Remote sensing retrieval method for vegetation chlorophyll content based on red edge vegetation index and extreme random tree
By combining Sentinel-2 multispectral remote sensing images and ground-based measured data, and employing an extreme random tree model, a leaf chlorophyll content inversion method was constructed by screening characteristic variables. This method solved the problem of insufficient inversion stability under complex forest structures and achieved high-precision monitoring of leaf chlorophyll content.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HANGZHOU NORMAL UNIVERSITY
- Filing Date
- 2026-03-27
- Publication Date
- 2026-07-14
AI Technical Summary
Existing remote sensing-based leaf chlorophyll content inversion methods lack stability and generalization ability under complex forest structure conditions, are easily affected by background factors, and are difficult to achieve accurate quantitative monitoring.
Using visible light, near-infrared, and red-edge band information from Sentinel-2 multispectral remote sensing images, combined with ground-measured data, a quantitative inversion relationship of leaf chlorophyll content was established through a machine learning model. The inversion model was constructed using the extreme random tree method, and the MERIS terrestrial chlorophyll index, modified simple ratio index, modified red-edge normalized difference vegetation index, and red-edge chlorophyll index were selected as feature variables to reduce the risk of overfitting and improve model stability.
It achieves accurate estimation and spatial distribution representation of leaf chlorophyll content in complex forest areas, improves inversion accuracy and model stability, and has a wide range of applications.
Smart Images
Figure CN122391905A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of ecological remote sensing and vegetation physiological parameter inversion technology, specifically a remote sensing inversion method for chlorophyll content in forest vegetation leaves based on multispectral remote sensing data and machine learning algorithms. This method utilizes Sentinel-2 multispectral remote sensing image data, combines red-edge bands to construct vegetation index features, and establishes a quantitative relationship between vegetation spectral features and leaf chlorophyll content through a machine learning model. This enables remote sensing inversion and spatial distribution estimation of chlorophyll content, and is primarily applicable to the quantitative remote sensing inversion and regional-scale spatial distribution monitoring of chlorophyll content in leaves of multiple tree species and other subtropical forest vegetation. Background Technology
[0002] Chlorophyll is the core pigment for photosynthesis in plants, and its content directly affects the efficiency of vegetation in absorbing and converting light energy. It is an important biochemical parameter characterizing the physiological state and primary productivity of vegetation. Leaf Chlorophyll Content (LCC) not only reflects the growth status and nutrient level of vegetation, but also exhibits a sensitive response to environmental stress, climate change, and anthropogenic disturbances. In the context of global climate change, quantitatively monitoring the spatiotemporal changes of LCC in forest ecosystems is of significant scientific importance for assessing regional ecosystem functions and revealing vegetation growth dynamics and carbon cycling processes.
[0003] Traditional LCC (Limited Chloride Coefficient) measurements primarily rely on field sampling and indoor chemical analysis. While these methods offer high accuracy, they suffer from limitations such as being highly destructive, labor-intensive, and lacking spatial representativeness, making them unsuitable for large-scale continuous monitoring. With the development of remote sensing technology, LCC retrieval based on vegetation spectral characteristics provides an effective approach for large-scale, non-destructive monitoring. Green vegetation exhibits significant chlorophyll absorption in the visible light band and high reflectivity in the near-infrared band, forming a typical "red edge" phenomenon at the boundary between red and near-infrared light. Numerous studies have shown that the red edge band is highly sensitive to LCC changes and is a crucial source of spectral information for constructing vegetation indices and retrieving biochemical parameters. In particular, the multiple red edge bands carried by the Sentinel-2 satellite provide richer spectral information support for the quantitative estimation of regional-scale vegetation LCC.
