Wheat yield prediction method based on spatial-temporal characteristics of unmanned aerial vehicle image

Abstract

Compared with the prior art, the wheat yield prediction method based on the spatial-temporal characteristics of the images of the unmanned aerial vehicle has the advantages that the unmanned aerial vehicle is used for shooting high-precision image data in the key growth period of wheat, a standardized sampling scheme is combined, and comprehensive and systematic image data collection is achieved. A prediction model is established through various models such as linear regression, random forest and partial least squares, the performance of the model is comprehensively evaluated by adopting indexes such as a decision coefficient, a root-mean-square error and an average relative error, and finally, an optimal model is selected to predict the wheat yield. According to the method, the data processing flow is simplified, the prediction precision and reliability are remarkably improved, and the technical use threshold is reduced. Through accurate yield prediction, scientific production decision support is provided for farmers, and agricultural resource allocation can be optimized.

Description

Technical Field

[0001] The present invention relates to the field of agricultural yield prediction, and in particular to a method for predicting wheat yield based on the spatiotemporal characteristics of unmanned aerial vehicle images. Background Art

[0002] Winter wheat yield is affected by a combination of factors, including genetic characteristics, water management, fertilization strategies, light conditions, pest and disease control, and meteorological disasters. Winter wheat varieties, soil conditions, and management measures vary from field to field, resulting in different sensitivities to the external environment, which in turn leads to significant differences in yield. Therefore, conducting winter wheat yield prediction research at the farmland scale can provide an important decision-making basis for precise field management, and is also of great significance for winter wheat breeding and the assessment of losses from agricultural meteorological disasters. At present, wheat yield prediction faces several key challenges. First, the acquisition and quality of data are inconsistent, which affects the accuracy of the model. Second, feature extraction methods are often affected by subjective factors and environmental changes, resulting in insufficient stability of the model under different conditions. In addition, existing models have limited capabilities in dealing with complex nonlinear relationships and are difficult to fully reflect various influencing factors. Finally, the lack of effective model fusion and optimization strategies reduces the reliability of prediction results. These issues still need to be studied and resolved in depth.

[0003] This experimental team has been browsing and researching a large amount of relevant records and materials on wheat planting related technologies for a long time. At the same time, it relies on relevant resources and conducts a large number of relevant experiments. After a lot of searches, it found existing prior arts such as CN202411149363.1, CN202410947348.5, and CN202410789777.4 disclosed in the prior art, such as a multi-model fusion method based on wheat yield estimation disclosed in the prior art. The method includes: obtaining the current environmental feature sequence, current biomass sequence, historical environmental feature sequence, and historical biomass sequence of each plot of land; obtaining the time-varying importance factor corresponding to each current environmental feature sequence; optimizing the current environmental feature sequence using the time-varying importance factor to obtain an optimized environmental feature sequence; obtaining the environmental sensitivity factor corresponding to each segment of data in the optimized environmental feature sequence; optimizing the optimized environmental feature sequence using the environmental sensitivity factor to obtain a first optimized environmental feature sequence; combining various first optimized environmental feature sequences and current biomass sequences of each plot of land, and classifying each plot of land according to the combination result; obtaining the yield prediction model of each type of land to predict wheat yield. However, this method has too many input parameters, requires complex data, and is difficult to use. The influencing factors considered are too complex, the required data is huge, and over-reliance on machine learning ignores the connection between related variables.

[0004] The present invention is made to solve the common problems in this field, such as high threshold for using existing model methods, inability of the prediction process to match actual crop growth, low prediction accuracy, etc. Summary of the invention

[0005] The purpose of the present invention is to propose a wheat yield prediction method based on the spatiotemporal characteristics of drone images in response to the current deficiencies in the field.

[0006] In order to overcome the shortcomings of the prior art, the present invention adopts the following technical solutions:

[0007] A method for predicting wheat yield based on the spatiotemporal characteristics of drone images, the method comprising the following steps:

[0008] S101: Setting up a ground test experiment, sampling and weighing wheat during the maturity period, and obtaining wheat yield data;

[0009] S102: acquiring an image dataset including images of each key growth period of wheat by using a drone, wherein the key growth period includes an early grain filling period, a middle grain filling period, and a mature period of wheat,

[0010] S103: Calculate the relevant parameters of all images in the image data set respectively, and the relevant parameters include the red channel mean R, the green channel mean G, the blue channel mean B, the gray mean Graym, the hue mean H, the saturation mean S, the brightness mean V, the correlation Corr, the angular second moment ASM, the homogeneity Hm, the contrast Cont, the dissimilarity Dss, the energy Ener, the one-dimensional entropy Ent1D, and the two-dimensional entropy Ent2D. The calculation of the relevant parameters of the image can be realized by the Python image processing library and the Python program developed based on PyCharm. The Python program integrates a variety of image processing and texture feature extraction algorithms. The user directly imports the image data and automatically outputs the relevant parameters.

