Spatio-temporal variation monitoring method for SPAD of winter wheat

Abstract

A spatio-temporal variation monitoring method for SPAD of winter wheat, relating to the technical field of crop monitoring. The monitoring method comprises: a multispectral sensor of an unmanned aerial vehicle acquiring canopy images of winter wheat in a heading stage, a flowering stage, and a late grain-filling stage; extracting from multispectral images spectral features based on the canopy reflectance of winter wheat and texture features based on a graylevel co-occurrence matrix; using a feature selection policy to preferably select sensitive remote-sensing features; and applying a feature fusion policy and an SVR algorithm to construct a winter-wheat SPAD estimation model. Spectral features combining a near-infrared band with other bands can fully capture the spectral difference in SPAD of winter wheat in a reproductive growth stage, and texture features in a red-light band and the near-infrared band are relatively sensitive to the SPAD of winter wheat. The stability and estimation accuracy of an SPAD model constructed by applying a feature selection policy and a feature fusion policy are both superior to those of a model applying a feature policy alone and a model not applying any policy.

Description

A method for monitoring spatiotemporal changes of winter wheat SPAD Technical Field

[0001] The invention belongs to the technical field of crop monitoring, and in particular relates to a method for monitoring spatiotemporal changes of SPADs in winter wheat. Background Art

[0002] Soil plant analysis development (SPAD) characterizes the relative chlorophyll content in leaves. Chlorophyll is an important indicator of crop growth and development and is closely related to crop nutritional status. Therefore, accurately and efficiently obtaining winter wheat SPAD is crucial for field management decisions and growth monitoring.

[0003] Traditional methods of obtaining crop chlorophyll content mainly rely on destructive field sampling and indoor chemical analysis, which is time-consuming, labor-intensive, and costly. Studies have shown that the data obtained by the handheld SPAD-502 chlorophyll meter is closely related to the chlorophyll content determined by chemical analysis in the laboratory. The SPAD acquisition method is more convenient than the traditional chlorophyll measurement method and can perform non-destructive sampling. However, due to the limitation of its measurement points, the measurement work still has a certain labor cost, a small operation scale, and low efficiency, making it difficult to meet the needs of obtaining crop chlorophyll content information over a large area. SPAD is an important indicator for assessing the nutritional status of crops and an important parameter for characterizing the reproductive growth status of winter wheat from the heading stage to the filling stage. Therefore, non-destructive, rapid, and accurate monitoring of winter wheat SPAD plays a vital role in ensuring stable grain production and guiding precise field management.

[0004] Spectral signatures (such as vegetation indices and band reflectance) calculated from drone imagery have become a common remote sensing tool in precision agriculture, demonstrating significant potential for estimating crop phenotypic traits such as LAI, AGB, SPAD, and nitrogen content. However, in the late stages of crop growth, when there is high canopy cover, vegetation indices can experience spectral saturation. This is due to the complex canopy background of wheat during the reproductive growth stage. UAV imagery includes wheat ears, stems, leaves, and a small amount of soil. This saturated image information reduces the sensitivity of vegetation indices. Therefore, it is difficult to establish a reliable SPAD estimation model for wheat during the reproductive growth stage using only spectral signatures.

[0005] Existing technologies primarily focus on SPAD estimation during the vegetative growth stage of wheat. However, due to the complex canopy background of winter wheat in its later stages of growth, the estimation model accuracy is low. Numerous studies have investigated the fusion of spectral and texture features, but the selection of feature variables for these massive fusion datasets is rarely considered. Data redundancy between spectral and texture features can reduce model performance. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for monitoring the spatiotemporal variation of SPAD in winter wheat. By taking winter wheat of different varieties and different nitrogen application rates as the research object, the multispectral sensor of unmanned aerial vehicles was used to obtain the canopy images of winter wheat at the heading stage, flowering stage and late filling stage. The spectral characteristics and texture characteristics of the images were extracted by image analysis technology. Based on the two feature selection methods of Boruta and Recursive Feature Elimination (RFE), remote sensing indicators sensitive to the SPAD of winter wheat were screened respectively. The support vector regression (SVR) machine learning algorithm was used to construct the SPAD prediction model of winter wheat in the reproductive growth stage. According to the coefficient of determination (R 2 ), root mean square error (RMSE) and residual prediction deviation (RPD) were used to evaluate 63 models and solve the existing problems.

[0007] To solve the above technical problems, the present invention is achieved through the following technical solutions:

[0008] A method for monitoring spatiotemporal changes of winter wheat SPAD comprises the following steps:

[0009] Step S001: using a drone to collect multispectral data of winter wheat during its growth period;

[0010] Step S002: using a SPAD-502 chlorophyll meter to measure the relative chlorophyll content (SPAD) of winter wheat in several sample plots;

[0011] Step S003: UAV image preprocessing: import the UAV multispectral images collected at each growth stage into PIX4Dmapper software for image stitching;

[0012] Step S004: Calculate the average of the relative chlorophyll content value of each plot and analyze it, and construct a SPAD prediction model for the spectral feature, texture feature and spectral texture feature fusion set respectively;

[0013] Step S005: Using the support vector machine regression (SVR) algorithm based on R language (version 4.1.3), the chlorophyll content of winter wheat based on the SPAD value was predicted.