[0004] Methods for chlorophyll retrieval based on remote sensing data mainly include empirical vegetation index methods and machine learning methods. Vegetation indices construct spectral feature parameters through combinations of different bands, offering advantages such as ease of calculation and clear physical meaning. However, a single index is easily affected by background soil, canopy structure differences, and spectral saturation effects, limiting model stability and regional applicability. In recent years, machine learning methods such as random forests, gradient boosting trees, support vector regression, and artificial neural networks have been widely applied to vegetation biochemical parameter retrieval research. By exploring the nonlinear relationships between multiple features, they have improved model fitting ability and prediction accuracy. However, different algorithms still exhibit differences in applicability and generalization ability under complex forest stand structures, particularly in their ability to extract red-edge signals independent of leaf optical properties and their comprehensive performance in assessing irradiance (seasonal) changes, which are somewhat lacking. Furthermore, the risk of overfitting is relatively high. Summary of the Invention
[0005] The purpose of this application is to address the problems in existing remote sensing-based leaf LCC inversion methods, such as the susceptibility of a single vegetation index to background interference and the insufficient stability and generalization ability of the model under complex forest structure conditions, and to provide a quantitative inversion method for forest leaf LCC that integrates ground-measured data and multispectral remote sensing data.
[0006] The main principles of this application are as follows: This application establishes a quantitative inversion relationship of leaf LCC (Leaf Canopy Color Value) based on the influence mechanism of chlorophyll on the spectral reflectance characteristics of vegetation leaves, using remote sensing spectral information and machine learning methods. Chlorophyll in green vegetation exhibits significant absorption characteristics in the visible light band, especially in the blue and red light regions, while possessing high reflectance in the near-infrared band and forming a significant red edge feature in the transition region between red and near-infrared light. As leaf LCC changes, the spectral reflectance characteristics and red edge position of the vegetation canopy will change accordingly. Therefore, multispectral information obtained from remote sensing imagery can be used to characterize changes in vegetation LCC.
[0007] Based on this, this application utilizes visible light, near-infrared, and red-edge band information contained in Sentinel-2 multispectral remote sensing imagery, such as the MERIS Land Chlorophyll Index (MTCI), Modified Simple Ratio Index (mSR705), Modified Red-Edge Normalized Difference Vegetation Index (MNDVIred), and Red-Band Chlorophyll Vegetation Index (Cired_edge), to form multidimensional spectral feature variables. A training sample set is constructed by combining this with ground-measured leaf LCC data. A machine learning model is used to learn and model the nonlinear relationship between the multidimensional spectral features and leaf LCC, thereby establishing a remote sensing inversion model for leaf LCC. By using the trained model to process remote sensing images, quantitative inversion of leaf LCC in the forest vegetation of the study area can be achieved, and its spatial distribution information can be obtained.
[0008] The method for quantitative inversion of vegetation chlorophyll content by remote sensing described in this application mainly includes the following steps: S1: Data Acquisition and Sample Construction in the Study Area We acquired measured chlorophyll content data of vegetation leaves and Sentinel-2 multispectral remote sensing image data in the study area and established the spatial matching relationship between the two, thereby constructing a sample dataset containing ground measured chlorophyll content and remote sensing spectral information. S11: Acquisition of Actual Measurement Data This application selected Chang Le Forest Farm in Hangzhou, Zhejiang Province as the study area. Multiple survey plots were established based on the vegetation type and forest stand structure of the study area, with each plot measuring no less than 90 m × 90 m to ensure spatial representativeness in remote sensing imagery. A central quadrat and four corner quadrats were set up within each survey plot. Canopy leaf samples were collected from representative plants, and the geographic coordinates of the sampling points were recorded. The collected leaf samples were brought back to the laboratory, and the leaf LCC (Leaf Canopy Calibrates) was determined using chemical analysis methods, thereby obtaining ground-measured leaf LCC data for the study area.
[0009] S12: Sentinel-2 Multispectral Remote Sensing Image Acquisition and Preprocessing Sentinel-2 Level-2A surface reflectance multispectral remote sensing imagery of the study area was acquired using the Google Earth Engine cloud computing platform. This data product has undergone preprocessing such as radiometric calibration and atmospheric correction. By querying all available images for each month and statistically analyzing image cloud cover, candidate images were screened for quality. Based on a comprehensive consideration of image cloud cover and actual image quality, remote sensing images with low cloud cover and good image quality were selected as representative images for each month. The selected images were then cropped to reflect the study area.
