A winter wheat LAI estimation method and system based on spectral feature clustering screening
By employing multi-dimensional correlation evaluation and hierarchical clustering to remove redundancy, a hyperspectral remote sensing model was constructed. This model addresses the issues of single and redundant feature selection in winter wheat LAI monitoring, achieving high-precision and stable LAI estimation and growth diagnosis. It is suitable for UAV hyperspectral remote sensing monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANDONG UNIV OF SCI & TECH
- Filing Date
- 2026-06-08
- Publication Date
- 2026-07-14
AI Technical Summary
Existing methods for monitoring the vegetation index (LAI) of winter wheat suffer from destructive sampling, cumbersome operation, and low efficiency. Multispectral remote sensing has low resolution and insufficient sensitivity to vegetation index responses. Hyperspectral remote sensing vegetation index screening methods are limited, suffer from severe multicollinearity, and have weak model generalization ability, making it difficult to achieve large-area real-time dynamic monitoring and high-precision estimation.
A vegetation index set was constructed using a multi-dimensional correlation evaluation system. The Pearson correlation coefficient, variable projection importance, random forest feature importance, and variable contribution were used for evaluation. Combined with hierarchical clustering to remove redundancy, a partial least squares regression model was constructed. Combined with UAV hyperspectral data and ground-based measured data, high-precision remote sensing inversion and growth diagnosis of LAI were achieved.
It achieves high-precision and stable remote sensing inversion and growth diagnosis of winter wheat LAI, improves monitoring efficiency and accuracy, reduces costs, and is suitable for UAV hyperspectral remote sensing monitoring, meeting the needs of precision agriculture.
Smart Images
Figure CN122391894A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of agricultural information technology and remote sensing monitoring technology, and relates to a method and system for estimating the leaf area index (LAI) of winter wheat by spectral feature clustering screening, and particularly to a hyperspectral remote sensing estimation method for winter wheat leaf area index (LAI) by spectral feature clustering screening. Background Technology
[0002] Winter wheat is an important food crop in my country. Leaf area index (LAI) is a core agronomic parameter characterizing crop growth status, photosynthetic efficiency, and yield potential. Accurate, rapid, and non-destructive acquisition of winter wheat LAI is of great significance for growth monitoring and yield prediction. Currently, winter wheat LAI monitoring mainly includes three types of methods: traditional measurement, multispectral remote sensing estimation, and hyperspectral remote sensing estimation. All of these methods have significant limitations in practical applications.
[0003] Traditional LAI measurements have high accuracy, but they require destructive sampling, are cumbersome to operate, and are inefficient. They cannot achieve large-area real-time dynamic monitoring and cannot meet the needs of efficient monitoring in modern agriculture.
[0004] Multispectral remote sensing data is convenient to acquire and has a wide coverage, but its spectral resolution is low, and vegetation indices are not sensitive enough to LAI (Labour Area Index), resulting in limited estimation accuracy. For example, the wheat LAI estimation method disclosed in existing patent CN116593419B only uses conventional multispectral vegetation indices to construct a linear model, which makes it difficult to distinguish the differences in growth at different growth stages, resulting in poor dynamic estimation of LAI.
[0005] Hyperspectral remote sensing offers abundant and high-resolution information, making it the mainstream approach for accurate LAI estimation. However, existing technologies have significant shortcomings: First, the vegetation index selection methods are simplistic, often relying solely on Pearson correlation analysis without constructing a multi-dimensional evaluation system or removing redundant features, leading to multicollinearity. For example, the wheat LAI estimation method disclosed in patent CN113159420A uses only a single indicator for feature selection without redundancy removal, resulting in weak model generalization ability. Second, the modeling methods have flaws. Linear models struggle to fit nonlinear relationships, and some machine learning models are prone to overfitting due to unreasonable feature selection, exhibiting poor spatiotemporal transferability. For instance, the remote sensing monitoring method disclosed in patent CN103942459A, while employing machine learning modeling, fails to remove redundancy and optimize the vegetation index system, resulting in insufficient model robustness.
[0006] Based on the above analysis, the problems and shortcomings of the existing technology are as follows: (1) Traditional winter wheat LAI measurement methods require destructive sampling, are cumbersome to operate, and are inefficient. They cannot achieve large-scale real-time dynamic monitoring and are difficult to meet the needs of efficient monitoring in modern agriculture.
[0007] (2) Existing multispectral remote sensing methods for winter wheat LAI have low spectral resolution, insufficient sensitivity of vegetation indices to LAI response, and limited estimation accuracy; and only conventional multispectral vegetation indices are used to construct linear models, making it difficult to distinguish growth differences at different growth stages (e.g., CN116593419B).
