Gingko canopy chlorophyll content estimation method based on airborne hyperspectral image

Abstract

A ginkgo canopy chlorophyll content estimation method based on an airborne hyperspectral image comprises the following steps: firstly, acquiring a ginkgo leaf canopy hyperspectral image and a chlorophyll content data set; and establishing a rapid prediction model for predicting the average content of chlorophyll in the canopy by using the hyperspectral image of the ginkgo leaf canopy, and predicting the corresponding chlorophyll content by using the model and taking the hyperspectral image of the ginkgo leaf canopy as input. The model establishment process comprises the following steps of: firstly, preprocessing an image by using multiple scatter correction (MSC); secondly, carrying out characteristic wave band selection on the preprocessed spectral data by adopting a competitive self-adaptive reweighted sampling algorithm CARS; and finally, substituting the screened characteristic wave band spectral data and the actually measured chlorophyll value into an XGBoost-RF machine learning model, and establishing a hyperspectral chlorophyll content estimation model. The method provided by the invention is a nondestructive rapid detection technology, and can monitor the plant health condition in real time.

Description

Technical Field

[0001] The invention belongs to the technical field of forestry, and in particular is a method for estimating chlorophyll content in a ginkgo canopy based on airborne hyperspectral images. Background Art

[0002] Ginkgo leaf is the leaf of Ginkgo biloba [1-2] According to modern pharmacological research, Ginkgo biloba has a promoting effect on memory loss and improving cardiovascular and cerebrovascular functions. [3-4] Chlorophyll plays an important role in the growth of plants and is an evaluation indicator that affects the growth of plants. [5-6] The current methods for testing chlorophyll mainly include spectrophotometry. [7-8] , HPLC [9-10] , electrochemical method

[11] However, such methods are limited to small sample experiments, and their operation steps are cumbersome and time-consuming. They pollute the environment during the processing and cause a certain degree of damage to the test sample itself. In view of the shortcomings of the above-mentioned sample detection methods, a simple, time-saving, pollution-free, large sample non-destructive detection method is proposed, that is, a method for predicting the chlorophyll content of ginkgo leaves using UAV hyperspectral canopy data.

[0003] The research on detecting chlorophyll content based on hyperspectral technology is gradually increasing. It mainly establishes chlorophyll content estimation models for different plants by preprocessing hyperspectral image data, selecting characteristic bands and combining machine learning model algorithms. For example, Song Zhenghua et al.

[12] Using the hyperspectral images of mosaic diseased apple leaves, a chlorophyll content estimation model in apple leaves infected with mosaic disease was established through band feature selection combined with support vector machine, with a prediction accuracy of R 2 =0.8665. Tenglong et al.

[13] The characteristic wavelength refinement method of interval band selection competitive adaptive reweighted sampling fusion (IBS-CARS-Fusing) was used to select the bands in the hyperspectral image of Dancong tea, and the kernel ridge regression (KRR) model was established to predict chlorophyll a and chlorophyll b in Dancong tea. The prediction accuracy was R 2 They are 0.9500 and 0.9481 respectively.

[14] Characteristic wavelengths were screened in the hyperspectral images of cotton leaves at different growth stages, and a prediction model for chlorophyll content in cotton leaves was established by combining the SSA-BPNN model. The prediction accuracy was R 2 =0.92. Zhao Ruomei et al.

[15] Using corn hyperspectral images at different growth stages, competitive adaptive weighted sampling (CARS) was used to select bands, and a CARS-LSTM corn dynamic growth effect chlorophyll content estimation model was established with a prediction accuracy of R 2= in the range of 0.76 to 0.96. Wen Shuangya et al.

[16] Taking rice as the research object, different fertilization treatments were set to explore the intrinsic relationship between chlorophyll fluorescence parameters and spectral reflectance. Spectral sensitive information was extracted by vegetation index and continuous wavelet transform (CWT), and a monitoring model of Fv / Fm in rice leaves was established using BPNN algorithm, with a prediction accuracy of R2 = 0.8236. Tang Weijie et al.

[17] The hyperspectral images of potato leaves were used to screen features using the continuous wavelet transform (CWT) spectrum and the convolutional neural network (CNN) of the deep learning algorithm, and a PLSR model for detecting chlorophyll content in potato leaves was established, with a detection accuracy of R2=0.8749.

[0004] The research of these scholars was all conducted in laboratory environments, and the operation process was complicated and time-consuming. Summary of the invention

[0005] Chlorophyll content is one of the key indicators to measure the growth and health of ginkgo leaves. By monitoring chlorophyll content, problems in the plant growth process can be discovered in time and corresponding management measures can be taken to promote the healthy growth of ginkgo leaves.

[0006] In order to quickly estimate the chlorophyll content of the canopy of individual trees in a ginkgo leaf forest over a large area, a UAV (Unmanned Aerial Vehicle) hyperspectral imager can be used to collect hyperspectral images of the canopy, and a portable chlorophyll meter can be used to sample and detect the chlorophyll content of several leaves in the canopy in situ. The hyperspectral and chlorophyll content datasets of the ginkgo leaf forest canopy are obtained, and a rapid estimation model for estimating the average chlorophyll content of the canopy using hyperspectral images of the ginkgo leaf forest canopy is established.

[0007] The model building process is as follows:

[0008] Firstly, the spectral preprocessing methods such as multivariate scattering correction (MSC), variable standardization (SNV), convolution smoothing (SG), first-order derivative (FD) and second-order derivative (SD) were used to establish the partial least squares detection model (PLSR), support vector machine model (SVR) and random forest model (RF) respectively to obtain the optimal preprocessing method MSC.

[0009] Secondly, the preprocessed spectral data are used to select characteristic bands using the successive projection method (SPA), competitive adaptive reweighted sampling algorithm (CARS), principal component analysis (PCA), etc. to obtain the optimal feature extraction method CARS.

[0010] Finally, the screened characteristic band spectral data and the measured chlorophyll values ​​were substituted into machine learning models such as random forest (RF), support vector machine (SVR), Extreme Gradient Boosting (XGBoost), PLSR, and XGB-RF to establish a hyperspectral chlorophyll content estimation model.

[0011] The results showed that MSC pretreatment had an excellent effect and its PLSR model is 0.8776, RMSE p The chlorophyll estimation accuracy is the highest when the characteristic bands selected by CARS are combined with the XGB-RF model. is 0.9033, RMSE p The MSC-CARS-(XGB-RF) model is an innovation in the field of hyperspectral remote sensing data processing and machine learning.

[0012] The method and model proposed in the present invention for estimating the chlorophyll content of a single plant based on UAV hyperspectral canopy images is an accurate, non-destructive, rapid detection technology that can help farmers monitor the health status of plants in real time.

[0013] The present invention is specifically:

[0014] A method for estimating chlorophyll content in a ginkgo canopy based on airborne hyperspectral imaging comprises the following steps: first, collecting a hyperspectral image of a ginkgo leaf canopy; then, using the collected image as input, estimating the corresponding chlorophyll content using a ginkgo leaf chlorophyll content prediction model; the method for collecting the hyperspectral image of the ginkgo leaf canopy is to use a hyperspectral imager carried by an unmanned aerial vehicle to obtain the hyperspectral image of the ginkgo leaf;

[0015] The process of establishing the prediction model for chlorophyll content in ginkgo leaf is as follows:

[0016] S1 hyperspectral image correction and ROI extraction

[0017] S1.1 Hyperspectral image correction: perform radiation correction, geometric correction and atmospheric correction on hyperspectral images. The specific steps include:

[0018] S1.1.1. Radiometric Correction

[0019] The purpose of radiation correction is to calibrate the raw data in remote sensing images to eliminate the influence of sensor errors, noise and other environmental factors on the images, and ensure that the images reflect the true radiation information of the objects. Steps:

[0020] a) Sensor calibration: The sensor may be affected by many factors when acquiring images, such as sensor sensitivity, temperature changes, etc. First, the image needs to be radiometrically corrected according to the sensor's calibration coefficients, usually using known reflectivity or radiometric calibration data.

[0021] b) Conversion from digital values ​​to radiometric values: Remote sensing images are usually expressed in digital values ​​(DN values), which have a certain relationship with the ground reflectivity (or radiometric brightness). These digital values ​​are converted into physical quantities, such as radiometric brightness or reflectivity.

