Unmanned aerial vehicle remote sensing crop planting structure image recognition method and system

Abstract

The invention discloses an unmanned aerial vehicle remote sensing crop planting structure image recognition method and system, and relates to the technical field of agricultural remote sensing. Preprocessing the remote sensing image data to obtain a farmland ortho-image map and a distortion risk area; performing crop type identification on the farmland orthophoto map to obtain a crop type identification result; according to the distortion risk area, carrying out confidence coefficient evaluation on the crop type identification result to obtain a comprehensive confidence coefficient; and generating a crop planting structure image according to the crop type identification result and the comprehensive confidence coefficient. According to the method, the confidence coefficient of the crop type can be evaluated in combination with the distortion risk region, so that crop planting structure image recognition is realized, and the accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing technology, and in particular to a method and system for recognizing crop planting structure images from unmanned aerial vehicle (UAV) remote sensing images. Background Technology

[0002] In modern smart agriculture, drones equipped with multispectral sensors are typically used for remote sensing data collection to quickly and accurately identify crop planting structures in large areas of farmland. These drones collect data above the farmland, acquiring reflectance spectral information in different bands to distinguish and identify different crop types and their planting areas. However, in actual farmland environments, drones may encounter localized, short-term airflow disturbances during flight, which can momentarily affect their flight attitude. This can distort the acquired images, leading to inaccurate stitched images of the planting structure and low identification accuracy.

[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention

[0004] The main objective of this invention is to propose a method and system for recognizing crop planting structure images using UAV remote sensing. This method combines distortion risk areas to assess the confidence level of crop types, thereby achieving crop planting structure image recognition and improving accuracy.

[0005] On one hand, embodiments of the present invention provide a method for recognizing crop planting structure images from unmanned aerial vehicle (UAV) remote sensing, comprising the following steps: Acquire remote sensing image data; The remote sensing image data is preprocessed to obtain farmland orthophoto maps and distortion risk areas; The crop type identification was performed on the orthophoto map of the farmland to obtain the crop type identification results; Based on the distortion risk area, the confidence level of the crop type identification results is evaluated to obtain a comprehensive confidence level; Based on the crop type identification results and the overall confidence level, a crop planting structure image is generated.

[0006] On the other hand, embodiments of the present invention provide a UAV remote sensing image recognition system for crop planting structures, comprising: The data acquisition module is used to acquire remote sensing image data; The preprocessing module is used to preprocess the remote sensing image data to obtain farmland orthophoto maps and distortion risk areas; The type identification module is used to identify the crop type in the orthophoto of the farmland and obtain the crop type identification result; The confidence assessment module is used to assess the confidence of the crop type identification results based on the distortion risk area to obtain a comprehensive confidence score. The map generation module is used to generate a crop planting structure image based on the crop type identification results and the comprehensive confidence level.

[0007] The embodiments of this application include at least the following beneficial effects: First, remote sensing image data is acquired. Then, the remote sensing image data is preprocessed to obtain farmland orthophoto maps and distortion risk areas. Next, crop type identification is performed on the farmland orthophoto maps to obtain crop type identification results. Based on the distortion risk areas, the confidence level of the crop type identification results is evaluated to obtain a comprehensive confidence level. Finally, based on the crop type identification results and the comprehensive confidence level, a crop planting structure image is generated. This allows for the evaluation of the confidence level of crop types by combining distortion risk areas, thereby achieving crop planting structure image identification and improving accuracy.

[0008] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the description and the drawings. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0010] Figure 1 This is a flowchart of a method for recognizing crop planting structure images using UAV remote sensing, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of an image recognition system for crop planting structure obtained by remote sensing from an unmanned aerial vehicle (UAV) according to an embodiment of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.

[0012] In modern smart agriculture, to quickly and accurately identify crop planting structures in large areas of farmland, drones equipped with multispectral sensors are typically used for remote sensing data collection. However, during drone flight, due to local airflow disturbances, a brief desynchronization occurs between the sensor attitude information recorded by the drone's onboard navigation system and the actual physical attitude of the sensor. This desynchronization leads to localized pixel misalignment, stretching, or blurring in overlapping areas during subsequent image stitching. These geometric distortions severely interfere with the accurate judgment of crop boundaries by crop identification algorithms, and may even misjudge pseudo-boundaries introduced by stitching as actual crop boundaries, thus affecting the accuracy and reliability of the final crop planting structure map and misleading precision agriculture decisions.

[0013] For example, in modern smart agriculture management, the rapid and accurate identification of crop planting structures in large areas of farmland is a crucial prerequisite for optimizing agricultural activities such as precision fertilization, irrigation, and pest and disease control. This typically relies on remote sensing data collection using drones equipped with multispectral sensors. Following a pre-set flight path and sensor parameters, the drones perform routine data collection operations above the farmland, acquiring reflectance spectral information in different bands to distinguish and identify different crop types and their planting areas. These raw multispectral images are then transmitted to a ground processing system, undergoing a series of complex image processing steps to ultimately generate a high-precision map of the crop planting structure.

[0014] However, in actual farmland operations, drones may encounter localized, short-lived airflow disturbances during flight. For example, when a drone flies over the boundary between different crop types, the difference in absorption and reflection of solar radiation between bare soil, low-lying crops, and tall crop canopies creates a local surface temperature gradient, which in turn triggers minute thermal convection, generating rising or sinking airflows. Furthermore, windbreaks or tall buildings at the edge of farmland may also create complex vortices or shear winds on their leeward side under specific wind conditions, causing the drone to suddenly encounter a crosswind or sinking airflow. These airflow disturbances are usually instantaneous and localized; their intensity may not be enough to trigger the flight control system's emergency obstacle avoidance or return-to-home mechanism, but it is sufficient to momentarily affect the drone's flight attitude, potentially causing distortion in the acquired images. This results in inaccurate stitched images of the planting structure and low recognition accuracy.

[0015] The embodiments of this application will be explained in detail below with reference to the accompanying drawings: Figure 1 This is an optional flowchart of an image recognition method for crop planting structure from UAV remote sensing provided in this application embodiment. Figure 1 The method may include, but is not limited to, steps S101 to S105.

[0016] Step S101: Acquire remote sensing image data; Step S102: Preprocess the remote sensing image data to obtain farmland orthophoto maps and distortion risk areas; Step S103: Perform crop type identification on the farmland orthophoto map to obtain the crop type identification results; Step S104: Based on the distortion risk area, assess the confidence level of the crop type identification results to obtain the comprehensive confidence level; Step S105: Generate a crop planting structure image based on the crop type identification results and the overall confidence level.

[0017] Steps S101 to S105 as shown in the embodiments of this application can combine the distortion risk area to assess the confidence level of crop type, so as to realize crop planting structure image recognition and improve accuracy.

[0018] In some embodiments, steps S101-S105 may involve acquiring remote sensing image data first. This can be done using a drone equipped with a multispectral or hyperspectral sensor for aerial photography to acquire remote sensing image data of farmland areas. This data typically contains spectral information across multiple bands, reflecting the growth status and type of crops. Alternatively, remote sensing image data can be acquired via a satellite remote sensing platform, or near-ground remote sensing data can be collected using a ground-based mobile platform equipped with sensors.

[0019] The remote sensing image data is then preprocessed to obtain an orthophoto map of farmland and distortion risk areas. The purpose is to eliminate noise and errors in the image and correct geometric distortions. For example, radiometric correction can be performed on the remote sensing image data to eliminate the influence of atmospheric effects, sensor response inhomogeneities, and other factors on the image radiometric values. After radiometric correction, image stitching techniques can be used to stitch multiple remote sensing images into an orthophoto map covering the entire farmland area. During the stitching process, it is necessary to identify overlapping areas between multiple radiometrically corrected remote sensing images and perform geometric distortion analysis on these overlapping areas to identify distortion risk areas. Distortion risk areas refer to areas where geometric distortion or misalignment may occur during image stitching; the confidence level of crop type identification results in these areas is low.