[0008] (3) The existing vegetation index screening methods of winter wheat LAI hyperspectral remote sensing methods are simple, mostly using only Pearson correlation analysis, without constructing a multi-dimensional evaluation system, and without removing redundant features, which easily leads to multicollinearity (e.g., the generalization ability of CN113159420A model is weak); linear models are difficult to fit nonlinear relationships, and some machine learning models are prone to overfitting due to unreasonable feature selection, with poor spatiotemporal mobility (e.g., CN103942459A did not remove redundancy and optimize the vegetation index system, and the model is not robust enough).
[0009] In summary, existing wheat LAI remote sensing estimation techniques suffer from problems such as simplistic feature selection methods, severe information redundancy, weak model generalization ability, and insufficient estimation accuracy, making it difficult to meet the needs of modern agriculture for precise monitoring of winter wheat growth dynamics. This application addresses the shortcomings of existing technologies by proposing a winter wheat LAI estimation method based on spectral feature clustering. Its core improvements are reflected in: (1) Improved feature selection method: Breaking through the limitations of traditional single Pearson correlation analysis, a multi-dimensional correlation evaluation system is constructed. At the same time, hierarchical clustering is introduced to remove redundancy from vegetation indices, effectively solving the multicollinearity problem and improving the scientificity and effectiveness of feature selection. (2) Improved model robustness: By combining features after clustering to remove redundancy, the redundancy of the model input is reduced, the generalization ability and spatiotemporal transferability of the model are improved, and the problems of overfitting and insufficient robustness of existing machine learning models are solved. (3) Optimization of estimation accuracy: Based on the combination of UAV hyperspectral data and ground measured data, a reproductive period allocation model is constructed to realize dynamic estimation and growth diagnosis of LAI, thereby improving the accuracy and applicability of LAI estimation at different reproductive periods.
[0010] This invention, through the aforementioned improvements, effectively overcomes the shortcomings of existing technologies, such as single feature selection, information redundancy, weak model generalization ability, and insufficient estimation accuracy, providing new technical support for precise monitoring of winter wheat growth in modern agriculture. Therefore, it is urgent to propose a new hyperspectral remote sensing estimation method for winter wheat LAI to address the aforementioned technical bottlenecks. Summary of the Invention
[0011] To overcome the problems of single vegetation index screening methods, feature redundancy, poor model generalization, and insufficient LAI estimation accuracy in related technologies, this invention discloses a method and system for estimating winter wheat LAI using spectral feature clustering screening. The technical solution is as follows: This invention is implemented as follows: A method for estimating the LAI (Label Area Index) of winter wheat using spectral feature clustering includes the following steps: S1. Based on hyperspectral remote sensing data of winter wheat, construct a set of vegetation indices that are highly correlated with the leaf LAI of winter wheat; S2. A multi-method correlation evaluation system is used to perform multi-dimensional correlation quantification calculation on the vegetation index set to obtain the four evaluation index values corresponding to each vegetation index: Pearson correlation coefficient, variable projection importance, random forest feature importance, and variable contribution. S3. Based on the four correlation evaluation index values, the vegetation index is standardized within the class, and the hierarchical clustering algorithm is used to perform clustering redundancy removal and feature optimization. Small clusters are merged, and the vegetation index closest to the cluster center is selected as the representative in each cluster to obtain the optimal combination of vegetation indices. S4. Divide the sample data into a modeling set and a validation set according to the proportion. With the optimal combination of vegetation indices as the independent variable and the measured value of winter wheat leaf LAI as the dependent variable, use cross-validation to select the model components, construct a partial least squares regression estimation model, and realize the hyperspectral remote sensing inversion of winter wheat leaf LAI based on the trained model. S5. Based on the measured LAI data of winter wheat throughout its entire growth period, an LAI level allocation model was constructed. The LAI remote sensing inversion results were compared and judged with the corresponding growth period thresholds to complete the diagnosis of winter wheat LAI growth.
[0012] In step S1, the winter wheat remote sensing data is remote sensing data acquired by a hyperspectral sensor carried by an unmanned aerial vehicle, and the vegetation index set contains 20-50 hyperspectral vegetation indices related to winter wheat LAI.
[0013] In step S2, the multi-method correlation evaluation system includes four methods: Pearson correlation coefficient analysis, variable projection importance analysis, random forest feature importance assessment, and variable contribution assessment, which quantify the characterization ability of each vegetation index on winter wheat LAI from multiple dimensions.