[0022] c) Radiation correction: It includes absolute radiation correction, which is to convert image data into physical reflectivity or radiation brightness of ground objects by comparing with the true reflectivity of the ground; relative radiation correction, which is to use multi-phase images for relative comparison and adjust the brightness difference between images.

[0023] d) Noise Removal: Remote sensing images may be affected by noise. Radiometric correction also includes removing sensor noise or other interfering signals to improve the quality of the data.

[0024] S1.1.2. Geometric Correction

[0025] The purpose of geometric correction is to align the spatial coordinate system of the remote sensing image with the coordinate system of the real world to eliminate geometric distortion caused by factors such as the curvature of the earth, sensor angle, platform motion, etc. After geometric correction, the image can accurately reflect the real position and shape of ground objects. Steps:

[0026] a) Geometric model selection: First, we need to select an appropriate geometric model to describe the geometric deformation of the image. Commonly used geometric models include projection model, affine transformation model and polynomial transformation model.

[0027] b) Ground Control Point (GCP) Selection: Select easily identifiable ground control points in the image and match them with actual geographic coordinates. Ground control points usually come from ground survey data or existing maps.

[0028] c) Geometric transformation: Based on the selected geometric model and ground control points, the image is geometrically corrected using transformation algorithms (such as affine transformation, perspective transformation, etc.). Common transformation methods are:

[0029] Affine transformation: considers factors such as translation, scaling, and rotation.

[0030] Higher-order transformations: consider more complex nonlinear deformations.

[0031] Resampling: After the geometric correction process, each pixel in the image may no longer correspond to the pixel position of the original image, and resampling is required. Common resampling methods include nearest neighbor interpolation, bilinear interpolation, and cubic convolution interpolation.

[0032] S1.1.3. Atmospheric Correction

[0033] The purpose of atmospheric correction is to remove the influence of water vapor, aerosol, cloud and other substances in the atmosphere on remote sensing images, so as to accurately extract the surface information of the ground objects from the images. Particles and gases in the atmosphere will scatter and absorb solar radiation, thereby changing the radiation value reaching the ground and affecting the accuracy of remote sensing data. Steps:

[0034] a) Application of radiation transfer model: Use radiation transfer models (such as MODTRAN, 6S, etc.) to simulate the influence of the atmosphere on ground radiation. Through these models, the influence of the atmosphere on each band in the remote sensing image can be calculated and then corrected.

[0035] b) Meteorological data acquisition: In order to perform accurate atmospheric correction, it is necessary to obtain meteorological data corresponding to the time and location of remote sensing image acquisition, including atmospheric temperature, humidity, air pressure, wind speed, etc.

[0036] c) Calculation of ground reflectivity: Through atmospheric correction, the radiation values ​​in the remote sensing image can be converted into the reflectivity of the ground object (or ground radiation brightness), thereby eliminating the influence of the atmosphere on radiation.

[0037] d) Scattering and absorption correction: Utilize ground reflectivity and atmospheric conditions to quantitatively correct atmospheric effects such as aerosol scattering and gas absorption, thereby eliminating the interference of atmospheric effects on image data.

[0038] S1.2 Hyperspectral image ROI extraction

[0039] First, ROI is extracted from the hyperspectral image; then, the hyperspectral dataset is constructed using the average method; the spectral characteristics of the region are represented by calculating the average value of all pixels in the selected region (ROI) in each band. The average method assumes that the spectrum of the region can be effectively described by its average value, which is suitable for some situations where the spectrum changes are relatively stable or the objects in the region are consistent.

[0040] The formula for the mean value method

[0041] Assume that the size of the hyperspectral image is M×N×B, where:

[0042] M is the number of rows in the image

[0043] N is the number of columns in the image

[0044] B is the number of bands of the image

[0045] For the selected region of interest (ROI), extract the values ​​of each pixel in the region in all bands and calculate the average value of each band. Assume that the ROI contains P pixels, the number of bands is B, and the spectral value of each pixel in the bth band is Q b (n), where n = 1, 2, ..., P represents the nth pixel in the ROI.

[0046] Then, the average value of the bth band is Calculated by the following formula:

[0047] in:

[0048] is the average value of the bth band

[0049] Q b (n) is the value of the nth pixel in the ROI area on b

[0050] P is the number of pixels in the ROI area

[0051] S2 Hyperspectral data preprocessing.

[0052] The present invention uses multiple scattering correction (MSC) to preprocess the hyperspectral data. The preprocessing process of MSC is as follows:

[0053] 1. Spectral data standardization and de-averaging

[0054] For spectral data, we first need to perform mean removal to eliminate the DC offset component in the spectral data. Suppose there is a set of spectral data Y = [y1, y2, …, y n ], where each y i is a p-dimensional spectral vector representing the spectrum of the i-th sample. First, the mean of each sample spectrum is removed:

[0055] y′ i =y i -μ y

[0056] Among them, μ y is the mean vector of all sample spectra:

[0057]

[0058] y i ' is the spectrum vector after removing the mean.

[0059] 2. Fitting the baseline scattering model

[0060] The MSC method assumes that there is a global scattering effect in the spectral data of the sample, and assumes that the scattering effect can be described by a baseline model. In order to correct this scattering effect, a scattering model is first established. The MSC method fits the scattering model through a linear regression model.

[0061] In order to correct the scattering effect of each spectrum, a suitable baseline model needs to be found. In MSC, this baseline model can be assumed to be:

[0062] y i '=β0+β1y i +ε i

[0063] Among them, β0 is the bias term, β1 is the regression coefficient of the spectral data, and ε i is the error term. Through regression analysis, these coefficients are estimated to obtain the scattered part of the spectrum.

[0064] 3. Calculate the principal components (regression coefficients) of the spectrum

[0065] The core of the MSC method is to use multiple linear regression to estimate the scattering effect of each spectrum. Through linear regression, the relationship between the spectral data and a reference (which can be a benchmark sample or a spectral baseline model) can be calculated. At this time, the regression coefficient β1 represents the correlation between the sample spectrum and the baseline model.

[0066] A regression model for the scattering effect can be fit using ordinary least squares (OLS):

[0067] y=Xβ

[0068] Where y is the scattering component predicted by the model, X is the original data matrix (the mean-devalued spectra), and β is the regression coefficient.

[0069] 4. Correction of spectral data

[0070] The scattering baseline model obtained by regression can be used to correct the original spectral data. The core idea of ​​MSC is to subtract the fitted scattering component from the spectral data of each sample:

[0071]

[0072] Where y is the prediction of the regression model

[0073] S3 feature band extraction

[0074] The data was processed using the competitive adaptive reweighted sampling algorithm (CARS) to extract characteristic bands;

[0075] S3.1. Data preprocessing

[0076] Denoising and standardization: Data preprocessing is critical, including denoising, removing outliers, smoothing, normalization, and other steps to improve data quality and usability. For spectral data, common preprocessing methods include smoothing, standardization, and normalization.

[0077] Normalization: To eliminate the scale differences between different features, the data is standardized or normalized so that the mean of each feature is 0 and the variance is 1.

[0078] S3.2. Initialize feature set

[0079] A selected set of features, typically all features extracted from the raw data. For example, for spectral data, the features might be the intensity of reflection at different wavelengths.

[0080] S3.3. Adaptive reweighting

[0081] Weighting process: The core idea of ​​CARS is to dynamically re-evaluate the weight of each feature through a weighting mechanism. Through this weighting, important features can be given higher weights while reducing the impact of irrelevant features. Weighting is achieved by setting the re-weighting coefficient of the feature, and the size of the coefficient reflects the contribution of the feature to the model.

[0082] Weighted calculation: Weighting is based on certain adaptive criteria, and the CARS algorithm continuously adjusts these weights according to the importance of the features. For each round of weighting, the algorithm adaptively selects the features that contribute more to the objective function (such as prediction error) to be retained.

[0083] S3.4. Feature selection process

[0084] The feature set will keep changing through multiple iterations of weighting. In each iteration, CARS optimizes the feature selection by reselecting a subset of features and weighting them.