[0011] S104: Further calculate and obtain the derived parameters of each image: NRI, NGI, NBI, NDVI, GRRI, GBRI, BRRI, NPCI, NGRDI, NGBDI, PPE, RGBVI, MGRVI,

[0012] S105: Screen all parameters in the derived parameters and related parameters in each key growth period to determine the image feature parameters and image texture parameters of the vegetation index, wherein the image feature parameters include GRR, NGBDI and RGBVI, and the image texture parameters include H,

[0013] S106: Using the selected key growth period, image feature parameters, image texture feature parameters and measured yield data of the wheat sample plot as model construction parameters, a preliminary model is established using three methods: linear regression, random forest and partial least squares method.

[0014] S107: Use measured wheat yield data to verify the wheat yield prediction model, using the determination coefficient Rr 2 , root mean square error RMSE, mean relative error MER to evaluate the performance of the prediction model, and determine the coefficient Rr 2 The most reasonable model is selected as the main basis, and the root mean square error RMSE and the mean relative error MER are used as auxiliary comprehensive selection.

[0015] S108: Determine the final model and make predictions.

[0016] Furthermore, the specific implementation steps of S102 are as follows:

[0017] S1021: When wheat is in the early filling, middle filling and mature stages, set five points A, B, C, D and E at the four corners and the center of the plot to be measured, and take photos and collect data at three different heights of 10m, 20m and 30m above the ground and at a -90° camera pitch angle.

[0018] S1022: After the shooting is completed, the images are stored and sorted. The image storage format is JPG, the horizontal resolution is 72DP, and the vertical resolution is 72DP.

[0019] S1023: Preprocess the acquired images to remove blurry images or images with poor lighting effects, so as to form an image dataset with temporal and spatial differences.

[0020] Among them, in order to ensure the accuracy of data information and avoid cloud obstruction and sunlight, the shooting time is scheduled from 9:00 to 15:00 when the weather is clear and the wind is light. Before the flight, the shooting parameters and route planning are set, and the automatic exposure mode is selected to reduce the inconsistency of parameters between different flight missions. In order to efficiently process and analyze the wheat image data collected by drone remote sensing, the same image size should be selected when shooting. At the same time, roads and buildings should be avoided.

[0021] Further, S104 is implemented by:

[0022]

[0023]

[0024]

[0025] The initial grouting stage, the middle grouting stage, and the mature stage are represented by serial numbers 1, 2, and 3 respectively. i represents the key growth period, i∈{1, 2, 3}, i=1 represents the initial grouting stage, i=2 represents the middle grouting stage, and i=3 represents the mature stage. k represents the shooting position of the plot to be measured, K∈{A, B, C, D, E}, R i k represents the mean value of the image R at the k shooting positions of the plot to be tested in the key growth period i, G i k represents the mean value of the image G at the k-shot position of the plot to be tested in the key growth period i, B i k represents the mean value of the image B at the k-shot position of the plot to be tested in the key growth period i, and

[0026]

[0027] Furthermore, the specific steps of constructing the preliminary model in S106 are as follows:

[0028] S1061: Generate a feature matrix X containing image feature parameters and texture parameters at different key growth periods i, and use Y as the target variable of the model output;

[0029] Where X = {GRRI i ,NGBDI i ,RGBVI i ,H i ,i},i∈{1,2,3}, the initial grouting stage, the middle grouting stage and the mature stage are represented by the serial numbers 1, 2 and 3 respectively, i represents the key growth period, i=1 represents the initial grouting stage, i=2 represents the middle grouting stage, i=3 represents the mature stage, GRRI i Represents the GRRI obtained at the shooting position A, B, C, D or E at the key growth period i in the image dataset, NGBDI i Represents the NGBDI, RGBVI obtained at the shooting position A, B, C, D or E at the key growth period i in the image dataset i represents the RGBVI obtained at the shooting position A, B, C, D or E at the key growth period i in the image dataset, H i represents H obtained at the shooting position A, B, C, D or E during the key growth period i in the image data set, and Y represents the measured wheat yield data under different feature matrices.

