Rice yield prediction method combining rice spectral information and three-dimensional structure information

Abstract

The invention relates to a rice yield prediction method combining rice spectral information and three-dimensional structure information. The rice yield prediction method comprises the steps of obtaining the rice spectral information and the three-dimensional structure information; identifying feature points in the remote sensing digital image or the multispectral remote sensing image by using a motion structure algorithm, estimating the positions of the feature points in a real scene of the image by detecting the motion of the feature points in a plurality of image sequences in combination with a forward intersection method, and rendering the feature points into a sparse three-dimensional point cloud; encrypting the sparse three-dimensional point cloud through a multi-view stereoscopic vision algorithm to reconstruct a dense three-dimensional point cloud reflecting most of scene details of the set; basic features of the rice point cloud are extracted to serve as parameters for operation of rice three-dimensional structure information; and constructing a rice yield prediction model, evaluating the prediction model, and predicting the rice yield through the evaluated prediction model. According to the method, the information amount of independent variables in the model is increased, and the precision and stability of the prediction model are improved.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing technology, and in particular to a rice yield prediction method combining rice spectral information and three-dimensional structure information. Background Art

[0002] At present, simple empirical prediction models such as the univariate linear regression model or the univariate nonlinear regression model constructed from the multispectral vegetation index extracted from multispectral remote sensing images or the visible light index in remote sensing digital images are widely used for rice yield prediction. Specifically, the signal value of each band in the collected multispectral remote sensing images is converted into the reflectance of the band through the empirical linear correction method based on the reflectivity of the radiation calibration plate. The radiation calibration plate is generally a diffuse reflector with a relatively constant reflectivity for different bands, while digital images are generally not calibrated for radiation, but normalized to map the data to a specific range to eliminate the dimensional differences between different features; after radiation correction, the multispectral image and the digital image are spliced ​​using the motion structure (SFM) algorithm and the multi-view stereo vision (MVS) algorithm; in general, this step can be completed by existing software, and then the images of each band in the spliced ​​multispectral image and the digital image are subjected to different combination operations to obtain different multispectral vegetation indices and visible light vegetation indices, and finally the obtained vegetation indices are used as independent variables, and the rice yield data is used as the dependent variable to construct a prediction model for rice yield; however, for the use of spectral index or visible light index alone to predict rice yield, since there are many factors affecting rice yield, the use of a single type of indicator for prediction has a strong limiting effect on the prediction results of rice, and these technologies also only use the optical information of rice, and the amount of available information is small. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a rice yield prediction method combining rice spectral information and three-dimensional structure information, thereby solving the shortcomings of the prior art.

[0004] The object of the present invention is achieved by the following technical solution: a rice yield prediction method combining rice spectral information and three-dimensional structure information, the prediction method comprising:

[0005] Step 1: Obtaining spectral information and three-dimensional structural information of rice;

[0006] Step 2: Use the motion structure algorithm to identify feature points in remote sensing digital images or multispectral remote sensing images, and estimate the position of the feature points in the real scene of the image by detecting the motion of the feature points in multiple image sequences combined with the forward intersection method, and render them into a sparse three-dimensional point cloud;

[0007] Then, the sparse 3D point cloud is encrypted through a multi-view stereo vision algorithm to reconstruct a dense 3D point cloud that reflects most of the details of the collection scene;

[0008] Extract the basic features of rice point cloud as parameters for calculating the three-dimensional structure information of rice;

[0009] Step 3: Based on the obtained rice spectral information and three-dimensional structure information, a rice yield prediction model is constructed in combination with the partial least squares method;

[0010] Step 4: Evaluate the prediction model and predict the rice yield using the evaluated prediction model.

[0011] The method of using a structure-from-motion algorithm to identify feature points in a remote sensing digital image or a multispectral remote sensing image, and estimating the position of the feature points in the real scene of the image by detecting the motion of the feature points in multiple image sequences combined with a forward intersection method, and rendering the feature points into a sparse three-dimensional point cloud includes:

[0012] Use the motion structure algorithm to detect key points such as corners, edge points or other significant texture points in the image;

[0013] The key points detected initially are accurately located. For each key point, a descriptor is extracted to describe the neighborhood around the key point for subsequent feature matching.

