A chlorophyll visualization method and system based on hyperspectral ensemble learning
By using a hyperspectral ensemble learning method to enhance spectral data and weighted merge models, the problems of poor generalization ability and high computational resource consumption in chlorophyll prediction with small sample sizes are solved, achieving high-precision and stable chlorophyll content prediction.
Patent Information
- Application Number
- CN202610308469.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-13
- Publication Date
- 2026-06-23
AI Technical Summary
Existing empirical/statistical models and deep learning methods based on feature bands have poor generalization ability, high computational resource requirements, and are prone to overfitting and have poor stability in small sample cases.
The hyperspectral ensemble learning method is adopted. By adding Gaussian noise and distortion data to one-dimensional spectral data, the data is divided into a modeling set and a prediction set. XGBOOST, PLS and RIDGE models are trained and weighted and merged to generate the optimal weight optimization ensemble model, which determines the region of interest and predicts the chlorophyll content.
It improves the accuracy and stability of chlorophyll prediction in small sample cases, reduces computational resource requirements, and avoids the problems of insufficient generalization ability and overfitting of a single model.
Smart Images

Figure CN122265625A_ABST
Abstract
Description
Technical Field
[0001] This disclosure belongs to the field of instrumentation technology, and in particular relates to a chlorophyll visualization method and system based on hyperspectral integrated learning. Background Technology
[0002] Existing empirical / statistical model methods based on characteristic bands are currently the most widely used approach, directly establishing the statistical relationship between spectral characteristics and chlorophyll content from the data. Vegetation index methods: These methods utilize combinations of two or more bands sensitive to chlorophyll to construct vegetation indices (such as NDVI, TCARI / OSAVI, MTCI, REP, etc.). Then, linear or nonlinear regression is used to establish the relationship between the indices and chlorophyll content. Feature selection + regression methods: These methods select several characteristic bands with the highest correlation to chlorophyll from the full spectrum, and then use multiple linear regression, partial least squares regression, support vector machine regression, etc., to build a predictive model. However, these models suffer from poor generalization ability, insufficient information utilization, and the risk of model singularity.
[0003] Existing deep learning models based on full-band data employ either one-dimensional convolutional neural networks to directly process spectral curves or two-dimensional CNNs to process feature maps derived from the spectrum, automatically extracting deep features for prediction. However, these methods require massive amounts of data, are prone to overfitting with small sample sizes, and suffer from high computational resource requirements and poor stability. Summary of the Invention
[0004] To address the aforementioned technical problems, this disclosure provides a chlorophyll visualization method and system based on hyperspectral ensemble learning. It solves the technical issues of poor generalization ability and model homogeneity risk in existing models, achieving high accuracy even with small samples and requiring low computational resources.
[0005] The purpose of this disclosure is achieved through the following technical solution: This paper discloses a chlorophyll visualization method based on hyperspectral ensemble learning. The method measures chlorophyll physicochemical indices from modeled one-dimensional spectral data and performs Gaussian noise and distortion data augmentation. The data is divided into a modeling set and a prediction set in a 7:3 ratio, ensuring the modeling set includes the prediction set. The XGBOOST, PLS, and RIDGE models are trained and their hyperparameters optimized using the divided modeling set, resulting in three optimal XGBOOST, PLS, and RIDGE models. These models are then weighted and merged to obtain an ensemble model. After iterative optimization, the best weighted ensemble model is obtained. By setting thresholds for normalized vegetation index and reflectance, the location of the region of interest in the hyperspectral data to be measured is determined and its corresponding spectral data is extracted. The optimal weights are input to optimize the ensemble model prediction and complete the model inversion to obtain the chlorophyll content. The predicted results and corresponding spatial locations are saved. On the same hyperspectral image to be measured or its visualization base map, the predicted chlorophyll value of each pixel is displayed according to its spatial coordinates, and a heat map legend is generated.