[0010] S13: Spatial matching, sample dataset construction Based on the geographic coordinate information of the sample points, this application spatially matches the ground-measured leaf LCC samples with the corresponding Sentinel-2 remote sensing image pixels, extracts the multispectral reflectance data of the corresponding pixels at the matched sample point locations, including reflectance information in the visible light, red edge and near-infrared bands, and constructs a sample dataset containing LCC observation values and multispectral reflectance feature variables.
[0011] S2: Vegetation Index Feature Construction and Feature Analysis Multispectral reflectance data related to vegetation chlorophyll content were obtained from Sentinel-2 multispectral remote sensing images. Through calculation and analysis, the two independent dimensions that can express red edge spectral changes, as well as the combination of vegetation indices that can extract red edge signals without relying on leaf optical properties and irradiance changes were selected. S21: Vegetation Index Calculation Based on Sentinel-2 multispectral reflectance data, various vegetation indices related to LCC were calculated, including the Simple Ratio Index (SR), Normalized Difference Vegetation Index (NDVI), MERIS Terrestrial Chlorophyll Index (MTCI), Modified Red Edge Normalized Difference Vegetation Index (MNDVIred), Red Band Chlorophyll Vegetation Index (CIred-edge), Modified Difference Ratio Index (MDRI), Reciprocal Reflectance Index (IR700), Scatter Corrected Reflectance Ratio Index (SaRR), Pigment Specific Normalized Difference Index (PSND650), Reciprocal Reflectance Difference Index (DRP550), New Three-Band Index (NTBI), Modified Simple Ratio Index (mSR705), Normalized Green-Red Difference Index (NGRDI), Normalized Red Edge Vegetation Index (NDRE), Red Edge Ratio Index (RERVI), and Enhanced Vegetation Index (EVI). Through multi-band combination and scatter correction mechanisms, the response to LCC changes under complex canopy conditions was improved.
[0012] S22: Construction of Feature Variable Set and Correlation Analysis This application combines the calculated vegetation indices with the corresponding ground-measured leaf LCC data from the sampling points to construct a set of characteristic variables. Statistical analysis methods are then used to calculate the correlation coefficients between each vegetation index and the LCC, establishing a correlation matrix among the vegetation indices.
[0013] S23: Feature Selection and Sensitivity Index Determination This application screens feature variables by assessing the correlation between vegetation indices and leaf chlorophyll content (LCC), their expressive power for various aspects of chlorophyll content, and the information redundancy between indices. Ultimately, the following combination of feature variables with high sensitivity to changes in leaf LCC is selected: the MERIS Terrestrial Chlorophyll Index (MTCI, formula MTCI = (B6-B5) / (B5-B4)), the Modified Simple Ratio Index (mSR705, formula mSR705 = (B6-B1) / (B5-B1)), the Modified Red Edge Normalized Difference Vegetation Index (MNDVIred, formula MNDVIred = (B6-B5) / (B6+B5-B1)), and the Red Edge Chlorophyll Index (CIred-edge, formula CIred-edge = B6 / B5-1). The specific screening process is as follows: S231: Candidate index grouping and initial screening based on spectral mechanisms This application first groups the vegetation indices based on their band combination methods and their characterization mechanisms of chlorophyll absorption characteristics, aiming to cover all relevant dimensions of chlorophyll content in vegetation leaves. The 16 candidate vegetation indices are divided into four functional groups: (1) Red edge position sensitive group, including MTCI, NDRE, and RERVI, which mainly use the reflectance changes of the red edge region (705–783 nm) to characterize the red edge displacement characteristics caused by chlorophyll; (2) Red edge amplitude sensitive group, including mSR705, MNDVIred, CIred-edge, and SR, which characterize the red edge reflectance amplitude changes through the ratio of the red edge to the reference band or normalization calculation; (3) Broadband integrated group, including NDVI, EVI, and NGRDI, which mainly use the visible light and near-infrared broadband information; (4) Scattering / reciprocal correction group, including IR700, SaRR, PSND650, DRP550, NTBI, and MDRI, which eliminate the influence of canopy scattering through reciprocal transformation or multi-band correction.