[0085] In each iteration, CARS builds models based on the weighted feature subsets and evaluates the performance of the model. If the weights of some features are low, it means that these features contribute little and can be removed or down-weighted.

[0086] S3.5. Reweighting Iterations

[0087] After each iteration, CARS will make adaptive adjustments based on the weight of each feature. Features with low weight values ​​may be reduced or discarded, while features with higher weights will be retained. This process helps to screen out the most representative features.

[0088] In each round of iteration, the feature subset is continuously adjusted until the optimal feature set is found.

[0089] S3.6. Model evaluation and selection

[0090] Performance evaluation: In each iteration, the prediction ability of the model is evaluated using the training set and the test set. You may choose methods such as cross-validation to perform model evaluation. The goal of the CARS algorithm is to maximize the prediction ability of the model by selecting the most relevant features.

[0091] Stopping criteria: Iterations usually continue until a certain stopping criterion is met, such as reaching a predetermined number of iterations, or the results of feature selection no longer significantly improve prediction accuracy.

[0092] S3.7. Final feature set

[0093] At the end of the iteration, CARS returns an optimal subset of features that are weighted, selected, and iteratively optimized to most effectively represent the variability of the data.

[0094] The final selected feature set can be used for subsequent modeling, analysis or prediction tasks.

[0095] S4 training prediction model

[0096] The characteristic band spectral data obtained by S3 and the corresponding measured chlorophyll values ​​were substituted into the XGB-RF hybrid model to obtain the final prediction model;

[0097] Concept: The XGB-RF hybrid model is mainly an integration of the XGBoost machine learning model and the RF machine learning model.

[0098] The basic idea and implementation process of XGBoost and RF hybrid model

[0099] The basic idea of ​​the hybrid model is to combine the prediction results of XGBoost and RF to improve the final prediction performance of Ginkgo biloba chlorophyll. This integrated method reduces the bias and variance problems that may be caused by a single model by making full use of the advantages of both - the efficiency of XGBoost and the stability of RF.

[0100] Steps to implement the hybrid model

[0101] S4.1 Train XGBoost and RF models separately

[0102] Train the XGBoost and RF models separately to get their respective prediction results. Cross-validation can be used to select the optimal hyperparameters for each.

[0103] a. For XGBoost, you can use the Gradient Boosted Tree (GBDT) algorithm to perform weighted combination through multiple rounds of weak classifiers (trees).

[0104] b. For RF, the final decision is made by building multiple decision trees for voting or averaging.

[0105] S4.2 Combination of prediction results

[0106] The prediction results of XGBoost and RF can be combined in the following ways:

[0107] a. Weighted Averaging: The prediction results of Ginkgo biloba chlorophyll are weighted averaged according to the performance of the models, giving priority to the results of models with higher accuracy.

[0108] b. Simple Averaging: Directly average the prediction results of the two models. This method is suitable when the performance of the models is relatively balanced.

[0109] c. Voting: If it is a classification problem, the majority voting method can be used, that is, the category with the most prediction results is selected as the final output.

[0110] d. Stacking: The prediction results of the two models are input as new features into another model (which can be logistic regression or other simple models) for learning and prediction. Stacking can achieve better performance, especially when the base learners are significantly different.

[0111] S4.3 Model fusion

[0112] In practical applications, XGBoost and RF can be used as base learners and further integrated using the stacking method. The specific steps are as follows:

[0113] a. Train XGBoost and RF models and obtain their prediction outputs (predicted values ​​for regression problems and probability distributions or categories for classification problems).

[0114] b. These predictions are used as features to feed a secondary learner (such as linear regression or logistic regression), whose task is to make a final decision based on these predictions.

[0115] S4.4 Evaluation and Adjustment

[0116] Use cross-validation to evaluate the performance of the hybrid model and adjust the weighting parameters or learners in the stacked model based on the results to optimize the final model effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0117] Figure 1 It is the overall process diagram of the experiment, which illustrates the relationship between the obtained hyperspectral canopy ginkgo leaf reflectance, canopy ginkgo leaf chlorophyll content and the prediction model;

[0118] Translation in the picture:

[0119] Ginkgo canopy: Ginkgo canopy

[0120] Airborne hyperspectral imager: hyperspectral imager

[0121] canopy ginkgo leaves: canopy ginkgo leaves

[0122] portable chlorophyll meter: portable chlorophyll meter

[0123] Hyperspectral canopy ginkgo leaf data acquisition: Hyperspectral canopy ginkgo leaf data acquisition

[0124] Hyperspectral data preprocessing: hyperspectral data preprocessing

[0125] Measurement of chlorophyll content in canopy ginkgo leaves

[0126] Band selection of hyperspectral chlorophyll data set: Band selection of hyperspectral chlorophyll data set

[0127] UVA hyperspectral inversion model of chlorophyll content in ginkgobiloba leaves: UVA hyperspectral inversion model of chlorophyll content in ginkgobiloba leaves;

[0128] Figure 2 It stands for Ginkgo Research Base;

[0129] Figure 3(a) and Figure 3(b) show the hyperspectral original reflectance and the optimal pre-processed reflectance, respectively;

[0130] Figure 4 Represents the bands screened by the continuous projection method;

[0131] Figure 5Represents the bands selected by the competitive adaptive reweighted sampling algorithm;

[0132] Figure 6 Represents the bands selected by the principal component analysis algorithm;

[0133] Figure 7(a) and Figure 7(b) show the input and output heat map and bar graph of each model, respectively;

[0134] Figure 8 It indicates the distribution of chlorophyll content in Ginkgo biloba leaves;

[0135] Fig. 9 It indicates the best effect of predicting the true value based on A3 and XGB-RF models. DETAILED DESCRIPTION

[0136] In view of the research characteristics of the prior art, the present invention adopts UAV hyperspectral equipment and chlorophyll meter to collect ginkgo leaf hyperspectral images and chlorophyll content, which is simple to operate, time-saving and has efficient result output.

[0137] Currently, there is a relative lack of research on combining hyperspectral technology with machine learning models to predict the chlorophyll content of ginkgo leaves.

[0138] Figure 1 This is a diagram of the overall implementation process of the estimation (prediction) model of the present invention. This schematic diagram illustrates the relationship between the obtained hyperspectral canopy ginkgo leaf reflectance, canopy ginkgo leaf chlorophyll content and the prediction model.

[0139] The contributions of the present invention mainly include:

[0140] 1. UAV hyperspectral imagers can perform time-saving, labor-saving, and non-destructive hyperspectral imaging of large areas of ginkgo leaf canopies.

[0141] 2. The XGB-RF hybrid model is superior to a single model in terms of prediction accuracy. The hybrid models can complement each other, reduce the risk of overfitting, and improve the robustness of the model.

[0142] 1 Materials and methods

[0143] 1.1 Test location and time

[0144] On May 18, 2024, from 14:00 to 16:00 in the afternoon, the hyperspectral image and chlorophyll data collection experiment of ginkgo leaves was carried out at the research base of Xuzhou Changrong Agricultural Development Co., Ltd. (34°8'42"N, 117°2'45"E). The weather was clear, the temperature was 28°, the humidity was 75%, and the wind speed was 8.6km / h. The experimental subjects were conducted on three-year-old ginkgo trees at the research base, with a row spacing of 1m and a plant spacing of 0.4m (Figure 3). The research base consists of 800 trees, covering a plot size of 0.8×4 square meters.

[0145] 1.2 Hyperspectral image acquisition of canopy ginkgo leaves

[0146] This study takes the ginkgo leaves in the Pizhou scientific research base as the research object, and uses the hyperspectral imager (GaiaSky-Mini3-VN, Shuanglihepu) carried by the UAV (M350, DJI) to obtain the hyperspectral images of the ginkgo leaves. The spectral range of the hyperspectral imager is 400-1000nm, the spectral resolution is 5nm, the number of spatial channels is 1024, the number of spectral channels is 224, and the image resolution is 1024×1003. The resolution of the imaging lens is 16nm, the power is 45W, the shooting height is 50 meters, and the shooting method is UAV hovering and through the built-in push-broom mode. At the same time, the hyperspectral images of the ginkgo leaves in the canopy of the three-year-old ginkgo trees in the scientific research base are collected every half month. The hyperspectral image collection time of the canopy ginkgo leaves is from 14:00 to 15:00 on May 18, 2024.

