growth status covering the entire field.

[0123] Based on the spatial superposition of plant sampling points and the horizontal projection range of the canopy, the crop row direction is scanned. For each plant sampling point, it is checked whether the point falls within the closed contour of any of the aforementioned horizontal projection ranges of the canopy. If a plant sampling point does not fall within the closed contour of any of the aforementioned horizontal projection ranges of the canopy, the point is marked as an empty spot. If two or more consecutive empty spots appear along the crop row direction, the row segment connecting the consecutive empty spots is identified as the location of the missing seedlings and broken rows, and the coordinates of the starting and ending plant sampling points and the row segment length are recorded. The field is divided into 10m × 10m grid units, and the coefficient of variation (CV) of plant height in each grid is calculated. The proportion of grid area with a CV less than 0.15 is the growth consistency rate. The data output for missing seedlings and broken rows is a set of geometric coordinates of the empty area, and the growth consistency rate is output as a value between 0 and 1, with the value closer to 1 indicating more balanced growth in the field.

[0124] In one possible implementation, image processing algorithms are used to identify planted crop rows in the field. An extended centerline is obtained by fitting the center position of each crop row, and multiple extended centerlines and their spacing relationships constitute a stable spatial layout. Plant sampling points are set with the extended centerlines of the crop rows in the stable spatial layout as reference axes. Along each extended centerline, starting from the beginning of the crop row, plant sampling points are sequentially placed forward with the plant spacing as the step size. Each plant sampling point corresponds to the desired planting position of a plant. If the plant spacing of a crop row is 25cm, then a plant sampling point is placed every 25cm along the extended centerline of that row.

[0125] It should be noted that the determination of the location of missing seedlings and broken rows relies on the spatial superposition of the plant sampling point and the horizontal projection range of the canopy. For each plant sampling point, it is checked whether the point falls inside the closed contour of any of the horizontal projection ranges of the canopy. If a plant sampling point does not fall inside the closed contour of any of the horizontal projection ranges of the canopy, then the point is marked as an empty position; if two or more empty positions appear consecutively along the crop row direction, then the row segment where the line connecting the consecutive empty positions is located is identified as the location of missing seedlings and broken rows, and the coordinates of the starting and ending plant sampling points and the row segment length are recorded.

[0126] Specifically, the core variables for growth consistency rate are leaf color sample value and plant height sample value. All observation points to which each plant belongs are retrieved from the observation point attribute table, and the average pixel intensity of the red, green, and blue channels in the leaf color appearance record of that plant is read as the leaf color sample value.

[0127] For example, the green channel component of normal leaf color samples during the tasseling stage of maize is significantly higher than that of the red and blue channels, while the green channel component of nitrogen-deficient plants is lower. The plant height of the same plant is used as the plant height sample value. The leaf color sample value reflects the nutritional status of the plant canopy, and the plant height sample value reflects the vertical growth status of the plant; both together characterize the growth performance of an individual plant.

[0128] Preferably, the consistency difference is synthesized using the median of all plant height samples and the median of all leaf color samples within the field as reference benchmarks. The median is less sensitive to individual abnormal plants compared to the mean, and its use as a reference benchmark reflects the typical growth performance of the entire field. For each plant, the absolute value of the difference between the plant height sample value and the median plant height is calculated as the plant height deviation, and the absolute value of the difference between the green channel component of the leaf color sample value and the green channel component of the median leaf color is calculated as the leaf color deviation. The plant height deviation and the leaf color deviation are then summed using a preset weighting coefficient to obtain the consistency difference for that plant. The value of the weighting coefficient reflects the relative importance of plant height and leaf color in growth assessment.

[0129] In one embodiment, the weight of plant height deviation is 0.6, and the weight of leaf color deviation is 0.4.

[0130] Specifically, the consistency difference is compared with a preset consistency threshold. The preset consistency threshold is set according to crop type and growth stage: 0.15 for rice seedling stage, 0.20 for tillering stage, and 0.25 for heading stage. This threshold is derived from statistical analysis of extensive field trial data and represents the reasonable fluctuation range of plant growth within the same field. If the consistency difference of a particular plant is lower than the preset consistency threshold, the plant's growth is considered consistent with the overall field growth; if it is higher than the preset consistency threshold, the plant's growth is considered to deviate from the overall level. The total number of plants in the entire field is obtained using UAV remote sensing imagery combined with image recognition technology. The number of plants with a consistency difference lower than the preset consistency threshold is counted, and the ratio of this number to the total number of plants in the entire field is the growth consistency rate.