[0020] Next, crop type identification is performed on the orthophoto map of the farmland to obtain the crop type identification results. The purpose is to distinguish and identify different crop types in the farmland. For example, image classification algorithms based on machine learning or deep learning can be used to classify the orthophoto map of the farmland at the pixel level or region level, thereby obtaining the distribution area of ​​each crop. Specifically, spectral features, texture features, geometric features, etc., can be extracted from the image, and these features can be used to train classification models, such as support vector machines (SVM), random forests, or convolutional neural networks (CNN).

[0021] Based on the distortion risk area, a confidence level assessment is performed on the crop type identification results to obtain a comprehensive confidence level. The purpose is to quantify the reliability of the crop type identification results. For example, different confidence levels can be assigned to the crop type identification results within the distortion risk area based on its size, location, and degree of distortion. Furthermore, other auxiliary data, such as UAV attitude fluctuation data and multispectral sensor spectral response drift data, can be combined to conduct a more comprehensive confidence level assessment of the identification results. For instance, when the UAV attitude fluctuation is large or the sensor spectral response drift is significant, the confidence level of the identification results will decrease accordingly.

[0022] Finally, based on the crop type identification results and overall confidence level, a crop planting structure image is generated. This image visually displays the crop type identification results and their confidence levels, helping users quickly identify high-risk areas. For example, different colored areas can be overlaid on a farmland orthophoto map to represent different crop types, with the intensity or transparency of the color indicating the level of confidence. High-risk areas (i.e., areas with lower confidence) can be highlighted with darker colors for users to focus on.

[0023] By introducing the concept of distortion risk regions and combining multi-source data to assess the confidence level of crop type identification results, this embodiment effectively solves the problem of image geometric distortion caused by instantaneous asynchrony in UAV flight attitude in existing technologies. Traditional methods, when processing stitched images containing local geometric deformations, may misjudge geometric distortions introduced by stitching errors as actual boundaries between different crops, thus affecting the accuracy of crop planting structure maps. This embodiment, by identifying distortion risk regions in the preprocessing stage and fully considering the impact of these risk regions in subsequent confidence assessments, can more accurately evaluate the reliability of crop type identification results. This embodiment can significantly improve the accuracy and reliability of crop planting structure identification, providing stronger technical support for precision agriculture management.

[0024] In some embodiments, step S102 involves preprocessing the remote sensing image data to obtain an orthophoto map of farmland and areas at risk of distortion. This may include, but is not limited to, the following steps: Radiometric correction of remote sensing image data; The radiometrically corrected remote sensing images are stitched together to obtain an orthophoto map of farmland. During the stitching process, overlapping areas between multiple radiometrically corrected remote sensing images are identified; Geometric distortion analysis of overlapping regions is performed using the scale-invariant feature transformation algorithm to identify distortion risk areas.

[0025] In some embodiments, if the preprocessing method fails to fully consider the radiometric differences, geometric deformations, and stitching errors that may exist in the remote sensing image data during the acquisition process, it may result in insufficient accuracy of the generated farmland orthophoto map or inaccurate identification of distortion risk areas, thereby affecting the reliability assessment of subsequent crop type identification results.

[0026] Therefore, radiometric correction can be performed on remote sensing image data first, aiming to eliminate or reduce radiometric errors caused by factors such as atmospheric scattering, inconsistent sensor responses, and changes in lighting conditions. Image data acquired at different times, with different sensors, or under different lighting conditions are converted into comparable physical quantities, such as surface reflectance, thereby ensuring the accuracy of subsequent image analysis.

[0027] The radiometrically corrected remote sensing images are then stitched together to obtain an orthophoto map of the farmland. Multiple images with overlapping areas can be combined using image registration and fusion techniques to generate a seamless, high-resolution orthophoto map covering a larger area. The purpose is to provide an accurate geographic reference base map, facilitating subsequent crop type identification.

[0028] During the stitching process, overlapping regions between multiple radiometrically corrected remote sensing images are identified. Overlapping regions refer to the geographical areas shared between adjacent images; these regions contain rich feature information that can be used for image registration and geometric distortion analysis.

[0029] Finally, geometric distortion analysis of the overlapping regions is performed using the scale-invariant feature transform (SIFT) algorithm to identify distortion risk areas. The purpose is to accurately detect and quantify the geometric deformation that may occur during image acquisition and stitching. SIFT is a local feature descriptor that maintains the stability of feature points under conditions such as image scaling, rotation, and brightness changes. It is suitable for identifying matching points between images and calculating the geometric transformation parameters of the images using these matching points. By performing feature matching and geometric transformation analysis on the overlapping regions using SIFT, areas with significant geometric distortion or registration errors can be identified; these areas are defined as distortion risk areas. Identifying distortion risk areas is crucial for subsequent confidence assessment because the crop type identification results in these areas may have lower reliability.

[0030] This embodiment introduces radiometric correction to unify the radiometric characteristics of remote sensing image data, eliminating interference from the environment and sensors, and laying a high-quality foundation for subsequent image processing. Based on this, the radiometrically corrected images are stitched together. This not only expands the observation range but also, during the stitching process, identifies overlapping regions between images and uses a scale-invariant feature transform (SMT) algorithm to perform fine geometric distortion analysis on these overlapping regions. The robustness of the SMT algorithm to image transformations allows for accurate detection of local geometric deformations even in complex farmland scenes, thereby precisely identifying distortion risk areas. This refined preprocessing workflow ensures the geometric accuracy and radiometric consistency of the farmland orthophoto map and provides reliable distortion risk information for the subsequent confidence assessment of crop type identification results.

[0031] To illustrate this technical solution more clearly, a specific example is used below. Suppose a drone equipped with a multispectral sensor conducts remote sensing operations on a farmland, acquiring multiple raw remote sensing images with overlapping areas. First, these raw image data are processed by a radiometric correction module to eliminate inconsistencies in brightness and color caused by atmospheric conditions, changes in the sun's angle, or differences in the sensor's own response. For example, by performing atmospheric correction and relative radiometric correction on the images, the pixel values ​​are converted to surface reflectance. Next, the radiometrically corrected images are input into an image stitching module. In this module, the system automatically identifies overlapping areas between adjacent images. For example, it uses feature point matching algorithms (such as scale-invariant feature transform algorithms) to find corresponding points within the overlapping areas. Once overlapping areas are identified, the scale-invariant feature transform algorithm is further applied to these areas to analyze and quantify the geometric deformation between the images, such as local distortion, stretching, or misalignment. Based on these analysis results, the system generates a distortion risk area layer, where high-risk areas are clearly marked, for example, by using color coding or numerical values ​​to represent their degree of distortion. Ultimately, these radiometrically corrected, precisely stitched farmland orthophotos with information on areas at risk of distortion will be used for subsequent crop type identification and confidence assessment.

[0032] Through the above technical solution, this embodiment effectively solves the problems of radiometric inconsistency and geometric distortion that may exist when processing complex remote sensing image data. Radiometric correction significantly improves image data quality; refined stitching and geometric distortion analysis based on scale-invariant feature transformation algorithms accurately generate high-precision orthophoto maps of farmland and reliably identify distortion risk areas in the image. This not only provides more accurate and reliable input data for subsequent crop type identification but also, through clear distortion risk area information, makes the confidence assessment of crop type identification results more targeted and accurate, thereby significantly improving the overall reliability and practicality of the entire UAV remote sensing crop planting structure image recognition method.