[0014] Filter the key points, remove edge points or low-contrast points, and retain the most stable key points as feature points;

[0015] The feature points are represented in the form of two-dimensional coordinates and corresponding descriptors, and are used to match the feature points in a series of images to form a sparse three-dimensional point cloud that reflects the three-dimensional structure of the scene and restore the camera's posture.

[0016] The method of encrypting the sparse 3D point cloud by a multi-view stereo vision algorithm to reconstruct a dense 3D point cloud reflecting most of the details of the scene includes:

[0017] The sparse 3D point cloud is encrypted by a multi-view stereo vision algorithm to reconstruct a dense 3D point cloud that reflects most of the geometric scene details;

[0018] Dense three-dimensional point clouds are used to generate grids and texture maps to obtain spliced ​​orthophotos and digital surface models. Then, the orthophotos of each band are calculated to obtain the spectral vegetation index and RGB vegetation index that reflect the growth status of rice. The digital surface models of different periods are subtracted to obtain the rice plant height, which is then compared with the plant height obtained through the point cloud.

[0019] The parameters for extracting basic features of rice point cloud as the three-dimensional structure information of rice for calculation include:

[0020] Compared with the actual situation on site, the abnormal points are filtered and removed. When the point clouds of all periods have the same neighborhood radius, the point clouds are analyzed by principal components and the plane and surface fitting is performed on the point clouds using the least squares method. The points in the point clouds are used as coordinate points in the three-dimensional coordinates in mathematical operations to calculate the eigenvalues ​​and eigenvectors of the point clouds.

[0021] Eight basic features of point clouds are extracted, including the average height of point cloud, the first eigenvalue of point cloud principal component analysis, the second eigenvalue of point cloud principal component analysis, the third eigenvalue of point cloud principal component analysis, the sum of eigenvalues ​​of point cloud principal component analysis, the average curvature of points on the neighborhood best fitting surface, the distance between points in the neighborhood and the best fitting plane, and the number of point clouds in the neighborhood divided by the neighborhood volume;

[0022] Several basic features of the eight point clouds are selected and introduced into the rice yield prediction model as parameters for calculating the three-dimensional structure information of rice.

[0023] The step three comprises:

[0024] Let p indicators be independent variables x1, x2, x3, …, x p , rice yield is the dependent variable y, thus forming the data tables X and Y of the independent and dependent variables;

[0025] The partial least squares method extracts components t1 and u1 in X and Y respectively, where t1 is x1, x2, x3, …, x p The linear combination of y1,y2,y3,…,y q A linear combination of , and regress X on t1 and Y on u1;

[0026] If the regression equation reaches the preset accuracy, the process is terminated. Otherwise, the residual information of X explained by t1 and the residual information of Y explained by u1 are used to extract the second round of components, and this process is repeated until the preset accuracy is reached.

[0027] If a total of m components t1, t1, t1, …, t m , the partial least squares method will be implemented by implementing y for t1,t1,t1,…,t m Then, x1,x2,x3,…,x p Substitute into t1,t1,t1,…,t m In the end, the rice yield y is expressed as the relationship between the various indicators x1, x2, x3, …, x p The regression equation of .

[0028] The spectral information includes a multispectral vegetation index and a visible light vegetation index;

[0029] The multispectral vegetation index includes normalized difference vegetation index, green normalized difference vegetation index, normalized difference red edge index and triangle vegetation index;

[0030] The visible light vegetation index includes a visible atmospheric resistance index, a green-red vegetation index, and a red-green-blue vegetation index.