[0006] Furthermore, the Gaussian noise enhancement of the chlorophyll physicochemical index is ,in: : indicates that the original one-dimensional spectral data is processed at the 1st... The spectral intensity values obtained after applying Gaussian noise enhancement to each band channel; The original spectrum at the 1st Intensity values for each wavelength / wavenumber channel; : A Gaussian distribution with a mean of 0 and a standard deviation of σ(i).
[0007] Furthermore, the distortion enhancement of the chlorophyll physicochemical indicators is divided into two types, including baseline drift enhancement and wavelength / intensity scaling and distortion enhancement. Baseline drift enhancement: , in: : indicates that the original one-dimensional spectral data is processed at the 1st... The spectral intensity values obtained after applying baseline drift enhancement at each band channel; It is a smooth baseline, generated by a low-order polynomial or a smooth spline curve: , K is the order of the polynomial, with a value between 1 and 4. These are the coefficients from random sampling, and N is the spectral length; Wavelength / intensity scaling and distortion enhancement: , in: : indicates that the original one-dimensional spectral data is processed at the 1st... The spectral intensity values obtained after applying wavelength / intensity scaling and distortion enhancement at each band channel; s is the scaling factor, and t is the translation amount, both of which are obtained by random sampling.
[0008] Furthermore, the method for dividing the modeling set and the prediction set is as follows: the spectra corresponding to the enhanced chlorophyll physicochemical indicators are arranged in ascending order. From every 10 spectra, the first 7 spectra are selected and assigned to the modeling set, and the last 3 spectra are assigned to the prediction set. This process is repeated. The chlorophyll content range of the modeling set is 1-100, and the chlorophyll content range of the prediction set is 10-90. That is, the chlorophyll content range of the modeling set includes the chlorophyll content range of the prediction set.
[0009] Furthermore, the training and hyperparameter optimization methods for the XGBOOST, PLS, and RIDGE models are as follows: XGBOOST, PLS, and RIDGE models are constructed using the partitioned modeling set; candidate hyperparameter combinations are set for each model type, and five-fold cross-validation is used to train and predict each group of candidate hyperparameters within the modeling set, with the five-fold average R² as the optimization score for that group of candidate hyperparameters; a Bayesian optimization algorithm is used to iterate 30 times within the preset hyperparameter search space, selecting the candidate hyperparameter group with the largest average R², and retraining on the entire modeling set to obtain the corresponding optimal XGBOOST, PLS, and RIDGE models; wherein, the hyperparameters of the XGBOOST model include the number of trees, the maximum tree depth, the learning rate, and the subsample sampling rate; the hyperparameter of the PLS model is the number of principal factors; and the hyperparameter of the RIDGE model is the regularization coefficient.
[0010] Furthermore, the weight-optimized ensemble model is obtained by weighting the three optimal XGBOOST, PLS and RIDGE models and merging them to output a single prediction result. The loss function of the ensemble model is defined as root mean square error. Five-fold cross-validation and Bayesian optimization are used to iterate 30 times to obtain the optimal weight-optimized ensemble model.
[0011] Furthermore, the location of the region of interest is determined by initially determining the chlorophyll pixel locations of vegetation using the calculated normalized vegetation index (NZI). Chlorophyll pixel regions with NZI values greater than the NZI threshold are then extracted for secondary screening. Chlorophyll pixel locations with reflectance values greater than the 560 nm visible light band reflectance threshold are then selected and extracted to eliminate shadow, background, and non-leaf pixel regions, retaining only the real leaf regions as the region of interest. The NZI threshold is 0.6, and the 560 nm visible light band reflectance threshold is 0.15.
[0012] Furthermore, the normalized vegetation index (NDI) is used to quantify vegetation cover and health status, and is calculated as follows: , Wherein: NDVI represents the Normalized Difference Vegetation Index, NIR is the band closest to 840nm, and RED is the band closest to 668nm.
[0013] Furthermore, the spectral data used for modeling is the average spectrum of the region of interest selected from the hyperspectral image dataset.
