A rice comprehensive growth monitoring method based on unmanned aerial vehicle hyperspectral remote sensing

Abstract

The present application relates to the field of agricultural remote sensing technology, and more particularly to a rice comprehensive growth monitoring method based on unmanned aerial vehicle hyperspectral remote sensing. It comprises: obtaining unmanned aerial vehicle hyperspectral images and field growth parameters of rice at key growth stages; obtaining spectral reflectance data by preprocessing; clustering normalized growth parameters to generate pseudo labels, determining weights by fusing information gain rate and intra-group stability, and constructing comprehensive growth monitoring indicators; screening vegetation indices to establish an estimation model; establishing a function relationship between the indicators and grain yield, dividing the growth grades according to the yield decline proportion threshold strategy; and outputting the growth grade distribution map. The present application takes into account the contribution and stability of parameter information, improves the reliability of growth evaluation and the accuracy of yield estimation, and is suitable for precise nitrogen management of rice.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing technology, and in particular to a method for comprehensive monitoring of rice growth based on UAV hyperspectral remote sensing. Background Technology

[0002] Timely, accurate, and non-destructive monitoring of rice growth is a crucial technical means to assess rice productivity, optimize field nitrogen fertilizer management, and increase grain yield. Hyperspectral remote sensing using unmanned aerial vehicles (UAVs) can rapidly acquire continuous spatial information in the field, enabling dynamic monitoring of rice canopy morphology and physiological and biochemical parameters, and has been widely applied in crop growth evaluation.

[0003] Current rice growth assessments primarily rely on remote sensing data to retrieve individual growth parameters such as plant height, aboveground biomass, plant water content, photosynthetic pigments, and nitrogen content. However, a single growth parameter can only characterize a localized aspect of rice, making it difficult to comprehensively evaluate its overall growth status. Furthermore, retrieving multiple parameters separately requires significant computational resources and is constrained by inconsistent evaluation standards.

[0004] The Comprehensive Growth Monitoring Indicator (CGMI) combines multi-source information such as morphological traits, physiological and biochemical parameters, and canopy structure to comprehensively evaluate crop growth. Currently, commonly used calculation methods for CGMI include objective weighting methods such as the Entropy Weight Method (EWM) and the Coefficient of Variation (CV). The Entropy Weight Method determines weights based on the information entropy of the indicator, while the CV method determines weights based on the degree of data dispersion.

[0005] However, the above methods have the following shortcomings: the entropy weight method relies excessively on the degree of data fluctuation, which can easily lead to weight shifts when the sample size is small or the data distribution is uneven; the coefficient of variation method is greatly affected by outliers and lacks stability. Neither method can simultaneously consider the information contribution of the indicators and intra-group stability under unsupervised conditions. Furthermore, existing combinations of CGMI and UAV hyperspectral remote sensing data are mostly limited to single-parameter inversion, lacking a comprehensive growth estimation system that combines CGMI, vegetation index screening, and machine learning regression models. They also fail to establish a quantitative functional relationship between CGMI and grain yield, or a growth level classification strategy based on the yield decline ratio, thus limiting the reliability of growth level judgment and yield estimation.

[0006] Therefore, there is a need for a comprehensive rice growth monitoring method that can simultaneously consider the contribution of indicator information and intragroup stability under unsupervised conditions and deeply integrate with UAV hyperspectral imagery, in order to improve the reliability of growth level identification and grain yield estimation, and provide quantitative basis for precision nitrogen fertilizer management of rice. Summary of the Invention

[0007] To achieve the above objectives, this application discloses a method for comprehensive rice growth monitoring based on UAV hyperspectral remote sensing, comprising the following steps:

[0008] The drones acquired hyperspectral images of rice during key growth stages, simultaneously collected field growth parameters, and obtained grain yield at maturity.

[0009] The hyperspectral imagery of the UAV was preprocessed to obtain spectral reflectance data of the rice canopy;

[0010] The field growth parameters are normalized, pseudo-labels are generated based on the clustering results, the information gain rate and intra-group stability of each growth parameter are calculated, the information gain rate and intra-group stability are fused to determine the weight of each growth parameter, and a comprehensive rice growth monitoring index is constructed based on the normalized growth parameters and the weights.

[0011] Vegetation indices related to the comprehensive rice growth monitoring indicators were screened, and an estimation model for the comprehensive rice growth monitoring indicators was established based on the screened vegetation indices.