[0014] Furthermore, the UAV includes five 12.08 million pixel monochrome sensors, namely blue, green, red, red-edge and near-infrared. The central wavelengths of the monochrome sensors are 450nm, 560nm, 650nm, 730nm and 840nm, respectively. The bandwidths of the monochrome sensors are ±16nm, ±16nm, ±16nm, ±16nm and ±26nm, respectively. The UAV includes an RTK (real-time kinematic) system. The vertical accuracy of the RTK (real-time kinematic) system is ±1.5 cm and the horizontal accuracy is ±1 cm.

[0015] Furthermore, in step S003, the image stitching method includes the following steps:

[0016] Step S0311: aligning histograms using a feature point matching algorithm to obtain a dense point cloud and texture mesh based on the drone image and position data;

[0017] Step S0312: Correcting the difference between bands;

[0018] Step S0313: Use ArcGIS 10.2 to create a vector (shapefile) file, divide the sample area according to the orthophoto map of the experimental area, generate several vector cells superimposed on the winter wheat image, and assign ID information to each vector cell.

[0019] Furthermore, the step S003 further includes the following steps:

[0020] Step S031: spectral feature extraction: obtaining the reflectance of several original bands of the winter wheat canopy during the growth period from the multispectral image;

[0021] Step S032: Texture feature extraction: Grey level co-occurrence matrices (GLCM) are used to extract texture information of the blue, green, red, red edge, and near-infrared bands in the multispectral image of the winter wheat canopy.

[0022] Furthermore, in step S0312, the method for correcting the difference between the bands is:

[0023] Place several radiation calibration panels with known reflectivity on the ground;

[0024] The multispectral image is radiometrically corrected using the image information of a ground-based radiation calibration plate with known reflectivity based on the empirical line method.

[0025] Furthermore, in step S031, 30 spectral features are selected for crop growth monitoring and parameter evaluation, and the spectral features are divided into 5 categories, namely, 5 original bands, 7 vegetation indices composed only of visible light bands, 5 vegetation indices composed of red edge bands and no near infrared bands, 8 vegetation indices composed of near infrared bands and no red edge bands, and 5 vegetation indices composed of near infrared and red edge bands.

[0026] Furthermore, in step S032, the texture information includes eight indicators, namely, mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment and correlation. The maximum value, minimum value, mean and standard deviation of the texture (GLCM) indicator of each sample plot are calculated to obtain 5×8×4 texture indicators.

[0027] Furthermore, in step S004, when constructing the SPAD prediction model, the RFE algorithm and the Boruta algorithm are used to screen different feature parameter sets, and the optimal feature combination is obtained by fusing different feature parameter sets and comparing and analyzing the model performance with the unscreened feature parameter sets;

[0028] The feature parameter set variables selected by the RFE algorithm are prefixed with R-, and the feature parameter sets selected by the Boruta algorithm are represented by C- and CT- respectively;

[0029] Among them, the feature parameter sets screened by the Boruta algorithm include two categories: "Confirmed" (C) and "Confirmed+Tentative" (CT).

[0030] Furthermore, in step S005, the method for predicting the chlorophyll content of winter wheat based on the SPAD value includes the following steps:

[0031] Step S0051: dividing the SPAD feature parameter set of winter wheat at each growth stage into a training set and a validation set;

[0032] Step S0052: Use three statistical indicators to test the machine learning model. The three statistical indicators are: coefficient of determination (R 2 ), root mean square error (RMSE), and performance to deviation ratio (RPD); where:

[0033] Where, and are the observed and measured values ​​of SPAD, is the average value of SPAD observations, is the number of samples, and SD is the standard deviation of the observations.

[0034] The present invention has the following beneficial effects:

[0035] This method, through the combination of spectral features involving the near-infrared band, can fully capture the spectral differences of winter wheat SPAD during the reproductive growth stage. The texture features of the red and near-infrared bands extracted by the Grey Level Co-occurrence Matrices (GCLM) method are more sensitive to wheat SPAD. In addition, the feature selection strategy combined with the feature fusion strategy significantly improves the robustness and accuracy of the prediction model.

[0036] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time.

[0037] BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0039] FIG1 is a flow chart of a method for monitoring spatiotemporal changes of winter wheat using SPAD according to the present invention;

[0040] FIG2 is a diagram showing the spatiotemporal variation of the spectral characteristics and SPAD of winter wheat at three growth stages according to the present invention;

[0041] Figure 3 is a graph showing the spectral characteristics and SPAD curves of wheat at four different nitrogen application levels;

[0042] Figure 4 shows the spatiotemporal variation of texture characteristics and SPAD under different nitrogen treatments during the three growth periods;

[0043] Figure 5 is a graph showing the changes in wheat texture characteristics and SPAD under four nitrogen application levels;

[0044] Figure 6 shows the optimal number of variables selected by Boruta and RFE feature selection methods for different feature sets at different reproductive stages;

[0045] Figure 7 is a scatter plot of the optimal model for estimating SPAD in the late growth period of winter wheat.

