Rice yield prediction method and system based on multi-source time sequence remote sensing data

Abstract

The invention provides a rice yield prediction method and system based on multi-source time sequence remote sensing data, and relates to the field of agricultural informatization and remote sensing technology application, and the method comprises the steps: collecting multi-spectral data and laser radar data, and carrying out the precise geographic positioning of the data through a high-precision satellite positioning system, obtaining multi-source unmanned aerial vehicle remote sensing data; performing camera internal orientation, wave band registration, orthoimage splicing and embedding, radiation correction and remote sensing index calculation on the multispectral data to obtain processed multispectral data; and performing point cloud reconstruction and point cloud filtering on the laser radar data, and calculating the height of the rice plant to obtain processed laser radar data. According to the method, multi-source remote sensing data and spatio-temporal data are fused, and remote sensing characteristics of different phenological periods of rice are dynamically analyzed and filtered, so that the precision and reliability of rice field yield prediction are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of agricultural informatization and remote sensing technology, and in particular to a rice yield prediction method and system based on multi-source time series remote sensing data. Background Art

[0002] The fusion of UAV multispectral image data and LiDAR data requires professional data processing technology and algorithm support. Its application areas are also becoming increasingly extensive, including agriculture, forestry, geology, meteorology and other aspects. In the agricultural field, remote sensing technology can provide an important basis for crop yield prediction by obtaining information such as the spectral characteristics and growth status of crops. However, there are still some challenges in the current data fusion technology, such as insufficient data registration accuracy, data redundancy and information loss, which will reduce the accuracy and reliability of the fused data. Combining spatiotemporal characteristics to build a high-precision yield prediction model requires professional knowledge background and algorithm support. However, due to the complexity and diversity of the rice growth process, the model construction process may face greater challenges, such as parameter tuning, model verification and result interpretation. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a rice yield prediction method and system based on multi-source time-series remote sensing data, which integrates multi-source remote sensing data and spatiotemporal data, and significantly improves the accuracy and reliability of rice field yield prediction by dynamically analyzing and filtering the remote sensing characteristics of rice in different phenological periods.

[0004] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0005] In a first aspect, a rice yield prediction method based on multi-source time series remote sensing data comprises:

[0006] Acquire multi-source UAV remote sensing data by collecting multispectral data and LiDAR data and using satellite positioning for accurate geolocation of the data;

[0007] Perform camera orientation, band registration, orthophoto stitching and mosaicking, radiation correction and remote sensing index calculation on multispectral data to obtain processed multispectral data;

[0008] Perform point cloud reconstruction and point cloud filtering on the LiDAR data and calculate the height of rice plants to obtain processed LiDAR data;

[0009] The processed lidar data is registered with the multispectral data to achieve preprocessing of multi-source UAV remote sensing data;

[0010] Based on the phenological period, multi-source remote sensing data of time series are obtained, and the time series data are optimized to obtain optimized time series data;

[0011] According to the optimized time series data, extract the feature features at the sub-region level to construct a feature feature set of the time series;

[0012] According to the feature set of the constructed time series, the measured rice yield data in the sub-region is obtained;

[0013] A training data set is constructed based on the preprocessed multi-source remote sensing data and the measured rice yield data in the sub-region;

[0014] The random forest regression model is trained according to the training data set, and the trained model is used to predict rice yield for new multi-source time series remote sensing data.

[0015] Furthermore, the multispectral data is subjected to camera orientation, band registration, orthophoto stitching and mosaicking, radiation correction and remote sensing index calculation to obtain processed multispectral data, including:

[0016] The image is geometrically corrected using the camera calibration parameters to obtain a corrected image and determine the internal orientation parameters;

[0017] According to the camera's internal orientation parameters, an accurate registration algorithm is used to register images of different bands to obtain multi-band images after band registration.

[0018] Perform feature matching on the overlapping areas of the multi-band images after band registration and apply an automated stitching algorithm to generate an orthophoto map with unified geographic coordinates;

[0019] Performing radiation correction on the orthophoto to obtain a radiation-corrected orthophoto;

[0020] According to the orthophoto map after radiation correction, Calculate the Normalized Difference Vegetation Index to get the NDVI image. NIR is the near infrared band reflectance, RED is the red band reflectance, and NDVI is the Normalized Difference Vegetation Index.

[0021] Furthermore, the LiDAR data is reconstructed and filtered, and the height of the rice plants is calculated to obtain processed LiDAR data, including:

[0022] By flying scans, the original point cloud data of the rice field area is collected to obtain the original point cloud data;

[0023] Splice and integrate the point cloud data and eliminate redundant points in overlapping areas to obtain reconstructed point cloud data;

[0024] According to the reconstructed point cloud data, height threshold filtering and density threshold filtering are performed to obtain filtered point cloud data;

[0025] According to the filtered point cloud data, the actual height image of the vegetation is calculated through CHM=DSM-DEM to obtain the processed lidar data, where DSM represents the combined height of the terrain and the object, DEM represents the pure terrain height, and CHM represents the actual height of the vegetation.

[0026] Furthermore, the processed lidar data is registered with the multispectral data to achieve the preprocessing of multi-source UAV remote sensing data, including:

[0027] Reprojecting the lidar data and the multispectral data to the same coordinate system, and analyzing the reprojections of the lidar data and the multispectral data to obtain reprojection analysis results;

[0028] According to the reprojection analysis results, the lidar data is resampled to a resolution consistent with the multispectral data to obtain the resampled data; the resampled data is vector-clipped to obtain image data covering the same geographic range:

[0029] The image data is aligned with the actual geographic reference data, and the position and angle of the image are adjusted for correction so that the residual error of the corrected image is controlled within ±2cm to achieve the preprocessing of multi-source UAV remote sensing data.