[0030] S1062: Standardize the feature matrix X to eliminate the dimensional differences between different features and obtain X c , X c As model input,

[0031] X c={GRRI i c ,NGBDI i c ,RGBVI i c ,H i c ,i},

[0032]

[0033]

[0034] Among them, μGRRI, μNGBDI, μRGBVI and μH represent the mean of GRRI, the mean of NGBDI, the mean of RGBVI and the mean of H in the entire image dataset respectively, σGRRI, σNGBDI, σRGBVI and σH represent the standard deviation of GRRI, the standard deviation of NGBDI, the standard deviation of RGBVI and the standard deviation of H in the entire image dataset respectively,

[0035] S1063: Use X c The model is trained as the input parameter of the model so that it can predict the target variable Y, and the first preliminary model is established using the linear regression method, the second preliminary model is established using the random forest method, and the third preliminary model is established using the partial least squares method.

[0036] Furthermore, the specific implementation steps of S107 are as follows:

[0037] S1071: Calculate the coefficient of determination Rr of each preliminary module 2 , root mean square error RMSE, and mean relative error MER,

[0038]

[0039]

[0040]

[0041] m∈{1,2,3}, the test set includes N data samples that are not involved in model training, each data sample includes X extracted from the actual wheat image c and the actual wheat yield corresponding to the actual wheat image, Y j Represented as the actual wheat yield of the jth data sample, Y1 j Y2 represents the wheat yield prediction result of the jth data sample output by the first preliminary model, j Y3 represents the wheat yield prediction result of the j-th data sample output by the second preliminary model. jrepresents the wheat yield prediction result of the j-th data sample output by the third preliminary model, Y avg represents the mean of the actual wheat yield of all samples in the test set, N represents the total number of all data samples in the test set, Rr1 2 ,Rr2 2 ,Rr3 2 denote the coefficient of determination of the first, second and third preliminary models, respectively; RMSE1, RMSE2 and RMSE3 denote the root mean square errors of the first, second and third preliminary models, respectively; MER1, MER2 and MER3 denote the mean relative errors of the first, second and third preliminary models, respectively;

[0042] S1072: Set a threshold th1, if the maximum determination coefficient Rr among the three preliminary models 2 first With the second largest Rr 2 sec The numerical difference between 2 first -Rr 2 sec When greater than th1, select Rr 2 first The corresponding preliminary module is taken as the optimal model, where the threshold th1 ranges from 0.01 to 0.05.

[0043] If Rr 2 first -Rr 2 sec When the value is below th1, further proceed to S1073.

[0044] S1073: Set a threshold th2 if the minimum root mean square error RMSE among the three preliminary models min With the second largest RMSE sec The RMSE of the numerical difference between min -RMSE sec When it is greater than th2, select RMSE min The corresponding preliminary module is used as the optimal model, where the threshold th2 ranges from 5% to 10%.

[0045] If RMSE min -RMSE sec When the value is below th2, further proceed to S1074.

[0046] S1074: Select the preliminary model with the smallest MER as the optimal model.

[0047] The beneficial effects achieved by the present invention are:

[0048] 1. Data acquisition standards: By formulating a standardized data collection process, the influence of subjective factors can be effectively reduced, thereby reducing model errors and improving prediction accuracy. In addition, standardized data sources also facilitate cross-regional comparative analysis.

[0049] 2. Few input parameters, simple and easy to understand process: The model structure is concise, making it easy to operate and interpret. This not only reduces the learning cost of users, but also makes it easier for farmers and decision makers to quickly understand and apply the prediction results, promoting the development of precision agriculture.

[0050] 3. Combining the actual growth conditions of crops with simulation: By combining the model with the actual growth conditions, the dynamic process of crop growth can be more realistically reflected. The results of experimental verification not only enhance the credibility of the model, but also provide data support for further optimization.

[0051] 4. Strong adaptability: The model can adapt to different varieties, soil types and climatic conditions, which increases its application range in diverse agricultural environments and makes it more practical. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The present invention can be further understood from the following description in conjunction with the accompanying drawings. The components in the figures are not necessarily drawn to scale, but the emphasis is placed on illustrating the principles of the embodiments. In different views, the same reference numerals designate corresponding parts.

[0053] Figure 1 The figure is a flow chart of the wheat yield prediction method based on the spatiotemporal characteristics of UAV images of the present invention.