[0046] DETAILED DESCRIPTION

[0047] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present application with unnecessary detail.

[0048] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0049] It will also be understood that the term "and / or" used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0050] References to "one embodiment" or "some embodiments" in this specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in one or more embodiments of the present application. Thus, phrases such as "in one embodiment," "in some embodiments," "in other embodiments," and "in other embodiments" appearing in various places in this specification do not necessarily refer to the same embodiment, but rather mean "one or more but not all embodiments," unless otherwise specifically emphasized. The terms "including," "comprising," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0051] Example 1:

[0052] The present invention provides a method for monitoring the spatiotemporal changes of winter wheat SPAD. The winter wheat experiment was carried out in Chuzhou City, Anhui Province, China from 2020 to 2021 (32°48′52″N, 117°46′7″E). The region is located in the middle and lower reaches of the Yangtze River and has a subtropical monsoon climate with a humid climate, four distinct seasons, an average annual temperature of 15.4°C, an annual precipitation of 1000-1100 mm, an average annual rainfall of 144 days, and a frost-free period of about 210 days throughout the year.

[0053] There are 36 plots in this field experiment, each plot has 16m 2A 2×8 m plot was developed, involving three winter wheat varieties with stable high-yield potential (V1: Huaimai 44, V2: Yannong 999, and V3: Ningmai 13) and four nitrogen fertilizer application levels (N0-N3: 0, 100, 200, and 300 kg / ha), with three replicates per treatment. The fertilization regimen consisted of a pre-sowing basal application of phosphorus fertilizer (P=90 kg / ha) and potassium fertilizer (K=135 kg / ha), with a nitrogen fertilizer ratio of 6:4 applied before sowing and during the jointing stage. Winter wheat was sown in drills on November 7, 2020, with a row spacing of 30 cm. Harvested on June 3, 2021, all other field management practices were in accordance with local high-yield cultivation practices. No pests, diseases, or weeds occurred throughout the winter wheat growth period, and the field environment was good, with no drought or waterlogging.

[0054] As shown in FIG1 , the present invention is a method for monitoring the spatiotemporal changes of winter wheat using SPAD, comprising the following steps:

[0055] Step S001: Data collection based on drones;

[0056] Step S011: Multispectral (MS) data were collected using a Phantom 4 Multispectral RTK (DJI Technology Co., Shenzhen, China) (P4M) drone from 11:00 to 13:00 during the growing season (wheat heading stage, April 18, 2021; flowering stage, April 29, 2021; and late grain filling stage, May 24, 2021).

[0057] As an embodiment provided by the present invention, preferably, P4M has five 2.08 million pixel monochrome sensors: Blue, Green, Red, Red_edge and NIR, with central wavelengths of 450nm, 560nm, 650nm, 730nm and 840nm, and bandwidths of ±16nm, ±16nm, ±16nm, ±16nm, ±26nm respectively. It is also equipped with a real-time kinematic (RTK) system. RTK can achieve a vertical positioning accuracy of ±1.5 cm and a horizontal positioning accuracy of ±1 cm, so that P4M can obtain high-precision spectral and texture information.

[0058] As an embodiment provided by the present invention, preferably, in order to avoid the loss of spectrum and texture information of the drone image due to cloud cover, the flight is carried out between 11:00 and 13:00 on the same day under weather conditions of clear sky, no wind and clouds, and stable sunlight radiation intensity.

[0059] As an embodiment provided by the present invention, preferably, four radiation calibration plates with known reflectivity are arranged on the ground to facilitate radiation calibration during subsequent data processing.

[0060] As an embodiment of the present invention, preferably, to improve image accuracy, the UAV equipped with the MS sensor automatically adjusts exposure during flight based on the environment. The flight route was planned in advance using DJI GS PRO software (https: / / www.dji.com / cn / ground-station-pro / ), and the drone's autopilot system executed a predefined flight plan. Each flight lasted 20 minutes, with an altitude of 30 meters and a speed of 2 meters per second. The heading and lateral overlaps were 90% and 85%, respectively, and the image resolution was 1600 × 1300 pixels.

[0061] Step S002: field data collection;

[0062] Step S021: Field measurements for each growth stage were conducted prior to each day's drone flight. Non-destructive SPAD data for winter wheat were measured at 36 plots using a SPAD-502 chlorophyll meter (Konica Minolta Optics Inc., Osaka, Japan). Three plants of average growth were selected from each plot, and SPAD readings were taken at 1 / 6, 3 / 6, and 5 / 6 of the flag leaf length. The average of these nine SPAD readings was used as the measured SPAD value for that plot. A total of 324 SPAD readings were collected for each growth stage, and the SPAD data for the 36 plots were calculated.