[0012] Establish a functional relationship between the comprehensive growth monitoring index of rice and grain yield, determine the growth level threshold according to the yield decline ratio threshold strategy, and classify the comprehensive growth of rice into weak growth, suitable growth and excessive growth based on the threshold.

[0013] The estimation model is used to invert the spectral reflectance data and output a comprehensive growth level distribution map of rice.

[0014] Furthermore, the key reproductive periods include the tillering stage, the jointing stage, and the heading stage.

[0015] Furthermore, the preprocessing includes image segmentation, geometric correction, radiometric correction, georegistration, image stitching, and reflectance conversion.

[0016] Furthermore, the field growth parameters include plant height, aboveground biomass, plant water content, chlorophyll content, carotenoid content, leaf nitrogen content, and leaf carbon content.

[0017] Furthermore, the process of determining the weights of each growth parameter by fusing the information gain rate and intra-group stability includes: introducing a fusion factor to perform weighted fusion of the information gain rate and intra-group stability to obtain a fusion strength index, and normalizing the fusion strength index to obtain the fusion weights of each growth parameter.

[0018] Furthermore, the fusion factor is determined through Bootstrap cross-validation: the clustering consistency index under different fusion factor values ​​is calculated, and the fusion factor corresponding to the maximum value of the clustering consistency index is taken as the optimal fusion factor.

[0019] Furthermore, the step of screening vegetation indices related to the comprehensive rice growth monitoring index includes: calculating the correlation between the band random combination vegetation index and the comprehensive rice growth monitoring index, screening vegetation indices whose correlation meets preset conditions as candidate vegetation indices, and using a permutation importance algorithm to screen important vegetation indices from the candidate vegetation indices.

[0020] Furthermore, the estimation model is established using partial least squares regression, random forest regression, or support vector regression algorithms.

[0021] Furthermore, the functional relationship is a quadratic function relationship; the step of determining the growth level threshold based on the yield decline ratio threshold strategy includes: calculating the theoretical maximum yield based on the quadratic function relationship, multiplying the theoretical maximum yield by the yield decline ratio threshold to obtain the yield threshold, and substituting the yield threshold into the quadratic function relationship to find the root, thereby obtaining the lower limit and upper limit of the growth level.

[0022] Furthermore, when the comprehensive growth monitoring index of rice is less than the lower limit, it is determined to be weak growth; when it is between the lower limit and the upper limit, it is determined to be suitable growth; and when it is greater than the upper limit, it is determined to be excessive growth.

[0023] During the tillering stage, the lower limit is 0.405 and the upper limit is 0.860;

[0024] During the jointing stage, the lower limit is 0.437 and the upper limit is 0.948;

[0025] During the heading stage, the lower limit is 0.497 and the upper limit is 0.956.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] This invention introduces cluster pseudo-labels, information gain ratio, and intra-group stability to fuse and weight multi-source growth parameters, enabling the constructed comprehensive growth monitoring index to simultaneously reflect the intra-group consistency and category discrimination ability of growth parameters. This overcomes the weight shift problem caused by the entropy weight method due to data fluctuations and the instability problem caused by the coefficient of variation method due to outliers, thus improving the stability of rice comprehensive growth assessment.

[0028] This invention achieves spatial continuous inversion from UAV hyperspectral imagery to rice comprehensive growth monitoring indicators by performing correlation analysis between randomly combined vegetation indices and comprehensive growth monitoring indicators, using a permutation importance algorithm to screen sensitive features, and combining machine learning regression algorithms to establish an estimation model. This improves the spatial resolution and physiological interpretability of comprehensive growth monitoring.

[0029] This invention establishes a quadratic function relationship between comprehensive growth monitoring indicators and grain yield, and uses a yield decline ratio threshold strategy to determine the growth level threshold. It quantitatively classifies the comprehensive growth of rice into weak growth, suitable growth, and excessive growth, making the growth level judgment directly related to grain yield estimation, thereby improving the reliability of yield estimation and providing a quantifiable and operable decision-making basis for precise nitrogen fertilizer management and differentiated regulation in the field. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart illustrating the model construction process of the present invention.

[0032] Figure 2 To verify the accuracy of rice CGMI estimation models based on entropy weight method (EWM), coefficient of variation method (CV), and information fusion stability method (IFS) (PLSR, RF, and SVR in the figure represent three machine learning algorithms: partial least squares regression, random forest regression, and support vector regression, respectively).