[0014] S232: Correlation-based screening and elimination of intragroup redundancy and intergroup complementarity analysis In the red-edge position sensitive group, the MERIS terrestrial chlorophyll index, which best represents the red-edge position, was selected based on characteristic analysis and correlation matrix analysis. The correlation coefficient between RERVI and CIred-edge was 0.97, forming a highly redundant pair. In the case of selecting CIred-edge in the red-edge amplitude sensitive group, the RERVI index was removed. In the red-edge amplitude sensitive group, the modified simple ratio index, by introducing the blue light band B1, corrects the leaf scattering effect, effectively eliminating the systematic bias of the red-edge ratio caused by specular reflection due to the leaf wax layer and epidermal trichomes. The modified red-edge normalized difference vegetation index adopts a normalized form, which can effectively suppress the influence of irradiance changes caused by differences in solar altitude angle and observation geometry, and contributes to cross-period (seasonal) consistency. The red-edge chlorophyll index and chlorophyll content maintain an approximately linear proportional relationship in the medium-high concentration range (35–60 μg / cm²), while SR has shown a clear saturation trend in this range. Therefore, based on characteristic analysis and correlation analysis, this group retains the modified simple ratio index, the modified red-edge normalized difference vegetation index, and the red-edge chlorophyll index, and removes SR. In the broadband composite group, the Normalized Difference Vegetation Index (NDVI) exhibited a severe saturation effect under high vegetation cover conditions in this study area, with a correlation coefficient of only 0.41 with the LCC. While the Enhanced Vegetation Index (EVI) partially mitigated the effects of atmospheric scattering by introducing the blue light band, its design was intended to characterize canopy structure parameters, resulting in limited sensitivity to changes in chlorophyll content at the leaf scale, with a correlation coefficient of only 0.35 with the LCC. The Normalized Green-Red Difference Index (NDRI), utilizing only the visible green and red bands, was susceptible to atmospheric aerosol and water vapor absorption, with a correlation coefficient of only 0.22 with the LCC. Therefore, all three indices in the broadband composite group were excluded.
[0015] In the scattering / reciprocal correction group, the reciprocal reflectance indices were redundant because their correlation coefficient with LCC was lower than that of the selected indices; the new three-band index, the scattering-corrected reflectance ratio index, and the modified simple ratio index had high correlation coefficients, indicating information redundancy; the pigment-specific normalized difference index did not adequately represent the overall expression of chlorophyll; and the modified difference ratio index had a high correlation coefficient with the red-edge chlorophyll index, indicating redundancy. Therefore, all six indices in this group were removed.
[0016] Spectral complementarity and synergistic effect analysis of the final feature combination After the above screening, the optimal feature variable combination was determined to be composed of four vegetation indices: the MERIS terrestrial chlorophyll index, the modified simple ratio index, the modified red-edge normalized difference vegetation index, and the red-edge chlorophyll index. This combination comprehensively and simultaneously represents the two independent dimensions of red-edge spectral changes, as well as the ability to extract red-edge signals independently of leaf optical properties (multi-species integrated expression) and the comprehensive expression of irradiance (seasonal) changes. Furthermore, it has less redundant information, a simple data structure, low computational cost, and low risk of overfitting. This combination possesses the following spectral complementarity characteristics: (a) The MERIS terrestrial chlorophyll index uses the difference ratio of three red edge bands to mainly capture the red edge position shift information caused by chlorophyll changes, which belongs to the red edge "displacement type" feature; the red edge chlorophyll index uses the simple ratio of B6 / B5 to mainly reflect the amplitude change of the red edge reflectance steep slope segment, which belongs to the red edge "amplitude type" feature; the two represent two independent dimensions of red edge spectral change (displacement direction and amplitude direction), and their information cross-correlation coefficient is 0.61, which is lower than the intragroup redundancy threshold of 0.85, confirming that they are complementary rather than redundant. (b) The modified simple ratio index provides the ability to extract red edge signals without relying on the optical properties of the leaf surface by correcting leaf scattering in the blue light band. It has cross-species stability advantages for mixed forests with significant differences in leaf structure in the study area. (c) The normalization operation form of the red-edge normalized difference vegetation index is modified to make it robust to irradiance changes. In the monthly multi-time re-evolution scenario, it can effectively suppress the false signal caused by seasonal solar altitude angle changes and complement the other three indices in the time dimension. S3: Construction of a vegetation chlorophyll content inversion model based on the extreme random tree method S31: Training Sample Construction This application uses the selected combination of highly sensitive vegetation indices as the model input and the ground-measured leaf LCC of the corresponding sample points as the model output to construct LCC inversion training samples. S32: Model Training and Accuracy Evaluation Based on the constructed training samples, an inversion model for vegetation chlorophyll content is constructed using the Extra Trees method.