[0063] Step S003: UAV image preprocessing, the UAV MS multispectral images collected in each growth period are imported into PIX4Dmapper software (4.4.12 version, Pix4D SA, Prilly, Switzerland) for image stitching, firstly aligning the images using the feature point matching algorithm, and obtaining dense point clouds and texture grids based on the UAV images and position data. In order to improve the data quality, the present invention uses the image information of the ground radiation calibration plate with known reflectivity to carry out radiation correction on the MS image based on the empirical linear method. ArcGIS10.2 (Environmental Systems Research Institute, Inc, RedLands, CA, USA) is used to create a shapefile file, and the sample area is divided according to the orthophoto map of the experimental area. 36 sample plots are generated and superimposed on the wheat image, and ID information is assigned one by one;

[0064] Specifically, field measurements for each growing season were conducted prior to each day's drone flight. Non-destructive SPAD data for winter wheat were measured at 36 plots using a SPAD-502 chlorophyll meter (Konica Minolta Optics Inc., Osaka, Japan). Three plants of average growth were selected from each plot, and SPAD readings were taken at 1 / 6, 3 / 6, and 5 / 6 of the flag leaf length. The average of these nine SPAD readings was used as the measured SPAD value for that plot. A total of 324 SPAD readings were collected for each growing season, and the SPAD data for all 36 plots were calculated.

[0065] Step S031: Spectral Feature Extraction: This method uses vegetation indices as spectral features (SFs). These indices are derived from characteristic bands through computation or combination, yielding a strong vegetation information factor, which enhances the expressive power of remote sensing data. Reflectances of five original bands of the winter wheat canopy during key growth stages (heading, flowering, and late grain filling) were obtained from MS multispectral images.

[0066] Step S032: Texture feature extraction: Texture is a common method for characterizing image information. It reflects important information such as surface structure and spatial arrangement in an image independently of brightness. It generally uses high-energy narrow peaks in the spectrum to detect image periodicity. This paper uses the grayscale co-occurrence matrix (GLCM) method to extract texture information in the blue, green, red, red-edge, and near-infrared bands of the multispectral MS image of winter wheat canopies.

[0067] Step S004: Calculate the average SPAD value of each plot and analyze it to directly characterize the growth status of winter wheat. Then, the spectral features, texture features and image information spectrum texture are fused and a SPAD prediction model is constructed respectively;

[0068] Step S005: The support vector machine regression (SVR) algorithm was used to predict the SPAD information of winter wheat based on R language version 4.1.3 (R Foundation, Vienna, Austria).

[0069] The drone includes five 2.08-megapixel monochrome sensors, namely: Blue, Green, Red, Red-edge, and NIR. The central wavelengths of the monochrome sensors are 450nm, 560nm, 650nm, 730nm, and 840nm, respectively, and the bandwidths of the monochrome sensors are 16nm, 16nm, 16nm, 16nm, and 26nm, respectively. The drone includes a real-time kinematic (RTK) system with a vertical and horizontal accuracy of ±0.1m, enabling P4M to obtain highly accurate spectral and texture information.

[0070] In step S003, the image stitching method includes the following steps:

[0071] Step S0311: aligning histograms using a feature point matching algorithm to obtain a dense point cloud and texture mesh based on the drone image and position data;

[0072] Step S0312: To improve data quality, correct the differences between bands;

[0073] Step S0313: Use ArcGIS 10.2 (Environmental Systems Research Institute, Inc, RedLands, CA, USA) to create a shapefile file, divide the sample area according to the orthophoto map of the experimental area, generate 36 cells superimposed on the winter wheat image, and assign ID information to each cell.

[0074] In step S0312, the method for correcting the difference between the bands is:

[0075] Place four radiation calibration plates with known reflectivity on the ground;

[0076] The MS multispectral image is radiometrically corrected using the image information of a ground-based radiation calibration plate with known reflectivity based on the empirical linear method.

[0077] As an embodiment provided by the present invention, preferably, in the step S031, the band reflectance and vegetation index are used as spectral features (SF). The vegetation index is a strong vegetation information factor obtained by calculating or combining the characteristic bands, which increases the expressive power of the remote sensing data to a certain extent. The reflectance of five original bands of the winter wheat canopy at the heading stage, flowering stage and late filling stage are obtained from the MS image to calculate the vegetation index. In order to study the influence of nitrogen level and growth stage on the SPAD value based on leaf chlorophyll, the present invention selects a total of 30 spectral features that have been widely used in crop growth monitoring and parameter evaluation. The spectral features are divided into five categories, namely, five original bands (I), seven vegetation indices consisting only of visible light bands (II), five vegetation indices consisting of red edge bands without near infrared bands (III), eight vegetation indices consisting of near infrared bands without red edge bands (IV), and five vegetation indices consisting of both near infrared and red edge bands (V), as shown in Table 1.

[0078] As an embodiment provided by the present invention, preferably, in step S032, the texture information includes eight indicators, namely Mean (Me), Variance (Va), Homogeneity (Ho), Contrast (Cn), Dissimilarity (Di), Entropy (En), Second moment (Se) and Correlation (Cr), as shown in Table 2. Texture does not depend on brightness but reflects important information such as surface structure and spatial arrangement in the image. Generally, the periodicity of the image is detected by using high-energy narrow peaks in the spectrum. The present invention adopts the Grey Level Co-occurrence Matrices (GLCM) method to extract texture information of blue, green, red, red edge and near-infrared bands in the MS image of winter wheat canopy.

