Crop cultivation mode recommendation and yield prediction method based on unmanned aerial vehicle image and machine learning

Abstract

The invention discloses a crop cultivation mode recommendation and yield prediction method based on unmanned aerial vehicle images and machine learning, and the method comprises the steps: carrying out the collection and key cell segmentation of an unmanned aerial vehicle multi-source image (RGB + multispectral) in a key growth period of a crop, and then extracting the common character indexes of visible light and multispectral images by taking a cell as a unit; and a data basis is provided for constructing a crop growth parameter estimation model fusing visible light and multispectral characteristics. Based on the prediction result of the model, two models are constructed, including a crop comprehensive growth monitoring model constructed by adopting multi-method weight optimization and a crop yield prediction model constructed in different growth periods. According to the method, through strategies of feature screening, staged modeling, weight optimization and the like, the growth parameter estimation precision and the yield prediction reliability are effectively improved, lossless and accurate estimation of crop growth parameters, scientific evaluation of comprehensive growth vigor and reliable yield prediction are realized, and technical support is provided for crop cultivation mode optimization and management decision in arid and semi-arid regions.

Description

Technical Field

[0001] This invention relates to the fields of precision agriculture and remote sensing technology, specifically to a method for crop cultivation pattern recommendation and yield prediction based on UAV imagery and machine learning. Background Technology

[0002] Traditional crop growth monitoring relies on manual measurement, which requires a large investment of manpower and has problems such as destructive sampling, poor data timeliness, and limited coverage, making it difficult to meet the needs of large-scale, high-frequency precision management. In existing remote sensing technologies, although satellite remote sensing can cover large areas, it is limited by cloud cover and low temporal resolution (such as long revisit period), making it unable to capture dynamic growth changes during key growth periods of crops; while fixed sensor platforms in the field have the disadvantages of high deployment cost and poor flexibility, making it difficult to adapt to experimental fields with different planting patterns.

[0003] With its advantages of high resolution, low flight cost, and flexible operation, UAV remote sensing technology has become an important means of crop phenotypic data collection. However, current UAV-based crop research still has shortcomings: First, growth parameter (chlorophyll content, leaf area index, plant height) estimation models mostly rely on single spectral features and do not fully integrate complementary information from visible light and multispectral data, leading to decreased prediction accuracy under complex environments (such as drought stress); Second, comprehensive growth evaluation models do not quantify the weight ratio of different growth parameters, making it difficult to objectively reflect the suitability of varieties and cultivation models; Third, yield prediction models do not consider the dynamic characteristics of crop growth stages, mixing data from the entire growth period for training, which fails to accurately capture the correlation between each stage and yield, and limits generalization ability. Therefore, there is an urgent need for a method that integrates multi-source data from UAVs, optimizes feature selection, and builds models to achieve non-destructive and accurate estimation of crop growth parameters, scientific evaluation of overall growth status, and reliable prediction of yield, providing technical solutions for optimizing crop cultivation patterns in arid and semi-arid regions. Summary of the Invention To address the aforementioned issues, this invention proposes a method for monitoring maize growth and predicting yield based on UAV imagery and machine learning. By employing strategies such as multi-source data fusion, phased modeling, and weight optimization, the method improves the accuracy of growth parameter estimation and the reliability of yield prediction.

[0004] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows: This invention provides a method for crop cultivation pattern recommendation and yield prediction based on UAV imagery and machine learning, including: S1: Acquire RGB and multispectral images of crops during critical growth periods using drones and perform key cell segmentation processing; S2: Based on the common trait indices extracted from visible light and multispectral images at the plot level, a crop growth parameter estimation model integrating visible light and multispectral features is constructed. S3: Construct a comprehensive crop growth monitoring model. Input the plant height, pH and chlorophyll content SPAD and leaf area index LAI obtained by the crop growth parameter estimation model into the comprehensive crop growth monitoring model to obtain the comprehensive growth monitoring index CGI of each plot. Select the cultivation mode corresponding to the plot with the highest CGI as the recommended cultivation mode for the target area. S4: Construct crop yield prediction models for different growth stages. Use plant height, pH, and chlorophyll content (SPAD) and leaf area index (LAI) estimated by crop growth parameter estimation models as input data, and use the prediction model corresponding to the growth stage of the input data to predict yield.