[0014] In step S2, the random forest feature importance assessment includes: constructing a random forest regression model, setting the number of decision trees to 500, enabling out-of-bag error verification, selecting the maximum feature using the square root strategy, and calculating the feature importance score of each vegetation index based on the permutation method; The expression for feature importance scoring is: ; In the formula, For feature importance scoring, This represents the total number of samples. Let be the mean square error value corresponding to the i-th feature in the m-th sample group; The variable contribution assessment employs a combination of non-information variable elimination and partial least squares method to calculate the variable contribution coefficient; the expression for variable contribution is: ; In the formula, Let be the variable contribution coefficient of the j-th independent variable. This represents the average regression coefficient of the j-th variable in partial least squares modeling. Let be the standard deviation of the regression coefficient of the j-th variable; The expression for Pearson correlation coefficient analysis is: ; In the formula, The Pearson correlation coefficient is used. For the two sets of research variable sequences to be analyzed, Let X be the variance of the independent variable. Let Y be the variance of the dependent variable. The expression for variable projection importance analysis is: ; In the formula, Let j be the variable projection importance value of the j-th variable. The total number of all independent variables involved in the modeling. The linear correlation coefficient, The dependent variable y and the h-th principal component The coefficient of determination on Let be the weight coefficient of the j-th variable on the h-th principal component. This represents the total number of principal components extracted by the partial least squares model. Principal component number.
[0015] In step S3, vegetation indices are hierarchically clustered and redundancy-removing optimized based on four correlation evaluation index values, dividing the vegetation indices into several clusters; within each cluster, the vegetation index closest to the cluster center is selected as the representative of the cluster, thereby achieving efficient redundancy removal and feature optimization of vegetation indices.
[0016] In step S3, the condition for merging and optimizing the small clusters in the clustering results is: merging clusters with fewer than 3 samples after clustering into the nearest neighboring cluster.
[0017] In step S3, the optimal combination of vegetation indices includes: modified normalized vegetation index, vegetation pigment ratio index, normalized red edge index, and optimal soil-adjusted vegetation index.
[0018] In step S4, the measured value of winter wheat LAI is obtained by synchronous measurement on the ground using a canopy analyzer.
[0019] In step S4, when constructing the machine learning estimation model, 10-fold cross-validation is used to optimize the number of principal components of the model, and the sample dataset is divided into a modeling set and a validation set in a 7:3 ratio.
[0020] In step S5, the LAI allocation model is constructed based on the measured LAI data of winter wheat throughout its entire growth period, and includes LAI level thresholds for each growth period; the growth diagnosis determines the winter wheat growth as weak, suitable, or vigorous by comparing the inverted LAI with the corresponding growth period thresholds.
[0021] In this embodiment, based on statistical analysis of measured LAI data from key growth stages of winter wheat in the experimental area, the LAI thresholds for each growth stage were determined as follows: Greening stage: weak threshold ≤ 1.5, suitable threshold 1.5~3.0, excessive threshold > 3.0; Jointing stage: weak threshold ≤ 2.5, suitable threshold 2.5~4.5, excessive threshold > 4.5; Heading stage: weak threshold ≤ 3.5, suitable threshold 3.5~5.5, excessive threshold > 5.5; Grain-filling stage: weak threshold ≤ 3.0, suitable threshold 3.0~5.0, excessive threshold > 5.0. The above thresholds represent the preferred ranges for this embodiment and can be adaptively adjusted according to the actual growth conditions of different regions and varieties of winter wheat.
[0022] Another objective of this invention is to provide a hyperspectral remote sensing estimation system for winter wheat LAI (Label Intake) based on spectral feature clustering. This system is used to implement the aforementioned method for estimating winter wheat LAI based on spectral feature clustering. The system includes: The vegetation index construction module is used to construct a set of vegetation indices related to the winter wheat LAI based on winter wheat remote sensing data. The multi-dimensional evaluation module is used to perform multi-dimensional correlation quantification calculation on the vegetation index set using a multi-method correlation evaluation system that includes four methods: Pearson correlation coefficient analysis, variable projection importance analysis, random forest feature importance assessment, and variable contribution assessment, to obtain the four correlation evaluation index values corresponding to each vegetation index. The clustering optimization module is used to perform intra-cluster standardization of vegetation indices based on four correlation evaluation index values, and to perform hierarchical clustering using cosine distance combined with Ward's method. Small clusters in the clustering results are merged and optimized. In each final cluster, the vegetation index closest to the cluster center is selected as the cluster representative index to obtain the optimal combination of vegetation indices. The model inversion module is used to construct a partial least squares regression estimation model by using the optimal combination of vegetation indices as input features and the measured value of winter wheat LAI as output variables, and to output the remote sensing inversion result of winter wheat LAI based on the trained model. The growth diagnosis module is used to construct a growth period allocation model based on measured LAI data of key growth periods of winter wheat, match the LAI results obtained by inversion to the corresponding growth period allocation model, determine the LAI level by threshold comparison, and complete the LAI growth diagnosis of winter wheat.
[0023] Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, the hyperspectral remote sensing estimation method for winter wheat leaf area index (LAI) provided by this invention constructs a vegetation index set based on winter wheat remote sensing data. It then builds a multi-dimensional correlation evaluation system using various methods and calculates a comprehensive score. Cluster analysis is combined to remove redundancy and optimize features in the vegetation indices. A machine learning model is constructed using the optimal combination of vegetation numbers as input features to achieve high-precision remote sensing inversion and growth diagnosis of winter wheat LAI. This invention solves the problems of unscientific vegetation index selection, feature redundancy, and poor model generalization ability in existing technologies, enabling rapid, lossless, and large-area monitoring of winter wheat growth information.