[0147] 1.3 Numerical collection of chlorophyll content in ginkgo leaves

[0148] At 15:10-16:00 on the afternoon of May 18, 2024, after the hyperspectral image of the ginkgo leaves in the canopy was collected, a portable chlorophyll meter (SPAD-502Plus, Konica Minolta, Japan) was immediately used to randomly test the ginkgo leaves in the canopy of the ginkgo trees at the research base, and the average chlorophyll content of the five leaves in the canopy of a single tree was taken as the average chlorophyll content of the canopy. The chlorophyll meter can quickly measure the relative content of chlorophyll in the current leaves by measuring the absorption rate of the leaves in two wavelength bands without destroying the crops.

[18] .

[0149] 1.4 Hyperspectral image correction and ROI extraction

[0150] 1.4.1 Hyperspectral Image Correction

[0151] Hyperspectral images are usually composed of a large amount of band information, which may be affected by various factors, such as environmental conditions, atmospheric effects, and sensor errors. Therefore, it is necessary to perform radiation correction, geometric correction, and atmospheric correction on hyperspectral images. The corrected images can effectively eliminate these effects, thereby improving the quality of the data.

[19] .

[0152] The radiation correction, geometric correction and atmospheric correction include the following specific steps:

[0153] 1. Radiometric Correction

[0154] The purpose of radiation correction is to calibrate the raw data in remote sensing images to eliminate the impact of sensor errors, noise and other environmental factors on the images, and ensure that the images reflect the true radiation information of the objects.

[0155] step:

[0156] a. Sensor calibration: The sensor may be affected by many factors when acquiring images, such as sensor sensitivity, temperature changes, etc. First, the image needs to be radiometrically corrected according to the sensor's calibration coefficients, usually using known reflectivity or radiometric calibration data.

[0157] b. Conversion from digital values ​​to radiometric values: Remote sensing images are usually represented by digital values ​​(DN values), which have a certain relationship with the ground reflectivity (or radiance). The first step of radiometric correction is to convert these digital values ​​into physical quantities, such as radiance or reflectance.

[0158] c. Noise removal: Remote sensing images may be affected by noise. Radiometric correction also includes removing sensor noise or other interfering signals to improve the quality of the data.

[0159] Calibration method:

[0160] Absolute radiation correction: By comparing with the true reflectivity of the ground, the image data is converted into the physical reflectivity or radiation brightness of the ground object.

[0161] Relative radiometric correction: Use multi-temporal images for relative comparison and adjust for brightness differences between images.

[0162] 2. Geometric Correction

[0163] The purpose of geometric correction is to align the spatial coordinate system of the remote sensing image with the coordinate system of the real world to eliminate geometric distortion caused by factors such as the curvature of the earth, sensor angle, platform motion, etc. After geometric correction, the image can accurately reflect the real position and shape of ground objects. Steps:

[0164] a. Geometric model selection: First, you need to select an appropriate geometric model to describe the geometric deformation of the image. Commonly used geometric models include projection models, affine transformation models, and polynomial transformation models.

[0165] b. Ground Control Point (GCP) Selection: Select easily identifiable ground control points in the image and match them with actual geographic coordinates. Ground control points usually come from ground survey data or existing maps.

[0166] c. Geometric transformation: Based on the selected geometric model and ground control points, use transformation algorithms (such as affine transformation, perspective transformation, etc.) to geometrically correct the image. Common transformation methods are:

[0167] Affine transformation: considers factors such as translation, scaling, and rotation.

[0168] Higher-order transformations: consider more complex nonlinear deformations.

[0169] Resampling: After the geometric correction process, each pixel in the image may no longer correspond to the pixel position of the original image, and resampling is required. Common resampling methods include nearest neighbor interpolation, bilinear interpolation, and cubic convolution interpolation.

[0170] 3. Atmospheric Correction

[0171] The purpose of atmospheric correction is to remove the influence of water vapor, aerosol, cloud and other substances in the atmosphere on remote sensing images, so as to accurately extract the surface information of the ground objects from the images. Particles and gases in the atmosphere will scatter and absorb solar radiation, thereby changing the radiation value reaching the ground and affecting the accuracy of remote sensing data. Steps:

[0172] a. Application of radiation transfer model: Use radiation transfer models (such as MODTRAN, 6S, etc.) to simulate the influence of the atmosphere on ground radiation. Through these models, the influence of the atmosphere on each band in the remote sensing image can be calculated and then corrected.

[0173] b. Meteorological data acquisition: In order to perform accurate atmospheric correction, it is necessary to obtain meteorological data corresponding to the time and location of remote sensing image acquisition, including atmospheric temperature, humidity, air pressure, wind speed, etc.

[0174] c. Calculation of ground reflectivity: Through atmospheric correction, the radiation value in the remote sensing image can be converted into the reflectivity of the ground object (or ground radiation brightness), thereby eliminating the influence of the atmosphere on radiation.

[0175] d. Scattering and absorption correction: Utilize ground reflectivity and atmospheric conditions to quantitatively correct atmospheric effects such as aerosol scattering and gas absorption, and eliminate the interference of atmospheric effects on image data.

[0176] 1.4.2 Hyperspectral Image ROI Extraction

[0177] Hyperspectral images usually contain a large amount of data. In order to obtain the information required for the experiment, data extraction is required for the hyperspectral images. The ROI tool in ENVI 5.3 software is used to extract the hyperspectral images, and the average method is used to construct the hyperspectral dataset of canopy ginkgo leaves.

[0178] 1.5 Hyperspectral data preprocessing

[0179] Due to differences in temperature, humidity, light intensity, and light angle, hyperspectral data are mixed with varying degrees of noise, and hyperspectral data need to be preprocessed.

[0180] The present invention uses multiple scattering correction (MSC) to preprocess the hyperspectral data. The preprocessing process of MSC is as follows:

[0181] 1. Spectral data standardization and de-averaging

[0182] For spectral data, we first need to perform mean removal to eliminate the DC offset component in the spectral data. Suppose there is a set of spectral data Y = [y1, y2, …, y n ], where each y i is a p-dimensional spectral vector representing the spectrum of the i-th sample. First, the mean of each sample spectrum is removed:

[0183] y′ i =y i -μ y

[0184] Among them, μ y is the mean vector of all sample spectra:

[0185]

[0186] y i ' is the spectrum vector after removing the mean.

[0187] 2. Fitting the baseline scattering model

[0188] The MSC method assumes that there is a global scattering effect in the spectral data of the sample, and assumes that the scattering effect can be described by a baseline model. In order to correct this scattering effect, a scattering model is first established. The MSC method fits the scattering model through a linear regression model.

[0189] In order to correct the scattering effect of each spectrum, a suitable baseline model needs to be found. In MSC, this baseline model is assumed to be:

[0190] y i '=β0+β1y i +ε i

[0191] Among them, β0 is the bias term, β1 is the regression coefficient of the spectral data, and ε i is the error term. Through regression analysis, these coefficients are estimated to obtain the scattered part of the spectrum.

[0192] 3. Calculate the principal components (regression coefficients) of the spectrum

[0193] The core of the MSC method is to use multiple linear regression to estimate the scattering effect of each spectrum. Through linear regression, the relationship between the spectral data and a reference (which can be a benchmark sample or a spectral baseline model) can be calculated. At this time, the regression coefficient β1 represents the correlation between the sample spectrum and the baseline model.

[0194] A regression model for the scattering effect can be fit using ordinary least squares (OLS):

[0195] y=Xβ

[0196] Where y is the scattering component predicted by the model, X is the original data matrix (the mean-devalued spectra), and β is the regression coefficient.

[0197] 4. Correction of spectral data

[0198] The scattering baseline model obtained by regression can be used to correct the original spectral data. The core idea of ​​MSC is to subtract the fitted scattering component from the spectral data of each sample:

[0199]

[0200] Here, y is the prediction from the regression model.

[0201] 1.6 Feature band extraction

[0202] The extraction of characteristic bands is an effective data dimensionality reduction technique, which aims to reduce redundancy and complexity by reducing the dimension of data.