[0005] Furthermore, the crops include corn, rice, and wheat; The key growth periods for maize include the tasseling and silking stage, the grain-filling stage, and the maturity stage; the key growth periods for rice include the heading and booting stage, the grain-filling and grain-setting stage, and the full maturity stage; and the key growth periods for wheat include the heading and booting stage, the grain-filling stage, and the waxy maturity stage.

[0006] Furthermore, step S1 specifically includes the following steps: S1.1: During the critical growth period of crops, drones equipped with visible light cameras and multispectral cameras are used to take aerial photos of the experimental fields to obtain visible light RGB images and multispectral images covering the crop experimental plots; S1.2: After the UAV data is collected, it is stitched and corrected, and then aerial triangulation is performed and ground control point calibration is carried out to generate two-dimensional orthophotos and three-dimensional point cloud data. S1.3: Use Airmeasurer to segment cells from the stitched visible light data. Use Airmeasurer to generate a digital surface model (DSM) and a digital elevation model (DEM) from the generated 2D orthophoto and 3D point cloud data. Calculate the canopy height model CHM = DSM - DEM. Use Hough transform to detect cell boundaries and combine it with a mask for automated cell segmentation. Obtain the visible light cell image corresponding to each cell and output a polygon shape file (ROI region of interest) describing the boundaries of each cell. S1.4: Using the polygon shape file obtained during the processing of visible light images, the corresponding multispectral image region is found through perspective transformation, rotation, and registration operations. Then, the same mask is applied to achieve cell segmentation of the multispectral image (specifically, the polygon and the mask are used to perform cell cropping and segmentation on the multispectral orthophoto). The cell segmentation results of the multispectral image in five bands of red, green, blue, near-infrared, and red edge are obtained, that is, each cell has a multispectral cell image in at least five bands of red, green, blue, near-infrared, and red edge.

[0007] Furthermore, in step S1.2, if the number of feature points extracted from a single image is ≥1000, the proportion of calibrated images is ≥95%, and the difference in camera parameter optimization is <5%, the image is deemed qualified.

[0008] Furthermore, in step S1.1, the image acquisition uses a DJI M300 drone equipped with a Zenmuse P1 camera (RGB) and a multispectral camera. The flight path adopts a dual-grid mode to ensure that the image overlap is ≥80%. The visible light image acquisition height is 20m, and the multispectral image acquisition height is 25m.

[0009] Furthermore, step S2 specifically includes the following steps: S2.1: Based on the visible light and multispectral images of each plot, extract image features to characterize crop phenotypes; the image features include visible light features and multispectral features; The visible light features are extracted from RGB orthophotos, including canopy coverage (CC), average plant height (PH), visible light atmospheric impedance index (VARI), normalized differential yellowness index (NDYI), excess green (ExG), excess red (ExR), normalized vegetation index (NVI), and green leaf index (GLI); the average plant height (PH) is obtained directly from CHM. The multispectral features extract soil conditioning index OSAVI, red-edge chlorophyll index CIRE, ratio vegetation index RVI, green normalized vegetation index GNDVI, leaf chlorophyll index LCI, normalized red-edge vegetation index NDRE, and crop nitrogen reflectance index NRI from multispectral images in blue, green, red, red-edge, and near-infrared bands. The image features extracted from each cell are associated with the corresponding artificially measured growth parameter samples of the cell. The artificially measured growth parameters include at least chlorophyll content SPAD, leaf area index LAI, and plant height PH. S2.2: The Pearson correlation coefficient is used to analyze the correlation between each image feature and SPAD and LAI, and a preset number of feature subsets are selected according to the correlation strength as the input features of the model. Using the selected image features as independent variables and chlorophyll content SPAD, leaf area index LAI, and plant height PH as dependent variables, decision tree regression model, random forest regression model, and support vector machine regression model were constructed respectively to obtain estimation models from image features to crop growth parameters.

[0010] Furthermore, step S3 specifically includes the following steps: S3.1: Chlorophyll content SPAD, leaf area index LAI, and plant height PH were selected as growth evaluation indicators. The weights were calculated and the average values ​​were taken using the coefficient of variation method, entropy weight method, and fuzzy comprehensive evaluation method. The calculation formula is as follows: Coefficient of variation method: ; Entropy weight method: ; Fuzzy comprehensive evaluation method: ; The final weights are used to calculate the Comprehensive Growth Monitoring Index (CGI) based on the selected data. The specific calculation formula is as follows:

[0011] S3.2: Calculate the Comprehensive Growth Index (CGI) for each plot according to the above formula; where each plot corresponds to a different combination of crop variety, planting pattern and planting density. Sort the CGI of each plot and select the variety, planting pattern and planting density combination corresponding to the plot with the highest CGI as the recommended cultivation pattern for the target area.