[0014] This disclosure provides a chlorophyll visualization system based on hyperspectral ensemble learning, used to perform the method, including: The data augmentation module is used to augment the chlorophyll physicochemical indicators obtained from the one-dimensional spectrum with Gaussian noise and distortion data, and expand the sample size to obtain a sample set. A data partitioning module, connected to the data augmentation module, is used to partition the sample set into a modeling set and a prediction set in a 7:3 ratio, and to make the data distribution of the modeling set cover the prediction set. The model training and optimization module is connected to the data partitioning module and is used to train and optimize the hyperparameters of the XGBOOST model, PLS model and RIDGE model respectively using the modeling set, and generate the optimal XGBOOST model, optimal PLS model and optimal RIDGE model respectively. An ensemble learning module, connected to the model training and optimization module, is used to perform weighted fusion of the outputs of the optimal XGBOOST model, the optimal PLS model, and the optimal RIDGE model, and to determine the weights of each model through an iterative optimization algorithm to generate the best weight-optimized ensemble model. The region of interest extraction module is used to acquire the hyperspectral image to be tested, and determine the location of the region of interest in the hyperspectral image to be tested based on the set normalized vegetation index threshold and reflectance threshold, and extract the spectral data corresponding to the region of interest. The model inversion module is connected to the region of interest extraction module and the ensemble learning module. It is used to input the extracted spectral data into the optimal weighted ensemble model to predict chlorophyll content and obtain chlorophyll prediction values. The visualization module, connected to the model inversion module, is used to overlay and display the chlorophyll prediction value and its corresponding spatial location on the same hyperspectral image to be measured or its visualization base map to generate a heat map legend.
[0015] The beneficial effects of this disclosure are as follows: 1. This disclosure solves the technical problems of poor generalization ability and model singularity risk of existing single models by merging multiple models into an integrated model through weighted merging.
[0016] 2. This disclosure alleviates the technical problems of small samples by using Gaussian noise and distorted data augmentation, and by stratified modeling based on chlorophyll content. High accuracy can be obtained even with small samples, and the requirements for computing resources are low.
[0017] 3. This disclosure determines the location of the region of interest (ROI) of the image under test by setting constraints, removes shadow, background and non-leaf pixel areas, and retains only the real leaf area as the ROI, thus avoiding interference from invalid information. Attached Figure Description
[0018] Figure 1 This is a flowchart illustrating the process disclosed herein.
[0019] Figure 2 This is an iterative graph showing the optimal parameters for the three models in this disclosure.
[0020] Figure 3 This is the weight optimization graph for the integrated model in this disclosure.
[0021] Figure 4 for Figure 3 The curve corresponding to the weight optimization of the integrated model.
[0022] Figure 5 Curves are extracted for the regions of interest in this disclosure.
[0023] Figure 6 for Figure 5 The corresponding region of interest extraction result image.
[0024] Figure 7 for Figure 5 The corresponding model inversion visualization effect of chlorophyll is shown in the figure.
[0025] Figure 8 This is a system block diagram of the present disclosure. Detailed Implementation
[0026] The present disclosure will now be described in detail with reference to the accompanying drawings and embodiments.
[0027] Example: Figure 1 As shown, this disclosure presents a chlorophyll visualization method based on hyperspectral ensemble learning, comprising the following steps: First, chlorophyll physicochemical indices were measured from the one-dimensional spectral data used for modeling, and Gaussian noise and distortion data augmentation were applied to expand the sample size. The data was then divided into a modeling set and a prediction set in a 7:3 ratio, ensuring that the modeling set included the prediction set. Next, the hyperparameters of the XGBOOST, PLS, and RIDGE models were optimized using the divided modeling set, resulting in three optimal XGBOOST, PLS, and RIDGE models. These models were then weighted and merged to obtain an ensemble model. After iterative optimization, the optimal weighted ensemble model was obtained. By setting thresholds for normalized vegetation index and reflectance, the location of the region of interest in the hyperspectral data to be measured is determined and its corresponding spectral data is extracted. The data is then input into the optimal weighted ensemble model for prediction to complete model inversion, obtain chlorophyll content, and save the prediction results and corresponding spatial locations. On the same hyperspectral image to be measured or its visualization base map, the predicted chlorophyll value of each pixel is displayed according to its spatial coordinates, and a heat map legend is generated.