[0033] Figure 3 This is a schematic diagram illustrating the growth level classification based on the yield decline ratio threshold strategy of the present invention.

[0034] Figure 4 This is an inversion diagram of the overall growth of rice based on the information fusion stabilization method of this invention. Blue pixels represent weak growth, yellow pixels represent suitable growth, and red pixels represent excessive growth. Detailed Implementation

[0035] To make the technical problems solved, the technical solutions, and the beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0036] like Figure 1As shown, this embodiment constructs a comprehensive rice growth monitoring process based on UAV hyperspectral imagery, seven field growth parameters, and grain yield. The process includes: acquiring UAV hyperspectral images of rice during key growth stages, simultaneously collecting field growth parameters, and obtaining grain yield at maturity; preprocessing the UAV hyperspectral images to obtain rice canopy spectral reflectance data; normalizing the field growth parameters, generating pseudo-labels based on clustering results, calculating the information gain rate and intra-group stability of each growth parameter, fusing the information gain rate and intra-group stability to determine the weight of each growth parameter, and constructing a comprehensive rice growth monitoring index (CGMI) based on the normalized growth parameters and the weights; screening vegetation indices related to the CGMI, and establishing a CGMI estimation model based on the screened vegetation indices; establishing a functional relationship between the CGMI and grain yield, determining growth level thresholds based on a yield decline ratio threshold strategy, and classifying rice comprehensive growth into weak growth, suitable growth, and excessive growth based on the thresholds; and using the estimation model to invert the spectral reflectance data and output a comprehensive rice growth level distribution map.

[0037] In this embodiment, UAV hyperspectral images of rice during key growth stages are acquired, and field growth parameters are simultaneously collected. Grain yield is obtained at maturity. The key growth stages include tillering, jointing, and heading. Field growth parameters include plant height, aboveground biomass, plant water content, chlorophyll content, carotenoid content, leaf nitrogen content, and leaf carbon content. Plant height is measured from the base of the plant to the highest leaf or panicle tip; aboveground biomass and plant water content are calculated by collecting the fresh and dry weights of stems, leaves, and panicles; chlorophyll and carotenoid content are calculated spectrophotometrically using absorbance at 470 nm, 645 nm, and 663 nm; leaf nitrogen and leaf carbon content are obtained by measuring total nitrogen and total carbon content using an elemental analyzer. Grain yield is obtained through field harvesting at maturity. It should be noted that the key growth periods are not limited to the tillering stage, jointing stage, and heading stage, but can be extended to the booting stage or grain-filling stage according to monitoring needs; the growth parameters are also not limited to the above seven items, and parameters such as leaf area index can be added according to the monitoring objectives.

[0038] In a further embodiment, the hyperspectral imagery of the UAV is preprocessed to obtain rice canopy spectral reflectance data. The preprocessing includes image segmentation, geometric correction, radiometric correction, georegistration, image stitching, and reflectance conversion. Specifically, firstly, the original image data is segmented using flight trajectory information extracted from POS data; then, geometric and radiometric corrections are performed using professional remote sensing processing software to eliminate geometric distortions and radiometric errors caused by the UAV's flight attitude and lighting conditions; the corrected hyperspectral image sub-files are imported into geographic information system software for georegistration, assigning geographic coordinates to the images; image stitching is performed using remote sensing image processing software to generate a hyperspectral hypercube; finally, band resampling is performed, and a linear empirical method is used to convert the data into reflectance data. It should be noted that the software tools in the above preprocessing steps are only implementation examples; geometric and radiometric corrections can also be implemented using other commercial or open-source software with hyperspectral processing capabilities; the reflectance conversion method can also be replaced by an atmospheric correction method based on a physical model.

[0039] In a further embodiment, the field growth parameters are normalized, pseudo-labels are generated based on the clustering results, the information gain ratio and intra-group stability of each growth parameter are calculated, and the weights of each growth parameter are determined by fusing the information gain ratio and intra-group stability. Based on the normalized growth parameters and the weights, a comprehensive rice growth monitoring index is constructed. The normalization uses the range normalization method to convert the original values ​​into dimensionless values. The formula is:

[0040] ;

[0041] in, Let be the value of the j-th indicator for the i-th sample. and These are the maximum and minimum sample values ​​for the j-th indicator, respectively. It should be noted that the normalization method is not limited to range normalization; Z-score standardization or fractional scaling standardization can also be used.