[0033] In some embodiments, in step S104, the confidence level of the crop type identification result is assessed based on the distortion risk area to obtain a comprehensive confidence level, which may include, but is not limited to, the following steps: Step S201: Obtain the attitude fluctuation data of the UAV; Step S202: Obtain the spectral response drift data of the multispectral sensor; Step S203: Extract geometric features, spectral features, texture features and morphological features from the distortion risk area. Geometric features include the local curvature and straightness of the crop boundary. Spectral features include the normalized difference vegetation index and reflectance values ​​of each band. Texture features include contrast, homogeneity and entropy. Morphological features include the shape, size and arrangement pattern of the crop canopy. Step S204: Based on the crop type identification results, attitude fluctuation data, and spectral response drift data, evaluate the first confidence level of geometric features, the second confidence level of spectral features, the third confidence level of texture features, and the fourth confidence level of morphological features; Step S205: The first confidence level, the second confidence level, the third confidence level and the fourth confidence level are fused to obtain the comprehensive confidence level.

[0034] In some embodiments, assessing the confidence level of crop type identification results solely based on the geometric distortion risk areas of remote sensing images may not fully reflect the reliability of the identification results. For example, drones may experience attitude fluctuations during flight, and multispectral sensors may exhibit spectral response drift. These factors can affect the quality of remote sensing images and the accuracy of subsequent crop type identification, leading to inaccurate confidence assessments. If these issues are not addressed, the generated crop planting structure images may not accurately reflect the actual risk status of the crop planting structure, impacting the scientific basis of agricultural decision-making.

[0035] To address this, we can first acquire attitude fluctuation data of the UAV. Sensors mounted on the UAV, such as an inertial measurement unit (IMU), can monitor real-time attitude changes during flight, including pitch, roll, and yaw. These attitude fluctuations affect the shooting angle and ground projection accuracy of remote sensing images, thus introducing recognition errors. Simultaneously, we can acquire spectral response drift data from multispectral sensors. By periodically or in real-time calibrating the multispectral sensors, we can detect changes in their spectral response characteristics under different operating conditions. This drift may be caused by factors such as temperature, humidity, and sensor aging, resulting in different spectral values ​​for the same ground feature at different times or on different sensors, affecting the accuracy of crop type identification.

[0036] Then, geometric, spectral, textural, and morphological features were extracted from the distortion-risk areas to quantify the characteristics of crop images from multiple dimensions. Geometric features include the local curvature and straightness of crop boundaries; spectral features include the normalized difference vegetation index and reflectance values ​​for each band; textural features include contrast, homogeneity, and entropy; and morphological features include the shape, size, and arrangement pattern of the crop canopy. Geometric features can reflect the regularity of crop planting or the degree of distortion. Spectral features are important indicators of crop health and growth stage. Textural features describe the fine structure and uniformity of image regions. Morphological features reflect the physical morphology and planting structure of crops.

[0037] Based on the crop type identification results, attitude fluctuation data, and spectral response drift data, the first confidence level of geometric features, the second confidence level of spectral features, the third confidence level of texture features, and the fourth confidence level of morphological features are then evaluated. This means that when evaluating the reliability of each feature, not only the feature itself is considered, but also the flight stability of the UAV (attitude fluctuation data) and the data quality of the sensors (spectral response drift data). For example, when attitude fluctuations are large, the first confidence level of geometric features may be reduced; when spectral response drift is significant, the second confidence level of spectral features will also be affected.

[0038] Finally, the first, second, third, and fourth confidence levels are fused to obtain the comprehensive confidence score. The purpose is to integrate multi-source and multi-dimensional confidence information to form a more comprehensive and robust final confidence assessment result.

[0039] This embodiment overcomes the limitations of relying solely on geometric distortion risk assessment by introducing attitude fluctuation data from a UAV and spectral response drift data from a multispectral sensor, and extracting multi-dimensional crop features from distortion risk areas. Specifically, attitude fluctuation data reflects the stability of the UAV during image acquisition, directly affecting the geometric accuracy of the image; spectral response drift data reveals the reliability of the sensor data itself. By combining these external influencing factors with the geometric, spectral, texture, and morphological features of the crop image itself, the confidence level of each feature's recognition result is independently assessed. For example, when the UAV's attitude fluctuation is large, even if the crop boundary appears regular, the first confidence level of its geometric features will be appropriately reduced to reflect potential geometric errors. Similarly, when spectral drift exists in the sensor, the second confidence level of the spectral features will also be corrected. Finally, by fusing these multi-dimensional confidence levels, the reliability of the crop type recognition results can be quantified more comprehensively and accurately, thus providing a more solid data foundation for subsequent crop planting structure image generation.

[0040] To illustrate this technical solution more clearly, a specific example is used below. Suppose that during a UAV remote sensing mission, the UAV encounters gusts of wind, causing slight fluctuations in its attitude. Simultaneously, the multispectral sensor, due to prolonged operation, experiences slight drift in the spectral response of some bands. After crop type identification, the system first acquires the UAV's attitude fluctuation data and the multispectral sensor's spectral response drift data. Next, it extracts the crop's geometric features (e.g., boundary curvature), spectral features (e.g., normalized difference vegetation index), texture features (e.g., contrast), and morphological features (e.g., crop canopy shape) from the distortion-risk areas of the farmland orthophoto map. When evaluating the confidence level of these features, the system incorporates the attitude fluctuation data and spectral response drift data for correction. For example, due to attitude fluctuations, the first confidence level of the geometric features of a crop boundary that was originally identified as straight will be appropriately reduced; due to spectral drift, the normalized difference vegetation index values ​​of some crops may be slightly off, leading to a corresponding reduction in the second confidence level of their spectral features. Finally, these revised first, second, third, and fourth confidence scores are weighted and fused to obtain a comprehensive confidence score. This comprehensive confidence score will more accurately reflect the true reliability of the crop type identification results. For example, if the comprehensive confidence score is low, the area will be marked as high-risk on the crop planting structure image, prompting the user to further verify or take action.

[0041] Through the above technical solution, this embodiment comprehensively considers the impact of external factors such as UAV attitude fluctuations and spectral response drift of multispectral sensors on the recognition results, and conducts a detailed evaluation of crop characteristics from multiple dimensions such as geometry, spectrum, texture, and morphology. This multi-source, multi-dimensional information fusion makes the confidence assessment results more comprehensive and robust, effectively avoiding the bias that may be caused by single-factor assessment. As a result, the generated crop planting structure image can more realistically and accurately reflect the actual risk status of crop planting structure, providing a more reliable basis for agricultural production management and decision-making, thereby improving the level of intelligence and precision in agricultural production.

[0042] In some embodiments, in step S201, acquiring the attitude fluctuation data of the UAV may include, but is not limited to, the following steps: The attitude information of the UAV body is obtained by the first inertial measurement unit installed on the UAV body; The attitude information of the airborne sensor gimbal is obtained by a second inertial measurement unit installed on the airborne sensor gimbal. The attitude information of the UAV body and the attitude information of the onboard sensor gimbal are compared to determine the dynamic attitude deviation of the onboard sensor gimbal relative to the UAV body. Attitude fluctuation data is generated based on dynamic attitude deviation and the attitude information of the UAV body.

[0043] In some embodiments, the attitude information of the UAV can be obtained first through a first inertial measurement unit (IMU) installed on the UAV body. The first IMU refers to a sensor assembly installed on the main structure of the UAV, whose main function is to monitor and output the UAV's attitude (such as pitch, roll, and yaw angles) and motion state (such as acceleration and angular velocity) information in real time during flight. This unit typically includes sensors such as a three-axis gyroscope, a three-axis accelerometer, and a three-axis magnetometer, providing high-precision UAV attitude data through data fusion algorithms.