[0028] If there are multiple images to be tested, chlorophyll thermograms should be generated independently for each image.
[0029] The one-dimensional spectral data used for modeling is the average spectrum of the region of interest in the hyperspectral image dataset. By averaging the spectrum within the region of interest, a stable one-dimensional spectral vector representing the overall spectral characteristics of the leaf can be obtained for modeling, thereby reducing the impact of local heterogeneity and random noise on the model. In this example, the chlorophyll physicochemical index is chlorophyll content.
[0030] The Gaussian noise enhancement of the chlorophyll physicochemical indicators is achieved by adding a certain amount of random noise (10% in this example) to the original spectrum to simulate the small errors in chlorophyll content measurement. ,in: : indicates that the original one-dimensional spectral data is processed at the 1st... The spectral intensity values obtained after applying Gaussian noise enhancement to each band channel; The intensity value of the original spectrum in the i-th wavelength / wavenumber channel; : A Gaussian distribution with a mean of 0 and a standard deviation of σ(i).
[0031] The distortion data enhancement involves applying a slight, non-linear distortion to the spectral curve to simulate morphological changes that may occur in the spectrum under natural conditions; this includes baseline drift enhancement, wavelength / intensity scaling, and distortion enhancement; baseline drift enhancement: , in, : indicates that the original one-dimensional spectral data is processed at the 1st... The spectral intensity values obtained after applying baseline drift enhancement at each band channel; It is a smooth baseline, generated by a low-order polynomial or a smooth spline curve: K is the order of the polynomial, with a value between 1 and 4. These are the coefficients from random sampling, and N is the spectral length; The baseline drift is used to simulate overall baseline shift due to instrument background or scattering; this allows the trained model to learn to ignore non-target changes such as "overall rise / fall" and focus more on spectral differences related to chlorophyll, thereby improving robustness across devices / batch: Wavelength / intensity scaling and distortion enhancement: , in, : indicates that the original one-dimensional spectral data is processed at the 1st... The spectral intensity values obtained after applying wavelength / intensity scaling and distortion enhancement at each band channel; s is the scaling factor, and t is the translation amount, both of which are obtained by random sampling; Wavelength / intensity scaling and distortion enhancement are used to perform amplitude scaling, baseline shifting, and slight distortion of the wavelength axis on the original one-dimensional spectral data, resulting in multiplicative bias and slight deformation of the wavelength axis in the enhanced spectral vector. The wavelength / intensity scaling and distortion enhancement can be applied to the original spectrum alone, or combined with Gaussian noise enhancement and baseline drift enhancement, to simulate the overall amplitude scaling caused by exposure changes, distance / angle changes, and changes in leaf surface reflection conditions, as well as slight peak position shifts or local stretching caused by instrument wavelength calibration errors, spectral resampling / stitching errors, and differences in spectral resolution.
[0032] Baseline drift primarily simulates additive systematic bias in the spectrum, while scaling and distortion primarily simulate multiplicative bias and wavelength axis geometric deformation. Together, they cover the most common types of systematic errors during spectral acquisition, thereby improving the model's generalization ability under different acquisition conditions.
[0033] The original spectral data, the data enhanced with Gaussian noise, and the data enhanced with distortion were stacked vertically to form a sample set that was three times the size of the original data.
[0034] The modeling set and prediction set are divided based on the obtained sample set. In this example, the method for dividing the modeling set and prediction set is as follows: the enhanced chlorophyll content range is 1-100, with a total of 100 spectra. They are arranged in ascending order. From every 10 spectra, the first 7 spectra are selected and assigned to the modeling set, and the last 3 spectra are assigned to the prediction set. This process is repeated. The chlorophyll content range of the modeling set is 1-100, and the chlorophyll content range of the prediction set is 10-90. That is, the chlorophyll content range of the modeling set includes the chlorophyll content range of the prediction set, thus improving the accuracy of the chlorophyll content in the prediction set.