[0030] Furthermore, based on the phenological period, multi-source remote sensing data of time series are obtained, and the time series data are optimized to obtain optimized time series data, including:

[0031] According to the phenological period of rice growth, the analysis time nodes are obtained, including tillering period, jointing period and heading period;

[0032] According to the analysis time nodes, the corresponding UAV flight plan is formulated to collect remote sensing data to build a complete time series data set;

[0033] According to the time series data set, NDVI and CHM images are fused to obtain a time series data set;

[0034] According to the time series data set, set the upper and lower limits of NDVI and To achieve adaptive threshold filtering of NDVI, where NDVI is the normalized vegetation index of multispectral images, NDVI(x, y, t) is the NDVI value at position (x, y) at phenological period t, T NDVI,min (t) and T NDVI,max (t) are the minimum and maximum values ​​of NDVI in phenological period t, respectively. f (x, y, t) is the NDVI value after adaptive threshold filtering;

[0035] According to the time series data set, set the upper and lower limits of CHM and To achieve adaptive threshold filtering of canopy height, where CHM is the canopy height of the lidar data, CHM(x, y, t) is the CHM value of the position (x, y) at phenological period t, and is the adaptive threshold range of CHM in phenological period t, CHM f (x, y, t) is the CHM value after adaptive threshold filtering;

[0036] The time series data are optimized through adaptive threshold filtering of NDVI and adaptive threshold filtering of CHM to obtain optimized time series data.

[0037] Furthermore, based on the optimized time series data, the feature features at the sub-region level are extracted to construct a feature feature set of the time series. The feature features include vegetation index, terrain height, and include:

[0038] According to the optimized time series data, the research area is divided into multiple sub-areas, each of which contains a number of pixels. All the pixels in each sub-area are marked and classified to achieve sub-area division;

[0039] According to the sub-regional division, Calculate the average value of the normalized difference vegetation index after adaptive threshold filtering, through Calculate the average of the canopy height after adaptive threshold filtering, where and They represent the normalized vegetation index and the mean canopy height of the sub-region under phenological period t, n is the number of valid pixels in the sub-region, and p represents the sub-region;

[0040] The feature set of time series was constructed based on the normalized difference vegetation index and the mean canopy height of the phenological period sub-regions.

[0041] Furthermore, based on the feature set of the constructed time series, the measured rice yield data in the sub-region is obtained, including:

[0042] The time series feature sets of all sub-regions are combined to form a multi-time series feature set for the entire study area;

[0043] According to the multi-time series feature set, the feature features in each sub-region are matched with the yield data of the sub-region through the corresponding geographical location and timestamp to obtain the measured yield data of rice in the sub-region.

[0044] In the second aspect, a rice yield prediction system based on multi-source time series remote sensing data comprises:

[0045] The acquisition module is used to acquire multi-source UAV remote sensing data by collecting multispectral data and LiDAR data and using satellite positioning to accurately locate the data; perform in-camera orientation, band registration, orthophoto stitching and mosaicking, radiation correction and remote sensing index calculation on the multispectral data to obtain processed multispectral data; perform point cloud reconstruction and point cloud filtering on the LiDAR data and calculate the height of rice plants to obtain processed LiDAR data; register the processed LiDAR data with the multispectral data to achieve preprocessing of multi-source UAV remote sensing data;

[0046] The processing module is used to obtain multi-source remote sensing data of time series based on phenological period, optimize the time series data to obtain optimized time series data; extract sub-region level feature features according to the optimized time series data to construct a feature feature set of the time series, the feature features include vegetation index, terrain height; obtain the measured yield data of rice in the sub-region according to the constructed feature feature set of the time series; construct a training data set according to the pre-processed multi-source remote sensing data and the measured yield data of rice in the sub-region; train the random forest regression model according to the training data set, and use the trained model to predict the rice yield of new multi-source time series remote sensing data.

[0047] According to a third aspect, a computing device includes:

[0048] one or more processors;

[0049] The storage device is used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method described.

[0050] In a fourth aspect, a computer-readable storage medium stores a program, and when the program is executed by a processor, the method described is implemented.

[0051] The above solution of the present invention includes at least the following beneficial effects:

[0052] By using drone platforms for data collection, drones can conduct refined monitoring at low altitudes, which is suitable for small-scale farmland, and can flexibly arrange flight missions to improve the real-time and operability of data collection. The drone multispectral image data and lidar data are integrated. Multispectral data captures the spectral characteristics of rice and reflects the growth status of crops; lidar data provides accurate three-dimensional structural information, realizes the acquisition of comprehensive crop phenotypic information, and thus improves the reliability and accuracy of yield prediction. Through spatiotemporal data fusion, especially the fusion of multi-source remote sensing data of key phenological periods, the changes in the rice growth process can be dynamically monitored. Combining spatiotemporal characteristics, a high-precision yield prediction model is constructed, which enhances the robustness of the model and improves the ability to predict yield fluctuations. The adaptive threshold filtering technology based on the phenological period can dynamically adjust the threshold of remote sensing data according to the characteristics of different rice growth periods. This method ensures the accuracy and effectiveness of data at each growth stage, making yield prediction more accurate. It not only has higher prediction accuracy, but can also be applied in large-scale agricultural production, adapting to various agricultural scenarios, and providing more flexible and reliable rice yield prediction services. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of a rice yield prediction method based on multi-source time series remote sensing data provided by an embodiment of the present invention.