[0024] This invention achieves quantitative characterization of vegetation indices through a multi-dimensional correlation evaluation system. It abandons the shortcomings of traditional methods that rely on single-indicator screening or subjective weighting to construct comprehensive scores for re-clustering. Instead, it directly uses multi-indicator raw data for hierarchical clustering to efficiently eliminate redundant features. A partial least squares regression model is used to fit the mapping relationship between features and the leaf area index (LAI), ultimately achieving high-precision and high-stability hyperspectral remote sensing inversion and growth diagnosis of winter wheat leaf LAI. Compared to traditional single correlation analysis screening methods, the model prediction accuracy is improved from 0.62 to 0.7, an increase of 10%-15%, with significantly enhanced stability and generalization ability. This ultimately achieves high-precision and high-stability hyperspectral remote sensing inversion and growth diagnosis of winter wheat leaf LAI.
[0025] Secondly, this invention constructs a multi-method correlation evaluation system, which comprehensively characterizes the degree of association between vegetation indices and leaf LAI from multiple dimensions such as correlation, variable importance, and feature contribution. Compared with the traditional method that only uses a single correlation index, it is more comprehensive and objective, and can effectively screen out characteristic vegetation indices that are sensitive to changes in LAI.
[0026] This invention directly performs hierarchical clustering to remove redundancy based on multi-dimensional evaluation indicators, avoiding the problems of manually setting weights and information compression loss in traditional methods. It can effectively reduce multicollinearity among features, optimize the model input structure, improve feature utilization efficiency, reduce the risk of model overfitting, and provide a more concise and effective feature set for subsequent modeling.
[0027] This invention employs a partial least squares regression model combined with cross-validation for modeling, which can better fit the mapping relationship between vegetation index and LAI. Experimental verification shows that the inversion model constructed by this method has higher accuracy and stronger stability. Compared with traditional feature selection methods, the model's prediction effect and generalization ability are significantly improved, and it shows good adaptability on winter wheat data in different growth stages and regions.
[0028] This invention enables non-destructive, rapid, and large-scale hyperspectral remote sensing estimation of the leaf light intensity (LAI) of winter wheat. The inversion accuracy can meet the actual needs of precision agriculture field growth monitoring and can provide data support for winter wheat growth diagnosis and field management.
[0029] Third, this invention can be directly applied to UAV hyperspectral remote sensing for monitoring the growth of winter wheat, eliminating the need for destructive manual sampling, significantly reducing field monitoring costs and manpower input, while improving monitoring efficiency and timeliness, providing low-cost, high-precision technical support for large-scale precision agricultural production. This invention overcomes the limitations of traditional single-indicator screening, constructing a multi-dimensional evaluation and clustering-based vegetation index optimization framework, solving the common industry problems of feature redundancy and insufficient model generalization ability in existing hyperspectral LAI estimation, and providing a new technical approach for crop hyperspectral remote sensing feature screening.
[0030] This invention solves the problems of insufficient model accuracy and stability caused by strong subjectivity in vegetation index selection and feature redundancy in the prior art, and achieves simultaneous improvement in LAI estimation accuracy and model robustness. Attached Figure Description
[0031] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the disclosure of this invention and, together with the description, serve to explain the principles of the disclosure of this invention. Figure 1 This is a flowchart of a method for estimating the LAI of winter wheat using spectral feature clustering, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the LAI inversion results provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the LAI inversion results for winter wheat using the traditional single-index VIP screening method provided in this embodiment of the invention; Figure 4 This is a schematic diagram of the LAI inversion results for winter wheat using the traditional single-index EVC screening method provided in this embodiment of the invention. Detailed Implementation
[0032] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0033] The innovation of this invention lies in its breakthrough from the limitations of traditional single-index screening, constructing an integrated framework of multi-dimensional evaluation, clustering for redundancy removal, and group modeling. By optimizing bands through multi-dimensional evaluation and then removing redundancy based on cosine distance clustering, multicollinearity is reduced. The group modeling strategy addresses the problems of overfitting and weak generalization ability inherent in traditional methods, significantly improving the accuracy and stability of winter wheat LAI hyperspectral inversion.