[0044] Then, the attitude information of the airborne sensor gimbal is acquired through a second inertial measurement unit (IMU) mounted on the airborne sensor gimbal. The second IMU is an independent sensor assembly mounted on the airborne sensor gimbal, whose function is to accurately measure and output the gimbal's own attitude information. Airborne sensor gimbals are typically used to stabilize and adjust the viewing angle of remote sensing sensors to counteract the shaking and attitude changes of the UAV body, ensuring the stability of the images acquired by the sensors.

[0045] Next, the attitude information of the UAV body and the attitude information of the onboard sensor gimbal are compared. The dynamic attitude deviation of the onboard sensor gimbal relative to the UAV body is calculated to quantify the dynamic changes of the gimbal relative to the UAV body during flight. This comparison is achieved through attitude calculation algorithms. For example, the two attitude information sets can be converted into quaternions or Euler angles, and then the relative rotation matrix or angle difference between them can be calculated to obtain the dynamic attitude deviation of the onboard sensor gimbal relative to the UAV body. This deviation reflects the relative motion between the gimbal and the UAV body during active stabilization or passive motion.

[0046] Finally, attitude fluctuation data is generated based on the dynamic attitude deviation and the UAV's attitude information. Attitude fluctuation data is a key indicator comprehensively reflecting the overall flight attitude stability of the UAV and the working status of the sensor gimbal. This data can include statistical quantities such as the mean, variance, maximum, and minimum values ​​of the attitude deviation, as well as its trend over time; it can also be attitude change curves after filtering and smoothing. Its purpose is to provide an accurate attitude stability reference for the subsequent confidence assessment of crop type identification results.

[0047] This embodiment acquires the attitude information of the UAV body and the onboard sensor gimbal separately, and calculates the dynamic attitude deviation between them. This allows for the precise quantification of attitude changes caused by UAV flight instability or gimbal jitter during remote sensing image acquisition. This refined attitude data acquisition enables the generation of more accurate attitude fluctuation data. This attitude fluctuation data directly reflects the platform stability during image acquisition, providing crucial input for assessing the geometric distortion risk of remote sensing images, thereby contributing to a more accurate evaluation of the confidence level of crop type identification results.

[0048] By employing the aforementioned technical solution, this embodiment, through separately measuring the attitude of the UAV body and the airborne sensor gimbal, and calculating their relative dynamic deviations, can more comprehensively and precisely capture the actual motion state of the image acquisition platform in the air. This precise attitude fluctuation data provides a more reliable basis for the subsequent confidence assessment of crop type identification results, effectively improving the accuracy and reliability of the confidence assessment, thereby indirectly enhancing the accuracy of the final crop planting structure image.

[0049] In some embodiments, in step S202, acquiring the spectral response drift data of the multispectral sensor may include, but is not limited to, the following steps: After setting multiple miniature spectral calibration modules inside the multispectral sensor, the miniature spectral calibration modules are controlled to emit light with target spectral characteristics toward the detector array. The miniature spectral calibration module contains a stable light source and multiple narrowband filters. Record the response data of the detector array; The response data is compared with a preset response reference to calculate the response deviation; Spectral response drift data are generated based on the response deviation and response data.

[0050] In some embodiments, the spectral response characteristics of multispectral sensors may drift during prolonged operation or when affected by environmental factors (such as changes in temperature and humidity), resulting in inaccurate spectral response drift data and consequently affecting the accuracy of confidence assessment of crop type identification results. Failure to address this issue could lead to misjudgments of crop growth status and compromise the scientific validity of agricultural production decisions.

[0051] To achieve this, multiple miniature spectral calibration modules can be integrated within the multispectral sensor. These modules then emit light with the target spectral characteristics towards the detector array. Each miniature spectral calibration module contains a stable light source and multiple narrowband filters. These miniature spectral calibration modules are small optical components integrated into the optical path of the multispectral sensor, designed to periodically or on-demand self-calibrate the sensor's spectral response. The stable light source can be a highly stable LED or laser diode to ensure that the intensity and spectral characteristics of the emitted light remain constant during calibration. The narrowband filters are used to select specific wavelengths of light from the broad spectrum emitted by the stable light source, thereby simulating light with the target spectral characteristics, which correspond to the center wavelengths of various bands in the multispectral sensor. During calibration, the miniature spectral calibration modules can be controlled to emit preset light with known spectral characteristics towards the detector array.

[0052] Then, the response data of the detector array is recorded. After the detector array receives light with the target spectral characteristics, it generates corresponding electrical signals, which are recorded as the detector array's response data. The response data is then compared with a preset response reference to calculate the response deviation. The preset response reference refers to the standard response value of the detector array to light with the same target spectral characteristics under ideal conditions; it is pre-labeled according to the device's performance at the sensor's factory. By comparison, the difference between the current response data and the standard response reference can be calculated, i.e., the response deviation. This response deviation directly reflects the degree of spectral response drift of the multispectral sensor. Finally, based on the response deviation and the response data, spectral response drift data is generated for subsequent confidence assessment.

[0053] This embodiment achieves real-time or near-real-time monitoring and quantification of sensor spectral response drift by integrating a miniature spectral calibration module within the multispectral sensor. When the multispectral sensor is operating in the field, its internal miniature spectral calibration module periodically emits light with known spectral characteristics. The detector array records the response data to these lights and compares it with a preset response reference obtained under ideal conditions. This allows for the precise calculation of the sensor's current spectral response deviation. This built-in calibration mechanism effectively avoids the cumbersome and delayed nature of external calibration, ensuring that the acquired spectral response drift data accurately reflects the sensor's true state under different environmental conditions. The high-precision spectral response drift data allows for a more accurate consideration of the impact of sensor performance variations when subsequently assessing the confidence level of crop type identification results, thereby improving the reliability of the assessment.

[0054] To illustrate this technical solution more clearly, a specific example is provided below. A miniature spectral calibration module can be integrated into the optical path of the multispectral sensor, for example, in a switchable location in front of the detector array. This module can contain a miniature LED array as a stable light source, with each LED corresponding to a narrowband filter to emit light of a specific wavelength, such as 450nm, 550nm, 670nm, and 800nm—bands commonly used in crop remote sensing. When calibration is required, the control unit activates these LEDs, causing them to emit light sequentially or simultaneously onto the detector array. The detector array records the response data for each band, for example, represented by digitally quantized grayscale values. Assuming that the preset response reference value for a certain band is 1000 during factory calibration, and the currently recorded response data is 980, then the response deviation is -20. This deviation value, along with the current response data, is used to generate spectral response drift data, which could be a correction coefficient or a drift curve. For example, if the response values ​​of a certain band are generally low, the reflectance data for that band will be adjusted upwards accordingly in subsequent crop type identification to eliminate the effects of sensor drift. In this way, even if the sensor operates at different temperatures, its spectral response drift can be accurately quantified and used for correction, thus ensuring the spectral consistency of remote sensing data.

[0055] Through the above technical solution, this embodiment effectively solves the problem of inaccurate data caused by spectral response drift in multispectral sensors. By setting a miniature spectral calibration module inside the sensor and using a stable light source and narrowband filters to emit light with the target spectral characteristics for self-calibration, accurate spectral response drift data can be acquired in real-time or near real-time. This significantly improves the accuracy and reliability of spectral response drift data, thereby making the confidence assessment of crop type identification results based on this data more accurate. This embodiment can effectively compensate for spectral response drift caused by sensor environmental changes or aging, ensuring the quality of remote sensing data, and thus improving the overall accuracy and reliability of crop planting structure image recognition, providing a more solid data foundation for precision agricultural management.