[0042] In a further implementation, generating pseudo-labels based on clustering results refers to using a clustering algorithm to divide samples into a predetermined number of growth categories, and using the clustering results as pseudo-labels in supervised learning. In this embodiment, the K-Means clustering algorithm is used to divide the samples into K=3 growth categories, corresponding to weak growth, suitable growth, and excessive growth, respectively. It should be noted that the clustering algorithm is not limited to K-Means; unsupervised clustering algorithms such as K-Means++, hierarchical clustering, DBSCAN, or Gaussian mixture models can also be used instead.

[0043] In a further implementation, calculating the within-class stability (Stab) of each growth parameter refers to calculating the within-class variation ratio (WCV) of each growth parameter as a stability indicator. The within-class variation ratio reflects the consistency of the indicator among samples of the same cluster as the ratio of the sample variance within each cluster category to the total variance of all samples. The smaller the within-class variation ratio, the higher the within-class stability of the indicator. The specific formula is:

[0044] ;

[0045] ;

[0046] in, This represents the within-group variation ratio of the j-th indicator. This represents the number of samples in the k-th class. This represents the sample variance of the j-th indicator in the k-th class. This represents the within-group stability of the j-th indicator. This represents the total variance of the indicator across all samples.

[0047] In a further embodiment, calculating the information gain ratio (IGR) of each growth parameter refers to discretizing each growth parameter according to its quantiles and then calculating its information gain ratio (IGR) relative to the pseudo-label, which serves as an indicator of discriminative ability. The information gain ratio measures the contribution of the growth parameter to the clustering category division, reflecting its category discrimination ability. The specific formula is:

[0048] ;

[0049] ;

[0050] ;

[0051] in, H(Y) represents the information gain ratio of the j-th index, and H(Y) represents the information entropy of the cluster label. Let K represent the conditional entropy of the cluster label of the j-th index, and K represent the number of clusters. This represents the probability of clustering k. Y represents the grouping probability of the index discretization, and Y represents the cluster label. Let B represent the j-th indicator, and let B represent the number of groups for the indicator discretization.

[0052] In a further embodiment, determining the weights of each growth parameter by fusing the information gain ratio and intra-group stability includes: introducing a fusion factor to weight and fuse the information gain ratio and intra-group stability to obtain a fusion strength index, and normalizing the fusion strength index to obtain the fusion weights of each growth parameter. Specifically, a fusion factor is introduced. Information gain ratio and intragroup stability Weighted fusion is performed to obtain the fusion strength index. : Then on Normalization is performed to obtain the fusion weights. The formula is:

[0053] ;

[0054] ;

[0055] in, This represents the fusion strength index of the j-th index. Let be the fusion weight of the j-th indicator.

[0056] The comprehensive growth monitoring index of rice is calculated based on the fusion weight. :

[0057] ;

[0058] in, This indicates a comprehensive rice growth monitoring index constructed based on the entropy weight method. This represents the value of the index after normalization.

[0059] In a further implementation, the fusion factor is determined through Bootstrap cross-validation: Bootstrap resampling is used to calculate the clustering consistency index under different fusion factor values, and the fusion factor corresponding to the maximum value of the clustering consistency index is taken as the optimal fusion factor. Specifically, multiple Bootstrap resampling subsets with replacement are constructed (preferably 30 times); clustering, stability calculation, information gain ratio calculation, weight calculation, and [other processes are repeated] on each subset. Calculation; prediction based on clustering algorithms The categories are determined; the consistency between the predicted categories and the original pseudo-labels is evaluated using the Adjusted Rand Index (ARI); finally, the candidate α with the largest average ARI is selected as the optimal fusion factor. The formula is:

[0060] ;

[0061] ;

[0062] ;

[0063] Where B represents the number of bootstrap resampling operations. This represents the average value of ARI. Let α represent the cluster consistency score of α under the b-th cross-validation.

[0064] It should be noted that the clustering consistency index is not limited to the adjusted RAND index; normalized mutual information, V-Measure, or Fowlkes-Mallows index can also be used as evaluation indicators.