[0012] Furthermore, step S4 specifically includes the following steps: S4.1: Based on the multi-temporal image features of each plot and the actual yield measurement data of the plot, the SelectKBest feature selection method based on f_regression is used to screen several image features with the strongest correlation with yield at different growth stages, and obtain the stage feature subsets corresponding to different growth stages (such as tasseling and silking stage, grain filling stage, and maturity stage). S4.2: Using the feature subsets corresponding to each growth stage as input and the actual yield of the corresponding plot as output, construct random forest regression models according to the growth stage to obtain crop yield prediction models for each growth stage. S4.3: Acquire UAV images of the target area during the key growth period, extract image features and estimate crop growth parameters through steps S1 and S2, and use a random forest regression model corresponding to the current growth period to predict yield and obtain the predicted yield value of the target area.

[0013] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art: This invention provides a method for recommending and predicting cereal crops based on UAV imagery and machine learning. This method includes the lossless, high-throughput acquisition of multi-source data (RGB + multispectral) during key crop growth stages using UAVs, improving efficiency by more than 50 times compared to manual measurement and avoiding data bias caused by destructive sampling. It integrates visible light and multispectral features with machine learning algorithms to meet the needs of precise monitoring. A comprehensive growth model is constructed through multi-method weight optimization, clearly defining LAI as the core indicator. The recommended optimal cultivation model has been verified as feasible in the field. Modeling by growth stage avoids interference from full-cycle data and can provide technical support for crop yield prediction and management decisions in arid, semi-arid, and other arid regions, making it suitable for further promotion and application. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0015] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the corn experimental field design; Figure 3 This is a diagram showing the results of the automatic segmentation of the residential area; Figure 4 and Figure 5 This is a screenshot of the UI of a multispectral image processing software. Figure 6 This is a comparison chart of plant height at three different stages; Figure 7 This is a heatmap of the comprehensive growth performance (CGI) of maize under different cultivation models; Figure 8 It is a comprehensive score chart of various varieties and planting patterns under different densities. Detailed Implementation

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] See attached document Figure 1 As shown, this invention provides a method for crop cultivation pattern recommendation and yield prediction based on UAV imagery and machine learning, specifically including the following steps: This invention provides a method for crop cultivation pattern recommendation and yield prediction based on UAV imagery and machine learning, including: S1: Acquire and perform critical cell segmentation processing on RGB and multispectral images obtained by drones during the key crop growth stages; specifically including the following steps: S1.1: During the critical growth period of crops, drones equipped with visible light cameras and multispectral cameras were used to conduct aerial photography of the experimental fields to obtain visible light RGB images and multispectral images covering the crop experimental plots; specifically, image acquisition was carried out using a DJI M300 drone equipped with a Zenmuse P1 camera (RGB) and a multispectral camera, and the flight path adopted a dual-grid mode to ensure that the image overlap was ≥80%; the visible light image acquisition height was 20m, and the multispectral image acquisition height was 25m.

[0018] The crops include corn, rice, and wheat; the key growth periods for corn include the tasseling and silking stage, the grain-filling stage, and the maturity stage; the key growth periods for rice include the heading and booting stage, the grain-filling stage, and the full maturity stage; and the key growth periods for wheat include the heading and booting stage, the grain-filling stage, and the waxy maturity stage.

[0019] S1.2: After the UAV data is collected, it is stitched and corrected, and then aerial triangulation and ground control point calibration are performed to generate two-dimensional orthophotos and three-dimensional point cloud data. If the number of feature points extracted from a single image is ≥1000, the proportion of calibrated images is ≥95%, and the difference in camera parameter optimization is <5%, the image is considered qualified.

[0020] S1.3: Use Airmeasurer to segment cells from the stitched visible light data. Use Airmeasurer to generate a digital surface model (DSM) and a digital elevation model (DEM) from the generated 2D orthophoto and 3D point cloud data. Calculate the canopy height model CHM = DSM - DEM. Use Hough transform to detect cell boundaries and combine it with a mask for automated cell segmentation. Obtain the visible light cell image corresponding to each cell and output a polygon shape file (ROI region of interest) describing the boundaries of each cell. S1.4: Using the polygon shape file obtained during the processing of visible light images, the corresponding multispectral image region is found through perspective transformation, rotation, and registration operations. Then, the same mask is applied to achieve cell segmentation of the multispectral image (specifically, the polygon and mask are used to perform cell cropping and segmentation on the multispectral orthophoto). The cell segmentation results of the multispectral image in five bands of red, green, blue, near-infrared and red edge are obtained, that is, the multispectral cell image of each cell in at least five bands of red, green, blue, near-infrared and red edge.