[0056] In some embodiments, in step S205, the first confidence level, the second confidence level, the third confidence level, and the fourth confidence level are fused to obtain a comprehensive confidence level, which may include, but is not limited to, the following steps: Step S301: Obtain the agronomic rule base, which contains characteristic rules for different crop types under different planting patterns; Step S302: Match the geometric features with the agronomic rule base and correct the first confidence level; Step S303: Match the spectral features with the agronomic rule base and correct the second confidence level; Step S304: Match the texture features with the agronomic rule base and correct the third confidence level; Step S305: Match the morphological features with the agronomic rule base and correct the fourth confidence level; Step S306: The corrected first confidence level, second confidence level, third confidence level and fourth confidence level are weighted and summed to obtain the comprehensive confidence level.

[0057] In some embodiments, relying solely on the direct evaluation and fusion of sensor data and image features may not adequately consider the biological characteristics, growth patterns, and unique performance under different planting patterns of crops, resulting in insufficient accuracy and reliability of the overall confidence level in certain complex or abnormal situations. For example, when certain image features exhibit similarity across different crop types or growth stages, simple feature matching and statistical fusion may not be able to effectively distinguish them, thus affecting the accuracy of the final confidence level assessment.

[0058] To this end, an agronomic rule base can be obtained first. This rule base contains characteristic rules for different crop types under different planting patterns. The agronomic rule base is a knowledge system that includes agricultural expert knowledge, crop growth models, environmental influencing factors, and the characteristics of crops under different planting patterns. The agronomic rule base aims to provide a reference framework based on biological and agronomic principles for verifying and correcting features extracted from remote sensing images. Specifically, characteristic rules refer to the typical patterns, ranges, or interrelationships that different crop types (e.g., maize, wheat, rice) should exhibit under specific planting patterns (e.g., row sowing, broadcast sowing, intercropping), including geometric features (e.g., plant spacing, row spacing, leaf shape), spectral features (e.g., reflectance in specific bands, vegetation index trends), textural features (e.g., canopy uniformity, roughness), and morphological features (e.g., crop height, canopy coverage). For example, the Normalized Difference Vegetation Index (NDVI) of a crop at a specific growth stage should fall within a certain range, or its canopy shape should conform to a certain geometric model.

[0059] To improve the accuracy of confidence scores, geometric features can be matched with an agronomic rule base to correct the first confidence score; spectral features can be matched with the agronomic rule base to correct the second confidence score; texture features can be matched with the agronomic rule base to correct the third confidence score; and morphological features can be matched with the agronomic rule base to correct the fourth confidence score. Geometric, spectral, texture, and morphological features can be matched with the agronomic rule base separately. This matching process can be implemented using machine learning classifiers, such as using support vector machines to identify the matching degree. The matching process determines whether the currently identified crop features conform to their corresponding agronomic rules, thus correcting the confidence score. The confidence score obtained from the initial assessment can be adjusted based on the matching degree. If a feature highly matches an agronomic rule, its corresponding confidence score can be increased; conversely, if there is a significant deviation, the confidence score may need to be decreased, or further analysis may need to be triggered. This correction mechanism makes confidence assessment no longer solely dependent on the statistical characteristics of the image itself, but incorporates professional knowledge of crop growth, improving the accuracy and reliability of the assessment.

[0060] The corrected first, second, third, and fourth confidence levels are then weighted and summed to obtain the overall confidence level. Different weights can be assigned to different features based on their importance or reliability in crop type identification. Each corrected confidence level is multiplied by its respective weight and then summed to obtain the overall confidence level. For example, in some crop identification scenarios, spectral features may be more decisive than texture features, so spectral features can be given a higher weight. These weights can be set based on expert experience, or each feature can be assigned the same weight.

[0061] This embodiment incorporates an agronomic rule base, integrating the biological laws and agronomic knowledge of crop growth and development into the confidence assessment process. Specifically, after initially assessing the first confidence level of geometric features, the second confidence level of spectral features, the third confidence level of texture features, and the fourth confidence level of morphological features, instead of simply fusing them directly, these features extracted from the image are first matched with preset feature rules in the agronomic rule base. This matching process can correct the original confidence level, that is, enhance or weaken the confidence level based on the degree of conformity between the features and agronomic laws. For example, if the identified crop spectral features are highly consistent with the typical spectral response of the crop at the current growth stage, the corresponding second confidence level will be increased; conversely, if abnormal features that contradict agronomic laws appear, the corresponding confidence level will be decreased. Through this correction based on agronomic knowledge, misidentification or low-confidence areas caused by factors such as image distortion, illumination changes, or sensor noise can be effectively filtered out, thus making the final comprehensive confidence level more biologically reasonable and practically reliable. Finally, by weighting and summing the confidence scores after agronomic rule corrections, the calculation of the overall confidence score can be further optimized based on the relative importance of different features in crop identification, ensuring that it can more accurately reflect the true credibility of crop type identification results.

[0062] To illustrate the technical solution more clearly, a specific example is used below. Suppose we need to identify corn and soybeans in a field. After initial identification, the system evaluates the geometric, spectral, textural, and morphological features of each region and calculates the corresponding first, second, third, and fourth confidence levels. At this point, an agronomic rule base is introduced. For example, the rule base may contain the following rules: (1) The NDVI value of corn during the jointing stage is usually between 0.7 and 0.9, and the leaves are long and narrow with relatively uniform canopy texture. (2) The NDVI value of soybeans during the flowering stage is usually between 0.6 and 0.8, the leaves are elliptical, and the canopy texture may be slightly uneven due to the distribution of flowers. (3) In the row planting mode, the row spacing of corn is usually 60-80 cm and the plant spacing is 20-30 cm.

[0063] When a region is initially identified as corn and its spectral features have a high second confidence level, but its geometric features (such as row spacing) deviate significantly from the typical row spacing of corn in the rule base, the system will appropriately reduce the first confidence level of the region based on the matching results of the agronomic rule base. Conversely, if all features of a region highly match the features of soybeans in the rule base, the corresponding confidence levels will be enhanced. After agronomic rule correction, the corrected confidence levels are weighted and summed according to preset weights (e.g., spectral feature weight 0.4, geometric feature weight 0.3, texture feature weight 0.15, morphological feature weight 0.15) to finally obtain the comprehensive confidence level of the region. For example, if the corrected first confidence level is 0.7, the second confidence level is 0.9, the third confidence level is 0.8, and the fourth confidence level is 0.85, then the overall confidence level could be calculated as 0.7*0.3 + 0.9*0.4 + 0.8*0.15 + 0.85*0.15 = 0.21 + 0.36 + 0.12 + 0.1275 = 0.8175. In this way, even when there is some uncertainty in the image data, a more accurate and reliable confidence level for crop type identification can be obtained with the assistance of agronomic knowledge.

[0064] Through the above technical solution, this embodiment can significantly improve the accuracy and robustness of the confidence assessment of UAV remote sensing crop planting structure image recognition results. This embodiment introduces an agronomic rule base, integrating professional agricultural knowledge and crop growth patterns into the confidence assessment process, effectively avoiding misjudgments that may result from purely data-driven approaches. Specifically, by matching geometric features, spectral features, texture features, and morphological features with agronomic rules and correcting the corresponding confidence levels, potential biases caused by environmental factors (such as uneven lighting, cloud cover), sensor errors, or image processing limitations can be effectively corrected, making the assessment results more consistent with the actual growth state of crops. Furthermore, weighted summation of the corrected confidence levels allows for flexible adjustments based on the importance of different features in specific recognition tasks, further optimizing the calculation accuracy of the overall confidence level.