[0131] For example, the growth consistency rate of fields with good growth is typically higher than 0.85. The plant height, canopy undulation, canopy coverage, location of missing seedlings / row breaks, and growth consistency rate are uniformly archived according to crop row area numbers to obtain an assessment index record covering the entire field's crop growth status. A unified coordinate system is established for the field using the BeiDou positioning system. Based on the agricultural machinery operation path and planting row spacing, the field is divided into several crop row areas. Each row area is the width of the seeder and 30 to 50 meters long, numbered sequentially from north to south and from west to east as R001, R002, etc. The plant height, canopy undulation, canopy coverage, location of missing seedlings / row breaks, and growth consistency rate are uniformly archived according to crop row area numbers to obtain an assessment index record covering the entire field's crop growth status. The evaluation index records are recorded with crop rows as the smallest summary unit. Each record is associated with the center point coordinates and boundary vertex coordinates of a crop row as spatial coordinates. The average plant height, standard deviation of plant height, canopy undulation amplitude, and canopy coverage are used as geometric indicators. The growth consistency rate and missing seedling rate are used as uniformity indicators, thus realizing a multi-dimensional quantitative characterization of crop growth status.

[0132] Example 2

[0133] This invention provides a crop growth status assessment system for low-altitude agricultural remote sensing image data analysis, comprising an acquisition module, an observation point processing module, a statistical module, an identification module, an extraction module, and an output module.

[0134] The acquisition module is used to deploy imaging units and inertial attitude measurement units on the low-altitude acquisition platform to acquire low-altitude images, record the shooting time and imaging pose, extract observation points on the plant surface and leaf color appearance, and calculate the initial spatial coordinates of the observation points in combination with the imaging pose.

[0135] The observation point processing module is used to detect the imaging position movement of each observation point between adjacent low-altitude images based on low-altitude images and shooting time, and to identify the observation points whose imaging position movement is greater than a preset movement threshold as a set of unstable plant points.

[0136] The statistics module is used to remove the set of unstable plant points from the initial spatial coordinates, construct a stable spatial layout of the farmland crop rows and the ground between rows with the remaining initial spatial coordinates, and calculate the plant spacing and row spacing by counting the distance between adjacent plants and the distance between crop rows.

[0137] The identification module is used to compare the spatial position with the initial spatial coordinates based on the stable spatial layout, identify observation points that deviate from the stable spatial layout and confirm them as a set of false plant points;

[0138] The extraction module is used to remove the set of false plant points from the initial spatial coordinates, and to reconstruct the three-dimensional morphology of the plants by combining the remaining initial spatial coordinates with low-altitude imagery. It also extracts the horizontal projection range of the canopy and the range of bare ground between rows within each crop row area.

[0139] The output module is used to extract the height of the top of the canopy and the height of the bare ground between rows from the three-dimensional morphology of the plant, determine the plant height and calculate the canopy undulation amplitude, and determine the canopy coverage by the area ratio of the horizontal projection range of the canopy to the area of ​​the bare ground between rows; check the continuous distribution of the canopy along the plant spacing and row spacing to mark the location of missing seedlings and broken rows, and determine the growth consistency rate by combining the numerical difference between the plant height and the leaf color appearance.

[0140] In summary, the embodiments of the present invention provide a method for assessing crop growth status by analyzing low-altitude agricultural remote sensing image data. By deploying imaging units and inertial attitude measurement units on a low-altitude acquisition platform to acquire image and pose data, the initial spatial coordinates of the observation points are calculated and the leaf color appearance is extracted.

[0141] This invention addresses the problems of dynamic interference and false point clouds in low-altitude remote sensing data by innovatively introducing a dual filtering mechanism of temporal motion analysis and spatial layout constraints. First, it identifies and removes unstable plant points caused by factors such as wind by detecting the movement of imaging positions of observation points between adjacent images. Second, it uses the remaining stable points to construct a spatial layout model of the crop rows and the ground between rows to statistically determine plant and row spacing. Third, it identifies and deletes false plant points that deviate from the layout through spatial position comparison. The reliable spatial coordinates obtained after dual filtering are used to reconstruct the three-dimensional morphology of the plants. The height of the canopy top and the height of the bare ground are extracted to calculate plant height and canopy undulation. Coverage is determined based on the horizontal projection range of the canopy and the range of bare ground between rows. Canopy continuity is checked along plant and row spacing, marking missing seedlings and broken rows. The difference between plant height and leaf color values ​​is combined to determine the growth consistency rate, achieving a high-precision assessment of crop growth status.