[0035] The modeling set is used to learn the mapping relationship between spectral features and chlorophyll physicochemical indicators, that is, to train the parameter fitting of each model, and to evaluate the candidate models in cross-validation and hyperparameter optimization. The parameters are obtained by training and fitting each model.
[0036] The prediction set is used to independently validate the performance of the best weighted ensemble model after training, measuring the generalization ability of the best weighted ensemble model on samples that were not trained, and preventing the best weighted ensemble model from fitting well only to the training samples and failing to fit new samples.
[0037] This disclosure employs a partitioned modeling set to train three complementary models: XGBoost, PLS, and Ridge. The resulting optimal weight-optimized ensemble model possesses the ability to fit nonlinear features (XGBoost model), robust dimensionality reduction modeling capability for high-dimensional collinear spectra (PLS model), and regular suppression capability for noise and multicollinearity (Ridge model), thereby improving the accuracy and stability of chlorophyll prediction.
[0038] The training and hyperparameter optimization methods for the XGBOOST, PLS, and RIDGE models are as follows: XGBOOST, PLS, and RIDGE models are constructed using a partitioned modeling set; candidate hyperparameter combinations are set for each model type, and five-fold cross-validation is used to train and predict each group of candidate hyperparameters within the modeling set, with the five-fold average R² as the optimization score for that group of candidate hyperparameters; a Bayesian optimization algorithm is used to iterate 30 times within a preset hyperparameter search space, selecting the candidate hyperparameter group with the largest average R², and retraining it on the entire modeling set to obtain the corresponding optimal XGBOOST, optimal PLS, and optimal RIDGE models. Figure 2 As shown; the hyperparameters of the XGBOOST model include the number of trees (between 10 and 140), the maximum depth of the trees (between 1 and 10), the learning rate (selected from [0.01, 0.05, 0.1, 0.15]), and the subsample sampling rate (1.0, i.e., using all samples). The hyperparameter of the PLS model is the number of principal factors (ranging from 1 to 70), and the hyperparameter of the RIDGE model is the regularization coefficient (selected from [0.01, 0.1, 1, 10, 100]).
[0039] The average R² score of the model on the training set was calculated using the five-fold cross-validation as an evaluation metric. R² is used to characterize the goodness of fit of the model's predicted values to the actual chlorophyll content. The closer the R² value is to 1, the better the fit. A value close to 0 indicates weak explanatory power. A negative value indicates that the model is inferior to the baseline model that uses the sample mean as the prediction. In this example, the R² values of the three optimized models are all close to 1.
[0040] The XGBOOST (eXtreme Gradient Boosting) model is an ensemble learning algorithm; The PLS (Partial Least Squares) model is a partial least squares method, also known as a regression modeling method. The RIDGE (Ridge Regression) model is a type of ridge regression.
[0041] In this example, Bayesian optimization (a current technique) is used to optimize the weights of the three optimal XGBoost, PLS, and RIDGE models. Specifically, the weighting coefficients are optimized on top of the outputs of the XGBoost, PLS, and RIDGE models, and the weighted results are then combined to produce a single prediction, resulting in an ensemble model. Figure 3 , Figure 4 As shown, the loss function of the ensemble model is defined as the root mean square error. The optimal weight-optimized ensemble model is obtained by iterating 30 times with the goal of minimizing the root mean square error.
[0042] The loss function is the root mean square error (RMSE), and its mathematical formula is as follows: , The number of samples in the dataset. , ,…, For the true value, , ,… These are predicted values.
[0043] The performance of the obtained optimal weighted ensemble model is evaluated using the following method: 1. Modeling set evaluation: R² and root mean square error of the modeling set are calculated directly on the modeling set to reflect the degree of fit of the best weighted ensemble model to the known data.
[0044] 2. Prediction Set Evaluation: Calculate R² and root mean square error on the prediction set that has never participated in training or weight optimization. At the same time, calculate the relative prediction bias, which is the ratio of the standard deviation of the prediction set to the root mean square error of the prediction set. A higher relative prediction bias usually indicates that the best weight-optimized ensemble model has a stronger predictive ability.