[0079] As an embodiment provided by the present invention, preferably, the feature selection technology is not limited to considering the relationship between features and SPAD, but should also take into account the relationship between features, and process the multispectral remote sensing data of the UAV to determine the best method to maximize the improvement of SPAD estimation in the late growth period. A feature selection strategy is used to respectively select spectral features and texture features that are sensitive to SPAD, and a SPAD prediction model is further constructed to examine the difference between implementing and not implementing the feature selection strategy. A feature fusion strategy is used to fuse different types of feature subsets together to construct a SPAD estimation model.

[0080] Table 1 Spectral characteristics

[0081]

[0082] Table 2 Texture features

[0083] Abbreviated texture feature formula MeMean VaVariance Homogeneity CnContrast DiDissimilarity EnEntropy SeSecond moment CrCorrelation

[0084] As an embodiment provided by the present invention, more preferably, in the step S004, when constructing the SPAD prediction model, the RFE (Recursive feature elimination) algorithm and the Boruta algorithm are used to screen different feature parameter sets, and compare and analyze the model performance with the unscreened feature sets to obtain the optimal feature combination. Specifically, the Recursive feature elimination (RFE) algorithm first uses all features to train the model, calculates the importance of each feature and sorts them, and uses each feature subset to train the model, compares the model results obtained for each subset, and then retrains the model based on specific features, and repeats this process until the optimal feature combination is screened out to obtain maximized model performance. RFE can run based on a variety of algorithm models, such as commonly used machine learning methods such as RF and SVR. In the present invention, RFE runs based on the RF algorithm, and will output the optimal feature set after the run is completed;

[0085] The Boruta algorithm is a wrapper around the random forest algorithm and an ensemble method that uses multiple independent decision trees to perform classification by voting. It classifies all trees based on a given attribute and calculates their importance, a Z score that reflects the fluctuation in accuracy between trees in the forest. During runtime, Boruta creates "shadow" attributes by reshuffling the original attributes and randomly shuffles the order of feature parameters. When calculating feature importance, it categorizes feature parameters into three categories: features with a Z score significantly higher than the "shadow" attribute are considered confirmed (important), features with a Z score similar to the "shadow" attribute are considered tentative (potentially important), and features with a Z score significantly lower than the "shadow" attribute are considered rejected (unimportant). This paper categorizes Boruta algorithm screening results into two categories: confirmed and confirmed + tentative. It analyzes the impact of the two feature parameter sets on model construction and compares their accuracy. This approach offers the advantages of the random forest algorithm, with low runtime cost and the ability to generate results without relying on parameter adjustments.

[0086] The feature parameter set variables selected by the RFE algorithm are prefixed with R-, and the feature parameter sets selected by the Boruta algorithm (the feature parameter sets are divided into two categories: Confirmed and Confirmed+Tentative) are represented by C- and CT- respectively. For example, the spectral features selected by the RFE algorithm are represented by R-SF;

[0087] Among them, the feature parameter sets screened by the Boruta algorithm include two categories: Confirmed and Confirmed+Tentative.

[0088] As an embodiment provided by the present invention, more preferably, in step S004, when constructing the SPAD prediction model, feature fusion is a method of constructing a model by fusing different types of remote sensing features together. After the feature fusion strategy is jointly executed with the feature selection strategy, the stability, accuracy and robustness of the winter wheat SPAD estimation model are in a leading position, and as the growth period progresses, the improvement in model accuracy gradually increases. The model accuracy is improved the most in the late filling period. Compared with the model that only uses the initial spectral features or the initial texture features, R 2 Val Improved RMSE from 0.092 to 0.202 Val Reduced by 0.076 to 4.916, RPD Val Improved by 0.237 to 0.960; In this invention, the feature fusion strategy is mainly divided into two parts:

[0089] 1. First, the spectral features and texture features are fused, and the SPAD estimation model is constructed after the feature selection method is optimized. The model is compared with the model constructed by the fusion features that do not participate in feature selection;

[0090] 2. Based on the feature selection strategy, a SPAD estimation model is constructed by integrating the selected feature subsets of different categories, in order to find the SPAD estimation model with the best performance in the late stage of wheat growth.

[0091] Example 2:

[0092] As an embodiment provided by the present invention, more preferably, the method for predicting the chlorophyll content of winter wheat based on SPAD values ​​comprises the following steps:

[0093] Step S0051: Divide the SPAD characteristic parameter set of winter wheat at each growth stage into a training set and a validation set, randomly sample the characteristic parameter set, take 2 / 3 for model training, and use 1 / 3 for model validation;

[0094] Step S0052: Use three statistical indicators to test the machine learning model. The three statistical indicators are: the coefficient of determination (R 2 ) coefficient of determination, root mean square error (RMSE), and the ratio of performance to deviation (RPD); where:

[0095] Where, and are the observed and measured values ​​of SPAD, is the average value of SPAD observations, is the number of samples, SD is the standard deviation of the observations; the descriptive statistics of SPAD for the calibration and validation sets are shown in Table 3;

[0096] Table 3

[0097]

[0098] As an embodiment provided by the present invention, more preferably, in order to obtain more texture information for feature selection, the maximum value (MAX), minimum value (MIN), mean (MEAN) and standard deviation (SD) of the GLCM index of each sample area are calculated, generating a total of 5 (bands) × 8 (GLCM indices) × 4 (statistical metrics) = 160 texture indices.