[0054] Figure 2 It is a schematic diagram of the process of S107 of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention is further described in detail in combination with its embodiments below; it should be pointed out that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present case. For those skilled in the art, after reviewing the following detailed description, other systems, methods and / or features of this embodiment will become apparent. And the terms describing the positional relationship in the accompanying drawings are only used for illustrative purposes and cannot be understood as limitations on this patent. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to the specific circumstances.

[0056] Embodiment 1: Combining with Figure 1 and attached Figure 2 This embodiment constructs a wheat yield prediction method based on the spatiotemporal characteristics of drone images. The wheat yield prediction method based on the spatiotemporal characteristics of drone images includes the following steps:

[0057] S101: Setting up a ground test experiment, sampling and weighing wheat during the maturity period, and obtaining wheat yield data;

[0058] S102: acquiring an image dataset including images of each key growth period of wheat by using a drone, wherein the key growth period includes an early grain filling period, a middle grain filling period, and a mature period of wheat,

[0059] S103: Calculate the relevant parameters of all images in the image data set respectively, and the relevant parameters include the red channel mean R, the green channel mean G, the blue channel mean B, the gray mean Graym, the hue mean H, the saturation mean S, the brightness mean V, the correlation Corr, the angular second moment ASM, the homogeneity Hm, the contrast Cont, the dissimilarity Dss, the energy Ener, the one-dimensional entropy Ent1D, and the two-dimensional entropy Ent2D. The calculation of the relevant parameters of the image can be realized by the Python image processing library and the Python program developed based on PyCharm. The Python program integrates a variety of image processing and texture feature extraction algorithms. The user directly imports the image data and automatically outputs the relevant parameters.

[0060] S104: Further calculate and obtain the derived parameters of each image: NRI, NGI, NBI, NDVI, GRRI, GBRI, BRRI, NPCI, NGRDI, NGBDI, PPE, RGBVI, MGRVI,

[0061] S105: Screen all parameters in the derived parameters and related parameters in each key growth period to determine the image feature parameters and image texture parameters of the vegetation index, wherein the image feature parameters include GRR, NGBDI and RGBVI, and the image texture parameters include H,

[0062] S106: Using the selected key growth period, image feature parameters, image texture feature parameters and measured yield data of the wheat sample plot as model construction parameters, a preliminary model is established using three methods: linear regression, random forest and partial least squares method.

[0063] S107: Use measured wheat yield data to verify the wheat yield prediction model, using the determination coefficient Rr 2 , root mean square error RMSE, mean relative error MER to evaluate the performance of the prediction model, and determine the coefficient Rr 2 The most reasonable model is selected as the main basis, and the root mean square error RMSE and the mean relative error MER are used as auxiliary comprehensive selection.

[0064] S108: Determine the final model and make predictions.

[0065] The specific implementation steps of S102 are as follows:

[0066] S1021: When wheat is in the early filling, middle filling and mature stages, set five points A, B, C, D and E at the four corners and the center of the plot to be measured, and take photos and collect data at three different heights of 10m, 20m and 30m above the ground and at a -90° camera pitch angle.

[0067] S1022: After the shooting is completed, the images are stored and sorted. The image storage format is JPG, the horizontal resolution is 72DP, and the vertical resolution is 72DP.

[0068] S1023: Preprocess the acquired images to remove blurry images or images with poor lighting effects, so as to form an image data set with time and space differences.

[0069] Among them, in order to ensure the accuracy of data information and avoid cloud obstruction and sunlight, the shooting time is scheduled from 9:00 to 15:00 when the weather is clear and the wind is light. Before the flight, the shooting parameters and route planning are set, and the automatic exposure mode is selected to reduce the inconsistency of parameters between different flight missions. In order to efficiently process and analyze the wheat image data collected by drone remote sensing, the same image size should be selected when shooting. At the same time, roads and buildings should be avoided.