[0031] The present invention has the following advantages: a rice yield prediction method combining rice spectral information and three-dimensional structural information, extracting the three-dimensional structural information of rice from remote sensing images, and combining multispectral vegetation index and visible light index to construct a rice yield prediction model, thereby increasing the information amount of independent variables in the model and improving the accuracy and stability of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic diagram of the process of the present invention;

[0033] Figure 2 This is a comparison chart of the correlation coefficient between the plant height calculated based on 3D point cloud computing, the plant height calculated based on DSM and the measured plant height;

[0034] Figure 3 Schematic diagram of the adjusted determination coefficient (Adjust R2) and root mean square error of prediction (RMSEP) of the PLSR rice yield prediction model constructed using the multispectral vegetation index alone, the visible light vegetation index alone, the point cloud basic features alone, the combined multispectral vegetation index and the visible light vegetation index, and the combined multispectral vegetation index, the visible light vegetation index and the point cloud basic features. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application provided below in conjunction with the drawings is not intended to limit the scope of protection of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application. The present invention is further described below in conjunction with the drawings.

[0036] like Figure 1As shown, the present invention specifically relates to a rice yield prediction method combining rice spectral information and three-dimensional structure information, which specifically includes the following contents:

[0037] First, the collected multispectral images are calibrated by radiation. Since the image data collected by general drones are continuous images collected by UAVs during long-term flight, it is difficult to ensure that the light intensity will not change during the flight of the UAV. Therefore, the light intensity correction method and the empirical linear correction method are combined to perform radiation correction on the multispectral images. Sunlight (incident light) shines on the ground and is reflected. At this time, the multispectral image obtains the reflection signal of the ground to generate an image. At the same time, the multispectral light intensity sensor obtains the signal value of the incident light. Therefore, the light intensity correction method can be used to obtain:

[0038]

[0039] X r Represents the reflectivity of the ground object in the X band, X refleced and X incident are respectively the reflected light and incident light in the X-band, DN camera is the image signal value obtained from the multispectral image in the X-band, DN LS is the light-sensing signal value obtained from the X-band multispectral light intensity sensor, ρ X It is a parameter that adjusts the conversion between the image signal and the multispectral light intensity sensor. Since the multispectral camera and the light intensity sensor have good linearity, their relationship can be linearly fitted with one parameter for each band. X We cannot get the specific value of . Using this method can only eliminate the influence of light intensity changes and cannot get the true reflectivity of the ground objects. Then, we use the empirical linear correction method on this basis:

[0040]

[0041] In the formula, R X represents the true surface reflectivity of the X band, G X Represents the gain factor of the X-band, B X Represents the bias coefficient of the X band.

[0042] Secondly, the digital image is normalized. In order to eliminate the influence of different illumination during each shooting, the signal values ​​of the R band, G band and B band need to be normalized according to the following formula.

[0043]

[0044] Where r, g, and b are the normalized values, and R, G, and B are the signal values ​​of the red, green, and blue channels, respectively.

[0045] Then the images are stitched: the motion structure (SFM) algorithm is used to identify the feature points in the UAV digital image or the UAV multispectral image, and the motion of these points in multiple image sequences is detected to estimate the position of these feature points in the image, and then the sparse 3D point cloud is encrypted by the multi-view stereo vision (MVS) algorithm to reconstruct a dense 3D point cloud that reflects most of the geometric scene details. Then the encrypted 3D point cloud is used to generate a grid and texture map to obtain the stitched orthophoto (DOM) and digital surface model (DSM); then the orthophotos of each band are calculated to obtain the spectral vegetation index and RGB vegetation index reflecting the growth status of rice; the digital surface models of different periods are subtracted to obtain the rice plant height and compare it with the plant height obtained by the point cloud. Usually this step can be implemented by existing software. Before the image is reconstructed, in order to provide geographic reference information for the UAV image, the coordinates of the image control points need to be imported as markers, the markers are manually aligned with the positions of the image control points in the image, and automatically projected onto the remaining images. The plane accuracy and elevation accuracy of the control points are less than a specific value. Therefore, based on this step, three-dimensional point clouds, digital orthophotos and multispectral orthophotos with precise geo-references can be obtained.