[0017] The sample data was divided into training and test sets in a 7:3 ratio, and stratified sampling was used. During model training, the key parameters of the model were automatically optimized based on the characteristics of the training data to obtain the optimal parameter configuration of the model, and the model training was completed under the optimal parameter conditions. The accuracy of the vegetation chlorophyll content inversion model was validated using a test set, and the coefficient of determination (R²) was selected. 2 R0 and normalized root mean square error (nRMSE) are used as evaluation metrics, where R0 2 nRMSE is used to characterize the model's ability to explain the measured LCC variation. It measures the relative magnitude of the model's prediction error, and the formula is: In the formula, Indicates the first i The measured LCC of a sample This represents the corresponding model prediction value. This is the average value of the measured LCC. n This represents the number of samples.
[0018] S4: Comparison of the inversion results of the vegetation chlorophyll content inversion model constructed in this application with other major models. Based on the training samples constructed using S31, leaf LCC inversion models were constructed using the remaining 8 machine learning algorithms. The 8 machine learning methods include: Random Forest (RF), Gradient Boosting Decision Tree (GBDT), Category Feature Gradient Boosting (CatBoost), Support Vector Regression (SVR), Artificial Neural Network (MLP), K Nearest Neighbor Regression (KNN), Adaptive Boosting (AdaBoost), and Bagging.
[0019] Based on the accuracy evaluation results of all models on the test set, the inversion performance of different models was comprehensively compared, and the Extra Trees inversion model was determined to be the optimal LCC inversion model.
[0020] S5: Inversion and Spatial Distribution Analysis of Vegetation Chlorophyll Content S51: Remote sensing inversion of vegetation chlorophyll content This application utilizes a vegetation chlorophyll content inversion model, taking vegetation index characteristic variables calculated from Sentinel-2 multispectral remote sensing images as input, to perform pixel-by-pixel calculations on remote sensing images of the study area for each month, obtaining leaf LCC prediction values for corresponding pixels, and generating multi-temporal LCC inversion result data.
[0021] S52: Analysis of Spatial Distribution and Change Characteristics of Vegetation Based on LCC inversion results, this application generates monthly LCC spatial distribution maps of the study area and analyzes their spatial distribution patterns, identifying the spatial differentiation characteristics between high-value and low-value areas. Statistical analysis of multi-temporal inversion results is performed to extract the temporal variation characteristics of LCC, revealing its seasonal variation patterns. The application also analyzes the magnitude of LCC variation and spatial heterogeneity at different growth stages, thereby achieving a quantitative characterization of the vegetation growth status and dynamic changes in the study area.
[0022] The Sentinel-2 multispectral remote sensing image mentioned in step S12 of this application is a Sentinel-2 Level-2A surface reflectance product, which includes visible light band, near-infrared band and red edge band, wherein the red edge band includes 705nm, 740nm and 783nm bands.
[0023] This application constructs a vegetation chlorophyll content inversion model based on the extreme random tree method to achieve regression prediction. During model construction, each decision tree is generated independently based on the training samples. At each node split, a split threshold is further randomly generated from a randomly selected feature subset, rather than using the optimal splitting criterion, thereby enhancing the overall randomness of the model. Finally, the inversion result is obtained by averaging the prediction results of all decision trees, effectively reducing model variance, mitigating the risk of overfitting, and improving the model's stability and generalization ability in modeling complex nonlinear relationships.
[0024] Parameter settings This application automatically optimizes the configuration of key parameters of the vegetation chlorophyll content inversion model. The number of decision trees is set to 500 to improve model stability; the maximum tree depth is set to 20 to enhance the model's ability to characterize complex feature relationships; the minimum number of node split samples is set to 5 to avoid overfitting caused by too small a sample split; and the minimum number of leaf node samples is set to 1 to ensure the model's flexibility in detail characterization.