[19] Specifically, this process involves selecting the most critical and representative bands for the analysis task from the original spectral data, improving the efficiency of data processing, and enhancing the accuracy of analysis, so that subsequent data analysis, classification or regression tasks become more efficient and reliable.

[20] .

[0203] Successive Projections Algorithm (SPA), Competitive Adaptive Reweighted Sampling (CARS), Principal Component Analysis (PCA) and other methods are used to select characteristic bands of the data.

[0204] It has been verified that CARS is the best. The steps of using CARS to extract characteristic fluctuations include:

[0205] 1. Data preprocessing

[0206] Denoising and standardization: Data preprocessing is critical, including denoising, removing outliers, smoothing, normalization, and other steps to improve data quality and usability. For spectral data, common preprocessing methods include smoothing, standardization, and normalization.

[0207] Normalization: To eliminate the scale differences between different features, the data is standardized or normalized so that the mean of each feature is 0 and the variance is 1.

[0208] 2. Initialize the feature set

[0209] The selected feature set is all the features extracted from the raw data. For example, for spectral data, the features may be the reflection intensity at different wavelengths.

[0210] 3. Adaptive reweighting

[0211] Weighting process: The core idea of ​​CARS is to dynamically re-evaluate the weight of each feature through a weighting mechanism. Through this weighting, important features can be given higher weights while reducing the impact of irrelevant features. Weighting is achieved by setting the re-weighting coefficient of the feature, and the size of the coefficient reflects the contribution of the feature to the model.

[0212] Weighted calculation: Weighting is based on certain adaptive criteria, and the CARS algorithm continuously adjusts these weights according to the importance of the features. For each round of weighting, the algorithm adaptively selects the features that contribute more to the objective function (such as prediction error) to be retained.

[0213] 4. Feature Selection Process

[0214] The feature set will keep changing through multiple iterations of weighting. In each iteration, CARS optimizes the feature selection by reselecting a subset of features and weighting them.

[0215] In each iteration, CARS builds models based on the weighted feature subsets and evaluates the performance of the model. If the weights of some features are low, it means that these features contribute little and can be removed or down-weighted.

[0216] 5. Reweighted Iteration

[0217] After each iteration, CARS will make adaptive adjustments based on the weight of each feature. Features with low weight values ​​may be reduced or discarded, while features with higher weights will be retained. This process helps to screen out the most representative features.

[0218] In each round of iteration, the feature subset is continuously adjusted until the optimal feature set is found.

[0219] 6. Model evaluation and selection

[0220] Performance evaluation: In each iteration, the prediction ability of the model is evaluated using the training set and the test set. Select a method such as cross-validation to perform model evaluation. The goal of the CARS algorithm is to maximize the predictive ability of the model by selecting the most relevant features.

[0221] Stopping criteria: Iterations usually continue until a certain stopping criterion is met, such as reaching a predetermined number of iterations, or the results of feature selection no longer significantly improve prediction accuracy.

[0222] 7. Final feature set

[0223] At the end of the iteration, CARS returns an optimal subset of features that are weighted, selected, and iteratively optimized to most effectively represent the variability of the data.

[0224] The final selected feature set can be used for subsequent modeling, analysis or prediction tasks.

[0225] 1.7 Machine Learning and Model Validation

[0226] The optimal model was selected from four machine learning regression methods: support vector regression (SVR), random forest (RF), extreme gradient boosting (XGBoost), and XGB-RF hybrid model.

[0227] It is verified that the CARS-XGB-RF hybrid model is the best.

[0228] The XGB-RF hybrid model is mainly an integration of the XGBoost machine learning model and the RF machine learning model.

[0229] The basic idea of ​​XGBoost and RF hybrid model

[0230] The basic idea of ​​the hybrid model is to combine the prediction results of XGBoost and RF to improve the final prediction performance. This integration method reduces the bias and variance problems that may be caused by a single model by making full use of the advantages of both - the efficiency of XGBoost and the stability of RF.

[0231] Steps to implement the hybrid model

[0232] Train XGBoost and RF models separately

[0233] Train the XGBoost and RF models separately to get their respective prediction results. Cross-validation can be used to select the optimal hyperparameters for each.

[0234] 1. For XGBoost, the gradient boosted tree (GBDT) algorithm is used to perform weighted combination through multiple rounds of weak classifiers (trees).

[0235] 2. For RF, the final decision is made by building multiple decision trees for voting or averaging.

[0236] Combination of prediction results

[0237] The prediction results of XGBoost and RF can be combined in the following ways:

[0238] 1. Weighted Averaging: The prediction results are weighted averaged according to the performance of the model, giving priority to the results of the model with higher accuracy.

[0239] 2. Simple Averaging: Directly average the prediction results of the two. This method is suitable when the performance of the models is relatively balanced.

[0240] 3. Voting: If it is a classification problem, the majority voting method can be used, that is, the category with the most prediction results is selected as the final output.

[0241] 4. Stacking: The prediction results of the two models are input as new features into another model (which can be logistic regression or other simple models) for learning and prediction. The stacking method can achieve better performance, especially when the base learners are significantly different.

[0242] Model Fusion

[0243] In practical applications, XGBoost and RF can be used as base learners and further integrated using the stacking method. The specific steps are as follows:

[0244] 1. Train XGBoost and RF models and obtain their prediction outputs (predicted values ​​for regression problems and probability distributions or categories for classification problems).

[0245] 2. These predictions are used as features and fed into a secondary learner (such as linear regression or logistic regression). The task of the secondary learner is to make a final decision based on these predictions.

[0246] Evaluate and adjust

[0247] Use cross-validation to evaluate the performance of the hybrid model and adjust the weighting parameters or learners in the stacked model based on the results to optimize the final model effect.

[0248] Extreme Gradient Boosting (XGBoost) is a gradient boosting model that can improve the accuracy of the model by boosting multiple decision trees. It is very powerful in handling complex nonlinear relationships and feature interactions.

[0249] Random Forest (RF) is an ensemble learning method that can effectively reduce variance and improve model stability by building multiple decision trees and averaging their prediction results. Combining these two models can achieve better results in prediction accuracy than a single model because they can complement each other in different aspects, reduce the risk of overfitting, and improve the robustness of the model.

[0250] The coefficient of determination R 2 , the root mean square error RMSE is used as the evaluation index of the regression model. 2 The larger the value, the smaller the RMSE. RMSE c is the result of the training set, RMSE p The result of the calibration set is better.

[21] , where the calculation formula is shown in formula (1) and formula (2)

[0251]

[0252] Note: N is the number of samples, y i is the actual value of the i-th sample, is the predicted value of the i-th sample, is the actual mean of all samples.

[0253] 2. Results and Discussion

[0254] 2.1 Spectral preprocessing

[0255] Using MSC, SNV, SG, FD, SD and other five spectral data preprocessing methods combined with partial least squares method (PLSR), support vector machine (SVR), random forest (RF), different regression models of chlorophyll content in ginkgo leaves were established. The prediction results of each model are shown in Table 1. and All of them are higher than the original spectrum and can detect the chlorophyll content of ginkgo leaves. Among these models, the MSC model has the highest accuracy, that is, is 0.8776, RMSE pThe value of the MSC preprocessing model is 0.4875, so the MSC preprocessing model is used as the basis for subsequent data processing. The hyperspectral raw reflectance and the optimal preprocessing reflectance are shown in Figure 3 (a) and Figure 3 (b). The weak absorption peak near 820nm is related to the stretching vibration of the NH triplet state. A smaller absorption peak appears near 930nm, which is mainly caused by the dual-frequency absorption of OH. The peaks and troughs near 560nm and 690nm are caused by chlorophyll and carotenoids contained in ginkgo leaves.

[0256] Table 1 Comparative analysis of different preprocessing methods and original spectra PLSR, SVR, RF models

[0257]

[0258] 2.2 Results of feature band extraction

[0259] The characteristic bands after MSC preprocessing were screened by continuous projection SPA method, and 23 characteristic bands were screened out. Figure 4 The distribution of these selected characteristic bands in the spectral range is shown. The specific characteristic band information and the number of bands are shown in Table 2. From the results of band screening, these 23 characteristic bands are mainly concentrated in the spectral range between 920 and 1000 nm, reducing redundant information and retaining the most representative spectral features, enhancing the accuracy and reliability of data analysis.