[0046] Then the index is extracted: the multispectral vegetation index and the visible light vegetation index are obtained by combining the values ​​of different bands according to a certain formula. The three-dimensional dense point cloud can retain the geometric scene detail information of the object to the greatest extent. Therefore, the height of the rice plant is calculated by subtracting the dense point cloud of the bare soil period from the dense point cloud of the rice growth period. Figure 2 As shown in the figure, the plant height (Hpoint) of rice in each plot based on three-dimensional point cloud computing and the plant height of rice in each plot calculated based on digital surface model were correlated with the measured plant height. Hpoint has a stronger correlation with the actual plant height and can better reflect the actual situation of rice plant height. Then, when the point clouds of all periods have the same neighborhood radius, the point clouds are subjected to principal component analysis and least squares plane fitting and surface fitting, and eight basic features of point clouds including plant height are extracted. The basic features of point clouds include the average height of point clouds, the first eigenvalue of principal component analysis of point clouds, the second eigenvalue of principal component analysis of point clouds, the third eigenvalue of principal component analysis of point clouds, the sum of eigenvalues ​​of principal component analysis of point clouds, the average curvature of points on the best fitting surface in the neighborhood, the distance between points in the neighborhood and the best fitting plane, and the number of point clouds in the neighborhood divided by the neighborhood volume. The average height of the point cloud is calculated by subtracting the average height of the point cloud in the bare soil period from the average height of the point cloud in the rice growth cycle. Since these indicators are not unique, we can select some of them to be introduced into the rice yield prediction model, and we need to ensure that the number of indicators of each type is not less than 1.

[0047] Finally, the prediction model is constructed: Compared with traditional linear and nonlinear regression, partial least squares method (PLSR) has many advantages, especially when the sample size is small, the types of independent variables are many, and there is strong collinearity between the independent variables. PLSR can realize regression modeling at the same time during the modeling process, such as multivariate linear regression; principal component analysis is used for data dimensionality reduction; data correlation analysis, etc., which can accurately extract components with better explanatory power for the dependent variable and construct a robust and accurate estimation model.

[0048] Let p indicators be independent variables x1, x2, x3, …, x p , rice yield is the dependent variable y. In order to study the statistical relationship between each indicator and yield, 36 sample points are observed, thus forming the data tables X and Y of the independent variable and the dependent variable. PLSR first extracts components t1 and u1 in X and Y respectively, that is, t1 is x1, x2, x3, …, x p The linear combination of y1,y2,y3,…,y q A linear combination of .

[0049] t1=a1x1+a2x2+a3x3+…+a p x p ,(a1,a2,a3,…a p is the combination coefficient)

[0050] u1=by, (b is the combination coefficient)

[0051] When extracting these two components, for the needs of regression analysis, there are two requirements:

[0052] (1) t1 and u1 should carry as much variation information as possible in their respective data tables.

[0053] (2) The correlation between t1 and u1 can reach the maximum.

[0054] After the first components t1 and u1 are extracted, PLSR performs regression of X on t1 and Y on u1 respectively. Then the accuracy of the model is tested. If the regression equation has reached a satisfactory accuracy, the algorithm terminates; otherwise, the residual information of X explained by t1 and the residual information of Y explained by u1 will be used for the second round of component extraction. This process is repeated until a satisfactory accuracy can be achieved.

[0055] If a total of m components t1, t1, t1, …, t m , PLSR will be implemented by implementing y for t1,t1,t1,…,t m The return of Substitute into t1,t1,t1,…,t mThe final expression is the rice yield y with respect to each index The regression equation of .