[0070] And in the S102 step, the DJI Phantom 4 Pro drone can be used as a data collection instrument. The DJI Phantom 4 Pro drone includes 6 visual sensors and a main camera, which ensures stable flight and precise obstacle avoidance in complex farmland environments. In addition, it is equipped with a GPS and GLONASS dual-mode satellite positioning system and IMU and compass dual redundant sensors, providing centimeter-level navigation and positioning accuracy, providing a solid foundation for low-altitude photogrammetry applications. The main camera uses a 1-inch CMOS sensor with an effective pixel of up to 20 million and a wide dynamic range of ISO100-3200. The lens parameter is FOV84 °, focal length is 8.8mm to 25mm, and aperture range is f2.8 to f11. These optical properties enable the UAV to obtain high-quality images, depth and positioning information in real time during flight. In addition, the UAV also has the ability to build a 3D map around the aircraft, which further enhances the accuracy of its spatial positioning, enabling it to be quickly deployed on farmland sites for efficient aerial survey operations. Its high-performance imaging system and precise navigation and positioning capabilities provide reliable data support for wheat growth monitoring and yield prediction. The specific model of the UAV in the present invention is selected by technicians in this field based on actual needs and is not limited here.

[0071] This scheme is based on the combination of multi-temporal images and multi-feature parameters of drones. Through the construction and evaluation of multiple models such as linear regression, random forest, and partial least squares, an efficient and accurate wheat yield prediction method is proposed. The scheme collects image data from the key growth period of wheat, extracts and screens multiple feature parameters such as color and texture, and ensures the stability and effectiveness of the features. Model selection is based on the determination coefficient Rr 2 The root mean square error (RMSE) and the mean relative error (MER) are used as the auxiliary comprehensive evaluation method to finally achieve accurate prediction of wheat yield, providing scientific support and practical application value for precision agricultural management.

[0072] Embodiment 2: Combining with Figure 1 and attached Figure 2 In addition to the contents of the above embodiments, S104 is implemented by:

[0073]

[0074]

[0075]

[0076] The initial grouting stage, the middle grouting stage, and the mature stage are represented by serial numbers 1, 2, and 3 respectively. i represents the key growth period, i∈{1, 2, 3}, i=1 represents the initial grouting stage, i=2 represents the middle grouting stage, and i=3 represents the mature stage. k represents the shooting position of the plot to be measured, K∈{A, B, C, D, E}, R i k represents the mean value of the image R at the k shooting positions of the plot to be tested in the key growth period i, G i k represents the mean value of the image G at the k-shot position of the plot to be tested in the key growth period i, B i k represents the mean value of the image B at the k-shot position of the plot to be tested in the key growth period i, and

[0077]

[0078] Specifically, the specific steps of screening the diffraction parameters and related parameters in each key growth period in step S105 and determining the image feature parameters and image texture parameters of the vegetation index are as follows:

[0079] Step 1: Extract all preliminary relevant parameters and derived parameters from each key growth stage;

[0080] Step 2: For the images taken by the drone at different flight altitudes, calculate the numerical variation of each parameter in the relevant parameters and derived parameters, and select the parameters with the smallest fluctuation in altitude to ensure that the change in flight altitude will not significantly affect the stability of the data;

[0081] Step 3: Compare the differences between the shooting positions of A, B, C, D and E for each parameter in the relevant parameters and derived parameters, and select the parameters that do not change significantly between the five positions to ensure that the selected parameters can represent the overall characteristics of the plot and avoid deviations caused by position differences;

[0082] Step 4: Through statistical analysis of sample data of different wheat varieties, the variance analysis method is used to calculate the differences between the relevant parameters and the parameters within the derived parameters among different varieties, and the parameters with significant distinguishing effects among varieties are screened out;

[0083] Step 5: Use the Pearson correlation coefficient to evaluate the correlation between the parameters, remove the parameters with strong correlation, and ensure that there is no multicollinearity between the screened parameters;

[0084] Step 6: According to the screening conditions of steps 1 to 5 above, the image feature parameters are finally determined to be GRR, NGBD and RGBV, and the texture feature parameter is H;

[0085] Step 7: Experimentally verify the screened parameters to ensure their stability and effectiveness. The final screening results are used to build the model and serve as input data for the yield prediction model.

[0086] The screening conditions in these steps 1 to 5 ensure that the selected image feature parameters and texture feature parameters are stable and discriminative, providing a reliable data basis for model construction. The specific selection of image feature parameters and image texture parameters is determined by technicians in this field according to actual needs and is not limited here.