[0054] Figure 2 It is a schematic diagram of a rice yield prediction system based on multi-source time series remote sensing data provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in a form and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0056] like Figure 1 As shown, an embodiment of the present invention proposes a rice yield prediction method based on multi-source time series remote sensing data, and the method comprises the following steps:

[0057] Step 11, by collecting multi-spectral data and lidar data, and using satellite positioning to accurately geolocate the data, to obtain multi-source UAV remote sensing data;

[0058] Step 12, performing camera orientation, band registration, orthophoto stitching and mosaicking, radiation correction and remote sensing index calculation on the multispectral data to obtain processed multispectral data;

[0059] Step 13, performing point cloud reconstruction and point cloud filtering on the laser radar data, and calculating the height of the rice plants to obtain processed laser radar data;

[0060] Step 14, registering the processed lidar data with the multispectral data to achieve preprocessing of multi-source UAV remote sensing data;

[0061] Step 15, based on the phenological period, obtaining multi-source remote sensing data of the time series, and optimizing the time series data to obtain optimized time series data;

[0062] Step 16, extracting sub-region level feature features based on the optimized time series data to construct a feature feature set of the time series;

[0063] Step 17, obtaining the measured yield data of rice in the sub-region according to the element feature set of the constructed time series;

[0064] Step 18, constructing a training data set based on the preprocessed multi-source remote sensing data and the measured rice yield data in the sub-region;

[0065] The random forest regression model is trained according to the training data set, and the trained model is used to predict rice yield for new multi-source time series remote sensing data.

[0066] In an embodiment of the present invention, by collecting multispectral data and laser radar data, and using satellite positioning to accurately locate the data, more accurate and comprehensive farmland information can be obtained, which is helpful to understand the growth status and environmental factors of rice, thereby improving the accuracy of yield prediction. A series of processing is performed on the multispectral data and the laser radar data, such as camera internal orientation, band registration, orthophoto stitching and mosaicking, and radiation correction, which can eliminate noise and errors in the data and improve data quality. The processed laser radar data is registered with the multispectral data to realize the preprocessing of multi-source UAV remote sensing data, which can comprehensively utilize the advantages of different data sources and provide richer and more comprehensive information. Based on the phenological period, multi-source remote sensing data of the time series is obtained, and these data are optimized, which can capture the dynamic changes in the rice growth process. According to the optimized time series data, the sub-region level feature features, such as vegetation index and terrain height, are extracted, which helps to improve the accuracy and reliability of yield prediction. The training data set is constructed using the preprocessed multi-source remote sensing data and the measured yield data of rice in the sub-region, and the random forest regression model is trained, which can take advantage of multi-source data and machine learning technology.

[0067] In a preferred embodiment of the present invention, the above step 11 includes:

[0068] Determine data collection areas and targets, and identify drones and remote sensing equipment, including multispectral cameras and lidar;

[0069] The drone is controlled to fly according to the planned flight route, multispectral data is collected through the multispectral camera, and elevation information of the terrain is collected through the lidar to generate point cloud data.

[0070] The satellite positioning system is used to accurately geo-locate the collected multispectral data and lidar data, and the collected multispectral data and lidar data are sorted and matched according to the geographic location information to obtain multi-source UAV remote sensing data.

[0071] In an embodiment of the present invention, by clarifying the data collection area and target, drones and remote sensing equipment can be deployed in a targeted manner. The drone has a high degree of flexibility and accessibility, and can easily cover complex terrain and difficult-to-reach areas. The multispectral camera can capture spectral information of different bands to reflect the various characteristics of the ground objects, such as vegetation health and soil composition. The laser radar can accurately obtain the elevation information of the terrain and generate high-resolution point cloud data. The satellite positioning system is used to accurately geolocate the collected data. The high-precision positioning service can effectively improve the accuracy of the drone route and the accuracy of data alignment, ensuring that the spatial data in the scanning area can be accurately mapped. Accurate geographic positioning enables multi-source data to be effectively integrated and compared in the same geographic framework, improving the availability and analysis accuracy of the data. The multispectral data and the laser radar data are sorted and matched according to the geographic location information to form a multi-source drone remote sensing data set, which can comprehensively utilize the advantages of different data sources.

[0072] In a specific embodiment of the present invention, the data collection area and target are determined, and a DJI Matrice350RTK drone equipped with a Changguang Yuchen multispectral camera and a DJI Zenmuse L1 laser radar camera is selected. Shooting is performed using a Changguang Yuchen multispectral camera, which has multi-band capture capabilities, covering visible light, near-infrared and red edge bands; the collected bands are set to 450nm@35nm, 555nm@27nm, 660nm@22nm, 720nm@10nm, 750nm@10nm and 840nm@30nm. This band combination can fully capture the spectral characteristics of crops, and the red edge and near-infrared bands provide sensitive information on the status of vegetation to obtain multispectral data, wherein the drone flight altitude is set to 50m, the heading overlap rate is set to 80%, and the lateral overlap rate is set to 70%. Scanning and recording was performed through the DJI Zenmuse L1 LiDAR camera, which uses high-precision LiDAR technology to capture accurate 3D point cloud data in large rice areas to obtain LiDAR data. The flight parameters were set as follows: the drone flight altitude was set to 50m, and the laser lateral overlap rate was set to 60%; to ensure that the visible light lateral overlap rate of the mapping camera was greater than 65%, and the visible light heading overlap rate was set to 80%. The positioning service uses Qianxun's custom network RTK to provide centimeter-level positioning accuracy.