[0056] y=c1x1+c2x2+c3x3+…c p x p ,(c1,c2,c3,…,c p is the combination coefficient)

[0057] The performance of the model was evaluated by using the adjusted coefficient of determination (Adjust R 2 ) and the root mean square error of prediction (RMSEP). In PLSR, if too many factors are selected, overfitting will easily occur. Therefore, an effective principle is needed to determine the number of factors. The "leave one out" method (LOO) in the cross-validation method is selected to calculate the root mean square error of prediction for each factor, and the number of factors with the smallest root mean square error or the root mean square error of prediction is almost unchanged as the number of factors selected for the final model. The "leave one out" method is to use k-1 data as training sets and the remaining one as test sets in the k data sets after standardization. This method is repeated k times until all samples are verified once, and finally the sum of the squares of the prediction errors of each sample is added to get the RMSEP.

[0058] The experimental site of the present invention is located in a modern agricultural industrial park in a city in Sichuan Province. The area is located in the western part of the Sichuan Plain in the middle and upper reaches of the Minjiang River in Sichuan Province. It has a subtropical humid monsoon climate with distinct four seasons, short spring and autumn, and long winter and summer, which is suitable for the growth of crops. The rice variety in the experimental area is Quanyou 1606. The same variety of rice is sown by mechanical planting. The fertilization method used is UAV aerial fertilization. About 20-25 kg of urea is applied per mu of land before sowing, and about 15-20 kg of urea is applied again at the early stage of division.

[0059] The experimental area was divided into 36 plots, each with a length of 8 meters, a width of 4 meters, and an area of ​​32 square meters. A total of 10 ground control points were evenly distributed on the concrete roads throughout the experimental area to provide geographic reference for UAV images collected at different growth stages and periods.

[0060] The data was collected using the DJI Phantom 4 Multispectral Edition aerial survey drone, which has an integrated multispectral imaging system that integrates 1 2.08 million pixel visible light camera and 5 2.08 million pixel multispectral cameras. The wavelengths of the 5 multispectral bands are blue (center wavelength 450nm, band width 32nm), green (center wavelength 560nm, band width 32nm), red (center wavelength 650nm, band width 32nm), red edge (center wavelength 730nm, band width 32nm) and near infrared (center wavelength 840nm, band width 52nm).

[0061] The rice was transplanted to the experimental field on May 28, 2023. On May 26, a flight mission was carried out to obtain the bare ground model of the experimental field to prepare for the calculation of rice plant height in the later stage. During the growth process, the collection of UAV flight data began on June 23, 2023, and was collected once every 5-10 days on average. Therefore, a total of 13 UAV image data of rice growth stages were collected during the experiment (Table 1). During the UAV image data collection process, the weather was clear and the wind speed was low. The time interval was from 10 am to 2 pm local time. The flight plan was constructed using the DJI GS PRO automatic flight control application. The flight altitude of the drone was set to 25 meters, the ground sampling distance (GCD) was 1.3 cm per pixel, the heading overlap rate was 80%, the side overlap rate was 70%, and the route angle was determined by the sun altitude angle when performing the aerial photography mission. The camera exposure time was automatically adjusted by the light intensity sensor on the top of the drone. Each flight mission takes about 30 minutes. Before taking off, the UAV is kept at an altitude of 0.8 to 1 meter to take a photo of three radiation calibration plates. The calibration plates are three 15-centimeter square diffuse reflection plates with reflectivities of 25%, 50%, and 75%, respectively, which are used for radiation correction of UAV multispectral images in the later stage.

[0062] Table 1. UAV image acquisition time and rice growth period

[0063] time Reproductive period time Reproductive period 2023 / 06 / 23 Mid-tillering period 2023 / 08 / 05 Booting stage 2023 / 06 / 28 Late tillering stage 2023 / 08 / 15 Heading period 2023 / 07 / 04 Late tillering stage 2023 / 08 / 21 Initial grouting 2023 / 07 / 09 Early stage of jointing 2023 / 08 / 26 Middle stage of grouting 2023 / 07 / 14 Early stage of jointing 2023 / 08 / 31 Late grouting 2023 / 07 / 20 Mid-jointing period 2023 / 09 / 06 Late grouting 2023 / 07 / 26 Late jointing stage