[0034] Example 1, as Figure 1 As shown in the figure, the method for estimating the LAI of winter wheat by spectral feature clustering screening provided in this embodiment of the invention includes the following steps: S1. Based on winter wheat remote sensing data, construct a set of vegetation indices related to winter wheat LAI; We acquired hyperspectral remote sensing data on key growth stages of winter wheat, including greening, jointing, heading, and grain filling. Based on existing research, we selected vegetation indices that are sensitive to leaf area index (LAI) and constructed a hyperspectral vegetation index set. This embodiment calculates a set of vegetation indices highly correlated with LAI based on the band reflectance of winter wheat hyperspectral remote sensing images. Specifically, these indices include: MTCI, TGI, NDDA, SAVI, RSI, NDRE, NDSI, REP, GM2, VOG, DCNI, VIopt, NDVIg-b, RVIinf-r, OSAVI, MSR, GI, NDREI, NDCI, PPR, RENDVI, Readone, RVI I, RVI II, RVI III, VogelmannREI, SR705, sLAIDI, DVI, DDn, CIgreen, GOSAVI, ARVI1, ARVI2, VARIred, MACI, GMI, ARI, CCRI, CIred-edge, NDMI, NDTI, CAI, TTVI, REIP, NNDVI, DATT, CI, WBI, and WBI / NDVI. To determine the appropriate size of the initial vegetation index set, this invention conducted multiple comparative experiments with different numbers of initial vegetation indices. The initial number of indices was set to 10, 15, 20, 30, 40, 50, and 55, respectively. Under the same multi-dimensional evaluation and clustering redundancy removal process, the impact of different initial index set sizes on the final model accuracy was compared. When the number of indices was less than 20, the initial features were insufficient, and the number of usable features was further reduced after clustering redundancy removal. The model R² was generally below 0.60, resulting in insufficient prediction accuracy. When the number of indices was greater than 50, the initial feature redundancy increased significantly, and the multicollinearity problem became severe, which not only increased the computational cost but also easily led to model overfitting and decreased stability. When the number of initial vegetation indices was in the range of 20-50, it could ensure the diversity and information coverage of the initial features while effectively avoiding severe feature redundancy problems. After multi-dimensional evaluation and clustering redundancy removal, the model R² reached 0.7, and the RMSE stabilized below 0.8, achieving optimal levels of prediction accuracy and stability. Therefore, the present invention preferably includes an initial vegetation index set containing 20-50 hyperspectral vegetation indices related to winter wheat LAI.
[0035] S2. A multi-method correlation evaluation system is used to perform multi-dimensional correlation quantification calculation on the vegetation index set. The multi-method correlation evaluation system includes four methods: Pearson correlation coefficient analysis, variable projection importance analysis, random forest feature importance assessment, and variable contribution assessment. The values of four correlation evaluation indicators corresponding to each vegetation index are calculated. A multi-method correlation evaluation system was adopted to perform multi-dimensional quantitative calculations on the vegetation index set, obtaining four correlation evaluation index values for each vegetation index. This method abandons the traditional subjective weighted summation method for calculating the comprehensive score, and fully preserves the multi-dimensional feature information. The specific calculation formulas for each evaluation index are shown in Table 1. This embodiment employs four evaluation methods to comprehensively characterize the correlation between various vegetation indices and the leaf LAI of winter wheat: a) Pearson correlation analysis: calculating the Pearson correlation coefficient |r| between vegetation indices and leaf LAI to characterize the degree of linear correlation between the two; b) Variable projection importance analysis: calculating the variable projection importance (VIP) value of each vegetation index to characterize its explanatory contribution to LAI; c) Random forest feature importance assessment: constructing a random forest regression model, setting the number of decision trees to 500, enabling out-of-bag (OOB) validation, using the square root strategy for selecting the largest feature, and fixing the random seed to 42; calculating the feature importance score (FIM) of each vegetation index based on the permutation method, and characterizing the importance of vegetation indices by the increase in model OOB error; d) Elimination of uninformative variables combined with variable contribution assessment: using the UVE-PLS method to calculate the variable contribution coefficient (EVC) to achieve effective feature identification and contribution quantification.
[0036] Table 1 Calculation Formulas for Evaluation Indicators
[0037] S3. Based on the four correlation evaluation index values, the vegetation index is standardized within the class, and hierarchical clustering is performed using cosine distance combined with Ward's method. Small clusters in the clustering results are merged and optimized. In each final cluster, the vegetation index closest to the cluster center is selected as the cluster representative index. This completes the clustering and redundancy removal and feature optimization of the vegetation index set, and obtains the optimal combination of vegetation indices. Based on the above four correlation evaluation indicators, vegetation indices are standardized within each cluster to eliminate dimensional differences. Cosine distance is used to calculate the similarity between indices, and hierarchical clustering is performed using the Ward method. Small clusters with fewer than three samples are merged for optimization to ensure clustering effectiveness. In each final cluster, the vegetation index closest to the cluster center is selected as the representative index within the cluster, completing feature redundancy removal and optimal vegetation index combination selection. In this embodiment, four vegetation index clusters are divided after hierarchical clustering, and the optimal vegetation index combination obtained in the final selection is: NNDVI, PPR, NDREE, and OSAVI. To reasonably determine the merging threshold for small clusters, this invention sets up multiple sets of comparative analyses for the clustering results: evaluating the clustering effect according to four methods: not merging small clusters, merging clusters with fewer than two samples, merging clusters with fewer than three samples, and merging clusters with fewer than four samples. If only clusters with fewer than 2 samples are merged, many scattered small clusters containing only 2 samples will remain, resulting in severe fragmentation of the clustering results and an increase in redundant features. If clusters with fewer than 4 samples are merged, effective feature clusters containing 3 samples will be forcibly merged, causing excessive merging of similar spectral features and reducing feature discrimination. However, when clusters with fewer than 3 samples are merged, scattered and invalid small clusters containing only 1 or 2 samples can be effectively removed, avoiding fragmentation of the clustering results, while retaining representative effective feature clusters. This optimizes the similarity within clusters and the discrimination between clusters, and achieves the highest correlation between the vegetation index within clusters and the winter wheat LAI. Therefore, after comprehensive comparison, this invention selects a sample size of less than 3 as the threshold for optimizing the merging of small clusters.