[0099] A method for monitoring the spatiotemporal changes of winter wheat using SPADs was proposed. A multispectral sensor mounted on an unmanned aerial vehicle (UAV) was used to acquire high-resolution images of the winter wheat canopy. Image processing techniques were used to extract the spectral and textural features of the wheat. Two feature selection methods were used to mine the spectral and textural features of winter wheat in the late growth stage. A SPAD monitoring model for winter wheat was developed based on a combination of feature selection and feature fusion strategies to address the following issues:

[0100] (1) Clarify the spectral and textural characteristics of winter wheat during its reproductive growth stage that are sensitive to SPAD;

[0101] (2) Evaluate the performance of the winter wheat SPAD prediction model under the feature selection strategy;

[0102] (3) Explore the potential of combining feature selection strategy with feature fusion strategy to estimate SPAD in the late growth stage of winter wheat.

[0103] Based on Examples 1 and 2, it can be concluded that the spectral characteristics of the winter wheat canopy are closely related to SPAD. Figure 2 shows the spatiotemporal changes of the spectral characteristics and SPAD of winter wheat in the three growth periods. The depth of the color of the sample plot intuitively shows the changes in the spectral characteristics NDVI, GCVI and CVI; NDVI, GCVI, and CVI represent the spectral characteristics of the sample plot, and a, b and c represent the heading period, flowering period and late filling period of winter wheat, respectively. Figure 3 is a broken line graph of the spectral characteristics and SPAD of wheat under four different nitrogen application levels, and the nodes represent the average value of the characteristics. The spectral characteristics and SPAD showed obvious differences under different nitrogen gradient treatments, and the overall trend showed a gradual increase with the increase of nitrogen application. As the growth period progressed, the changes in the spectral characteristics of Huaimai 44, Yannong 999 and Ningmai 13 were consistent with the overall trend of SPAD dynamic changes, with no significant differences;

[0104] Based on Examples 1 and 2, it can be concluded that the temporal and spatial changes of the texture characteristics and SPAD of the three growth stages under different nitrogen treatments are shown in FIG4 ( FIG4 Spatiotemporal variation of winter wheat canopy texture characteristics and SPAD. MEAN.R.Di, MEAN.NIR.Me, and MEAN.NIR.Cr represent plot textural characteristics, with a, b, and c representing the heading, anthesis, and late grain-filling stages of winter wheat, respectively. MEAN.NIR.Di and MEAN.NIR.Me follow the same spectral characteristics, showing a decreasing trend from heading to late grain-filling, while MEAN.R.Cr shows the opposite trend. As shown in Figure 5 (a line graph of wheat textural characteristics and SPAD under four nitrogen fertilizer levels, with nodes representing the average value of each characteristic), MEAN.NIR.Di and MEAN.NIR.Me generally show an increasing and then decreasing trend with increasing nitrogen fertilizer application, with peaks generally occurring at the N2 and N1 levels. MEAN.NIR.Cr shows a gradually decreasing trend with increasing nitrogen fertilizer application. From heading to late grain-filling, the trends of MEAN.NIR.Cr and SPAD are almost completely opposite across the three wheat varieties and four nitrogen fertilizer treatments.

[0105] Based on Examples 1 and 2, it can be concluded that the present invention extracts 30 spectral features and 160 texture features from drone multispectral imagery. Not all features contribute to wheat SPAD estimation, and redundant features may affect model accuracy. Radar images of wheat spectral feature screening for three different periods (obtained experimentally and not provided at this time) show that the three feature sets selected for the heading stage spectral features are identical, with all five spectral feature types selected. Spectral feature types IV and V perform best during the flowering stage, with C-SF and CT-SF selecting all spectral features of types IV and V, while R-SF primarily selects type IV as the optimal variable set. Types IV and V perform best during the late grain filling stage, with the three feature sets showing highly consistent trends in the preferred spectral feature types. Spectral feature types IV and V share the same characteristics: both are composed of near-infrared bands. The spectral features are divided into five categories: five original bands (I), seven vegetation indices composed solely of visible light bands (II), five vegetation indices composed of red-edge bands without near-infrared bands (III), eight vegetation indices composed of near-infrared bands without red-edge bands (IV), and five vegetation indices composed of both near-infrared and red-edge bands (V). Figure 6 shows the optimal number of variables selected by the Boruta and RFE feature selection methods for different feature sets at different growth stages. A, B, and C represent the heading, flowering, and late grain filling stages of winter wheat. SF represents spectral features. R-SF is the spectral feature dataset retained by the RFE method, C-SF is the "Confirmed" spectral feature dataset retained by the Boruta method, and CT-SF is the "Confirmed + Tentative" spectral feature dataset retained by the Boruta method. TF represents texture features, and R-TF, C-TF and CT-TF are the same as above. According to the screening results of texture feature band types for the three periods in the present invention, the texture performance of Red, Red edge and NIR bands in the heading period is better, much greater than that of Blue and Green bands. The texture feature band types of the three screening results in the flowering period have the same trend, and the Red band and NIR band perform better; the texture feature bands of the three screening results in the late filling period have the same performance, and there are no significant differences in the 5 texture feature band types. The texture performance of the Red and NIR bands is relatively stable, and the number of feature retention after screening is significantly higher than that of other bands except for the late filling period. The 5 bands in the late filling period have the same performance without significant differences.