[0065] In a further embodiment, vegetation indices related to the comprehensive rice growth monitoring index are screened, and an estimation model for the comprehensive rice growth monitoring index is established based on the screened vegetation indices. The screening includes: calculating the correlation between the randomly combined band vegetation indices and the comprehensive rice growth monitoring index; screening vegetation indices whose correlation meets preset conditions as candidate vegetation indices; and using a permutation importance algorithm to screen important vegetation indices from the candidate vegetation indices. The vegetation indices include at least one of the following: Difference Vegetation Index (DVI), Ratio Vegetation Index (RVI), Normalized Difference Vegetation Index (NDVI), Green Normalized Difference Vegetation Index (GNDVI), Enhanced Vegetation Index (EVI), Red Edge Chlorophyll Index (CI_{red edge}), Red Edge Carotenoid Index (CI_{red edge car}), Water Index (WI), Double-Peak Canopy Nitrogen Index (DCNI), and Photochemical Reflectance Index (PRI). It should be noted that the types of vegetation indices are not limited to the ten mentioned above; other narrow-band vegetation indices may be added based on the hyperspectral band range.

[0066] In a further implementation, the estimation model is built using Partial Least Squares Regression (PLSR), Random Forest Regression (RF), or Support Vector Regression (SVR) algorithms. It should be noted that the machine learning regression algorithm is not limited to the above three; gradient boosting trees (XGBoost, LightGBM), neural networks (multilayer perceptrons, convolutional neural networks), or other ensemble learning algorithms can also be used instead. All vegetation indices are standardized before modeling to eliminate dimensional differences.

[0067] In a further implementation, a functional relationship is established between the comprehensive rice growth monitoring index and grain yield. A growth level threshold is determined based on a yield decline ratio threshold strategy. Based on this threshold, the comprehensive rice growth is classified into weak growth, suitable growth, and excessive growth. The functional relationship is a quadratic function, specifically in the form of:

[0068] Where y represents grain yield, and x represents... .

[0069] The method of determining the growth level threshold based on the yield decline ratio threshold strategy includes: calculating the theoretical maximum yield based on the quadratic function relationship; multiplying the theoretical maximum yield by the yield decline ratio threshold to obtain the yield threshold; and substituting the yield threshold into the quadratic function relationship to find the root, thereby obtaining the lower and upper limits of the growth level. Specifically, the theoretical maximum yield corresponding to the maximum grain yield is determined by the vertex of the quadratic function. and optimal ; Combine the output threshold coefficient (e.g., 0.95) with Multiply to obtain the output threshold ;Will Substituting into the quadratic function to find the roots, we get and These values ​​serve as the lower and upper limits for the rice growth level, respectively. When the comprehensive growth monitoring index of rice is below the lower limit, it is determined to be weak growth; when it is between the lower and upper limits, it is determined to be suitable growth; and when it is above the upper limit, it is determined to be excessive growth. At the tillering stage, the lower limit is 0.405 and the upper limit is 0.860; at the jointing stage, the lower limit is 0.437 and the upper limit is 0.948; and at the heading stage, the lower limit is 0.497 and the upper limit is 0.956.

[0070] In a further embodiment, the estimation model is used to invert the spectral reflectance data and output a comprehensive growth level distribution map of rice. Specifically, the preprocessed UAV hyperspectral image is input pixel by pixel into the estimation model, and the corresponding value for each pixel is calculated. The value is then used to classify each pixel into weak growth, suitable growth, or excessive growth based on the growth level threshold, thereby generating a spatially continuous growth level distribution map.

[0071] The technical effects of the above-mentioned methods are as follows: By integrating cluster pseudo-labels, information gain ratio, and intra-group stability into a weighted system, the comprehensive growth monitoring index can simultaneously reflect the intra-group consistency and class discrimination ability of growth parameters, reducing the weight bias of traditional entropy weighting and coefficient of variation methods under high-dimensional fluctuations and outlier conditions; By combining band random combination vegetation indices with machine learning regression models, high-precision inversion from UAV hyperspectral imagery to comprehensive growth monitoring indicators is achieved; By establishing a quadratic function relationship between comprehensive growth monitoring indicators and grain yield, and a yield decline ratio threshold strategy, growth level discrimination and yield estimation are directly linked, improving the reliability of growth level classification and yield estimation, and providing a quantifiable and operable decision-making basis for precision nitrogen fertilizer management in rice.

[0072] The technical solutions and effects of this application will be described in detail below through specific embodiments.

[0073] Example 1: A Comprehensive Rice Growth Monitoring Method Based on UAV Hyperspectral Remote Sensing

[0074] Field data were collected in 2022 and 2023 at the Huihe Base of Sichuan Agricultural University in Chengdu, Sichuan Province. The experiment employed a three-factor randomized block design, with 10 nitrogen application treatments, each with three replicates, for a total of 30 plots. The rice varieties used were F You 498 and Jing Liang You 534.