[0038] The vegetation indices related to LAI in this invention are shown in Tables 2, 3, and 4.
[0039] Table 2
[0040] Table 3
[0041] Table 4
[0042] S4. Divide the sample dataset into a modeling set and a validation set according to the proportion. Use the optimal combination of vegetation indices as the input features and the measured value of winter wheat LAI as the output variable. Use cross-validation to select the optimal components of the model and construct a partial least squares regression estimation model. Complete the model training through the modeling set, and then input the vegetation index data of the validation set into the trained model to output the remote sensing inversion result of winter wheat LAI. S5. Based on the measured LAI data of winter wheat during the key growth stages of greening, jointing, heading, and grain filling, construct a growth stage allocation model and determine the LAI abundance or deficiency thresholds for each growth stage; match the LAI results obtained from step S4 to the corresponding growth stage allocation model, determine the LAI level by threshold comparison, and complete the winter wheat LAI growth diagnosis.
[0043] Using the optimal combination of vegetation indices obtained through screening as the independent input variable of the model, and the measured LAI value of winter wheat leaves as the dependent output variable, a partial least squares (PLS) regression estimation model is constructed to achieve hyperspectral remote sensing inversion of winter wheat LAI. The LAI inversion results of this invention are as follows: Figure 2 As shown.
[0044] In this embodiment, the technical features of the input optimized model include: Multiple representative vegetation indices were obtained after redundancy removal through hierarchical clustering. Each vegetation index has been screened through a multi-method correlation evaluation system and has the characteristics of strong correlation with LAI, complementary information, and low redundancy among them. At the same time, the input features have been standardized within the class, eliminating the influence of differences in the scale of different indicators, and the feature structure is more stable.
[0045] The model building process is as follows: The sample dataset is divided into a modeling set and a validation set in a 7:3 ratio. 10-fold cross-validation is used to select the optimal number of principal components in the model to avoid overfitting. The model is trained using the modeling set and the accuracy is evaluated using the validation set.
[0046] The optimal vegetation number combination calculated from the winter wheat hyperspectral data of the area to be inverted is input into the trained PLS model. The model performs prediction calculations based on the learned mapping relationship and directly outputs the winter wheat LAI inversion results, completing a large-scale, lossless, and rapid remote sensing inversion.
[0047] Example 2, the winter wheat LAI hyperspectral remote sensing estimation system based on spectral feature clustering screening provided in this embodiment of the invention includes: The vegetation index construction module is used to construct a set of vegetation indices related to the winter wheat LAI based on winter wheat remote sensing data. The multi-dimensional evaluation module is used to perform multi-dimensional correlation quantification calculation on the vegetation index set using a multi-method correlation evaluation system that includes four methods: Pearson correlation coefficient analysis, variable projection importance analysis, random forest feature importance assessment, and variable contribution assessment, to obtain the four correlation evaluation index values corresponding to each vegetation index. The clustering optimization module is used to perform intra-cluster standardization of vegetation indices based on four correlation evaluation index values, and to perform hierarchical clustering using cosine distance combined with Ward's method. Small clusters in the clustering results are merged and optimized. In each final cluster, the vegetation index closest to the cluster center is selected as the cluster representative index to obtain the optimal combination of vegetation indices. The model inversion module is used to construct a partial least squares regression estimation model by using the optimal combination of vegetation indices as input features and the measured value of winter wheat LAI as output variables, and to output the remote sensing inversion result of winter wheat LAI based on the trained model. The growth diagnosis module is used to construct a growth period allocation model based on measured LAI data of key growth periods of winter wheat, match the LAI results obtained by inversion to the corresponding growth period allocation model, determine the LAI level by threshold comparison, and complete the LAI growth diagnosis of winter wheat.