[0073] In a preferred embodiment of the present invention, the above step 12 includes:

[0074] The image is geometrically corrected using the camera calibration parameters to obtain a corrected image and determine the internal orientation parameters;

[0075] According to the camera's internal orientation parameters, an accurate registration algorithm is used to register images of different bands to obtain multi-band images after band registration.

[0076] Perform feature matching on the overlapping areas of the multi-band images after band registration and apply an automated stitching algorithm to generate an orthophoto map with unified geographic coordinates;

[0077] Performing radiation correction on the orthophoto to obtain a radiation-corrected orthophoto;

[0078] According to the orthophoto map after radiation correction, Calculate the Normalized Difference Vegetation Index to get the NDVI image. NIR is the near infrared band reflectance, RED is the red band reflectance, and NDVI is the Normalized Difference Vegetation Index.

[0079] In the embodiment of the present invention, geometric orientation is performed through camera calibration parameters, which can eliminate image distortion and ensure that the geographical location of each pixel is accurate. Accurate band registration can ensure that the information of the same geographical location in different bands is consistent, eliminating small offsets and distortions between bands, which helps to accurately analyze the spectral characteristics of objects and improve the interpretation ability of remote sensing data. Automated stitching technology can perform feature matching based on overlapping areas to generate orthophotos covering the entire study area, improve work efficiency, and ensure the continuity and seamlessness of the image; radiation correction can eliminate the influence of external factors on the radiation value of the image, so that the spectral information in the image more truly reflects the reflective characteristics of surface objects. By calculating remote sensing indices, such as the normalized vegetation index, the growth status and coverage of vegetation can be quantitatively evaluated.

[0080] In a specific embodiment of the present invention, the Changguang Yuchen multispectral preprocessing software is used to perform in-camera orientation correction on the collected images, and the camera calibration parameters, such as focal length, principal point offset, and lens distortion parameters, are used to perform geometric correction on the images to eliminate image distortion. The different band images taken by the multispectral camera are accurately aligned by using the Changguang Yuchen multispectral preprocessing software, and the spatial positions between the bands are aligned to eliminate the slight offsets and distortions that may exist between the bands. Professional drone mapping software is used to splice and mosaic the multi-band images after band alignment to generate an orthophoto map covering the entire study area. The splicing process will use automated splicing and feature matching based on overlapping areas; the Changguang Yuchen multispectral preprocessing software is used to perform radiation correction on the images using standardized data from the on-site reflectance plate. The remote sensing index is calculated using a formula to preprocess the remote sensing index image.

[0081] In a preferred embodiment of the present invention, the above step 13 includes:

[0082] By flying scans, the original point cloud data of the rice field area is collected to obtain the original point cloud data;

[0083] Splice and integrate the point cloud data and eliminate redundant points in overlapping areas to obtain reconstructed point cloud data;

[0084] According to the reconstructed point cloud data, height threshold filtering and density threshold filtering are performed to obtain filtered point cloud data;

[0085] According to the filtered point cloud data, the actual height image of the vegetation is calculated through CHM=DSM-DEM to obtain the processed lidar data, where DSM represents the combined height of terrain and objects, DEM represents the pure terrain height, and CHM represents the actual height of vegetation.

[0086] In the embodiment of the present invention, the use of laser radar for flight scanning can capture the three-dimensional information of the rice field area with high precision and generate original point cloud data. By splicing and integrating the point cloud data and eliminating redundant points in the overlapping area, reconstructed point cloud data can be obtained, which improves the consistency and accuracy of the data; height threshold filtering and density threshold filtering can remove abnormal points and redundant points, purify the point cloud data, and not only optimize the accuracy and consistency of the data. Based on the filtered point cloud data, the actual height image of the vegetation can be accurately calculated.

[0087] In a specific embodiment of the present invention, the original point cloud data collected by the DJI Zenmuse L1 LiDAR camera is spliced ​​and integrated using DJI Zhitu software, and the standard point cloud data file is filtered using ENVI software to remove abnormal points and redundant points. The height threshold needs to be set according to the flight altitude of the drone to filter out points that exceed the range of ±2 meters; the density threshold is set by Set up, where n0 is the total number of points in the point cloud, N(A i ,r) is point A i The number of neighboring points within a sphere with a radius of r and a center represents the neighborhood density of the i-th point, te is the threshold percentage for setting the standard for screening density, and H is the density threshold. The filtered point cloud data is processed using ENVI software to obtain the combined height of terrain and plants and the terrain height. The precise plant height of rice can be extracted by combining the combined height of terrain and plants and the terrain height data, and the actual height of vegetation can be obtained through a formula to obtain the actual height image of vegetation after preprocessing.