[0021] S2: Based on common trait indices extracted from visible light and multispectral images at the plot level, a crop growth parameter estimation model integrating visible light and multispectral features is constructed; specifically, the following steps are included: S2.1: Based on the visible light and multispectral cell images of each cell, extract image features to characterize crop phenotypes; image features include visible light features and multispectral features; Visible light features were extracted from RGB orthophotos, including canopy coverage (CC), average plant height (PH), visible light atmospheric impedance index (VARI), normalized differential yellowness index (NDYI), excess green (ExG), excess red (ExR), normalized vegetation index (NVI), and green leaf index (GLI). Average plant height (PH) was obtained directly from CHM. Multispectral features were used to extract soil conditioning index OSAVI, red-edge chlorophyll index CIRE, ratio vegetation index RVI, green normalized vegetation index GNDVI, leaf chlorophyll index LCI, normalized red-edge vegetation index NDRE, and crop nitrogen reflectance index NRI from multispectral images in blue, green, red, red-edge, and near-infrared bands. Specifically, this invention utilizes a multispectral image processing software to achieve a complete process from multispectral image segmentation to automated extraction of common trait indices. After logging into the system, users upload three files: polygon shape file polygon.shp, mask image mask.png, and multispectral images of different bands image.tif, obtained from processing visible light images with Airmeasurer. The system then uses the ROI region (polygon.shp file) to crop the multispectral image and uses the mask to complete segmentation. After segmentation, the system uploads the segmented multispectral images of five bands. Each cell is numbered based on connectivity. Then, according to the formulas for extractable phenotypic parameters from common multispectral images, the system calculates the values ​​of these parameters for each pixel in each cell, averages them, and finally saves all extracted phenotypic parameters from each cell into a CSV file. This automated extraction and saving of common trait indices from multispectral images provides a data foundation for constructing a maize growth parameter estimation model.

[0022] The image features extracted from each cell were correlated with the corresponding artificially measured growth parameter samples. The artificially measured growth parameters included at least chlorophyll content SPAD, leaf area index LAI, and plant height PH. S2.2: The Pearson correlation coefficient is used to analyze the correlation between each image feature and SPAD and LAI, and a preset number of feature subsets are selected according to the correlation strength as the input features of the model. Using the selected image features as independent variables and chlorophyll content SPAD, leaf area index LAI, and plant height PH as dependent variables, decision tree regression model, random forest regression model, and support vector machine regression model were constructed respectively to obtain estimation models from image features to crop growth parameters.

[0023] S3: Construct a comprehensive crop growth monitoring model. Input the plant height, pH, and chlorophyll content (SPAD) and leaf area index (LAI) estimated by the crop growth parameter estimation model into the comprehensive crop growth monitoring model to obtain the comprehensive growth index (CGI) for each plot. Select the cultivation mode corresponding to the plot with the highest CGI as the recommended cultivation mode for the target area. Specifically, this includes the following steps: S3.1: Chlorophyll content SPAD, leaf area index LAI, and plant height PH were selected as growth evaluation indicators. The weights were calculated and the average values ​​were taken using the coefficient of variation method, entropy weight method, and fuzzy comprehensive evaluation method. The calculation formula is as follows: Coefficient of variation method: ; Entropy weight method: ; Fuzzy comprehensive evaluation method: ; The final weights are used to calculate the Comprehensive Growth Monitoring Index (CGI) based on the selected data. The specific calculation formula is as follows:

[0024] S3.2: Calculate the Comprehensive Growth Index (CGI) for each plot according to the above formula; where each plot corresponds to a different combination of crop variety, planting pattern and planting density. Sort the CGI of each plot and select the variety, planting pattern and planting density combination corresponding to the plot with the highest CGI as the recommended cultivation pattern for the target area.