[0106] Comparative Example:

[0107] 1. Based on the spectral and texture features extracted from UAV multispectral imagery and combined with the feature selection strategy, SPAD estimation models were established for the heading stage, flowering stage, and late grain filling stage, respectively. The original feature set was added to each stage for model construction. A total of 24 (4*2*3) SPAD regression prediction models were constructed using four feature sets, two feature types (spectral features and texture features), and three growth stages.

[0108] From the perspective of model accuracy before and after feature selection, the model accuracy of the filtered dataset is almost higher than that of the initial dataset, but there are some exceptions. For example, the flowering period SF (validation set: R 2 =0.805, RMSE=3.211, RPD=1.998) is higher than the accuracy of the SPAD estimation model constructed by the R-SF, C-SF, and CT-SF datasets. The accuracy of the three datasets formed after feature selection varies, and the performance is inconsistent in different growth stages and different feature types. For example, the accuracy of the R-TF dataset at the heading stage (validation set: R 2 =0.800, RMSE=3.732, RPD=2.080) is lower than C-TF and CT-TF, while the accuracy of the R-TF dataset during flowering period (validation set: R 2 =0.799, RMSE=2.941, RPD=2.181) are higher than C-TF and CT-TF. The model accuracy of the C-dataset and CT-dataset filtered out by the Boruta algorithm is not constant.

[0109] 2. Based on the spectral and texture features extracted from UAV multispectral imagery, a feature fusion strategy was applied followed by a feature selection strategy to establish SPAD estimation models for the heading, anthesis, and late grain filling stages. The original feature set was incorporated into the model construction for each stage. A total of 12 (4 x 3) SPAD regression prediction models were constructed using four feature sets, one feature type (spectral and texture feature fusion), and three growth stages.

[0110] Under this feature strategy, the winter wheat canopy SPAD estimation model (R-SFTF) constructed by using the RFE feature selection method combined with the SFTF dataset at the heading stage has the best accuracy. The performance indicators of the model are as follows: Validation set: R 2 =0.861, RMSE=3.604, RPD=2.154. The winter wheat flowering canopy SPAD estimation model (C-SFTF) constructed by using the Boruta feature selection method combined with the SFTF dataset has the best accuracy. The specific performance indicators are: Validation set: R 2=0.740, RMSE=3.256, RPD=1.971. The SPAD estimation model of winter wheat canopy in the late filling period (R-SFTF) constructed by using RFE feature selection method combined with SFTF dataset has the best performance, among which the validation set: R 2 =0.761, RMSE=6.250, RPD=1.983.

[0111] 3. Based on the spectral and texture features extracted from UAV multispectral imagery, a feature selection strategy was applied followed by a feature fusion strategy to establish SPAD estimation models for the heading, anthesis, and late grain filling stages. For each stage, 27 SPAD prediction models (9 x 3) were constructed using 9 (3 x 3) feature sets, 1 feature type (feature fusion set), and 3 growth stages. Unselected SFTF feature sets were added for comparison at each stage, for a total of 30 regression models.

[0112] Under this strategy, since the feature sets of R-SF, C-SF, and CT-SF are the same during the heading period, there are three optimal estimation models that are fused with R-TF data during this period, namely R-SF--R-TF, C-SF--R-TF, and CT-SF--R-TF. The performance indicators of the models are as follows: Validation set: R 2 =0.857, RMSE=3.134, RPD=2.477. The winter wheat flowering period canopy SPAD estimation model (R-SF--R-TF) constructed by integrating the RFE feature selection method with the SF and TF datasets has the best accuracy. The specific performance indicators are: Validation set: R 2 =0.807, RMSE=2.850, RPD=2.251. The winter wheat canopy SPAD estimation model (C-SF-CT-TF) constructed by combining the SF and TF datasets using the Boruta feature selection method performed best. 2 =0.809, RMSE=5.878, RPD=2.108.

[0113] To further analyze the accuracy of the SPAD estimation models, Figure 7 shows a scatter plot of the measured and predicted values ​​for the optimal models at the heading, anthesis, and late grain filling stages, among the 63 models developed in this paper. As can be seen from the figure, the optimal monitoring models for the late winter wheat growth period are all SPAD estimation models constructed using a dual-feature strategy. Data points for each period cluster near the 1:1 line, and the predicted values ​​output by the models agree well with the field measurements, with minimal error. These results demonstrate that these models can accurately estimate the SPAD of the winter wheat canopy.

[0114] Through further analysis of the statistical indicators of texture features, four different statistical indicators based on GLCM features were extracted for each sample area: mean (MEAN), standard deviation (SD), maximum (MAX), and minimum (MIN). The commonly used mean indicator MEAN feature retention number always maintains the leading position among the four statistical indicators. However, in the importance ranking, it shows a trend of gradually decreasing importance as the growth period progresses. The changing trend of the importance of other indicators is opposite to that of MEAN. From the heading stage to the late filling stage, the importance gradually increases and reaches a peak in the late filling stage. Among them, SD shows the best performance trend, which indicates that other statistical indicators also have the potential to estimate wheat SPAD, especially SD. The results of this invention show that the texture feature SD information indicator also has the potential to estimate crop SPAD.