[0260] The competitive adaptive reweighted sampling CARS method was used to select the characteristic bands of the data after MSC preprocessing. Finally, 78 characteristic bands were screened out from the preprocessed data, and the information of the selected characteristic bands is presented in Table 2. The screening results show that the selected characteristic bands are mainly concentrated between 400-430nm and 820-1000nm. The band between 400-430nm covers the key information related to the initial spectral characteristics of ginkgo leaves, while the band between 820-1000nm provides rich information about ginkgo leaves between the middle and high bands. The combination of these two band ranges ensures the key information of spectral data in the full-band spectral region and optimizes the data processing and feature extraction process. Figure 5 The distribution of the selected characteristic bands in the full-band spectrum is displayed, intuitively showing the selection results of the characteristic bands.

[0261] The characteristic bands of the data after MSC preprocessing were screened by principal component analysis (PCA), and 21 characteristic bands were finally screened out. The information of the screened characteristic bands is presented in Table 2. The screened characteristic bands are mainly concentrated between 400 and 450 nm. This band selection result is significantly different from the results of the SPA algorithm and CARS algorithm used above. The characteristic band selection of the SPA algorithm and the CARS algorithm is mainly concentrated in the spectral range of 820 to 1000 nm, while the PCA method mainly screens the characteristic bands between 400 and 450 nm, reflecting the different emphasis and sensitivity of different methods on the extraction of spectral data features. Figure 6 The distribution of characteristic bands screened by the PCA method within the spectral range is shown, providing an intuitive understanding of the characteristic band selection results.

[0262] Table 2 Feature bands selected by different algorithms

[0263]

[0264]

[0265] 2.3 Prediction results based on different machine learning methods and feature band selection methods

[0266] The full band and the bands selected by features are used as input variables for machine learning, where A1 is the full band (224 bands), A2, A3, and A4 are the feature bands selected by the continuous projection method, the competitive adaptive weighted sampling method, and the principal component analysis method, respectively. A1 to A4 are used as inputs of different machine learning models, and the chlorophyll content of ginkgo leaves is used as output. The results are shown in Table 3. As shown in Table 3, when the machine learning model is SVR, when the input variable is the full band A1, is 0.6927, RMSE p is 1.3879, compared with A2, A3, and A4, Reduce,RMSE p Improve, when the machine learning model is RF, and the input variable is full-band A1, RMSE p =1.2789, compared with A2, A3, A4, Reduce,RMSE p When A3 is the input variable, the prediction accuracy of the RF model is the highest. RMSE p=0.511. In order to more clearly and intuitively show the prediction accuracy of each model, A1~A4 are used as the input of different machine learning models, and the chlorophyll content of ginkgo leaves is used as the output to draw heat maps and bar graphs. The heat maps and bar graphs are shown in Figure 7(a) and Figure 7(b). As can be seen from the heat map of Figure 7(a), the darker the color and the closer the value is to 1, the better the prediction accuracy. The values ​​corresponding to the horizontal axis A3 and the vertical axis XGB-RF in the figure are the largest, with a maximum value of 0.9033, indicating that under the input of A3, the XGB-RF model has the best prediction accuracy. In the bar graph 7(b), the part corresponding to A3 of the model XGB-RF accounts for the largest proportion and the highest value relative to the parts corresponding to A1~A4 of other models, that is, R 2 The largest, indicating that the model XGB-RF has the best prediction accuracy. Combined with the prediction accuracy of the XGB-RF model known in Table 3 RMSE p =0.3652.

[0267] In order to visualize the distribution ratio of chlorophyll in Ginkgo biloba leaves, the XGB-RF model with better prediction performance was used, and the calibration results of A3 variable analysis were applied to each pixel of the image. Figure 8 The chlorophyll value predicted for each pixel of the ginkgo leaf is shown according to the color scale on the right, with the color gradually changing from blue to red, with red indicating a higher chlorophyll value. In addition, the color changes in the ginkgo leaves further indicate that the distribution of chlorophyll values ​​is uneven, which may be related to the height of the trees and the differences in nutrient supply to each tree.

[0268] In summary, when the A3 variable is combined with the XGB-RF model, the prediction accuracy of the chlorophyll content of ginkgo leaves is the highest, i.e. RMSE p =0.3652, the results of the optimal prediction model are as follows Fig. 9 shown.

[0269] Table 3 Results of the prediction model for chlorophyll content in Ginkgo biloba leaves

[0270]

[0271] 3 Conclusion

[0272] This paper takes ginkgo leaves as the research object, collects the average spectrum of ginkgo leaf samples at 400-1000nm through a hyperspectral imaging system, uses a portable SPAD-502Plus chlorophyll meter to measure the chlorophyll content of ginkgo leaves, and proposes a prediction method for ginkgo leaf chlorophyll content using hyperspectral data inversion. This method preprocesses and reduces the dimension of hyperspectral data, and uses machine learning to predict the chlorophyll content. The main results are as follows:

[0273] (1) The original spectrum of ginkgo leaves was preprocessed using the MSC, SG, SNV, FD, and SD methods. According to the results obtained by the PLSR regression model, it was found that the MSC preprocessing method was more effective. Subsequent processing was carried out based on this preprocessing method, and the best preprocessing algorithm was obtained through comparison.

[0274] (2) The MSC-preprocessed spectra were subjected to characteristic wavelength selection using SPA, CARS, PCA and other methods. The MSC-preprocessed spectra obtained 23, 78 and 21 characteristic wavelengths using the above methods, respectively, accounting for 9.5%, 3.7% and 17.8% of the full spectrum band.

[0275] Based on the above preprocessing and characteristic wavelength extraction methods, PLSR and SVR, RF, XGBoost and XGB-RF were used to establish the relationship models between the full band, characteristic wavelength and chlorophyll value of ginkgo leaves respectively.

[0276] The results show that the characteristic wavelength prediction model has less computational effort and better accuracy than the full-band prediction model.

[0277] Among them, the MSC-CARS-(XGB-RF) algorithm combination is the optimal combination. Compared with the PLSR, XGBoost, and SVR models of the characteristic band, the MSC-CARS-(XGB-RF) model has better prediction performance for the chlorophyll value of ginkgo leaves, and its determination coefficient of the calibration set is The root mean square error is 0.9325, RMSE c The determination coefficient of the prediction set is 0.3478. The root mean square error is 0.9033, RMSE p It is 0.3652.

[0278] In summary, the combination of MSC-CARS-(XGB-RF) model can achieve the best prediction of Ginkgo biloba chlorophyll value, providing a reliable and effective new non-destructive detection method for Ginkgo biloba chlorophyll value. Accurate monitoring of chlorophyll content in Ginkgo not only improves prediction accuracy, but also has a potential impact on its growth and production. This monitoring helps to deeply understand the growth status and nutritional needs of plants, guide rational fertilization and irrigation, and thus improve crop yield and quality. In addition, this study can establish an effective diagnostic system for the early prediction of Ginkgo biloba leaf chlorosis, and provide a theoretical basis and technical support for the long-term dynamic observation of chlorophyll content. Combined with hyperspectral technology and machine learning algorithms, the application of this model is expected to deepen the mining of plant physiological information and achieve a comprehensive assessment of plant health status. With the development of cloud computing and big data technology, larger data sets can be processed to further improve the generalization ability and prediction accuracy of the model.

[0279] References

[0280] [1].Lu F,Jiayi S,Guocai Z,et al.Protective Effect of Ginkgobiloba(Ginkgoaceae Family)Leaf Extracts on Female Drosophila melanogaster Sufferingfrom Ultraviolet Irradiation through Activating the Keap1-Nrf2 SignalingPathway[J].Biology Bulletin,2024,51(3):518-529.

[0281] [2].Yanxia L,Huawei X,Yunchao Z,et al.Leaves,seeds and exocarp ofGinkgo biloba L.(Ginkgoaceae):A Comprehensive Review of Traditional Uses,phytochemistry,pharmacology,resource utilization and toxicity.[J].Journal ofethnopharmacology,2022,298115645-115645.