[0025] S4: Construct crop yield prediction models for different growth stages, using plant height, pH, and chlorophyll content (SPAD) and leaf area index (LAI) estimated by crop growth parameter estimation models as input data. Then, use the prediction model corresponding to the growth stage of the input data to predict yield. Specifically, this includes the following steps: S4.1: Based on the multi-temporal image features of each plot and the actual yield measurement data of the plot, the SelectKBest feature selection method based on f_regression is used to screen several image features with the strongest correlation with yield at different growth stages, and obtain the stage feature subsets corresponding to different growth stages (such as tasseling and silking stage, grain filling stage, and maturity stage). S4.2: Using the feature subsets corresponding to each growth stage as input and the actual yield of the corresponding plot as output, construct random forest regression models according to the growth stage to obtain crop yield prediction models for each growth stage. S4.3: Acquire UAV images of the target area during the key growth period, extract image features and estimate crop growth parameters through steps S1 and S2, and use a random forest regression model corresponding to the current growth period to predict yield and obtain the predicted yield value of the target area.

[0026] The method of this invention will be further explained below based on maize experimental data from the arid zone of central Ningxia Hui Autonomous Region: Example 1 Step 1: UAV image data acquisition and preprocessing, the specific steps are as follows: Step 1.1: Field Trial Design. The experimental site was selected as the Modern Agricultural Water-Saving and High-Efficiency Demonstration Park in the Arid Central Zone of Ningxia Hui Autonomous Region (36°54′6″N, 105°58′17″E). The maize varieties tested were Ningdan 52, Ningdan 56, and Xianyu 1225 (recombinant inbred line population). Four planting patterns were set up (equal row spacing 55cm, one plant per hole with 70-40cm wide and narrow rows, two plants per hole with 70-40cm wide and narrow rows, and three plants per hole with 70-40cm wide and narrow rows) and four planting densities (5000 plants / mu, 5500 plants / mu, 6000 plants / mu, and 6500 plants / mu), for a total of 16 treatment combinations. The area of ​​each plot was 33m². 2 (3.3m × 10m), such as Figure 2 As shown.

[0027] Drone parameter settings: DJI M300 drone flight speed 5m / s, shutter speed 1 / 1000s, ISO 100; multispectral camera wavelength bands: blue (450nm), green (560nm), red (650nm), red edge (730nm), near-infrared (850nm).

[0028] Ground control points (GCPs) layout: 10 GCPs are evenly distributed in the experimental field. The coordinates are measured using RTK equipment (accuracy ±2cm) for image stitching calibration.

[0029] Step 1.2: Stitching and correction of UAV images. Using Pix4D Mapper software, aerial triangulation and ground control point (GCPs, acquired via RTK) calibration were performed on the original images to generate a 2D orthophoto (GeoTIFF format) and a 3D point cloud.

[0030] Pix4D Mapper software processing flow: Data import → Initialization processing (3D Maps template) → GCPs import → Point cloud generation (low density) → Orthophoto output (resolution 1.32cm / pixel).

[0031] Step 1.3: Use Airmeasurer to segment the visible light data obtained by stitching together the data using Pix4D Mapper. Use Airmeasurer to generate a Digital Surface Model (DSM) and a Digital Elevation Model (DEM) based on the 3D point cloud data from the generated 2D orthophoto map and 3D point cloud data, and calculate the canopy height model (CHM = DSM - DEM). Use Hough transform to detect cell boundaries, and combine this with a mask to achieve automatic cell segmentation, extracting independent image data for each cell (e.g., ...). Figure 3 (As shown).

[0032] Step 1.4: Airmeasurer can only process visible light data. This step adds the function of cell segmentation for multispectral images. Using polygon shape files (ROI regions) obtained during visible light image processing, the corresponding multispectral image regions are found through perspective transformation, rotation, and other operations. Then, the same mask is applied to achieve cell segmentation of the multispectral image, resulting in cell segmentation results for the five bands (red, green, blue, nir, and red edge) of the multispectral image.

[0033] Step 2: Construct a corn growth parameter estimation model. The specific steps are as follows: Step 2.1: Extract common characteristic indices for each plot. Visible light characteristics: Canopy cover (CC), average plant height (PH, directly obtained from CHM), visible light atmospheric impedance index (VARI), normalized differential yellowness index (NDYI), excess green (ExG), excess red (ExR), normalized vegetation index (NVI), and green leaf index (GLI) are extracted from RGB orthophotos. The calculation formulas are shown in Table 1 below. Table 1. Phenotypic indices that can be extracted from common visible light images

[0034] Multispectral features: Soil adjustment index (OSAVI), red edge chlorophyll index (CIRE), ratio vegetation index (RVI), green normalized vegetation index (GNDVI), leaf chlorophyll index (LCI), normalized red edge vegetation index (NDRE), and crop nitrogen reflectance index (NRI) were extracted from multispectral images (blue, green, red, red edge, and near-infrared bands).