[0064] The rice yield data was collected on September 26, 2023, the day before the rice harvest. The rice ears of 20 rice plants were collected in each area of ​​interest, transferred to the laboratory and dried until the rice grains had no weight change. Finally, the average value of the rice ears of the 20 rice plants was taken as the representative of the single rice yield in each area of ​​interest. Since the rice in the experimental field is sown mechanically, the rice planting density in each area of ​​interest is not the same, which is also a common problem in modern agriculture. Therefore, on June 23, we counted the number of rice in each area of ​​interest on site. Finally, the average weight of the rice ears of the 20 collected rice plants was multiplied by the statistical number to obtain the measured total rice yield in each area of ​​interest.

[0065] In this experiment, four commonly used multispectral vegetation indices (Table 2), three commonly used visible light vegetation indices (Table 3) and eight point cloud basic features (Table 4) were selected to jointly construct a PLSR model to predict the total yield of rice. The average value of each index in each area of ​​interest was taken as the representative of the area of ​​interest, so there were a total of 36 samples in the experimental area.

[0066] Table 2. Multispectral vegetation index table

[0067] <![CDATA[Multispectral Vegetation Index (VI MS )]]> formula Normalized Difference Vegetation Index (NDVI) (NIR-R) / (NIR+R) Green Normalized Difference Vegetation Index (GNDVI) (NIR-G) / (NIR+G) Normalized Difference Red Edge Index (NDRE) (NIR-RE) / (NIR+RE) Triangular Vegetation Index (TVI) 60×(NIR-G)-100×(RG)

[0068] Table 3. Visible light vegetation index

[0069] <![CDATA[Visible Vegetation Index (VI RGB )]]> formula Visible Atmospheric Resistance Index (VARI) (gr) / (g+rb) Green-Red Vegetation Index (GRVI) (gr) / (g+r) Red-Green-Blue Vegetation Index (RGBVI) <![CDATA[(g 2 -r×b) / (g 2 +r×b)]]>

[0070] Table 4. Point cloud basic features table

[0071]

[0072]

[0073] Comparative Example 1: A PLSR model was constructed using the multispectral vegetation index alone for prediction, and the other steps were the same as those in the embodiment.

[0074] Comparative Example 2: A PLSR model was constructed using the visible light vegetation index alone for prediction, and other steps were the same as those in the embodiment.

[0075] Comparative Example 3: A PLSR model was constructed using only the basic features of the point cloud for prediction, and the other steps were the same as those in the embodiment.

[0076] Comparative Example 4: A PLSR model was constructed by combining the multispectral vegetation index and the visible light vegetation index for prediction, and the other steps were the same as those in the embodiment.

[0077] Comparative Example 5: A PLSR model was constructed by combining the multispectral vegetation index, the visible light vegetation index and the point cloud features for prediction, and the other steps were the same as those in the embodiment.

[0078] from Figure 3 It can be seen that the PLSR model with spectral index as the independent variable has an Adjusted R 2 All of them were greater than 0.6, with the maximum value occurring on July 20, the middle period of rice jointing (Adjusted R 2 =0.82, RMSEP = 987.62 g / m2). Compared with the models based on the other two single-type indicators, this model has higher estimation accuracy and smaller RMSEP in most time periods. The PLSR model with RGB index as the independent variable has a higher accuracy than the model based on the jointing of rice on July 25. 2 All of them were greater than 0.6, with the maximum value occurring on July 9, when rice just entered the jointing stage (Adjusted R 2 =0.8, RMSEP = 980.41 g / m2), after July 25 Adjusted R 2 Almost all of them dropped below 0.6, except for the early stage of rice filling on August 21.2 The PLSR model with point cloud features as independent variables was used to calculate the value of Adjusted R before the rice jointing on July 25, except for the middle tillering period on June 23. 2 All of them were greater than 0.6, with the maximum value occurring on July 9, when rice just entered the jointing stage (Adjusted R 2 =0.9, RMSEP=966.06g / m2). After the rice jointing on July 25, the estimation accuracy of the model decreased. However, after the rice entered the filling stage on August 21, the estimation accuracy of the model was higher and the RMSEP was smaller than that of the models constructed by the other two types of indicators.