[0025] Model advantages Extremely randomized tree models introduce strong randomness during node partitioning, significantly increasing the differences between decision trees. This effectively reduces the overall model variance, mitigates overfitting risk, and enhances the stability of modeling complex nonlinear relationships. Furthermore, this model does not require strict distribution assumptions about input features, enabling it to directly handle multi-dimensional and correlated vegetation index data. It also weakens the influence of outliers and noise through multi-tree ensemble averaging, thereby enhancing the model's stability and generalization ability.
[0026] Given the nonlinear relationship between vegetation chlorophyll content and multispectral features, the limited number of samples, and the certain correlation between features in this application, the extreme random tree model can achieve stable prediction while ensuring computational efficiency, and is suitable for remote sensing inversion scenarios of vegetation chlorophyll content.
[0027] Compared with the prior art, this application has the following advantages and effects: it can realize the accurate estimation and spatial distribution expression of the LCC of vegetation leaves in complex forest areas, and has the advantages of high inversion accuracy, good model stability and wide applicability. Attached Figure Description
[0028] Figure 1 This is a schematic diagram showing the geographical location of the study area and sampling points, illustrating the scope of the study area and the spatial distribution of sampling points located in Changle Forest Farm, Hangzhou City, Zhejiang Province (30.19°N, 119.51°E).
[0029] Figure 2A quadrat configuration map is provided for each survey area. The survey area is a square area of 90 m × 90 m. Four corner quadrats are set up 3 m inside the four corner points along the adjacent side, and one central quadrat is set up at the center of the area. The quadrats are numbered as follows: Southwest quadrat (1), Northwest quadrat (2), Northeast quadrat (3), Southeast quadrat (4) and Central quadrat (5).
[0030] Figure 3 This is a schematic diagram of the spectral reflectance curves of leaves of different vegetation types. The curves shown include the reflectance changes of Chinese fir, bamboo, oak, and sweetgum in the wavelength range of 400 nm to 1000 nm.
[0031] Figure 4 This is a schematic diagram illustrating the correlation between chlorophyll content and vegetation index. Figure 4 (a) is a distribution diagram of the correlation coefficients between various vegetation indices and chlorophyll content; Figure 4 (b) is a heatmap of the correlation matrix between various vegetation indices, reflecting the correlation between different vegetation indices.
[0032] Figure 5 A diagram comparing the results of LCC inversion using different machine learning models. Figure 5 (a) ~ Figure 5 (i) represents the correspondence between predicted and measured values for Random Forest (RF), Gradient Boosting Regression (GBDT), CatBoost, Support Vector Regression (SVR), Multilayer Perceptron (MLP), K-Nearest Neighbors (KNN), AdaBoost, Bagging, and Extra Trees models, respectively. The x-axis of each subplot represents the measured LCC, and the y-axis represents the predicted LCC. A 1:1 reference line and a fitting trend line are indicated, along with the coefficient of determination (R²) of the model on the training and test sets. 2 ).
[0033] Figure 6 This is a schematic diagram of the monthly spatial distribution of chlorophyll content (LCC) in the study area in 2024 based on an extreme random tree model. Each sub-map represents the LCC inversion results for different months. Sentinel-2 multispectral imagery was used as input data, and the LCC content of the vegetation cover area in the study area was obtained pixel-by-pixel inversion. The figure shows the spatial distribution of LCC for each month. Inversion was not performed for some months due to image limitations, and some areas were affected by cloud cover.
[0034] Figure 7 This is a box plot showing the monthly LCC temporal variation in the study area in 2024. Each box plot corresponds to a different month and represents the median, interquartile range, and distribution of LCC, reflecting the characteristics of LCC variation over time and its degree of dispersion. Detailed Implementation
[0035] The method for quantitative inversion of chlorophyll content in forest vegetation leaves using remote sensing in this application mainly includes the following steps: S1: Data Acquisition and Sample Construction in the Study Area S11: Acquisition of Actual Measurement Data This application selected Chang Le Forest Farm in Hangzhou, Zhejiang Province as the study area. Multiple survey plots were established based on the vegetation type and forest stand structure of the study area, with each plot measuring no less than 90 m × 90 m to ensure spatial representativeness in remote sensing imagery. A central quadrat and four corner quadrats were set up within each survey plot. Canopy leaf samples were collected from representative plants, and the geographic coordinates of the sampling points were recorded. The collected leaf samples were brought back to the laboratory, and the leaf LCC (Leaf Canopy Calibrates) was determined using chemical analysis methods, thereby obtaining ground-measured leaf LCC data for the study area.