[0035] Step 2.2: VARI, NDRE, and other indices are calculated and CHMs are generated using Python open-source libraries (OpenCV, Rasterio, WhiteboxTools). A complete interactive system was also developed, encompassing multispectral image cell segmentation and automated extraction of common trait indices from multispectral images. After logging into the system, users upload three files: a polygon shape file (polygon.shp) obtained from processing visible light images with Airmeasurer, a mask image (mask.png), and multispectral images (image.tif) of different bands. The system then allows cropping of the multispectral image using the ROI region (polygon.shp file) and cell segmentation using the mask. After segmentation, the five-band multispectral images of the segmented cells are uploaded, and the "Calculate Vegetation Indices" button is clicked (e.g., ...). Figure 4-5 As shown in the figure, the system will number each cell according to connectivity, and then calculate the value of each of the seven phenotypic parameters for each pixel in each cell according to the formulas for the seven phenotypic parameters that can be extracted from common multispectral images in Table 2 below. The average value of the seven parameters extracted in each cell will be obtained by taking the average value. Finally, the system will save all the phenotypic parameters extracted from each cell into a CSV file, automatically completing the extraction and saving of common trait indices from multispectral images, providing a data foundation for building a maize growth parameter estimation model.

[0036] Table 2 Phenotypic Indices Extractable from Common Multispectral Images

[0037] Model training environment: Python 3.9, Scikit-learn library to build DT, RF and SVM models, and grid search (GridSearchCV) for hyperparameter optimization.

[0038] Step 2.3: Feature Selection and Model Training. The Pearson correlation coefficient was used to select the top 6 features with the strongest correlation to chlorophyll content (SPAD) and leaf area index (LAI) (SPAD: ExR, NRI, NDRE, CIRE, LCI, GNDVI; LAI: ExR, NRI, NDRE, CIRE, LCI, CC), as shown in Table 3 below. Using the selected features as independent variables and manually measured values ​​(SPAD: 24 plants per plot, measured 3 times each at the leaf tip / middle / base, and the average was taken; LAI: length-width coefficient method, empirical coefficient 0.75; PH: 10 plants per plot, measured and the average was taken) as dependent variables, decision tree (DT), random forest (RF), and support vector machine (SVM) regression models were constructed. Mean squared error (MSE) and coefficient of determination (R²) were used. 2 Evaluate model performance.

[0039] Table 3. Correlation between various phenotypic indices and chlorophyll content and leaf area index

[0040] Step 2.4: Verification of growth parameter estimation results. Plant height: The average plant height extracted from CHM and the manually measured value R. 2 =0.83, MSE=0.0096 (e.g.) Figure 6 (As shown); Chlorophyll content: Random forests showed the best estimation performance during the grain-filling stage (R0). 2 =0.81, MSE=2.40); Leaf area index: Decision trees showed the best estimation performance at maturity (R 2 =0.91, MSE=0.09).

[0041] The prediction results of each model for chlorophyll content are shown in the table below:

[0042] The prediction results of each model for leaf area index are shown in the table below:

[0043] Step 3: Construct a comprehensive corn growth monitoring model. The specific steps are as follows: Step 3.1: Select chlorophyll content (SPAD), leaf area index (LAI), and plant height (PH) as growth indicators. Calculate the weights using the coefficient of variation method, entropy weight method, and fuzzy comprehensive evaluation method, and take the average value. The calculation formula is as follows: Coefficient of variation method: .

[0044] Entropy weight method: .

[0045] Fuzzy comprehensive evaluation method: .

[0046] The weights of the various growth indicators for corn are determined as shown in the table below:

[0047] The final weights are: LAI (0.6121) > PH (0.2227) > SPAD (0.1652).

[0048] The final weights are multiplied by the data to obtain the comprehensive growth monitoring index. The specific calculation formula is as follows: .

[0049] Obtain the overall growth status of each community .