[0142] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementations. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for assessing crop growth status based on low-altitude agricultural remote sensing image data analysis, characterized in that, The method includes: An imaging unit and an inertial attitude measurement unit are deployed on a low-altitude acquisition platform to acquire low-altitude images, record the shooting time and imaging pose, extract observation points on the plant surface and leaf color appearance, and calculate the initial spatial coordinates of the observation points by combining the imaging pose. Based on the low-altitude image and the shooting time, the imaging position movement of each observation point between adjacent low-altitude images is detected, and the observation points whose imaging position movement is greater than the preset movement threshold are identified as the unstable plant point set. The unstable plant point set is removed from the initial spatial coordinates, and the remaining initial spatial coordinates are used to construct a stable spatial layout of the farmland crop rows and the ground between rows. The plant spacing and row spacing are obtained by counting the distance between adjacent plants and the distance between crop rows. Based on the comparison of the stable spatial layout and the initial spatial coordinates, the observation points that deviate from the stable spatial layout are identified and confirmed as a set of false plant points. False plant point sets are removed from the initial spatial coordinates. The three-dimensional morphology of the plants is reconstructed by combining the remaining initial spatial coordinates with low-altitude imagery. The horizontal projection range of the canopy and the range of bare ground between rows are extracted from each crop row area. The height of the canopy top and the height of the bare ground between rows are extracted from the three-dimensional morphology of the plant to determine the plant height and calculate the canopy undulation amplitude. The canopy coverage is determined by the area ratio of the horizontal projection range of the canopy to the area of ​​the bare ground between rows. Check the continuous distribution of the canopy along the plant spacing and row spacing to mark the locations of missing seedlings and broken rows, and determine the growth consistency rate by combining the numerical difference between plant height and leaf color appearance.

2. The method for assessing crop growth status based on low-altitude agricultural remote sensing image data analysis according to claim 1, characterized in that, The deployment of an imaging unit and an inertial attitude measurement unit on a low-altitude acquisition platform to acquire low-altitude images and record the shooting time and imaging pose includes: The imaging unit uses an area array visible light sensor to continuously acquire low-altitude images of farmland along the flight path. The inertial attitude measurement unit combines a three-axis gyroscope, a three-axis accelerometer, and a global navigation satellite receiver to read the attitude angle and position coordinates of the platform at the moment of shooting. Continuous exposure is triggered according to the preset heading overlap rate. The exposure time is synchronized by the platform's internal clock to the attitude reading of the inertial attitude measurement unit and written into the image metadata field.

3. The method for assessing crop growth status based on low-altitude agricultural remote sensing image data analysis according to claim 1, characterized in that, The extraction of plant surface observation points and leaf color appearance, combined with the calculation of the initial spatial coordinates of the observation points by imaging pose, includes: For the image region where the crop canopy is located in the low-altitude image, the scale-invariant feature transform algorithm is used to extract the leaf tip, leaf vein bifurcation and leaf edge turning point as plant surface observation points. For the same observation point, the correspondence between corresponding points is established between adjacent low-altitude images through descriptor matching. The average pixel intensity of the red, green and blue channels in the neighborhood of the observation point is read as the leaf color appearance. The initial spatial coordinates are obtained by collinear equation intersection solution based on the imaging pose and the focal length, principal point and distortion parameters of the imaging unit.

4. The method for assessing crop growth status based on low-altitude agricultural remote sensing image data analysis according to claim 1, characterized in that, The method of detecting the imaging position shift of each observation point between adjacent low-altitude images based on low-altitude imagery and the shooting time includes: According to the chronological order of the shooting time, the corresponding relationship of the same point is retrieved between adjacent low-altitude images. The image-side coordinates of the same observation point in the low-altitude images at the previous and subsequent times are combined into temporal image-side coordinate pairs. According to the imaging pose, the image-side coordinates at the previous time are back-projected to the image plane at the subsequent time using the collinearity equation to obtain the image-side coordinates after pose compensation. The Euclidean distance between the image-side coordinates and the actual image-side coordinates at the subsequent time is used as the imaging position movement range. Observation points that exceed the preset movement threshold are classified into the unstable plant point set.

5. The method for assessing crop growth status based on low-altitude agricultural remote sensing image data analysis according to claim 1, characterized in that, The process of removing unstable plant point sets from the initial spatial coordinates and constructing a stable spatial layout of crop rows and inter-row ground using the remaining initial spatial coordinates includes: The remaining initial spatial coordinates are divided into stable parts of the plant canopy and stable parts of the ground according to the height value. Based on the planar projection distribution of the stable parts of the plant canopy and stable parts of the ground on the horizontal plane, the alternating rows of farmland crops and the ground between the rows are identified. The planar distance between the center projection points of the stable parts of the canopy of adjacent plants in the same crop row is calculated as the plant spacing along the extension direction of the crop row. The vertical distance between the center lines of the extension of two adjacent crop rows is used as the row spacing.