[0282] [3].Sen D,Rathee S,Pandey V,et al.Comprehensive Insights intoPathophysiology of Alzheimer's Disease:Herbal Approaches for MitigatingNeurodegeneration.[J].Current Alzheimer research,2024,

[0283] [4].Vitor C A D,Gabriela A S,Maria S C.Standardized extract of Ginkgobiloba treatment and novelty on the weak encoding of spatial recognitionmemory in rats.[J].Learning&memory(Cold Spring Harbor,N.Y.),2023,30(4):85-95.

[0284] [5].Shi H.Exploring vegetation chlorophyll fluorescence with leafdorsiventrality[J].Remote Sensing of Environment,2024,313114342-114342.

[0285] [6].Wang Z,Xu H,Wang F,et al.EMS-induced missense mutation in TaCHLI-7D affects leaf color and yield-related traits in wheat[J].Theoretical andApplied Genetics,2024,137(10):223-223.

[0286] [7].Bylappa Y,Nag A.A statistical approach to study anatomicalchanges of pink guava cultivar(Psidium guajava L.cv Arka Kiran)during itsripening at the room temperature storage[J].Kuwait Journal of Science,2025,52(1):100295-100295.

[0287] [8].Custureri G M I,GiuffrèM A,Loizzo R M,et al.Bergamot flavouredolive oil:Comparison between enrichment processes,evaluation of shelf-lifeand health properties[J].Applied Food Research,2024,4(1):100400-.

[0288] [9].Marius L,Elvyra J,Jurgita K,et al.Studies of the Variability ofBiologically Active Compounds and Antioxidant Activity in Organically,Biodynamically,and Naturally Grown and Fermented Fireweed(Chamerionangustifolium(L.)Holub)Leaves.[J].Plants(Basel,Switzerland),2023,12(12):

[0289]

[10] .Delpino-Rius A,Cosovanu D,Eras J,et al.A fast and reliableultrahigh-performance liquid chromatography method to assess the fate ofchlorophylls in teas and processed vegetable foodstuff[J].Journal ofChromatography A,2018,156869-79.

[0290]

[11] .Yuanzhe W,Lifu C,G R C.Selective Voltammetric Detection ofChlorophylls Using a Semi-circular Potential Wave[J].Food Chemistry,2020,323(prepublish):126844.

[0291]

[12] .Song Z,Liu Y,Yu J,et al.Estimation of Chlorophyll Content inApple Leaves Infected with Mosaic Disease by Combining Spectral and TexturalInformation Using Hyperspectral Images[J].Remote Sensing,2024,16(12):2190-2190.

[0292]

[13] .Long T,Tang X,Liang C,et al.Detecting bioactive compoundcontents in Dancong tea using VNIR-SWIR hyperspectral imaging and KRR modelwith a refined feature wavelength method[J].Food Chemistry,2024,460(P2):140579-140579.

[0293]

[14] .Yu S,Bu H,Hu X,et al.Establishment and Accuracy Evaluation ofCotton Leaf Chlorophyll Content Prediction Model Combined with HyperspectralImage and Feature Variable Selection[J].Agronomy,2023,13(8):

[0294]

[15] .Ruomei Z,Lulu A,Weijie T,et al.Improving chlorophyll contentdetection to suit maize dynamic growth effects by deep features ofhyperspectral data[J].Field Crops Research,2023,297

[0295]

[16] .Shuangya W,Nan S,Junwei L,et al.Continuous Wavelet Transform andBack Propagation Neural Network for Condition Monitoring ChlorophyllFluorescence Parameters Fv / Fm of Rice Leaves[J].Agriculture,2022,12(8):1197-1197.

[0296]

[17] .Ruomei Z,Lulu A,Weijie T,et al.Deep learning assisted continuouswavelet transform-based spectrogram for the detection of chlorophyll contentin potato leaves[J].Computers and Electronics in Agriculture,2022,195.

[0297]

[18] .Sen L,Tingting Y,Shaowei Z,et al.A method for screening copper-tolerant rice(Oryza sativa L.)cultivars based on hydroponic experiments andcluster analysis.[J].International journal of phytoremediation,2017,19(12):1093-1099.

[0298]

[19] .Lu F,Gao X,Ma J,et al.Intelligent marine detection based onspectral imaging and neural network modeling[J].Ocean Engineering,2024,310(P1):118640-118640.

[0299]

[20] .Ali K M,Amin B,Maud R A,et al.Hyperspectral target detectionusing self-supervised background learning[J].Advances in Space Research,2024,74(2):628-646.

[0300]

[21] .Long T,Tang X,Liang C,et al.Detecting bioactive compoundcontents in Dancong tea using VNIR-SWIR hyperspectral imaging and KRR modelwith a refined feature wavelength method[J].Food Chemistry,2024,460(P2):140579-140579.

Claims

1. A method for estimating chlorophyll content in ginkgo canopy based on airborne hyperspectral imaging, characterized by the following steps: include: First, hyperspectral images of ginkgo leaves in the canopy were collected; Then, the collected images are used as input, and the corresponding chlorophyll content is estimated by using the ginkgo leaf chlorophyll content prediction model; the method for collecting the canopy ginkgo leaf hyperspectral image is to use a hyperspectral imager carried by an unmanned aerial vehicle to obtain the ginkgo leaf hyperspectral image, and the ginkgo leaf hyperspectral image is a remote sensing image; The ginkgo leaf chlorophyll content prediction model is the MSC-CARS-(XGB-RF) model, and its establishment process is as follows: S1 hyperspectral image correction and ROI extraction S1.1 Hyperspectral image correction: Perform radiation correction, geometric correction and atmospheric correction on hyperspectral images; S1.1.

1. Radiation correction Calibrate the raw data in remote sensing images to eliminate the impact of sensor errors, noise and other environmental factors on the images, and ensure that the images reflect the true radiation information of the objects; S1.1.

2. Geometric correction Align the spatial coordinate system of the remote sensing image with the coordinate system of the real world to eliminate geometric distortion caused by factors such as the curvature of the earth, sensor angle, platform motion, etc. After geometric correction, the image can accurately reflect the real position and shape of ground objects; S1.1.

3. Atmospheric correction Remove the influence of water vapor, aerosol, cloud and other substances in the atmosphere on remote sensing images, so as to accurately extract the surface information of the ground objects from the images; particles and gases in the atmosphere will scatter and absorb solar radiation, thereby changing the radiation value reaching the ground and affecting the accuracy of remote sensing data; S1.2 Hyperspectral image ROI extraction First, extract the selected area ROI of the hyperspectral image; then, construct the hyperspectral dataset; S2 Hyperspectral Data Preprocessing Multivariate scatter correction (MSC) is used to pre-process the hyperspectral data to eliminate the interference of spectral data caused by light scattering and avoid the deviation and deformation of spectral data caused by light scattering. S3 feature band extraction The competitive adaptive reweighted sampling algorithm CARS is used to extract characteristic bands from the data obtained from S2, and the characteristic band spectral data set is selected for subsequent modeling, analysis or prediction tasks; S4 training prediction model The XGB-RF hybrid model is used as the prediction model; the XGB-RF hybrid model integrates the extreme gradient boosting tree XGBoost machine learning model and the random forest RF machine learning model, and combines the prediction results of XGBoost and RF; The characteristic band spectral data obtained by S1 to S3 and the corresponding measured chlorophyll values ​​are substituted into the XGB-RF hybrid model to obtain the final prediction model. The steps include: S4.1 Train XGBoost and RF models separately Train the XGBoost and RF models separately to obtain their respective prediction results; use cross-validation to select their respective optimal hyperparameters; a. For the XGBoost model, the gradient boosting tree (GBDT) algorithm is used to perform weighted combination through multiple rounds of weak classifiers, i.e., trees; b. For the RF model, the final decision is made by building multiple decision trees for voting or averaging; S4.2 Combination of prediction results Methods for combining the prediction results of XGBoost and RF include: a. Weighted average method: weighted average the prediction results of ginkgo leaf chlorophyll according to the performance of the model, giving priority to the results of the model with higher accuracy; b. Simple average method: directly average the prediction results of the two models. This method is suitable when the performance of the models is relatively balanced; c. Voting method: select the category with the most prediction results as the final output; d. Stacking method: The prediction results of the two models are input as new features into another simple model for learning and prediction; S4.3 Model fusion XGBoost and RF are used as base learners and further integrated using the stacking method. The steps include: S4.3.