[0115] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0116] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for monitoring the spatio-temporal variation of SPAD in winter wheat, characterized in that, It includes the following steps: Step S001: Use a drone to collect multispectral data of winter wheat during the growth period; Step S002: Use a SPAD-502 chlorophyll meter to measure the relative chlorophyll content values of winter wheat in several sample plots; Step S003: Drone image preprocessing: Import the drone multispectral images collected in each growth period into the PIX4Dmapper software for image stitching; Step S004: Calculate the average value of the relative chlorophyll content values of each plot and conduct analysis, and construct SPAD prediction models for the spectral features, texture features, and spectral texture feature fusion sets respectively; Step S005: Based on the R language, use the support vector machine regression algorithm to predict the chlorophyll content of winter wheat based on the SPAD value.

2. The method for monitoring the spatio-temporal variation of SPAD of winter wheat according to claim 1, wherein, The drone includes 5 monochrome sensors with 12.08 million pixels each. The monochrome sensors are: blue, green, red, red edge, and near infrared. The central wavelengths of the monochrome sensors are 450nm, 560nm, 650nm, 730nm, and 840nm respectively. The bandwidths of the monochrome sensors are ±16nm, ±16nm, ±16nm, ±16nm, and ±26nm respectively. The drone includes an RTK system, and the vertical accuracy of the RTK system is ±1.5 cm and the horizontal accuracy is ±1 cm.

3. A method for monitoring the spatio-temporal variation of SPAD in winter wheat according to claim 1, characterized in that, In the step S003, the method of image stitching includes the following steps: Step S0311: Use the feature point matching algorithm to align the histograms, and obtain the dense point cloud and texture grid based on the drone image and position data; Step S0312: Correct the differences between bands; Step S0313: Use ArcGIS10.2 to create a vector file, divide the sample area according to the orthophoto map of the test area, generate several vector plots superimposed on the winter wheat image, and assign ID information to each vector plot one by one.

4. A method for monitoring the spatio-temporal variation of SPAD in winter wheat according to claim 1, characterized in that, The step S003 also includes the following steps: Step S031: Spectral feature extraction: Obtain the reflectances of several original bands of the winter wheat canopy during the growth period from the multispectral image; Step S032: Texture feature extraction: Use the gray-level co-occurrence matrix method to extract the texture information of the blue, green, red, red edge, and near infrared bands in the multispectral image of the winter wheat canopy.

5. A method for monitoring the spatio-temporal variation of winter wheat SPAD according to claim 3, characterized in that, In the step S0312, the method of correcting the differences between bands is: Deploy several radiometric calibration plates with known reflectances on the ground; Based on the empirical linear method, use the image information of the ground radiometric calibration plates with known reflectances to perform radiometric calibration on the multispectral image.

6. The method for monitoring the spatio-temporal variation of winter wheat SPAD according to claim 4, characterized in that, In the step S031, 30 spectral features for crop growth monitoring and parameter evaluation are selected. The spectral features are divided into 5 categories, namely 5 original bands, 7 vegetation indices composed only of visible light bands, 5 vegetation indices composed of red edge bands and without near infrared bands, 8 vegetation indices composed of near infrared bands and without red edge bands, and 5 vegetation indices composed of both near infrared and red edge bands.

7. A method for monitoring the spatio-temporal variation of winter wheat SPAD according to claim 4, characterized in that, In the step S032, the texture information includes eight indicators: mean, variance, homogeneity, contrast, dissimilarity, entropy, angular second moment, and correlation. Calculate the maximum value, minimum value, mean value, and standard deviation of each plot's texture indicators to obtain 5×8×4 texture indicators.

8. A method for monitoring the spatio-temporal variation of winter wheat SPAD according to claim 1, characterized in that, In the step S004, when constructing the SPAD prediction model, the RFE algorithm and the Boruta algorithm are used to screen different feature parameter sets. By fusing different feature parameter sets and comparing and analyzing with the model performance of the unscreened feature parameter sets, the optimal feature combination is obtained. The variables of the feature parameter sets screened by the RFE algorithm are prefixed with R-, and the feature parameter sets screened by the Boruta algorithm are represented by C- and CT- respectively. Among them, the feature parameter sets screened by the Boruta algorithm include two categories: "confirmed" and "to be determined".

9. A method for monitoring the spatio-temporal variation of winter wheat SPAD according to claim 7, characterized in that In the step S005, the method for predicting the chlorophyll content of winter wheat based on the SPAD value includes the following steps: Step S0051: Divide the SPAD feature parameter sets of each growth stage of winter wheat into a training set and a validation set. Step S0052: Use three statistical indicators to test the machine learning model. The three statistical indicators are: coefficient of determination R 2 , root mean square error RMSE, and performance-to-bias ratio RPD; where: In the formula, And are the observed and measured values of the SPAD, is the average value of the SPAD observations, is the number of samples, and SD is the standard deviation of the observed values.

Images