[0045] 3. Validation and evaluation: Use cross-validation prediction (cross_val_predict) to perform 3-fold cross-validation prediction on the best weighted ensemble model, and then calculate the R² and root mean square error of the best weighted ensemble model.
[0046] With R² close to 1 and the root mean square error minimized, the obtained optimal weighted ensemble model has a good fit and high prediction accuracy.
[0047] The actual moisture content values of the modeling set and the predicted value of the prediction set are saved separately for subsequent scatter plotting or error analysis.
[0048] The optimal weighted ensemble model obtained at the end was used to predict the water content of new tomato leaf spectral data.
[0049] In model inversion, the location of the region of interest in the hyperspectral data to be measured is determined and its corresponding spectral data is extracted. The optimal weights are then input into the ensemble model to perform prediction and complete the model inversion.
[0050] like Figure 5 , Figure 6 As shown, the location of the region of interest is determined by initially determining the chlorophyll pixel locations of vegetation using the calculated Normalized Difference Vegetation Index (NDVI). Chlorophyll pixel regions with values greater than the NDVI threshold are then extracted for secondary screening. Additionally, chlorophyll pixel locations with reflectance values greater than the 560 nm visible light band reflectance threshold are extracted to eliminate shadow, background, and non-leaf pixel regions, retaining only the real leaf regions as the region of interest. The threshold for the NDVI is 0.6, and the reflectance threshold for the 560 nm visible light band is 0.15.
[0051] The normalized vegetation index is used to quantify vegetation cover and health status, and is calculated as follows: , Wherein: NDVI represents the Normalized Difference Vegetation Index, corresponding to the pixel points in the hyperspectral data; NIR is the band closest to 840nm; and RED is the band closest to 668nm.
[0052] The method for determining the region of interest is as follows: The NDVI threshold is set to 0.6. Pixels with a reflectance greater than 0.6 are set to 1, and those less than or equal to 0.6 are set to 0. Then, a second, more precise extraction is performed on the positions set to 1: the reflectance at 560nm is checked against a threshold of 0.15. If the reflectance is greater than 0.15, it remains unchanged and is set to 1; otherwise, it is set to 0. Finally, the positions set to 1 represent the extracted region of interest. Figure 4 The white area shown is position 1, which represents the leaf part of the vegetation, i.e., the pixel position that may contain chlorophyll and needs to be inverted.
[0053] The chlorophyll content is obtained after model inversion, and the predicted results and corresponding spatial locations are saved. On the same hyperspectral image to be measured or its visualization base map, the predicted chlorophyll value of each pixel is displayed according to its spatial coordinates, and a heat map legend is generated.
[0054] As shown in the figure, this disclosure presents a chlorophyll visualization system based on hyperspectral ensemble learning, used to execute the method, including: The data augmentation module is used to augment the one-dimensional spectral data of the acquired chlorophyll physicochemical indicators with Gaussian noise and distortion data, thereby expanding the sample size to obtain a sample set. A data partitioning module, connected to the data augmentation module, is used to partition the sample set into a modeling set and a prediction set in a 7:3 ratio, and to make the data distribution of the modeling set cover the prediction set. The model training and optimization module is connected to the data partitioning module and is used to train and optimize the hyperparameters of the XGBOOST model, partial least squares regression (PLS) model and ridge regression (RIDGE) model using the modeling set, respectively, and generate the optimal XGBOOST model, optimal PLS model and optimal RIDGE model accordingly. An ensemble learning module, connected to the model training and optimization module, is used to perform weighted fusion of the outputs of the optimal XGBOOST model, the optimal PLS model, and the optimal RIDGE model, and to determine the weights of each model through an iterative optimization algorithm to generate the best weight-optimized ensemble model. The region of interest extraction module is used to acquire the hyperspectral image to be tested, and determine the location of the region of interest in the hyperspectral image to be tested based on the set normalized vegetation index threshold and reflectance threshold, and extract the spectral data corresponding to the region of interest. The model inversion module is connected to the region of interest extraction module and the ensemble learning module. It is used to input the extracted spectral data into the optimal weighted ensemble model to predict chlorophyll content and obtain chlorophyll prediction values. The visualization module, connected to the model inversion module, is used to overlay and display the chlorophyll prediction value and its corresponding spatial location on the same hyperspectral image to be measured or its visualization base map to generate a heat map legend.