[0048] The embodiments of this invention have achieved significant positive results in research and development and use, demonstrating outstanding advantages compared to existing technologies. Theoretically, existing hyperspectral inversion methods often involve direct full-band modeling, which suffers from high dimensionality, redundancy, overfitting, and poor generalization. This invention employs a combination of clustering feature optimization and partial least squares regression, effectively reducing information redundancy, mitigating multicollinearity, and improving model stability and estimation accuracy. This invention is designed and verified by referencing similar mature experimental paradigms. Through steps such as experimental area deployment, acquisition of UAV hyperspectral data and ground-based measurements, and model accuracy comparison, the technical advantages are verified using indicators such as R² and RMSE. Further intuitive demonstrations can be achieved using experimental data and charts.
[0049] To further demonstrate the positive effects of the above embodiments, the present invention conducts the following experiments based on the above technical solutions.
[0050] 1. Overview of the Experimental Area; This example uses the Xiaotangshan Winter Wheat Experimental Area in Beijing as the experimental area. The experimental area is 85m long from east to west and 70m long from north to south, with a total area of approximately 5950m². 2 A total of 32 experimental plots were set up in the experimental area, each with an area of 10m×9m. The plots were evenly distributed and could well reflect the differences in winter wheat growth in the region.
[0051] 2. Data acquisition methods; Remote sensing data: Using an unmanned aerial vehicle (UAV) hyperspectral imaging system, hyperspectral remote sensing images of the experimental area were acquired during the four key growth stages of winter wheat: greening, jointing, heading, and grain filling. Ground-based measured data: The LAI-2200 canopy analyzer was used to conduct LAI measurements on each of the 32 experimental plots at each growth stage. The measured values were used as reference values for model training and accuracy verification.
[0052] 3. Comparative Experiment Setup: To demonstrate the technical improvement effect of this invention, two existing conventional methods were set up as controls, such as... Figure 3 , Figure 4 As shown, the comparison results are shown in Table 5: Comparison Group A: The estimation model was directly constructed using only a single vegetation index such as NDVI, without conducting multi-index evaluation and feature screening; Comparison Group B: Only the Pearson correlation coefficient was used for vegetation index screening, without multi-dimensional evaluation or clustering to remove redundancy.
[0053] Table 5 Comparison of winter wheat inversion accuracy between the present invention and traditional feature screening methods
[0054] 4. Experimental results and positive significance; Experimental results show that the method of the present invention has significant advantages over control groups A and B: (1) The multi-dimensional correlation evaluation system is adopted, which provides a more comprehensive feature representation and avoids the problem of missing sensitive features caused by screening with a single indicator; (2) Hierarchical clustering is performed directly based on the four indicators, eliminating the need for subjective weighting and comprehensive score calculation, reducing human interference, and making the screening results more objective; (3) After clustering to remove redundancy, the multicollinearity among features is significantly reduced, the model is more stable, and the inversion results are more consistent with the field measurements. (4) It can quickly, non-destructively, and over a wide area realize LAI remote sensing monitoring of winter wheat. The inversion accuracy meets the needs of precision agriculture growth monitoring and can provide reliable data support for winter wheat field management, growth diagnosis and water and fertilizer regulation in Xiaotangshan Experimental Area.
[0055] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.
Claims
1. A method for estimating the LAI (Label Area Index) of winter wheat using spectral feature clustering, characterized in that, The method includes the following steps: S1. Based on winter wheat remote sensing data, construct a set of vegetation indices related to winter wheat LAI; S2. A multi-method correlation evaluation system is used to perform multi-dimensional correlation quantification calculation on the vegetation index set. The multi-method correlation evaluation system includes four methods: Pearson correlation coefficient analysis, variable projection importance analysis, random forest feature importance assessment, and variable contribution assessment. The values of four correlation evaluation indicators corresponding to each vegetation index are calculated. S3. Based on the four correlation evaluation index values, the vegetation index is standardized within the class, and hierarchical clustering is performed using cosine distance combined with Ward's method. Small clusters in the clustering results are merged and optimized. In each final cluster, the vegetation index closest to the cluster center is selected as the cluster representative index. This completes the clustering and redundancy removal and feature optimization of the vegetation index set, and obtains the optimal combination of vegetation indices. S4. Divide the sample dataset into a modeling set and a validation set according to the proportion. Use the optimal combination of vegetation indices as the input features and the measured value of winter wheat LAI as the output variable. Use cross-validation to select the optimal components of the model and construct a partial least squares regression estimation model. Complete the model training through the modeling set, and then input the vegetation index data of the validation set into the trained model to output the remote sensing inversion result of winter wheat LAI. S5. Based on the measured LAI data of winter wheat during the key growth stages of greening, jointing, heading, and grain filling, a growth stage allocation model is constructed. This model is used to classify the LAI level evaluation criteria corresponding to different growth stages and determine the LAI abundance or deficiency threshold for each growth stage. The LAI results obtained from step S4 are matched to the corresponding growth stage allocation model, and the LAI level is determined by threshold comparison to complete the diagnosis of winter wheat LAI growth.