[0088] In a preferred embodiment of the present invention, the above step 14 includes:

[0089] Reprojecting the lidar data and the multispectral data to the same coordinate system, and analyzing the reprojections of the lidar data and the multispectral data to obtain reprojection analysis results;

[0090] According to the reprojection analysis results, the lidar data is resampled to a resolution consistent with the multispectral data to obtain resampled data; the resampled data is vector-clipped to obtain image data covering the same geographic range;

[0091] By aligning the image data with the actual geographic reference data and adjusting the position and angle of the image for correction, the residual error of the corrected image is controlled within ±2cm to achieve preprocessing of multi-source UAV remote sensing data;

[0092] In an embodiment of the present invention, all data sets are reprojected to the same coordinate system, ensuring spatial consistency between different data sources, which is helpful for subsequent data fusion and analysis. Resampling the lidar data to a resolution consistent with the multispectral data can improve the comparability between the data. The matching of resolution enables information from different data sources to be compared and analyzed on the same scale, enhancing the availability and accuracy of the data. Vector clipping is performed on the resampled data to ensure that all data cover the same geographic range. By aligning with the actual geographic reference data and adjusting the position and angle of the image for correction, the geographic positioning accuracy of the remote sensing data can be significantly improved, and the corrected image residual is controlled within ±2cm, so that the lidar data and multispectral data can be effectively fused, and the preprocessed data has a unified coordinate system, resolution and geographic range.

[0093] In a specific embodiment of the present invention, the target coordinate system is determined to be WGS84, the original laser radar data and multispectral data are obtained, the laser radar data and multispectral data are opened using geographic information processing software, the target coordinate system is set to WGS84, the two data are reprojected respectively, the data before and after reprojection are compared, the corresponding relationship of the data in geographic space is determined, the errors and deformations that may occur during the reprojection process are analyzed, and the analysis results are recorded; the resolution of the multispectral data is checked and used as the target resolution of the laser radar data, and in the geographic information processing software, the target resolution is set to the resolution of the multispectral data according to the laser radar data, and the laser radar data before and after resampling are compared to check whether the resolution is consistent and whether the data quality is affected. According to the scope or actual needs of the multispectral data, a vector clipping frame is created, the resampled laser radar data and multispectral data, as well as the created vector clipping frame are input, and the clipping operation is performed to ensure that the two data are clipped to the same geographic scope. Ground control points are collected in the actual geographic area through GPS equipment, and the cropped image data and ground control point data are loaded. The position and angle of the image are adjusted to ensure that the corrected image residual is controlled within ±2cm, so as to obtain the corrected lidar data and multispectral data.

[0094] In a preferred embodiment of the present invention, the above step 15 includes:

[0095] According to the phenological period of rice growth, the analysis time nodes are obtained, including tillering period, jointing period and heading period;

[0096] According to the analysis time nodes, the corresponding UAV flight plan is formulated to collect remote sensing data to build a complete time series data set;

[0097] According to the time series data set, NDVI and CHM images are fused to obtain a time series data set;

[0098] According to the time series data set, set the upper and lower limits of NDVI and To achieve adaptive threshold filtering of NDVI, where NDVI is the normalized vegetation index of multispectral images, NDVI(x, y, t) is the NDVI value at position (x, y) at phenological period t, T NDVI,min (t) and T NDVI,max (t) are the minimum and maximum values ​​of NDVI in phenological period t, respectively. f (x, y, t) is the NDVI value after adaptive threshold filtering;

[0099] According to the time series data set, set the upper and lower limits of CHM and To achieve adaptive threshold filtering of canopy height, where CHM is the canopy height of the lidar data, CHM(x, y, t) is the CHM value of the position (x, y) at phenological period t, and is the adaptive threshold range of CHM in phenological period t, CHM f (x, y, t) is the CHM value after adaptive threshold filtering;

[0100] The time series data are optimized through adaptive threshold filtering of NDVI and adaptive threshold filtering of CHM to obtain optimized time series data.

[0101] In an embodiment of the present invention, through adaptive threshold filtering, outliers caused by environmental factors can be effectively removed, thereby improving the accuracy of the data; the optimized time series data is more stable and reliable, and the high-quality data set enables agricultural managers to make decisions more quickly. Accurate monitoring of key phenological periods in the growth process of rice helps to achieve precision agricultural management, improve resource utilization efficiency, and reduce environmental burden.

[0102] In a specific embodiment of the present invention, the phenological period information of rice in the study area is collected, and the tillering period, jointing period and heading period are selected as the key time nodes for analysis. According to the phenological period, a flight plan is formulated, and a drone is used to collect data at each key stage. Multispectral images and lidar data are collected at each stage to construct a complete time series data set. The normalized vegetation index of the multispectral image and the canopy height of the lidar data are fused to obtain a time series data set, and the data set includes 6 bands of data. According to the time series data set, the upper and lower limits of the normalized vegetation index are set, and the adaptive threshold filtering of the normalized vegetation index is realized. According to the time series data set, the upper and lower limits of the canopy height are set, and the adaptive threshold filtering of the canopy height is realized. The adaptive threshold filtering of the normalized vegetation index and the adaptive threshold filtering of the canopy height are optimized to obtain the optimized time series data.

[0103] In a preferred embodiment of the present invention, the above step 16 includes:

[0104] According to the optimized time series data, the research area is divided into multiple sub-areas, each of which contains a number of pixels. All the pixels in each sub-area are marked and classified to achieve sub-area division;

[0105] According to the sub-regional division, Calculate the average value of the normalized difference vegetation index after adaptive threshold filtering, through Calculate the average of the canopy height after adaptive threshold filtering, where and They represent the normalized vegetation index and the mean canopy height of the sub-region under phenological period t, n is the number of valid pixels in the sub-region, and p represents the sub-region;

[0106] The feature set of time series was constructed based on the normalized difference vegetation index and the mean canopy height of the phenological period sub-regions.