[0065] In some embodiments, after obtaining the agronomic rule base in step S301, the method may further include, but is not limited to, the following steps: Step S401: Obtain images of the experimental planting area; Step S402: Construct a crop feature template library; Step S403: Extract features from the image of the experimental planting area to obtain the features of the planting structure area; Step S404: Match the regional features of the planting structure with the crop feature template library to obtain the target feature template; Step S405: Based on the target feature template, perform a difference analysis on the features of the planting structure area to obtain local feature differences; Step S406: Generate supplementary rules for the characteristics of the planting structure region based on the differences in local features; Step S407: Update the agronomic rule base according to the supplementary rules.

[0066] In some embodiments, images of experimental planting areas can be acquired first. Remote sensing image data of specific experimental plots can be obtained through drone remote sensing. These images of experimental planting areas typically contain clear crop planting information and growth status, serving as a reliable data source for calibrating and validating agronomic rule bases. Simultaneously, a crop characteristic template library is constructed. For example, a template library containing typical characteristic patterns of various crops at different growth stages and under different environmental conditions can be established. This template library can cover various characteristics such as crop growth cycle characteristics, response characteristics to environmental stresses, family and genus classification, planting density range, pest and disease status, and nutritional status, for subsequent comparison with the characteristics of actual planting structure areas.

[0067] Then, feature extraction is performed on the images of the experimental planting area to obtain the planting structure area features. Image processing and pattern recognition techniques can be used to extract feature information reflecting the crop planting structure and growth status from the acquired images of the experimental planting area. This feature information forms the basis for matching with a crop feature template library. The planting structure area features are then matched with the crop feature template library to obtain target feature templates. By comparing the extracted planting structure area features with the various feature templates in the template library, the crop feature template that best matches the current experimental area can be identified.

[0068] Next, based on the target feature template, a difference analysis is performed on the characteristics of the planting structure region to obtain local feature differences. After determining the target feature template, subtle differences between the actual extracted planting structure region characteristics and the target template can be further analyzed. These differences may reflect local variations, abnormal growth, or special cases not covered by a general rule base in a specific region. Supplementary rules for the planting structure region characteristics are then generated based on the local feature differences. These supplementary rules, designed to address these local feature differences, can be automatically generated based on the difference analysis results. These supplementary rules aim to compensate for the deficiencies of the existing agronomic rule base, making it more targeted.

[0069] Finally, the agronomic rule base is updated based on the supplementary rules. The newly generated supplementary rules are integrated into the existing agronomic rule base, enabling the agronomic rule base to dynamically adapt to new planting environments and crop growth conditions, thereby improving its accuracy and applicability.

[0070] This embodiment achieves dynamic updates to the agronomic rule base by introducing the analysis of experimental planting area images and the construction of a crop feature template library. Specifically, by acquiring actual experimental planting area images and extracting planting structure area features, information on the current actual crop growth and planting patterns can be obtained. Matching these actual features with the pre-constructed crop feature template library can identify the closest typical patterns. Based on this, difference analysis reveals local differences between the actual situation and the typical patterns. These local differences are precisely the parts that the existing agronomic rule base may not fully cover or accurately describe. By transforming these local differences into supplementary rules and integrating them into the agronomic rule base, the rule base can continuously learn and adapt to new and specific planting environments and crop performance, thereby overcoming the limitations of static rule bases in the face of complex and ever-changing agricultural environments. This continuous, data-driven rule update mechanism enables the agronomic rule base to more accurately reflect the true state of crops, providing a more reliable basis for subsequent confidence assessments.

[0071] Through the above technical solution, this embodiment significantly improves the accuracy of the agronomic rule base and its adaptability to specific regions and planting patterns by introducing data from actual experimental planting areas for calibration and supplementation. Therefore, in the subsequent confidence assessment process, the corrected first, second, third, and fourth confidence levels will more accurately reflect the reliability of the crop type identification results, making the final comprehensive confidence level more realistic and reliable, and providing a more solid data foundation for generating crop planting structure images. This dynamic update mechanism effectively enhances the robustness and practicality of the entire identification method.

[0072] In some embodiments, step S402, constructing the crop feature template library, may include, but is not limited to, the following steps: Based on the growth cycle characteristics of crops, a first feature template is constructed; Based on the response characteristics of crops to environmental stress, a second feature template is constructed; Construct a third feature template based on the family and genus classification of crops; Based on the range of crop planting density, a fourth feature template is constructed; Based on the status of crop diseases and pests, construct the fifth feature template; Based on the nutritional status of crops, a sixth feature template is constructed; Based on the template granularity, the first feature template, the second feature template, the third feature template, the fourth feature template, the fifth feature template, and the sixth feature template are integrated to generate a crop feature template library.

[0073] In some embodiments, a first feature template can be constructed based on the growth cycle characteristics of crops to capture the unique spectral, geometric, textural, and morphological features exhibited by crops at different growth stages (e.g., seedling stage, vegetative growth stage, reproductive growth stage, and maturity stage). For example, during the seedling stage, the crop canopy is small, and the spectral reflectance may be greatly affected by the soil background; while during the maturity stage, the canopy coverage is high, and the spectral features mainly reflect the physiological state of the crop itself.

[0074] Then, based on the response characteristics of crops to environmental stress, a second feature template was constructed to record the physiological and morphological responses of crops when subjected to various environmental stresses (such as drought, flood, high temperature, low temperature, salinity, nutrient deficiency or excess). For example, drought stress may cause leaves to curl, darken or lighten in color, while diseases may manifest as specific spots, necrotic areas, or abnormal growth.

[0075] Next, a third characteristic template is constructed based on the family and genus classification of crops. Since crops from different families and genera exhibit significant differences in leaf shape, plant structure, and inflorescence characteristics, a third characteristic template can be constructed based on the biological classification of crops (such as Poaceae, Leguminosae, Brassicaceae, etc.). These family and genus classifications help distinguish different crop types, even if they exhibit similar spectral characteristics at certain growth stages.

[0076] Based on the range of crop planting densities, a fourth feature template is constructed to focus on the characteristic performance of crops under different planting densities. Planting density affects competition among crops, canopy structure, and light utilization efficiency, thereby influencing their spectral reflectance and texture features. For example, high-density planting may result in a denser canopy and a more uniform texture.

[0077] Based on the status of crop diseases and pests, a fifth feature template is constructed to describe the typical symptoms and characteristics of crops when they are attacked by different diseases and pests. This includes the color, shape, size, and distribution pattern of lesions, as well as the feeding marks, galls, or wilting caused by pests.

[0078] Based on the nutritional status of crops, a sixth characteristic template is constructed to reflect the physiological and morphological characteristics of crops under different nutritional conditions. For example, nitrogen deficiency may cause leaves to turn yellow, phosphorus deficiency may cause leaves to turn purple, and potassium deficiency may cause leaf margins to scorch.

[0079] Finally, based on the template granularity, the first, second, third, fourth, fifth, and sixth feature templates are integrated to generate a crop feature template library. Template granularity refers to the level of detail or abstraction in feature description. For example, features from different growth stages and under different stress types can be organized hierarchically, or features from different families and genera or planting densities can be cross-correlated. Through this integration, a multi-level, multi-dimensional crop feature template library can be formed. This library can comprehensively and systematically describe various states and characteristics of crops, thus providing richer and more accurate references for subsequent crop feature matching and anomaly detection.

[0080] This embodiment constructs a multi-dimensional, multi-level crop feature template library, enabling a more comprehensive capture of the intrinsic characteristics and extrinsic manifestations of crops. Specifically, the first to sixth feature templates model crop characteristics from multiple key dimensions, including growth cycle, environmental stress, family and genus classification, planting density, pest and disease status, and nutritional status. This multi-dimensional feature description allows the crop feature template library to cover a wider range of crop state variations, thereby enabling more accurate differentiation between normal and abnormal planting structure area features during subsequent feature extraction and matching. For example, when crops in an experimental planting area image exhibit a certain abnormality, matching with this multi-dimensional template library can more accurately pinpoint the specific cause of the abnormality (such as disease, malnutrition, or environmental stress), rather than simply identifying the vague result of "abnormality." Furthermore, by integrating based on template granularity, the template library possesses good organization and scalability, adapting to the needs of different precision levels and application scenarios.