[0087] The specific steps of constructing the preliminary model in S106 are as follows:

[0088] S1061: Generate a feature matrix X containing image feature parameters and texture parameters at different key growth periods i, and use Y as the target variable of the model output;

[0089] Where X = {GRRI i ,NGBDI i ,RGBVI i ,H i ,i},i∈{1,2,3}, the initial grouting stage, the middle grouting stage and the mature stage are represented by the serial numbers 1, 2 and 3 respectively, i represents the key growth period, i=1 represents the initial grouting stage, i=2 represents the middle grouting stage, i=3 represents the mature stage, GRRIi Represents the GRRI obtained at the shooting position A, B, C, D or E at the key growth period i in the image dataset, NGBDI i Represents the NGBDI, RGBVI obtained at the shooting position A, B, C, D or E at the key growth period i in the image dataset i represents the RGBVI obtained at the shooting position A, B, C, D or E at the key growth period i in the image dataset, H i H represents the data obtained at the shooting position A, B, C, D or E during the key growth period i in the image data set, and Y represents the measured wheat yield data under different feature matrices;

[0090] S1062: Standardize the feature matrix X to eliminate the dimensional differences between different features and obtain X c , X c As model input,

[0091] X c ={GRRI i c ,NGBDI i c ,RGBVI i c ,H i c ,i},

[0092]

[0093]

[0094] Among them, μGRRI, μNGBDI, μRGBVI and μH represent the mean of GRRI, the mean of NGBDI, the mean of RGBVI and the mean of H in the entire image dataset respectively; σGRRI, σNGBDI, σRGBVI and σH represent the standard deviation of GRRI, the standard deviation of NGBDI, the standard deviation of RGBVI and the standard deviation of H in the entire image dataset respectively;

[0095] S1063: Use X c The model is trained as the input parameter of the model so that it can predict the target variable Y, and the first preliminary model is established using the linear regression method, the second preliminary model is established using the random forest method, and the third preliminary model is established using the partial least squares method.

[0096] The specific implementation steps of S107 are as follows:

[0097] S1071: Calculate the coefficient of determination Rr of each preliminary module 2, root mean square error RMSE, and mean relative error MER,

[0098]

[0099]

[0100]

[0101] m∈{1,2,3}, the test set includes N data samples that are not involved in model training, each data sample includes X extracted from the actual wheat image c and the actual wheat yield corresponding to the actual wheat image, Y j Represented as the actual wheat yield of the jth data sample, Y1 j Y2 represents the wheat yield prediction result of the jth data sample output by the first preliminary model, j Y3 represents the wheat yield prediction result of the j-th data sample output by the second preliminary model. j represents the wheat yield prediction result of the j-th data sample output by the third preliminary model, Y avg represents the mean of the actual wheat yield of all samples in the test set, N represents the total number of all data samples in the test set, Rr1 2 ,Rr2 2 ,Rr3 2 denote the coefficient of determination of the first, second and third preliminary models, respectively; RMSE1, RMSE2, RMSE3 denote the root mean square errors of the first, second and third preliminary models, respectively; MER1, MER2, MER3 denote the mean relative errors of the first, second and third preliminary models, respectively;

[0102] S1072: Set a threshold th1, if the maximum determination coefficient Rr among the three preliminary models 2 first With the second largest Rr 2 sec The numerical difference between 2 first -Rr 2 sec When greater than th1, select Rr 2 first The corresponding preliminary module is taken as the optimal model, where the threshold th1 ranges from 0.01 to 0.05.

[0103] If Rr 2 first -Rr 2 sec When it is less than th1, further proceed to S1073;

[0104] S1073: Set a threshold th2 if the minimum root mean square error RMSE among the three preliminary models min With the second largest RMSE sec The RMSE of the numerical difference between min -RMSE sec When it is greater than th2, select RMSE min The corresponding preliminary module is used as the optimal model, where the threshold th2 ranges from 5% to 10%.

[0105] If RMSE min -RMSE sec When it is less than th2, further proceed to S1074;

[0106] S1074: Select the preliminary model with the smallest MER as the optimal model.

[0107] Higher Rr 2 A small RMSE value indicates that the model has a strong ability to explain the data and has a good fitting effect. A small RMSE indicates that the absolute error between the model prediction value and the actual value is small and the prediction accuracy is high. A low MERM means that the relative error of the model is small, indicating that the accuracy of the model is relatively consistent between different samples. 2 , RMSE and MERM help select the optimal model to ensure the reliability and accuracy of the prediction results.

[0108] And in S108, the optimal model selected in S107 is used as the final model, and the wheat yield is predicted by the final model.