[0055] The parts not described in detail in this application are all existing conventional technologies and will not be elaborated here.
[0056] It is understood that the above specific description of this disclosure is only for illustrating this disclosure and is not limited to the technical solutions described in the embodiments of this disclosure. Those skilled in the art should understand that modifications or equivalent substitutions can still be made to this disclosure to achieve the same technical effect; as long as the usage needs are met, they are all within the protection scope of this disclosure.
Claims
1. A chlorophyll visualization method based on hyperspectral ensemble learning, characterized in that: Chlorophyll physicochemical indices were measured from the one-dimensional spectral data used for modeling, and Gaussian noise and distortion data augmentation were applied. The data was divided into a modeling set and a prediction set in a 7:3 ratio, ensuring that the modeling set included the prediction set. The XGBOOST, PLS, and RIDGE models were trained and their hyperparameters optimized using the partitioned modeling set, resulting in three optimal XGBOOST, PLS, and RIDGE models. These models were then weighted and merged to obtain an ensemble model. After iterative optimization, the best weighted ensemble model was obtained. By setting thresholds for normalized vegetation index and reflectance, the location of the region of interest in the hyperspectral data to be measured is determined and its corresponding spectral data is extracted. The optimal weights are input to optimize the ensemble model prediction and complete the model inversion to obtain the chlorophyll content. The predicted results and corresponding spatial locations are saved. On the same hyperspectral image to be measured or its visualization base map, the predicted chlorophyll value of each pixel is displayed according to its spatial coordinates, and a heat map legend is generated.
2. The chlorophyll visualization method according to claim 1, characterized in that: The Gaussian noise enhancement of the chlorophyll physicochemical index ,in: : indicates that the original one-dimensional spectral data is processed at the 1st... The spectral intensity values obtained after applying Gaussian noise enhancement to each band channel; The original spectrum at the 1st Intensity values for each wavelength / wavenumber channel; : A Gaussian distribution with a mean of 0 and a standard deviation of σ(i).
3. The chlorophyll visualization method according to claim 1, characterized in that: The distortion enhancement of the chlorophyll physicochemical indicators is divided into two types, including baseline drift enhancement and wavelength / intensity scaling and distortion enhancement. Baseline drift enhancement: , in: : indicates that the original one-dimensional spectral data is processed at the 1st... The spectral intensity values obtained after applying baseline drift enhancement at each band channel; It is a smooth baseline, generated by a low-order polynomial or a smooth spline curve: , K is the order of the polynomial, with a value between 1 and 4. These are the coefficients from random sampling, and N is the spectral length; Wavelength / intensity scaling and distortion enhancement: , in: : indicates that the original one-dimensional spectral data is processed at the 1st... The spectral intensity values obtained after applying wavelength / intensity scaling and distortion enhancement at each band channel; s is the scaling factor, and t is the translation amount, both of which are obtained by random sampling.
4. The chlorophyll visualization method according to claim 1, characterized in that: The method for dividing the modeling set and the prediction set is as follows: the spectra corresponding to the enhanced chlorophyll physicochemical indicators are arranged in ascending order. From every 10 spectra, the first 7 spectra are selected and assigned to the modeling set, and the last 3 spectra are assigned to the prediction set. This process is repeated. The chlorophyll content range of the modeling set is 1-100, and the chlorophyll content range of the prediction set is 10-90. That is, the chlorophyll content range of the modeling set includes the chlorophyll content range of the prediction set.