[0107] In an embodiment of the present invention, by dividing the sub-regions, the changes in vegetation growth and canopy structure can be analyzed at a finer spatial scale, which helps to improve the accuracy and resolution of the research. Calculating the average value for each sub-region can reduce redundant information in the data. Using the average value as a characteristic representative of the sub-region can more accurately reflect the overall condition of the region and reduce the impact of individual abnormal pixels on the results. The constructed time series feature set can clearly show the changes in each sub-region in different phenological periods, which is convenient for time series analysis and comparison.

[0108] In a specific embodiment of the present invention, based on the optimized time series data, the entire study area is divided into multiple sub-areas according to the actual field test plan and geographical boundaries. Each sub-area should contain a number of pixels, and ensure that these pixels are continuous in geographic space to ensure that the pixel data in each sub-area can effectively represent the rice growth status of the specific area. GIS software or programming scripts are used to achieve automated regional division and pixel labeling, and each pixel will be assigned a specific sub-area identifier. After the sub-area division is completed, for each sub-area, the number of valid pixels in the sub-area is counted, and the average value of the normalized vegetation index and the mean value of the canopy height of these pixels after adaptive threshold filtering under a specific phenological period are calculated to construct a feature set of the time series.

[0109] In a preferred embodiment of the present invention, the above step 17 includes:

[0110] The time series feature sets of all sub-regions are combined to form a multi-time series feature set for the entire study area;

[0111] According to the multi-time series feature set, the feature features in each sub-region are matched with the yield data of the sub-region through the corresponding geographical location and timestamp to obtain the measured yield data of rice in the sub-region.

[0112] In the embodiment of the present invention, by merging the time series feature set of all sub-regions, the vegetation growth and canopy structure characteristics of the entire study area can be more comprehensively reflected, avoiding the one-sidedness of local data. The feature characteristics are associated with the measured yield data to ensure the accuracy of data analysis. Accurately associating sub-region feature characteristics with yield data provides strong support for precision agricultural management. The constructed multi-time series feature set and the correlation analysis with yield data provide valuable data resources for research in the fields of agricultural science, ecology, etc.

[0113] In a specific embodiment of the present invention, a data structure in a programming language is used to store a time series element feature set of each sub-region. By traversing all sub-regions, their feature sets are added one by one to the total multi-time series feature set for merging. In the merged multi-time series feature set, each sub-region corresponds to a feature subset, and each feature subset contains the normalized vegetation index and the mean canopy height of the sub-region under all key phenological periods. After constructing the multi-time series feature set, the geographical location information of each sub-region in the multi-time series feature set is matched with the field test records to find the actual yield data of the corresponding sub-region. By creating a data structure containing feature subsets and corresponding target variables in a programming environment, the measured yield data of each sub-region is associated with the corresponding feature subset to achieve a one-to-one correspondence between the feature data and the yield data of each sub-region. Finally, the measured yield data of all sub-regions are sorted to form a complete target variable set.

[0114] In a preferred embodiment of the present invention, the above step 18 includes:

[0115] According to the preprocessed multi-source remote sensing data and the measured rice yield data in the sub-region, the features related to rice yield are extracted to obtain the extracted features, which include the average value of the normalized vegetation index and the mean value of the canopy height in different phenological periods;

[0116] The extracted features are integrated with the measured yield data of the corresponding sub-region to form a training dataset containing input features and target variables. The input features include the average value of the normalized vegetation index in each phenological period and the remote sensing characteristics of the canopy height, and the target variable is the measured rice yield of the corresponding sub-region.

[0117] In the embodiment of the present invention, by extracting features directly related to rice yield and integrating measured yield data, a more accurate prediction model can be constructed. Accurate yield prediction helps decision makers optimize the allocation and utilization of agricultural resources. The extracted features and integrated data sets provide valuable data resources for research in the fields of agricultural science, ecology, etc. The prediction model constructed based on the training data set can provide a scientific basis for agricultural decision-making. Accurate yield prediction helps to assess the risks and uncertainties of agricultural production.

[0118] In a specific embodiment of the present invention, pre-processed multi-source remote sensing data is loaded from a storage medium. The data should include vegetation index and canopy height of different phenological periods. Different phenological periods, such as tillering period, jointing period and heading period, are determined according to the growth cycle of rice and the time of acquiring remote sensing data. For the remote sensing data of each phenological period, the average value of the normalized vegetation index and the average value of the canopy height are obtained. The measured rice yield data of each sub-region is obtained, and the remote sensing feature data of each sub-region is matched with the measured yield data of the same period. The extracted remote sensing features are used as input features, and the corresponding measured yield data are used as target variables. A training data set containing input features and target variables is created. This data set is a two-dimensional array or table, in which each row represents the data of a sub-region or field, and each column represents a feature or variable.

[0119] In a preferred embodiment of the present invention, the above step 19 includes:

[0120] Constructing a random forest regression model RF Yield =f(InputSet p ,Yield p ), InputSet p It is a feature set composed of normalized vegetation index and canopy height features at different time points. p is the actual yield data of the corresponding plot, RF Yield It is the final prediction model, where the random forest model parameters include the number of decision trees is 100 trees, the maximum depth is 10 layers, the number of features selected during splitting is to randomly select some features at each split, and the bootstrap method is to randomly sample from the training data set to generate multiple subsets, each subset is used to train a different decision tree.

[0121] According to the training data set, the time-series normalized vegetation index and canopy height characteristics of each plot were input into the model, and the yield was used as the output variable. The bootstrap method was used to generate multiple subsets, and the decision trees were trained separately. The prediction results of all trees were integrated by voting or averaging to generate the final prediction model.

[0122] The trained random forest model is used to input the time series feature data to obtain the yield prediction value.