1. Train XGBoost and RF models and obtain their prediction outputs; For regression problems, the prediction output is the predicted value; for classification problems, the prediction output is the probability distribution or category; S4.3.

2. These predicted outputs are input as features into a secondary learner, and the secondary learner makes the final decision based on the predicted outputs; S4.4 Evaluation and Adjustment Evaluate the performance of the hybrid model and adjust the weighting parameters or learners in the stacked model based on the results to optimize the final model effect; When estimating the chlorophyll content of the ginkgo leaf, the input of the prediction model is the hyperspectral data obtained by the hyperspectral imager carried by the drone to obtain the hyperspectral image of the ginkgo leaf and processed by step S1, and the output of the prediction model is the corresponding chlorophyll content of the ginkgo leaf.

2. The method for estimating chlorophyll content in ginkgo canopy based on airborne hyperspectral imaging according to claim 1, characterized in that The steps of S1.1.1 include: a) Sensor calibration: radiometric correction of the image based on the sensor's calibration coefficients, using known reflectivity or radiometric calibration data; b) Conversion from digital values ​​to radiation values: Remote sensing images are represented by digital values, which have a corresponding relationship with the ground reflectivity or radiation brightness. These digital values ​​are converted into the reflectivity or radiation brightness of physical quantities; c) Absolute radiation correction, that is, converting image data into physical reflectivity or radiation brightness of ground objects by comparing with the true reflectivity of the ground; d) Relative radiation correction, that is, using multi-phase images for relative comparison and adjusting the brightness difference between images; e) Remove sensor noise or other interfering signals.

3. The method for estimating chlorophyll content in ginkgo canopy based on airborne hyperspectral imaging according to claim 1, characterized in that The steps of S1.1.2 include: a) Geometric model selection: It is necessary to select an appropriate geometric model to describe the geometric deformation of the image; the geometric model includes a projection model, an affine transformation model or a polynomial transformation model; b) Ground control point (GCP) selection: select easily identifiable ground control points in the image and match them with actual geographic coordinates; ground control points come from ground measurement data or existing maps; c) Geometric transformation: Based on the selected geometric model and ground control points, the image is geometrically corrected using a transformation algorithm; the transformation method is affine transformation, high-order transformation or resampling.

4. The method for estimating chlorophyll content in ginkgo canopy based on airborne hyperspectral imaging according to claim 1, characterized in that The steps of S1.1.3 include: a) Application of radiation transfer models: Use radiation transfer models to simulate the influence of the atmosphere on ground radiation; through these models, calculate the influence of the atmosphere on each band in the remote sensing image, and then make corrections; b) Meteorological data acquisition: Acquisition of meteorological data corresponding to the time and location of remote sensing image acquisition, including atmospheric temperature, humidity, air pressure and wind speed; c) Ground reflectivity calculation: convert the radiation value in the remote sensing image into the reflectivity of the ground object or the ground radiation brightness, thereby eliminating the influence of the atmosphere on the radiation; d) Scattering and absorption correction: Quantitative correction of aerosol scattering and / or gas absorption atmospheric effects is performed using ground reflectivity and atmospheric conditions to eliminate the interference of atmospheric effects on image data.

5. The method for estimating chlorophyll content in ginkgo canopy based on airborne hyperspectral imaging according to claim 1, characterized in that In S1.2, the average method is used to construct the hyperspectral dataset; the spectral characteristics of the region are represented by calculating the average value of all pixels in the ROI in each band; If the ROI spectrum changes steadily or the objects in the ROI are consistent, the spectrum of the ROI is suitable for effective description by its average value, as follows: Assume that the size of the hyperspectral image is M×N×B, where M is the number of rows, N is the number of columns, and B is the number of bands. For ROI, extract the value of each pixel in the region in all bands and calculate the average value of each band; assuming that ROI contains P pixels, the number of bands is B, and the spectral value of each pixel in the bth band is Q b (n), where n = 1, 2, ..., P represents the nth pixel in the ROI; Then, the average value of the bth band is Calculated by the following formula: in: is the average value of the bth band, Q b (n) is the value of the nth pixel in the ROI region on b, and P is the number of pixels in the ROI region.

6. The method for estimating chlorophyll content in ginkgo canopy based on airborne hyperspectral imaging according to claim 1, The characteristic is that in S2, the steps of MSC preprocessing the hyperspectral data include: S2.

1. Spectral data standardization and deaveraging The spectral data is mean removed to eliminate the DC offset component in the spectral data. The method is: Assume that there is a set of sample spectral data Y = [y1, y2, ..., y n ], where each y i is a p-dimensional spectral vector, representing the spectrum of the i-th sample; Perform mean removal on the spectral data of each sample: and' i =and i -μ y Among them, y' i is the spectral vector of the sample after removing the mean, μ y is the mean vector of the spectral data of all samples: S2.

2. Fitting the baseline scattering Assuming that there is some global scattering effect in the spectral data of the sample, the scattering effect is described by a baseline model; In order to correct this scattering effect, a scattering model is first established and fitted by a linear regression model; In order to correct the scattering effect of each spectrum, a suitable baseline model is found, which is assumed to be: y i '=β0+β1y i +e i Among them, β0 is the bias term, β1 is the regression coefficient of the spectral data, and ε i is the error term; these coefficients are estimated through regression analysis to obtain the scattered part of the spectrum; S2.

3. Calculate the principal components of the spectrum, i.e. the regression coefficients Multiple linear regression is used to estimate the scattering effect of each spectrum. The relationship between the spectral data and a baseline model is calculated through linear regression. At this time, the regression coefficient β1 represents the correlation between the sample spectrum and the baseline model. Use ordinary least squares (OLS) to fit a regression model for the scattering effect: y=Xβ Where y is the scattering component predicted by the model, X is the original data matrix, i.e., the spectrum without the mean, and β is the regression coefficient; S2.

4. Correction of spectral data The scattering baseline model obtained by regression is used to correct the original spectral data; the spectral data of each sample is subtracted from its fitted scattering component:

7. The method for estimating chlorophyll content in ginkgo canopy based on airborne hyperspectral imaging according to claim 1, The characteristic is that in S3, CARS is used to extract characteristic bands from the data processed by S2, and the steps include: S3.

1. Data preprocessing Denoising and standardization: standardize or normalize the data so that the mean of each feature is 0 and the variance is 1; S3.

2. Initialize feature set The selected feature set is all the features extracted from the raw data. For spectral data, the features are the reflection intensity at different wavelengths; S3.

3. Adaptive reweighting Weighting process: Through the weighting mechanism, the weight of each feature is dynamically re-evaluated, important features are given higher weights, and the influence of irrelevant features is reduced; weighting is achieved by setting the re-weighting coefficient of the feature, and the size of the coefficient reflects the contribution of the feature to the model; Weighted calculation: Weighting is based on the adaptive criterion. CARS continuously adjusts these weights according to the importance of the features. For each round of weighting, the features that contribute most to the objective function are adaptively selected for retention. S3.

4. Feature selection process Through multiple iterative weighting, the feature set will continue to change; in each iteration, CARS optimizes the feature selection by reselecting a subset of features and weighting them. In each iteration, CARS builds models based on the weighted feature subsets and evaluates the performance of the model; features with low weights are removed or down-weighted; S3.

5. Reweighting Iterations After each iteration, CARS makes adaptive adjustments based on the weight of each feature. Features with low weights are reduced or discarded, while features with higher weights are retained. In each iteration, feature subsets are continuously adjusted until the optimal feature set is found. S3.

6. Model evaluation and selection Performance evaluation: In each iteration, the prediction ability of the model is evaluated using the training set and the test set. CARS maximizes the prediction ability of the model by selecting the most relevant features. Stopping criteria: When the predetermined number of iterations is reached or the result of feature selection no longer significantly improves the prediction accuracy, the iteration stops; S3.

7. Final feature set At the end of the iteration, CARS returns an optimal feature subset.

Images