[0079] The PLSR rice yield prediction model was constructed by combining the multispectral index and the RGB index. Except for August 31, when rice just entered the late filling period, the estimation accuracy was higher in other growth periods of rice. 2 All of them were greater than 0.65, with the maximum value occurring on July 20, the late jointing stage of rice (Adjusted R 2 =0.84, RMSEP = 982.89 g / m2), but the model did not show much advantage over the models based on spectral index and RGB index before the rice booting stage on August 5, and even the estimation accuracy in some growth stages was lower than that of the PLSR model with spectral index as the independent variable. After the rice entered the booting stage, the model's estimation accuracy was significantly improved compared with the models based on spectral index and RGB index respectively. The PLSR model constructed by combining the three types of parameters had a higher estimation accuracy throughout the rice growth period, and the Adjusted R 2 All of them were greater than 0.7, with the maximum value occurring on July 9, when rice just entered the jointing stage (Adjusted R 2 =0.91, RMSEP = 963.92g / m2). Compared with the other four models, this model has the highest prediction accuracy in most of the growth period of rice, and the RMSEP is smaller. Only on August 15, the estimation accuracy of the heading period of rice is slightly lower than that of the model constructed by combining the two optical indicators. According to the above comparison results, it can be seen that:

[0080] Among the models constructed based on a single type of index, the PLSR model constructed based on spectral index has a higher estimation accuracy for the total rice yield before the rice filling period, and the PLSR model constructed based on point cloud features has a higher estimation accuracy for the total rice yield after the rice enters the filling period. The PLSR rice total yield prediction model constructed by combining three types of parameters, optical index and point cloud features, has the best prediction accuracy during most of the rice growth period. 2The best time for yield estimation by this model is July 9, when rice just ends the tillering stage and enters the jointing stage (Adjusted R 2 =0.91, RMSEP = 963.92g / m2). During this period, rice completed tillering and entered the vertical growth stage. The growth trend of rice vegetation coverage slowed down. The acquired image data not only contained rice planting density information, but also contained rich rice growth information. Compared with the model built based on optical indicators, the prediction accuracy of the combined model built with three types of indicators has been improved to a certain extent throughout the rice growth cycle, but the improvement effect is small before the rice heading stage, and the improvement effect is significant after the rice enters the heading stage.

[0081] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used for various other combinations, modifications and improvements, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art do not deviate from the spirit and scope of the present invention, and should be within the scope of protection of the claims attached to the present invention.

Claims

1. A rice yield prediction method combining rice spectral information and three-dimensional structure information, characterized in that: The prediction method comprises: Step 1: Obtaining spectral information and three-dimensional structural information of rice; Step 2: Use the motion structure algorithm to identify feature points in remote sensing digital images or multispectral remote sensing images, and estimate the position of the feature points in the real scene of the image by detecting the motion of the feature points in multiple image sequences combined with the forward intersection method, and render them into a sparse three-dimensional point cloud; Then, the sparse 3D point cloud is encrypted through a multi-view stereo vision algorithm to reconstruct a dense 3D point cloud that reflects most of the details of the collection scene; Extract the basic features of rice point cloud as parameters for calculating the three-dimensional structure information of rice; Step 3: Based on the obtained rice spectral information and three-dimensional structure information, a rice yield prediction model is constructed in combination with the partial least squares method; Step 4: Evaluate the prediction model and predict the rice yield using the evaluated prediction model.