[0075] Step 1: Collect hyperspectral images and rice field growth parameters using a drone during the tillering, jointing, and heading stages. The drone used was a DJI M600 Pro, equipped with a pushbroom hyperspectral camera with a spectral range of 400-1000 nm, a spectral resolution of 3.3 nm, and a band spacing of 4 nm; the flight altitude was 50 m, and the flight path speed was 1.5 ms. -1 The flight path overlap rate and lateral overlap rate were 80% and 50%, respectively. Flight missions were conducted between 11:00 and 13:00 on clear, partly cloudy days. A 30% reflectance correction board was laid in the field for subsequent reflectance correction. Field growth parameters collected simultaneously included: plant height, aboveground biomass, plant water content, chlorophyll content, carotenoid content, leaf nitrogen content, and leaf carbon content. Yield was measured at maturity in each plot to obtain grain yield.

[0076] Step two, preprocessing, includes: segmenting the raw image data using flight trajectory information extracted from POS data; performing geometric and radiometric corrections using MegaCube software; importing hyperspectral image subfiles into ArcMap software for georegistration; stitching images using ENVI software; and generating a hyperspectral hypercube with a spatial resolution of 4 cm and a spectral range of 400-1000 nm in MegaCube software. Finally, band resampling is performed, and the data is converted to reflectance data using a linear empirical method.

[0077] Step 3: Normalize the range of the seven growth parameters. Use K-Means clustering to divide the samples into K=3 growth categories as pseudo-labels. Calculate the within-group variation ratio (WCV) and population variance for each growth parameter to obtain the within-group stability (Stab). Discretize each growth parameter according to its quantiles and calculate its information gain ratio (IGR) relative to the pseudo-label. Introduce a fusion factor α∈[0,1] to weight and fuse Stab and IGR to obtain the fusion strength G; normalize G to obtain the fusion weight. ;according to Formulas are used to calculate comprehensive growth monitoring indicators for rice.

[0078] To determine the optimal fusion factor A pseudo-supervised cross-validation framework is adopted: 30 bootstrap resampling cycles with replacement are constructed; clustering, stability calculation, information gain ratio calculation, weight calculation, and [other processes] are repeated on each subset. Calculation; using K-Means prediction based on The categories were adjusted to ensure consistency between the predicted categories and the original pseudo-labels in the RAND Index ARI evaluation; finally, the candidate α with the largest average ARI was selected as... .

[0079] Step 4: Calculate the following vegetation indices based on random band combinations: difference vegetation index, ratio vegetation index, normalized difference vegetation index, green light normalized difference vegetation index, enhanced vegetation index, red-edge chlorophyll index, red-edge carotenoid index, moisture index, bimodal canopy nitrogen index, and photochemical reflectance index. Calculate the above vegetation indices and... The correlation between vegetation indices was analyzed, and vegetation indices with high correlation were selected as candidate vegetation indices.

[0080] Step 5: Use the 2022 data as the training set and the 2023 data as the test set. Employ partial least squares regression, random forest regression, and support vector regression to build a training set. Estimation model: Before modeling, all vegetation indices are standardized, and a permutation importance algorithm is used to screen sensitive and important features from the standardized vegetation indices.

[0081] Model evaluation metrics include the coefficient of determination. Root mean square error (RMSE) and relative root mean square error (rRMSE). Figure 2 As shown, during the tillering, jointing, and heading stages, random forest regression-based methods... The estimation model demonstrated superior test accuracy, with RMSE values ​​of 0.030, 0.045, and 0.044, and rRMSE values ​​of 7.275%, 11.364%, and 9.318%, respectively. This is in contrast to the entropy-weighted method (…). ) and coefficient of variation method ( The CGMI estimation model constructed in this invention is based on the information fusion stabilization method. The estimation model showed higher [fertility rates] in all three fertility periods. The lower RMSE indicates that the method of the present invention improves the stability and estimation accuracy of the comprehensive growth assessment of rice.

[0082] Step Six: Establish Quadratic function relationship with grain yield Determine the optimal value corresponding to the maximum grain yield. The production threshold coefficient is set to 0.95. This is obtained by finding the root of a quadratic function. and The thresholds for weak and excessive growth during the tillering, jointing, and heading stages ( and The values ​​were 0.405 and 0.860; 0.437 and 0.948; and 0.497 and 0.956, respectively. Figure 3 As shown, the crop growth classification based on the yield decline ratio threshold strategy can intuitively reflect different growth levels. The corresponding production level and grade boundary of the interval.