[0081] Through the above technical solution, the template library constructed in this embodiment can more accurately reflect the actual growth status and potential problems of crops. Therefore, in the subsequent matching process between planting structure regional features and the crop feature template library, higher matching accuracy and stronger robustness can be achieved, effectively reducing the risk of misjudgment and missed judgment. This not only helps to more accurately identify crop types and growth states, but also provides a more reliable basis for generating supplementary rules for planting structure regional features, thereby optimizing and updating the agronomic rule base, and ultimately improving the overall accuracy and practicality of the UAV remote sensing crop planting structure image recognition method.

[0082] In some embodiments, step S403 involves extracting features from the experimental planting area image to obtain planting structure area features, which may include, but is not limited to, the following steps: The experimental planting area image was divided into multiple sub-regions; Extracting regional sub-features from sub-regions, including geometric sub-features, spectral sub-features, texture sub-features, and morphological sub-features; Multi-dimensional fusion of regional sub-features is performed to construct a multi-dimensional feature vector space; Identify cluster centers and boundaries in a multi-dimensional feature vector space; Identify local anomalous features in sub-regions that deviate from the cluster center and exceed the boundaries. These local anomalous features include irregular geometric shapes, anomalous spectral responses, anomalous texture patterns, or atypical morphological structures. Weight decay processing is applied to local abnormal features; Median filtering is applied to the regional sub-features and local anomaly features after weight decay processing to generate regional features of planting structures.

[0083] In some embodiments, the experimental planting area image can be first divided into multiple sub-regions. The entire experimental planting area image can be segmented into multiple smaller, independently analyzable sub-regions. For example, regular grid division, semantic segmentation based on image content, or segmentation based on specific crop row spacing can be used. The aim is to decompose complex images into more easily processed and analyzed units to facilitate the extraction of local features and anomaly detection.

[0084] Then, regional sub-features are extracted from the sub-regions. For each sub-region, key information inherent in the crop planting structure can be extracted. These regional sub-features include geometric sub-features, spectral sub-features, texture sub-features, and morphological sub-features. Geometric sub-features describe the spatial attributes of crops, such as their shape, size, and arrangement, including canopy area, perimeter, aspect ratio, row spacing, and plant spacing. Spectral sub-features refer to the reflectance or absorption characteristics of crops in different wavelengths, such as the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and reflectance values ​​in red, green, blue, and near-infrared bands. Texture sub-features refer to the regularity of pixel grayscale or color changes in an image, such as contrast, homogeneity, entropy, and second moment of angle; these features can reflect the roughness, smoothness, or complexity of the crop canopy. Morphological sub-features refer to the morphological structural characteristics of individual crops or groups, such as the shape (circular, elliptical, irregular), size, and arrangement pattern (regular arrangement, random distribution) of the crop canopy in the field. Its purpose is to comprehensively capture information on all aspects of crop planting structure.

[0085] Then, multi-dimensional fusion of regional sub-features is performed to construct a multi-dimensional feature vector space. Regional sub-features extracted from different dimensions (geometry, spectrum, texture, morphology) can be integrated into a unified representation to construct the multi-dimensional feature vector space. For example, all extracted sub-feature values ​​can be concatenated to form a high-dimensional vector, or dimensionality reduction techniques such as Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) can be used to map high-dimensional features to a lower-dimensional but more information-rich space. The aim is to comprehensively utilize multi-source information to improve the expressive power and discriminative power of features. Simultaneously, cluster centers and boundaries are identified in the multi-dimensional feature vector space. Clustering algorithms (such as K-means, DBSCAN, Gaussian mixture models, etc.) can be used to analyze the feature vectors, find typical feature sets representing different planting structure patterns (i.e., cluster centers), and determine the range of each cluster (i.e., boundaries). The purpose is to establish a benchmark model of normal planting structure characteristics, providing a reference for the subsequent identification of abnormal features.

[0086] This method identifies local anomalous features in sub-regions that deviate from cluster centers and extend beyond boundaries. Features that do not conform to normal planting structure patterns can be detected by comparing the feature vectors of each sub-region with the identified cluster centers and boundaries. Local anomalous features can include irregular geometric shapes (e.g., crop lodging, missing seedlings, gaps in rows), abnormal spectral responses (e.g., abnormal leaf color due to pests and diseases, changes in reflectance due to water stress), abnormal texture patterns (e.g., texture chaos due to weed invasion), or atypical morphological structures (e.g., abnormal crop growth). The goal is to accurately locate and identify anomalies in the image that may affect the accuracy of planting structure features. Simultaneously, weight attenuation is applied to local anomalous features. In feature fusion or subsequent analysis, the influence of these identified local anomalous features can be reduced. For example, a lower weight can be assigned to anomalous features, or their values ​​can be compressed using a nonlinear function. The aim is to reduce the interference of anomalous data on the overall planting structure feature representation and improve the robustness of the features.

[0087] Finally, median filtering is applied to the regional sub-features and the local anomaly features after weight decay processing to generate planting structure regional features. Median filtering can be used to smooth the processed feature data. Median filtering is a non-linear digital filtering technique that eliminates noise by replacing the feature value of each point with the median of all feature values ​​in its neighborhood. Its purpose is to further remove random noise and isolated outliers from the feature data, making the final generated planting structure regional features smoother and more stable.

[0088] This embodiment meticulously divides the experimental planting area image into multiple sub-regions and extracts sub-features from these regions from multiple dimensions, constructing a comprehensive feature vector space. Based on this, by identifying cluster centers and boundaries, local anomalous features deviating from the normal pattern can be effectively identified. By applying weight attenuation to these local anomalous features and combining them with median filtering, this embodiment significantly reduces the negative impact of anomalous data on overall feature extraction, thereby ensuring that the generated planting structure area features are purer, more accurate, and more representative. This refined processing method overcomes the limitations of being susceptible to noise and local anomalies when facing complex and ever-changing agricultural scenarios.

[0089] To illustrate this technical solution more clearly, a specific example is used below. Suppose that in an image of an experimental planting area, some crops exhibit abnormal leaf color (spectral feature abnormalities) and irregular growth patterns (morphological feature abnormalities) due to localized diseases, while a small number of weeds have invaded, causing localized texture disorder (texture feature abnormalities). First, the image is divided into multiple sub-regions. Next, geometric sub-features (such as crop canopy area), spectral sub-features (such as NDVI value), texture sub-features (such as contrast), and morphological sub-features (such as crop shape) are extracted from each sub-region. These sub-features are fused to construct a multi-dimensional feature vector space. Within this feature vector space, a clustering algorithm identifies the feature cluster centers and boundaries of healthy crop areas. Sub-regions affected by diseases and weeds will have feature vectors that significantly deviate from the cluster centers of healthy crops and extend beyond their boundaries, thus being identified as locally anomalous features.

[0090] Subsequently, these identified local anomalous features are subjected to weight attenuation processing, for example, reducing their weight in the feature vector by 50% to weaken their impact on the overall feature representation. Finally, median filtering is applied to the weight-attenuated local anomalous features and regional sub-features of normal areas to further smooth the data and remove residual noise, ultimately generating a feature that accurately reflects the main planting structure of the experimental planting area without being excessively disturbed by local diseases or weed anomalies.

[0091] Through the above technical solution, this embodiment can obtain more accurate and robust planting structure region features, effectively avoiding feature extraction deviations caused by image noise or local anomalies. This enables more accurate identification of target feature templates and more reliable difference analysis when subsequently matching planting structure region features with a crop feature template library, thereby generating more targeted and effective supplementary rules. Ultimately, this helps to update and improve the agronomic rule base, enhance the accuracy and adaptability of crop planting structure image recognition, and provide more reliable data support for precision agriculture management.