[0123] In an embodiment of the present invention, random forests can generally obtain higher prediction accuracy than a single model by integrating the prediction results of multiple decision trees. Each tree in the random forest is trained based on a randomly sampled subset of data, which helps to reduce the risk of the model overfitting the training data. The random selection of features at each split also increases the diversity of the model, and random forests can effectively handle data sets containing a large number of features. By randomly selecting features at each split, the model can automatically perform feature selection to identify the most important features for prediction. Random forests can calculate the contribution of each feature to the prediction results, thereby helping to understand which features are most critical to yield prediction, helping to guide future data collection, and optimizing agricultural management strategies.

[0124] In a specific embodiment of the present invention, a random forest regression model is constructed, the parameters of the random forest regression model are determined, the random forest regression model is initialized with the selected parameters, and the bootstrap method is applied to randomly sample and generate multiple subsets from the training data set, each subset will be used to train an independent decision tree, for each generated subset, use it as training data, take the time series normalized vegetation index and canopy height characteristics as input, and the actual yield of the corresponding plot as output, train a decision tree, during the training process, each tree independently builds its own model, is not affected by other trees, and splits the nodes according to the set maximum depth and feature selection number, after all decision trees are trained, they are integrated to form a random forest, when there is new input data, each tree will independently predict. Prepare the time series normalized vegetation index and canopy height feature data of the plot whose yield needs to be predicted, input the prepared prediction data into the trained random forest regression model, and output the yield prediction value of the corresponding plot through the integrated decision tree prediction results.

[0125] like Figure 2 As shown, an embodiment of the present invention further provides a rice yield prediction system 20 based on multi-source time series remote sensing data, comprising:

[0126] The acquisition module 21 is used to acquire multi-source UAV remote sensing data by collecting multi-spectral data and LiDAR data and using satellite positioning to accurately locate the data; perform in-camera orientation, band registration, orthophoto stitching and mosaicking, radiation correction and remote sensing index calculation on the multi-spectral data to obtain processed multi-spectral data; perform point cloud reconstruction and point cloud filtering on the LiDAR data and calculate the height of rice plants to obtain processed LiDAR data; and register the processed LiDAR data with the multi-spectral data to achieve pre-processing of multi-source UAV remote sensing data;

[0127] The processing module 22 is used to obtain multi-source remote sensing data of time series based on the phenological period, optimize the time series data to obtain optimized time series data; extract the element features at the sub-region level according to the optimized time series data to construct a feature set of the time series, and the element features include vegetation index and terrain height; obtain the measured yield data of rice in the sub-region according to the constructed feature set of the time series; construct a training data set according to the pre-processed multi-source remote sensing data and the measured yield data of rice in the sub-region; train the random forest regression model according to the training data set, and use the trained model to predict the rice yield of new multi-source time series remote sensing data.

[0128] It should be noted that the system is a system corresponding to the above method, and all implementation methods in the above method embodiment are applicable to this embodiment and can achieve the same technical effect.

[0129] The embodiment of the present invention further provides a computing device, comprising: a processor, a memory storing a computer program, wherein when the computer program is executed by the processor, the method described above is executed. All implementations in the above method embodiment are applicable to this embodiment and can achieve the same technical effect.

[0130] The embodiment of the present invention also provides a computer-readable storage medium storing instructions, which, when executed on a computer, enable the computer to execute the method described above. All implementations in the above method embodiment are applicable to this embodiment and can achieve the same technical effect.

[0131] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A rice yield prediction method based on multi-source time series remote sensing data, characterized in that: The method comprises: Acquire multi-source UAV remote sensing data by collecting multispectral data and LiDAR data and using satellite positioning for accurate geolocation of the data; Perform camera orientation, band registration, orthophoto stitching and mosaicking, radiation correction and remote sensing index calculation on multispectral data to obtain processed multispectral data; Perform point cloud reconstruction and point cloud filtering on the LiDAR data and calculate the height of rice plants to obtain processed LiDAR data; The processed lidar data is registered with the multispectral data to achieve preprocessing of multi-source UAV remote sensing data; Based on the phenological period, multi-source remote sensing data of time series are obtained, and the time series data are optimized to obtain optimized time series data; According to the optimized time series data, extract the feature features at the sub-region level to construct a feature feature set of the time series; According to the feature set of the constructed time series, the measured rice yield data in the sub-region is obtained; A training data set is constructed based on the preprocessed multi-source remote sensing data and the measured rice yield data in the sub-region; The random forest regression model is trained according to the training data set, and the trained model is used to predict rice yield for new multi-source time series remote sensing data.

2. The rice yield prediction method based on multi-source time series remote sensing data according to claim 1, characterized in that: Perform camera orientation, band registration, orthophoto stitching and mosaicking, radiation correction and remote sensing index calculation on multispectral data to obtain processed multispectral data, including: The image is geometrically corrected using the camera calibration parameters to obtain a corrected image and determine the internal orientation parameters; According to the camera's internal orientation parameters, an accurate registration algorithm is used to register images of different bands to obtain multi-band images after band registration. Perform feature matching on the overlapping areas of the multi-band images after band registration and apply an automated stitching algorithm to generate an orthophoto map with unified geographic coordinates; Performing radiation correction on the orthophoto to obtain a radiation-corrected orthophoto; According to the orthophoto map after radiation correction, Calculate the Normalized Difference Vegetation Index to obtain the NDVI image map; NIR is the near-infrared band reflectance, RED is the red band reflectance, and NDVI is the Normalized Difference Vegetation Index.