[0109] The present invention makes full use of UAV remote sensing technology and image processing technology, combined with multi-stage screening, comprehensive model construction and optimization methods, to achieve high-precision prediction of wheat yield. The flexible parameter selection mechanism and efficient data processing flow of the present invention not only improve the efficiency of data analysis, but also ensure the accuracy and robustness of the model, providing a scientific basis for refined management and yield estimation in agricultural production.

[0110] Although the present invention has been described above with reference to various embodiments, it should be understood that many changes and modifications may be made without departing from the scope of the present invention. That is, the methods, systems and devices discussed above are examples. Various configurations may appropriately omit, replace or add various processes or components. For example, in alternative configurations, the method may be performed in an order different from that described, and / or various components may be added, omitted and / or combined. Moreover, the features described with respect to certain configurations may be combined in various other configurations, such as different aspects and elements of the configuration may be combined in a similar manner. In addition, as technology develops, the elements therein may be updated, i.e., many elements are examples and do not limit the scope of the present disclosure or claims. And it should be understood that after reading the contents of the record of the present invention, the technician may make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A wheat yield prediction method based on the spatiotemporal characteristics of drone images, characterized in that: The wheat yield prediction method based on the spatiotemporal characteristics of drone images includes the following steps: S101: Setting up a ground test experiment, sampling and weighing wheat during the maturity period to obtain wheat yield data; S102: acquiring an image dataset including images of each key growth period of wheat by using a drone, wherein the key growth period includes an early grain filling period, a middle grain filling period, and a mature period of wheat, S103: Calculate the relevant parameters of all images in the image data set respectively, and the relevant parameters include the red channel mean R, the green channel mean G, the blue channel mean B, the gray mean Graym, the hue mean H, the saturation mean S, the brightness mean V, the correlation Corr, the angular second moment ASM, the homogeneity Hm, the contrast Cont, the dissimilarity Dss, the energy Ener, the one-dimensional entropy Ent1D, and the two-dimensional entropy Ent2D. The calculation of the relevant parameters of the image can be realized by the Python image processing library and the Python program developed based on PyCharm. The Python program integrates a variety of image processing and texture feature extraction algorithms. The user directly imports the image data and automatically outputs the relevant parameters. S104: Further calculate and obtain the derived parameters of each image: NRI, NGI, NBI, NDVI, GRRI, GBRI, BRRI, NPCI, NGRDI, NGBDI, PPE, RGBVI, MGRVI, S105: Screen all parameters in the derived parameters and related parameters in each key growth period to determine the image feature parameters and image texture parameters of the vegetation index, wherein the image feature parameters include GRR, NGBDI and RGBVI, and the image texture parameters include H, S106: Using the selected key growth period, image feature parameters, image texture feature parameters and measured yield data of the wheat sample plot as model construction parameters, a preliminary model is established using three methods: linear regression, random forest and partial least squares method. S107: Use measured wheat yield data to verify the wheat yield prediction model, using the determination coefficient Rr 2 , root mean square error RMSE, mean relative error MER to evaluate the performance of the prediction model, and determine the coefficient Rr 2 The most reasonable model is selected as the main basis, and the root mean square error RMSE and the mean relative error MER are used as auxiliary comprehensive selection. S108: Determine the final model and make predictions.

2. The wheat yield prediction method according to claim 1, characterized in that: The specific implementation steps of S102 are as follows: S1021: When wheat is in the early filling, middle filling and mature stages, set five points A, B, C, D and E at the four corners and the center of the plot to be measured, and take photos and collect data at three different heights of 10m, 20m and 30m above the ground and at a -90° camera pitch angle. S1022: After the shooting is completed, the images are stored and sorted. The image storage format is JPG, the horizontal resolution is 72DP, and the vertical resolution is 72DP. S1023: Preprocess the acquired images to remove blurry images or images with poor lighting effects, so as to form an image data set with temporal and spatial differences. Among them, in order to ensure the accuracy of data information and avoid cloud obstruction and sunlight influence, the shooting time is arranged from 9:00 to 15:00 when the weather is clear and the wind is light. Before the flight, the shooting parameters and route planning are set, and the automatic exposure mode is selected to reduce the inconsistency of parameters between different flight missions. In order to efficiently process and analyze the wheat image data collected by UAV remote sensing, the same image size should be selected when shooting. At the same time, roads and buildings should be avoided.