[0092] The beneficial effects of implementing the embodiments of the present invention include: First, remote sensing image data is acquired, then the remote sensing image data is preprocessed to obtain farmland orthophoto maps and distortion risk areas, then crop type identification is performed on the farmland orthophoto maps to obtain crop type identification results, and based on the distortion risk areas, the confidence level of the crop type identification results is evaluated to obtain a comprehensive confidence level, and finally, based on the crop type identification results and the comprehensive confidence level, a crop planting structure image is generated. This allows for the evaluation of the confidence level of crop types by combining distortion risk areas, thereby achieving crop planting structure image identification and improving accuracy.

[0093] like Figure 2 As shown, this embodiment of the invention also provides an unmanned aerial vehicle (UAV) remote sensing image recognition system for crop planting structures, comprising: Data acquisition module 501 is used to acquire remote sensing image data; Preprocessing module 502 is used to preprocess remote sensing image data to obtain farmland orthophoto maps and distortion risk areas; The type recognition module 503 is used to identify crop types in farmland orthophoto maps and obtain crop type recognition results. The confidence assessment module 504 is used to assess the confidence of crop type identification results based on the distortion risk area to obtain a comprehensive confidence score. The map generation module 505 is used to generate crop planting structure images based on crop type identification results and comprehensive confidence levels.

[0094] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0095] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

Claims

1. A method for recognizing crop planting structure images from unmanned aerial vehicle (UAV) remote sensing images, characterized in that, Includes the following steps: Acquire remote sensing image data; The remote sensing image data is preprocessed to obtain farmland orthophoto maps and distortion risk areas; The crop type identification was performed on the orthophoto map of the farmland to obtain the crop type identification results; Based on the distortion risk area, the confidence level of the crop type identification results is evaluated to obtain a comprehensive confidence level; Based on the crop type identification results and the overall confidence level, a crop planting structure image is generated.

2. The method according to claim 1, characterized in that, The preprocessing of the remote sensing image data to obtain farmland orthophoto maps and distortion risk areas includes: Radiometric correction is performed on the remote sensing image data; The radiometrically corrected remote sensing images are stitched together to obtain the orthophoto map of the farmland. During the stitching process, overlapping areas between multiple radiometrically corrected remote sensing images are identified; The overlapping region is analyzed for geometric distortion using a scale-invariant feature transformation algorithm to obtain the distortion risk region.

3. The method according to claim 1, characterized in that, The step of assessing the confidence level of the crop type identification results based on the distortion risk region to obtain a comprehensive confidence level includes: Acquire attitude fluctuation data of the drone; Acquire spectral response drift data from a multispectral sensor; Geometric features, spectral features, texture features, and morphological features are extracted from the distortion risk area. The geometric features include the local curvature and straightness of the crop boundary. The spectral features include the normalized difference vegetation index and reflectance values ​​of each band. The texture features include contrast, homogeneity, and entropy. The morphological features include the shape, size, and arrangement pattern of the crop canopy. Based on the crop type identification results, the attitude fluctuation data, and the spectral response drift data, the first confidence level of the geometric features, the second confidence level of the spectral features, the third confidence level of the texture features, and the fourth confidence level of the morphological features are evaluated. The first confidence level, the second confidence level, the third confidence level, and the fourth confidence level are fused together to obtain the comprehensive confidence level.

4. The method according to claim 3, characterized in that, The acquisition of the drone's attitude fluctuation data includes: The attitude information of the UAV body is obtained by the first inertial measurement unit installed on the UAV body; The attitude information of the airborne sensor gimbal is obtained by a second inertial measurement unit installed on the airborne sensor gimbal. The attitude information of the UAV body and the attitude information of the airborne sensor gimbal are compared to calculate the dynamic attitude deviation of the airborne sensor gimbal relative to the UAV body. The attitude fluctuation data is generated based on the dynamic attitude deviation and the attitude information of the UAV body.

5. The method according to claim 3, characterized in that, The acquisition of spectral response drift data from the multispectral sensor includes: After setting multiple miniature spectral calibration modules inside the multispectral sensor, the miniature spectral calibration modules are controlled to emit light with target spectral characteristics toward the detector array. The miniature spectral calibration modules include a stable light source and multiple narrowband filters. Record the response data of the detector array; The response data is compared with a preset response reference to calculate the response deviation; The spectral response drift data is generated based on the response deviation and the response data.

6. The method according to claim 3, characterized in that, The process of fusing the first confidence level, the second confidence level, the third confidence level, and the fourth confidence level to obtain the comprehensive confidence level includes: Obtain an agronomic rule base, which contains characteristic rules for different crop types under different planting patterns; The geometric features are matched with the agronomic rule base to correct the first confidence level; The spectral features are matched with the agronomic rule base to correct the second confidence level; The texture features are matched with the agronomic rule base to correct the third confidence level; The morphological features are matched with the agronomic rule base to correct the fourth confidence level; The weighted sum of the corrected first, second, third, and fourth confidence levels is used to obtain the overall confidence level.

7. The method according to claim 6, characterized in that, After acquiring the agronomic rule base, the method further includes: Acquire images of the experimental planting area; Construct a crop feature template library; Feature extraction was performed on the image of the experimental planting area to obtain the planting structure area features; The planting structure region features are matched with the crop feature template library to obtain the target feature template; Based on the target feature template, a difference analysis is performed on the features of the planting structure area to obtain local feature differences; Based on the differences in local features, supplementary rules for the features of the planting structure region are generated; Update the agronomic rule base according to the supplementary rules.

8. The method according to claim 7, characterized in that, The construction of the crop feature template library includes: Based on the growth cycle characteristics of crops, a first feature template is constructed; Based on the response characteristics of crops to environmental stress, a second feature template is constructed; Construct a third feature template based on the family and genus classification of crops; Based on the range of crop planting density, a fourth feature template is constructed; Based on the status of crop diseases and pests, construct the fifth feature template; Based on the nutritional status of crops, a sixth feature template is constructed; Based on the template granularity, the first feature template, the second feature template, the third feature template, the fourth feature template, the fifth feature template, and the sixth feature template are integrated to generate the crop feature template library.

9. The method according to claim 7, characterized in that, The step of extracting features from the image of the experimental planting area to obtain the features of the planting structure area includes: The image of the experimental planting area is divided into multiple sub-regions; Extract region sub-features from the sub-regions, the region sub-features including geometric sub-features, spectral sub-features, texture sub-features and morphological sub-features; The sub-features of the region are fused in multiple dimensions to construct a multi-dimensional feature vector space; In the multi-dimensional feature vector space, cluster centers and boundaries are identified; Identify local anomalous features in the sub-region that deviate from the cluster center and exceed the boundary, wherein the local anomalous features include irregular geometric shapes, anomalous spectral responses, anomalous texture patterns, or atypical morphological structures; The local anomaly features are subjected to weight decay processing; The planting structure region features are generated by performing median filtering on the sub-features of the region and the local anomaly features after weight decay processing.

10. A UAV remote sensing image recognition system for crop planting structures, characterized in that, include: The data acquisition module is used to acquire remote sensing image data; The preprocessing module is used to preprocess the remote sensing image data to obtain farmland orthophoto maps and distortion risk areas; The type identification module is used to identify the crop type in the orthophoto of the farmland and obtain the crop type identification result; The confidence assessment module is used to assess the confidence of the crop type identification results based on the distortion risk area to obtain a comprehensive confidence score. The map generation module is used to generate a crop planting structure image based on the crop type identification results and the comprehensive confidence level.

Images