3. The rice yield prediction method based on multi-source time series remote sensing data according to claim 2 is characterized in that: Perform point cloud reconstruction and point cloud filtering on the LiDAR data and calculate the height of rice plants to obtain processed LiDAR data, including: By flying scans, the original point cloud data of the rice field area is collected to obtain the original point cloud data; Splice and integrate the point cloud data and eliminate redundant points in overlapping areas to obtain reconstructed point cloud data; According to the reconstructed point cloud data, height threshold filtering and density threshold filtering are performed to obtain filtered point cloud data; According to the filtered point cloud data, the actual height image of the vegetation is calculated through CHM=dSM-DEM to obtain the processed lidar data, where DSM represents the combined height of the terrain and the object, DEM represents the pure terrain height, and CHM represents the actual height of the vegetation.

4. The rice yield prediction method based on multi-source time series remote sensing data according to claim 3 is characterized in that: The processed lidar data is registered with the multispectral data to achieve preprocessing of multi-source UAV remote sensing data, including: Reprojecting the lidar data and the multispectral data to the same coordinate system, and analyzing the reprojections of the lidar data and the multispectral data to obtain reprojection analysis results; According to the reprojection analysis results, the lidar data is resampled to a resolution consistent with the multispectral data to obtain resampled data; the resampled data is vector-clipped to obtain image data covering the same geographic range; The image data is aligned with the actual geographic reference data, and the position and angle of the image are adjusted for correction so that the residual error of the corrected image is controlled within ±2cm to achieve the preprocessing of multi-source UAV remote sensing data.

5. The rice yield prediction method based on multi-source time series remote sensing data according to claim 4, characterized in that: Based on the phenological period, multi-source remote sensing data of time series are obtained, and the time series data are optimized to obtain optimized time series data, including: According to the phenological period of rice growth, the analysis time nodes are obtained, including tillering period, jointing period and heading period; According to the analysis time nodes, the corresponding UAV flight plan is formulated to collect remote sensing data to build a complete time series data set; According to the time series data set, NDVI and CHM images are fused to obtain a time series data set; According to the time series data set, set the upper and lower limits of NDVI and To achieve adaptive threshold filtering of NDVI, where NDVI is the normalized vegetation index of multispectral images, NDVI(x,y,t) is the NDVI value at position (x,y) in phenological period t, T NDVI,min (t) and T NDVI,max (t) are the minimum and maximum values ​​of NDVI in phenological period t, respectively. f (x, y, t) is the NDVI value after adaptive threshold filtering; According to the time series data set, set the upper and lower limits of CHM and To achieve adaptive threshold filtering of canopy height, where CHM is the canopy height of the lidar data, CHM(x,y,t) is the CHM value of the position (x,y) at phenological period t, and is the adaptive threshold range of CHM in phenological period t, CHM f (x, y, t) is the CHM value after adaptive threshold filtering; The time series data are optimized through adaptive threshold filtering of NDVI and adaptive threshold filtering of CHM to obtain optimized time series data.

6. The rice yield prediction method based on multi-source time series remote sensing data according to claim 5, characterized in that: Based on the optimized time series data, the feature features at the sub-region level are extracted to construct a feature feature set of the time series. The feature features include vegetation index, terrain height, and include: According to the optimized time series data, the research area is divided into multiple sub-areas, each of which contains a number of pixels. All the pixels in each sub-area are marked and classified to achieve sub-area division; According to the sub-regional division, Calculate the average value of the normalized difference vegetation index after adaptive threshold filtering, through Calculate the average of the canopy height after adaptive threshold filtering, where and They represent the normalized vegetation index and the mean canopy height of the sub-region under phenological period t, n is the number of valid pixels in the sub-region, and p represents the sub-region; The feature set of time series was constructed based on the normalized difference vegetation index and the mean canopy height of the phenological period sub-regions.

7. The rice yield prediction method based on multi-source time series remote sensing data according to claim 6, characterized in that: According to the feature set of the constructed time series, the measured rice yield data in the sub-region is obtained, including: The time series feature sets of all sub-regions are combined to form a multi-time series feature set for the entire study area; According to the multi-time series feature set, the feature features in each sub-region are matched with the yield data of the sub-region through the corresponding geographical location and timestamp to obtain the measured yield data of rice in the sub-region.

8. A rice yield prediction system based on multi-source time series remote sensing data, characterized in that: include: The acquisition module is used to acquire multi-source UAV remote sensing data by collecting multispectral data and lidar data and using satellite positioning to accurately geolocate the data; Perform camera orientation, band registration, orthophoto stitching and mosaicking, radiation correction and remote sensing index calculation on multispectral data to obtain processed multispectral data; perform point cloud reconstruction and point cloud filtering on lidar data, and calculate rice plant height to obtain processed lidar data; register the processed lidar data with multispectral data to achieve preprocessing of multi-source UAV remote sensing data; A processing module is used to obtain multi-source remote sensing data of time series based on phenological period, optimize the time series data to obtain optimized time series data; extract sub-region level feature characteristics according to the optimized time series data to construct a feature feature set of the time series, the feature characteristics include vegetation index, terrain height; obtain the measured yield data of rice in the sub-region according to the constructed feature feature set of the time series; construct a training data set according to the pre-processed multi-source remote sensing data and the measured yield data of rice in the sub-region; The random forest regression model is trained according to the training data set, and the trained model is used to predict rice yield for new multi-source time series remote sensing data.

9. A computing device, characterized in that include: one or more processors; A storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a program, which, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images