3. The wheat yield prediction method according to claim 2, characterized in that: S104 is achieved by: The initial grouting stage, the middle grouting stage, and the mature stage are represented by serial numbers 1, 2, and 3 respectively. i represents the key growth period, i∈{1, 2, 3}, i=1 represents the initial grouting stage, i=2 represents the middle grouting stage, and i=3 represents the mature stage. k represents the shooting position of the plot to be measured, K∈{A, B, C, D, E}, R i k represents the mean value of the image R at the k shooting positions of the plot to be tested in the key growth period i, G i k represents the mean value of the image G at the k-shot position of the plot to be tested in the key growth period i, B i k represents the mean value of the image B at the k-shot position of the plot to be tested in the key growth period i, and 4. The wheat yield prediction method according to claim 3, characterized in that: The specific steps of constructing the preliminary model in S106 are as follows: S1061: Generate a feature matrix X containing image feature parameters and texture parameters at different key growth periods i, and use Y as the target variable of the model output; Where X = {GRRI i ,NGBDI i ,RGBVI i ,H i ,i},i∈{1,2,3}, the initial grouting stage, the middle grouting stage and the mature stage are represented by the serial numbers 1, 2 and 3 respectively, i represents the key growth period, i=1 represents the initial grouting stage, i=2 represents the middle grouting stage, i=3 represents the mature stage, GRRI i Represents the GRRI obtained at the shooting position A, B, C, D or E at the key growth period i in the image dataset, NGBDI i Represents the NGBDI, RGBVI obtained at the shooting position A, B, C, D or E at the key growth period i in the image dataset i represents the RGBVI obtained at the shooting position A, B, C, D or E at the key growth period i in the image dataset, H i represents H obtained at the shooting position A, B, C, D or E during the key growth period i in the image data set, and Y represents the measured wheat yield data under different feature matrices. S1062: Standardize the feature matrix X to eliminate the dimensional differences between different features and obtain X c , X c As model input, X c ={GRRI i c ,NGBDI i c ,RGBVI i c ,H i c ,i}, Among them, μGRRI, μNGBDI, μRGBVI and μH represent the mean of GRRI, the mean of NGBDI, the mean of RGBVI and the mean of H in the entire image dataset respectively, σGRRI, σNGBDI, σRGBVI and σH represent the standard deviation of GRRI, the standard deviation of NGBDI, the standard deviation of RGBVI and the standard deviation of H in the entire image dataset respectively, S1063: Use X c The model is trained as the input parameter of the model so that it can predict the target variable Y, and the first preliminary model is established using the linear regression method, the second preliminary model is established using the random forest method, and the third preliminary model is established using the partial least squares method.

5. The wheat yield prediction method according to claim 4, characterized in that: The specific implementation steps of S107 are as follows: S1071: Calculate the coefficient of determination Rr of each preliminary module 2 , root mean square error RMSE, and mean relative error MER, m∈{1,2,3}, the test set includes N data samples that are not involved in model training, each data sample includes X extracted from the actual wheat image c and the actual wheat yield corresponding to the actual wheat image, Y j Represented as the actual wheat yield of the jth data sample, Y1 j Y2 represents the wheat yield prediction result of the jth data sample output by the first preliminary model, j Y3 represents the wheat yield prediction result of the j-th data sample output by the second preliminary model, j represents the wheat yield prediction result of the j-th data sample output by the third preliminary model, Y avg represents the mean of the actual wheat yield of all samples in the test set, N represents the total number of all data samples in the test set, Rr12, Rr22, Rr3 2 denote the coefficient of determination of the first, second and third preliminary models, respectively; RMSE1, RMSE2 and RMSE3 denote the root mean square errors of the first, second and third preliminary models, respectively; MER1, MER2 and MER3 denote the mean relative errors of the first, second and third preliminary models, respectively; S1072: Set a threshold th1, if the maximum determination coefficient Rr among the three preliminary models 2 first With the second largest Rr 2 sec The numerical difference between |Rr 2 first -Rr 2 sec |When greater than th1, select Rr 2 first The corresponding preliminary module is taken as the optimal model, where the threshold th1 ranges from 0.01 to 0.

05. If |Rr 2 first -Rr 2 sec |When below th1, further proceed to S1073, S1073: Set a threshold th2 if the minimum root mean square error RMSE among the three preliminary models min With the second largest RMSE sec The numerical difference between |RMSE min -RMSE sec |When greater than th2, select RMSE min The corresponding preliminary module is used as the optimal model, where the threshold th2 ranges from 5% to 10%. If |RMSE min -RMSE sec |When below th2, further proceed to S1074, S1074: Select the preliminary model with the smallest MER as the optimal model.

Images