2. A rice yield prediction method combining rice spectral information and three-dimensional structure information according to claim 1, characterized in that: The method of using a structure-from-motion algorithm to identify feature points in a remote sensing digital image or a multispectral remote sensing image, and estimating the position of the feature points in the real scene of the image by detecting the motion of the feature points in multiple image sequences combined with a forward intersection method, and rendering the feature points into a sparse three-dimensional point cloud includes: Use the motion structure algorithm to detect key points such as corners, edge points or other significant texture points in the image; The key points detected initially are accurately located. For each key point, a descriptor is extracted to describe the neighborhood around the key point for subsequent feature matching. Filter the key points, remove edge points or low-contrast points, and retain the most stable key points as feature points; The feature points are represented in the form of two-dimensional coordinates and corresponding descriptors, and are used to match the feature points in a series of images to form a sparse three-dimensional point cloud that reflects the three-dimensional structure of the scene and restore the camera's posture.

3. The rice yield prediction method combining rice spectral information and three-dimensional structure information according to claim 2, characterized in that: The method of encrypting the sparse 3D point cloud by a multi-view stereo vision algorithm to reconstruct a dense 3D point cloud reflecting most of the details of the scene includes: The sparse 3D point cloud is encrypted by a multi-view stereo vision algorithm to reconstruct a dense 3D point cloud that reflects most of the geometric scene details; Dense three-dimensional point clouds are used to generate grids and texture maps to obtain spliced ​​orthophotos and digital surface models. Then, the orthophotos of each band are calculated to obtain the spectral vegetation index and RGB vegetation index that reflect the growth status of rice. The digital surface models of different periods are subtracted to obtain the rice plant height, which is then compared with the plant height obtained through the point cloud.

4. The rice yield prediction method combining rice spectral information and three-dimensional structure information according to claim 3, characterized in that: The parameters for extracting basic features of rice point cloud as the three-dimensional structure information of rice for calculation include: Compared with the actual situation on site, the abnormal points are filtered and removed. When the point clouds of all periods have the same neighborhood radius, the point clouds are analyzed by principal components and the plane and surface fitting is performed on the point clouds using the least squares method. The points in the point clouds are used as coordinate points in the three-dimensional coordinates in mathematical operations to calculate the eigenvalues ​​and eigenvectors of the point clouds. Eight basic features of point clouds are extracted, including the average height of point cloud, the first eigenvalue of point cloud principal component analysis, the second eigenvalue of point cloud principal component analysis, the third eigenvalue of point cloud principal component analysis, the sum of eigenvalues ​​of point cloud principal component analysis, the average curvature of points on the neighborhood best fitting surface, the distance between points in the neighborhood and the best fitting plane, and the number of point clouds in the neighborhood divided by the neighborhood volume; Several basic features of the eight point clouds are selected and introduced into the partial rice yield prediction model as parameters for calculating the three-dimensional structure information of rice.

5. The rice yield prediction method combining rice spectral information and three-dimensional structure information according to claim 1, characterized in that: The step three comprises: Let p indicators be independent variables x1, x2, x3, …, x p , rice yield is the dependent variable y, thus forming the data tables X and Y of the independent and dependent variables; The partial least squares method extracts components t1 and u1 in X and Y respectively, where t1 is x1, x2, x3, …, x p The linear combination of y1,y2,y3,…,y q A linear combination of , and regress X on t1 and Y on u1; If the regression equation reaches the preset accuracy, the process is terminated. Otherwise, the residual information of X explained by t1 and the residual information of Y explained by u1 are used to extract the second round of components, and this process is repeated until the preset accuracy is reached. If a total of m components t1, t1, t1, …, t m , the partial least squares method will be implemented by implementing y for t1,t1,t1,…,t m Then, x1,x2,x3,…,x p Substitute into t1,t1,t1,…,t m In the end, the rice yield y is expressed as the relationship between the various indicators x1, x2, x3, …, x p The regression equation of .

6. The rice yield prediction method combining rice spectral information and three-dimensional structure information according to claim 1, characterized in that: The spectral information includes a multispectral vegetation index and a visible light vegetation index; The multispectral vegetation index includes normalized difference vegetation index, green normalized difference vegetation index, normalized difference red edge index and triangle vegetation index; The visible light vegetation index includes a visible atmospheric resistance index, a green-red vegetation index, and a red-green-blue vegetation index.

Images