6. The method for assessing crop growth status based on low-altitude agricultural remote sensing image data analysis according to claim 1, characterized in that, The process of comparing spatial positions based on stable spatial layout and initial spatial coordinates to identify observation points deviating from the stable spatial layout and confirming them as a set of spurious plant points includes: Based on the distribution range of the extended center line of the crop row and the ground between the rows in the stable spatial layout, the permissible distribution area of ​​the plants is delineated on each crop row along the crop row direction with the plant spacing as the interval, and the permissible distribution area between adjacent crop rows is delineated with the row spacing as the interval, and the two constitute the permissible distribution map of the location. Compare the horizontal projection coordinates of each observation point in the initial spatial coordinates with the positional allowable distribution map, and mark the observation points that exceed the plant's allowable distribution area and the row's allowable distribution area as horizontal deviation observation points. The height difference between the horizontal deviation observation point and the average height of the stable ground part is read. Horizontal deviation observation points whose height difference is higher than the maximum height value of the stable part of the plant canopy are identified as false plant points and included in the set of false plant points.

7. The method for assessing crop growth status based on low-altitude agricultural remote sensing image data analysis according to claim 1, characterized in that, The process of deleting the set of false plant points from the initial spatial coordinates and reconstructing the three-dimensional morphology of the plants by combining the remaining initial spatial coordinates with low-altitude imagery includes: The observation point numbers of the false plant point set are traversed, and the corresponding three-dimensional coordinate records are deleted from the initial spatial coordinates to obtain the effective initial spatial coordinate set. The observation points in the effective initial spatial coordinate set are matched with dense corresponding points using multi-view stereo matching. The dense corresponding points are combined with the imaging pose and substituted into the collinear equation for forward intersection to obtain the dense plant surface point cloud. The dense plant surface point cloud and the effective initial spatial coordinate set are merged to form the three-dimensional shape of the plant.

8. The method for assessing crop growth status based on low-altitude agricultural remote sensing image data analysis according to claim 1, characterized in that, The process of extracting the canopy top height and the height of bare ground between rows from the three-dimensional morphology of the plant to determine the plant height and calculate the canopy undulation amplitude, and determining the canopy coverage by the area ratio of the horizontal projection range of the canopy to the area of ​​bare ground between rows, includes: The dense plant surface point cloud with a height higher than the preset canopy height threshold within the same crop row area in the three-dimensional morphology of the plant is classified into the canopy top point set and the top height value is read. The average height of the point cloud within the bare ground area between adjacent rows is taken as the average ground height. The difference between the top height value and the average ground height is taken as the plant height. The difference between the maximum and minimum top height values ​​within the canopy top point set is taken as the canopy undulation amplitude. The sum of the closed area of ​​the horizontal projection range of the canopy and the closed area of ​​the bare ground area between adjacent rows is taken as the reference area. The ratio of the closed area of ​​the horizontal projection range of the canopy to the reference area is taken as the canopy coverage.

9. The method for assessing crop growth status based on low-altitude agricultural remote sensing image data analysis according to claim 1, characterized in that, The method of checking the continuous distribution of the canopy along the plant spacing and row spacing to mark the locations of missing seedlings and gaps in the rows includes: Plant sampling points are set sequentially along the crop rows in the farmland at the plant spacing. Each plant sampling point is checked to see if it is covered by the horizontal projection range of the canopy. If two or more consecutive adjacent plant sampling points are not covered by the horizontal projection range of the canopy, the row segment where the line connecting the consecutive uncovered plant sampling points is located is marked as the location of the missing seedlings and broken rows.

10. The method for assessing crop growth status based on low-altitude agricultural remote sensing image data analysis according to claim 1, characterized in that, The determination of growth uniformity rate by combining the numerical difference between plant height and leaf color appearance includes: The average pixel intensity of the three channels corresponding to each plant is read as the leaf color sample value, and the plant height of the plant is read as the plant height sample value. The absolute value of the difference between the plant height sample value of each plant and the median of the plant height sample values ​​of the whole field is calculated. The absolute value of the difference between the leaf color sample value of each plant and the median of the leaf color sample values ​​of the whole field is calculated. The two are weighted and summed to obtain the consistency difference of the plant. The proportion of the number of plants with a consistency difference lower than the preset consistency threshold to the total number of plants in the whole field is determined as the growth consistency rate. Plant height, canopy undulation amplitude, canopy coverage, missing seedling and broken row position and growth consistency rate are summarized and archived according to crop row area number.

Images