5. The chlorophyll visualization method according to claim 1, characterized in that: The training and hyperparameter optimization methods for the XGBOOST, PLS, and RIDGE models are as follows: XGBOOST, PLS, and RIDGE models are constructed using a partitioned modeling set; candidate hyperparameter combinations are set for each model type, and five-fold cross-validation is used to train and predict each group of candidate hyperparameters within the modeling set, with the five-fold average R² as the optimization score for that group of candidate hyperparameters; a Bayesian optimization algorithm is used to iterate 30 times within a preset hyperparameter search space, selecting the candidate hyperparameter group with the largest average R², and retraining on the entire modeling set to obtain the corresponding optimal XGBOOST, PLS, and RIDGE models; The hyperparameters of the XGBOOST model include the number of trees, the maximum tree depth, the learning rate, and the subsample sampling rate; the hyperparameter of the PLS model is the number of principal factors; and the hyperparameter of the RIDGE model is the regularization coefficient.
6. The chlorophyll visualization method according to claim 1, characterized in that: The weighted ensemble model is obtained by weighting the three optimal XGBOOST, PLS and RIDGE models and merging them to output a single prediction result. The loss function of the ensemble model is defined as root mean square error. Five-fold cross-validation and Bayesian optimization are used to iterate 30 times to obtain the optimal weighted ensemble model.
7. The chlorophyll visualization method according to claim 1, characterized in that: The location of the region of interest is determined by initially determining the chlorophyll pixel locations of vegetation using the calculated normalized vegetation index (NZI). Chlorophyll pixel regions with NZI values greater than a threshold are then extracted for secondary screening. Chlorophyll pixel locations with reflectance values greater than a 560 nm visible light band threshold are further selected to eliminate shadow, background, and non-leaf pixel regions, retaining only the actual leaf regions as the region of interest. The threshold for the NZI is 0.6; the reflectance threshold for the 560 nm visible light band is 0.
15.
8. The chlorophyll visualization method according to claim 7, characterized in that: The normalized vegetation index (NDI) is used to quantify vegetation cover and health status, and is calculated as follows: , Wherein: NDVI represents the Normalized Difference Vegetation Index, NIR is the band closest to 840nm, and RED is the band closest to 668nm.
9. The chlorophyll visualization method according to claim 7, characterized in that: The spectral data used for modeling is the average spectrum of the region of interest selected from the hyperspectral image dataset.
10. A chlorophyll visualization system based on hyperspectral ensemble learning, used to perform the method according to any one of claims 1-9, characterized in that, include: The data augmentation module is used to augment the chlorophyll physicochemical indicators obtained from the one-dimensional spectrum with Gaussian noise and distortion data, and expand the sample size to obtain a sample set. A data partitioning module, connected to the data augmentation module, is used to partition the sample set into a modeling set and a prediction set in a 7:3 ratio, and to make the data distribution of the modeling set cover the prediction set. The model training and optimization module is connected to the data partitioning module and is used to train and optimize the hyperparameters of the XGBOOST model, PLS model and RIDGE model respectively using the modeling set, and generate the optimal XGBOOST model, optimal PLS model and optimal RIDGE model respectively. An ensemble learning module, connected to the model training and optimization module, is used to perform weighted fusion of the outputs of the optimal XGBOOST model, the optimal PLS model, and the optimal RIDGE model, and to determine the weights of each model through an iterative optimization algorithm to generate the best weight-optimized ensemble model. The region of interest extraction module is used to acquire the hyperspectral image to be tested, and determine the location of the region of interest in the hyperspectral image to be tested based on the set normalized vegetation index threshold and reflectance threshold, and extract the spectral data corresponding to the region of interest. The model inversion module is connected to the region of interest extraction module and the ensemble learning module. It is used to input the extracted spectral data into the optimal weighted ensemble model to predict chlorophyll content and obtain chlorophyll prediction values. The visualization module, connected to the model inversion module, is used to overlay and display the chlorophyll prediction value and its corresponding spatial location on the same hyperspectral image to be measured or its visualization base map to generate a heat map legend.