[0083] Step 7: Perform inversion using the threshold. For example... Figure 4 As shown, the heading stage is based on The inverted distribution map of rice comprehensive growth level can clearly identify the spatial differences in growth caused by different nitrogen application treatments. Blue pixels represent weak growth, yellow pixels represent suitable growth, and red pixels represent excessive growth. This shows that the method can identify abnormal growth caused by nitrogen deficiency stress and excessive fertilization, and provide a quantitative basis for field nitrogen management.

[0084] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for comprehensive rice growth monitoring based on UAV hyperspectral remote sensing, characterized in that, Includes the following steps: The drones acquired hyperspectral images of rice during key growth stages, simultaneously collected field growth parameters, and obtained grain yield at maturity. The hyperspectral imagery of the UAV was preprocessed to obtain spectral reflectance data of the rice canopy; The field growth parameters are normalized, pseudo-labels are generated based on the clustering results, the information gain rate and intra-group stability of each growth parameter are calculated, the information gain rate and intra-group stability are fused to determine the weight of each growth parameter, and a comprehensive rice growth monitoring index is constructed based on the normalized growth parameters and the weights. Vegetation indices related to the comprehensive rice growth monitoring indicators were screened, and an estimation model for the comprehensive rice growth monitoring indicators was established based on the screened vegetation indices. Establish a functional relationship between the comprehensive growth monitoring index of rice and grain yield, determine the growth level threshold according to the yield decline ratio threshold strategy, and classify the comprehensive growth of rice into weak growth, suitable growth and excessive growth based on the threshold. The estimation model is used to invert the spectral reflectance data and output a comprehensive growth level distribution map of rice.

2. The method according to claim 1, characterized in that, The key growth periods include the tillering stage, the jointing stage, and the heading stage.

3. The method according to claim 1, characterized in that, The preprocessing includes image segmentation, geometric correction, radiometric correction, georegistration, image stitching, and reflectance conversion.

4. The method according to claim 1, characterized in that, The field growth parameters include plant height, aboveground biomass, plant water content, chlorophyll content, carotenoid content, leaf nitrogen content, and leaf carbon content.

5. The method according to claim 1, characterized in that, The method of determining the weights of each growth parameter by fusing the information gain rate and intra-group stability includes: introducing a fusion factor to perform weighted fusion of the information gain rate and intra-group stability to obtain a fusion strength index, and normalizing the fusion strength index to obtain the fusion weights of each growth parameter.

6. The method according to claim 5, characterized in that, The fusion factor is determined through Bootstrap cross-validation: the clustering consistency index is calculated under different fusion factor values, and the fusion factor corresponding to the maximum value of the clustering consistency index is taken as the optimal fusion factor.

7. The method according to claim 1, characterized in that, The process of screening vegetation indices related to the comprehensive rice growth monitoring index includes: calculating the correlation between the band random combination vegetation index and the comprehensive rice growth monitoring index; screening vegetation indices whose correlation meets preset conditions as candidate vegetation indices; and using a permutation importance algorithm to screen important vegetation indices from the candidate vegetation indices.

8. The method according to claim 1, characterized in that, The estimation model is established using partial least squares regression, random forest regression, or support vector regression algorithms.

9. The method according to claim 1, characterized in that, The functional relationship is a quadratic function relationship; the step of determining the growth level threshold based on the yield decline ratio threshold strategy includes: calculating the theoretical maximum yield based on the quadratic function relationship, multiplying the theoretical maximum yield by the yield decline ratio threshold to obtain the yield threshold, and substituting the yield threshold into the quadratic function relationship to find the root, thereby obtaining the lower limit and upper limit of the growth level.

10. The method according to claim 9, characterized in that, When the comprehensive growth monitoring index of rice is less than the lower limit, it is judged as weak growth; when it is between the lower limit and the upper limit, it is judged as suitable growth; when it is greater than the upper limit, it is judged as excessive growth. During the tillering stage, the lower limit is 0.405 and the upper limit is 0.860; During the jointing stage, the lower limit is 0.437 and the upper limit is 0.948; During the heading stage, the lower limit is 0.497 and the upper limit is 0.956.

Images