2. The method for estimating the LAI of winter wheat by spectral feature clustering screening according to claim 1, characterized in that, In step S2, the random forest feature importance assessment includes: constructing a random forest regression model, setting the number of decision trees to 500, enabling out-of-bag error verification, selecting the maximum feature using the square root strategy, and calculating the feature importance score of each vegetation index based on the permutation method; The expression for feature importance scoring is: ； In the formula, For feature importance scoring, This represents the total number of samples. Let be the mean square error value corresponding to the i-th feature in the m-th sample group; The variable contribution assessment employs a combination of non-information variable elimination and partial least squares method to calculate the variable contribution coefficient; the expression for variable contribution is: ； In the formula, Let be the variable contribution coefficient of the j-th independent variable. This represents the average regression coefficient of the j-th variable in partial least squares modeling. Let be the standard deviation of the regression coefficient of the j-th variable; The expression for Pearson correlation coefficient analysis is: ； In the formula, The Pearson correlation coefficient is used. For the two sets of research variable sequences to be analyzed, Let X be the variance of the independent variable. Let Y be the variance of the dependent variable. The expression for variable projection importance analysis is: ； In the formula, Let j be the variable projection importance value of the j-th variable. The total number of all independent variables involved in the modeling. The linear correlation coefficient, The dependent variable y and the h-th principal component The coefficient of determination on Let be the weight coefficient of the j-th variable on the h-th principal component. This represents the total number of principal components extracted by the partial least squares model. Principal component number.
3. The method for estimating the LAI of winter wheat by spectral feature clustering screening according to claim 1, characterized in that, In step S3, the condition for merging and optimizing the small clusters in the clustering results is: merging clusters with fewer than 3 samples after clustering into the nearest neighboring cluster.
4. The method for estimating the LAI of winter wheat by spectral feature clustering screening according to claim 3, characterized in that, The optimal vegetation index combination is: after multi-dimensional correlation evaluation and hierarchical clustering to remove redundancy from the initial vegetation index set, the representative index set that has the highest correlation with winter wheat LAI and the lowest redundancy is selected from each index cluster.
5. The method for estimating the LAI of winter wheat by spectral feature clustering screening according to claim 1, characterized in that, In step S5, the LAI abundance / deficiency threshold includes a weak threshold, a suitable threshold, and a strong threshold. By comparing the LAI obtained from the inversion with the abundance / deficiency threshold of the corresponding growth stage, the growth of winter wheat is determined to be one of three levels: weak, suitable, or strong.
6. The method for estimating the LAI of winter wheat by spectral feature clustering screening according to claim 1, characterized in that, In step S1, the winter wheat remote sensing data is remote sensing data acquired by a hyperspectral sensor carried by an unmanned aerial vehicle.
7. The method for estimating the LAI of winter wheat by spectral feature clustering screening according to claim 1, characterized in that, In step S4, the measured value of winter wheat LAI is obtained by synchronous measurement on the ground using a canopy analyzer.
8. The method for estimating the LAI of winter wheat by spectral feature clustering screening according to claim 1, characterized in that, In step S1, the vegetation index set contains 20-50 hyperspectral vegetation indices associated with winter wheat LAI.
9. A hyperspectral remote sensing system for estimating the LAI (Label Intake) of winter wheat using spectral feature clustering screening, characterized in that, This system is used to implement the winter wheat LAI estimation method based on spectral feature clustering screening as described in any one of claims 1 to 8, the system comprising: The vegetation index construction module is used to construct a set of vegetation indices related to the winter wheat LAI based on winter wheat remote sensing data. The multi-dimensional evaluation module is used to perform multi-dimensional correlation quantification calculation on the vegetation index set using a multi-method correlation evaluation system that includes four methods: Pearson correlation coefficient analysis, variable projection importance analysis, random forest feature importance assessment, and variable contribution assessment, to obtain the four correlation evaluation index values corresponding to each vegetation index. The clustering optimization module is used to perform intra-cluster standardization of vegetation indices based on four correlation evaluation index values, and to perform hierarchical clustering using cosine distance combined with Ward's method. Small clusters in the clustering results are merged and optimized. In each final cluster, the vegetation index closest to the cluster center is selected as the cluster representative index to obtain the optimal combination of vegetation indices. The model inversion module is used to construct a partial least squares regression estimation model by using the optimal combination of vegetation indices as input features and the measured value of winter wheat LAI as output variables, and to output the remote sensing inversion result of winter wheat LAI based on the trained model. The growth diagnosis module is used to construct a growth period allocation model based on measured LAI data of key growth periods of winter wheat, match the LAI results obtained by inversion to the corresponding growth period allocation model, determine the LAI level by threshold comparison, and complete the LAI growth diagnosis of winter wheat.