[0050] Step 3.2: Growth Evaluation and Cultivation Pattern Recommendation. Each plot has different planting methods (variety, planting pattern, and planting density), and vegetation indices and three growth parameters (SPAD, LAI, and PH) were previously extracted for each plot. The CGI score for each plot was calculated using the formula above (e.g., ...). Figure 7 As shown), all plots were sorted based on CGI, and the combination with the highest CGI score was selected: Variety 1 Ningdan 52 + two plants per hole + 5500 plants / acre (e.g. Figure 8 (As shown). CGI was significantly correlated with key phenotypic indices such as ExR and NRI (P < 0.05), consistent with the results recommended by breeding experts from the Ningxia Academy of Agricultural and Forestry Sciences, thus verifying the reliability of the model.

[0051] The correlation between various growth monitoring indicators and vegetation indices was tested, and the results are shown in the table below:

[0052] Step 4: Construct a maize yield prediction model, specifically involving feature selection and phased modeling. Feature selection uses SelectKBest(f_regression) to select the top features most strongly correlated with yield (tasseling and silking stage: GLI, PH, VARI, CC, NDYI, ExG, ExR, NRI; grain-filling stage: PH, VARI, CC, NDYI, ExG, ExR, CIRE, NDRE; maturity stage: GLI, PH, VARI, CC, NDYI, ExG, ExR, NRI). Using the selected features as independent variables and manually measured yield (plot yield measurement) as the dependent variable, construct decision tree, BP neural network (BPNN), and random forest regression models, training and validating them for each growth stage (tasseling and silking stage, grain-filling stage, maturity stage).

[0053] Yield prediction data showed that the actual yield (unit: kg / mu) of the 16 plots in the experimental field ranged from 580 to 820 kg / mu. Comparing the yield prediction results of different models at different growth stages, the random forest model was optimal at all stages, with the highest yield at the tasseling and silking stage (R0). 2 =0.79, MSE=6.92), grouting period (R 2 =0.83, MSE=5.42), maturity period (R 2 =0.87, MSE=4.26), significantly outperforming traditional full-cycle modeling (R² = 0.87, MSE = 4.26). 2 =0.57).

[0054] The prediction results of the random forest model based on the yield at different growth stages are shown in the table below:

[0055] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for crop cultivation pattern recommendation and yield prediction based on UAV imagery and machine learning, characterized in that, include: S1: Acquire RGB and multispectral images of crops during critical growth periods using drones and perform key cell segmentation processing; S2: Based on the common trait indices extracted from visible light and multispectral images at the plot level, a crop growth parameter estimation model integrating visible light and multispectral features is constructed. S3: Construct a comprehensive crop growth monitoring model. Input the plant height, pH and chlorophyll content SPAD and leaf area index LAI obtained by the crop growth parameter estimation model into the comprehensive crop growth monitoring model to obtain the comprehensive growth monitoring index CGI of each plot. Select the cultivation mode corresponding to the plot with the highest CGI as the recommended cultivation mode for the target area. S4: Construct crop yield prediction models for different growth stages. Use plant height, pH, and chlorophyll content (SPAD) and leaf area index (LAI) estimated by crop growth parameter estimation models as input data, and use the prediction model corresponding to the growth stage of the input data to predict yield.

2. The method for crop cultivation pattern recommendation and yield prediction based on UAV imagery and machine learning according to claim 1, characterized in that, The crops include corn, rice, and wheat; The key growth periods for maize include the tasseling and silking stage, the grain-filling stage, and the maturity stage; the key growth periods for rice include the heading and booting stage, the grain-filling and grain-setting stage, and the full maturity stage; and the key growth periods for wheat include the heading and booting stage, the grain-filling stage, and the waxy maturity stage.

3. The method for crop cultivation pattern recommendation and yield prediction based on UAV imagery and machine learning according to claim 1, characterized in that, Step S1 specifically includes the following steps: S1.1: During the critical growth period of crops, drones equipped with visible light cameras and multispectral cameras are used to take aerial photos of the experimental fields to obtain visible light RGB images and multispectral images covering the crop experimental plots; S1.2: After the UAV data is collected, it is stitched and corrected, and then aerial triangulation is performed and ground control point calibration is carried out to generate two-dimensional orthophotos and three-dimensional point cloud data. S1.3: Use Airmeasurer to segment cells in the stitched visible light data. Use Airmeasurer to generate a digital surface model (DSM) and a digital elevation model (DEM) based on the 3D point cloud data. Calculate the canopy height model CHM = DSM - DEM. Use Hough transform to detect cell boundaries and combine it with a mask for automated cell segmentation. Obtain the visible light cell image corresponding to each cell and output a polygon shape file describing the boundaries of each cell. S1.4: Using the polygon shape file obtained during the processing of visible light images, the corresponding multispectral image region is found through perspective transformation, rotation, and registration operations. Then, the same mask is applied to achieve cell segmentation of the multispectral image, and the cell segmentation results of the multispectral image in the five bands of red, green, blue, near-infrared and red edge are obtained.

4. The method for crop cultivation pattern recommendation and yield prediction based on UAV imagery and machine learning according to claim 3, characterized in that, In step S1.2, if the number of feature points extracted from a single image is ≥1000, the proportion of calibrated images is ≥95%, and the difference in camera parameter optimization is <5%, the image is deemed qualified.

5. The method for crop cultivation pattern recommendation and yield prediction based on UAV imagery and machine learning according to claim 3, characterized in that, In step S1.1, image acquisition was performed using a DJI M300 drone equipped with a Zenmuse P1 camera and a multispectral camera. The flight path adopted a dual-grid mode to ensure that the image overlap was ≥80%. The visible light image acquisition height was 20m, and the multispectral image acquisition height was 25m.

6. The method for crop cultivation pattern recommendation and yield prediction based on UAV imagery and machine learning according to claim 1, characterized in that, Step S2 specifically includes the following steps: S2.1: Based on the visible light and multispectral images of each plot, extract image features to characterize crop phenotypes; the image features include visible light features and multispectral features; The visible light features are extracted from RGB orthophotos, including canopy coverage (CC), average plant height (PH), visible light atmospheric impedance index (VARI), normalized differential yellowness index (NDYI), excess green (ExG), excess red (ExR), normalized vegetation index (NVI), and green leaf index (GLI); the average plant height (PH) is obtained directly from CHM. The multispectral features extract soil conditioning index OSAVI, red-edge chlorophyll index CIRE, ratio vegetation index RVI, green normalized vegetation index GNDVI, leaf chlorophyll index LCI, normalized red-edge vegetation index NDRE, and crop nitrogen reflectance index NRI from multispectral images in blue, green, red, red-edge, and near-infrared bands. The image features extracted from each cell are associated with the corresponding artificially measured growth parameter samples of the cell. The artificially measured growth parameters include at least chlorophyll content SPAD, leaf area index LAI, and plant height PH. S2.2: The Pearson correlation coefficient is used to analyze the correlation between each image feature and SPAD and LAI, and a preset number of feature subsets are selected according to the correlation strength as the input features of the model. Using the selected image features as independent variables and chlorophyll content SPAD, leaf area index LAI, and plant height PH as dependent variables, decision tree regression model, random forest regression model, and support vector machine regression model were constructed respectively to obtain estimation models from image features to crop growth parameters.

7. The method for crop cultivation pattern recommendation and yield prediction based on UAV imagery and machine learning according to claim 1, characterized in that, Step S3 specifically includes the following steps: S3.1: Chlorophyll content SPAD, leaf area index LAI, and plant height PH were selected as growth evaluation indicators. The weights were calculated and the average values ​​were taken using the coefficient of variation method, entropy weight method, and fuzzy comprehensive evaluation method. The calculation formula is as follows: Coefficient of variation method: ; Entropy weight method: ; Fuzzy comprehensive evaluation method: ; The final weights are used to calculate the Comprehensive Growth Monitoring Index (CGI) based on the selected data. The specific calculation formula is as follows: S3.2: Calculate the Comprehensive Growth Index (CGI) for each plot according to the above formula; where each plot corresponds to a different combination of crop variety, planting pattern and planting density. Sort the CGI of each plot and select the variety, planting pattern and planting density combination corresponding to the plot with the highest CGI as the recommended cultivation pattern for the target area.

8. The method for crop cultivation pattern recommendation and yield prediction based on UAV imagery and machine learning according to claim 1, characterized in that, Step S4 specifically includes the following steps: S4.1: Based on the multi-temporal image features of each plot and the actual yield measurement data of the plot, the SelectKBest feature selection method based on f_regression is used to select several image features with the strongest correlation with yield at different growth stages to obtain the stage feature subsets corresponding to different growth stages. S4.2: Using the feature subsets corresponding to each growth stage as input and the actual yield of the corresponding plot as output, construct random forest regression models according to the growth stage to obtain crop yield prediction models for each growth stage. S4.3: Acquire UAV images of the target area during the key growth period, extract image features and estimate crop growth parameters through steps S1 and S2, and use a random forest regression model corresponding to the current growth period to predict yield and obtain the predicted yield value of the target area.

Images