[0036] S12: Sentinel-2 Multispectral Remote Sensing Image Acquisition and Preprocessing Sentinel-2 Level-2A surface reflectance multispectral remote sensing imagery of the study area was acquired using the Google Earth Engine cloud computing platform. This data product has undergone preprocessing such as radiometric calibration and atmospheric correction. By querying all available images for each month and statistically analyzing image cloud cover, candidate images were screened for quality. Based on a comprehensive consideration of image cloud cover and actual image quality, remote sensing images with low cloud cover and good image quality were selected as representative images for each month. The selected images were then cropped to reflect the study area.
[0037] S13: Spatial matching, sample dataset construction Based on the geographic coordinate information of the sample points, this application spatially matches the ground-measured leaf LCC samples with the corresponding Sentinel-2 remote sensing image pixels, extracts the multispectral reflectance data of the corresponding pixels at the matched sample point locations, including reflectance information in the visible light, red edge and near-infrared bands, and constructs a sample dataset containing LCC observation values and multispectral reflectance feature variables.
[0038] S2: Vegetation Index Feature Construction and Feature Analysis S21: Vegetation Index Calculation Based on Sentinel-2 multispectral reflectance data, several vegetation indices related to LCC were calculated, including the Simple Ratio Index (SR), which amplifies chlorophyll change signals by comparing the ratio of red-edge or near-infrared to visible light reflectance; the Modified Simple Ratio Index (mSR705), used to enhance the responsiveness of the red-edge band to chlorophyll changes; the Normalized Difference Vegetation Index (NDVI), used to reduce the influence of light conditions and background; the Normalized Green-Red Difference Index (NGRDI), which characterizes vegetation color changes by comparing the differences between green and red light bands; the Normalized Difference Red Edge Index (NDRE), used to enhance sensitivity to chlorophyll changes; the MERIS Terrestrial Chlorophyll Index (MTCI); and the Modified Red Edge Normalized Difference Vegetation Index (NDRE). The Red-edge Normalized Difference Vegetation Index (MNDVIred), the Chlorophyll Index Red-edge (CIred-edge), and the Red Edge Ratio Vegetation Index (RERVI) are used to characterize changes in vegetation LCC by utilizing the high sensitivity of the red-edge band to chlorophyll changes. The Modified Difference Ratio Index (MDRI) further enhances spectral response information by transforming the form of related indices. Other indices include Inverse Reflectance (IR700; Difference of the Reciprocal Reflectance (DRP550), the New Three Bands Index (NTBI), the Scatter-adjusted Reflectance Ratio (SaRR), the Pigment-Specific Normalized Difference Index (PSND650), and the Enhanced Vegetation Index. Index (EVI) enhances the response to LCC changes under complex canopy conditions through multi-band combination and scattering correction mechanisms.
[0039] S22: Construction of Feature Variable Set and Correlation Analysis This application combines the calculated vegetation indices with the corresponding ground-measured leaf LCC data from the sampling points to construct a set of characteristic variables. Statistical analysis methods are then used to calculate the correlation coefficients between each vegetation index and the LCC, establishing a correlation matrix among the vegetation indices.
[0040] S23: Feature Selection and Sensitivity Index Determination Figure 4 a shows the correlation between different vegetation indices and LCC. The results indicate that most vegetation indices constructed based on the red edge band are significantly positively correlated with LCC. Among them, the MTCI, mSR705, MNDVIred, and Cired_edge indices have high correlation coefficients, showing strong sensitivity to chlorophyll changes. However, some indices have weak correlations, indicating that their ability to indicate LCC is limited. Figure 4 b further reveals the correlation structure among the vegetation indices, showing that there is a strong correlation between some indices, reflecting a certain degree of information redundancy, especially among indices constructed by combining similar wavebands.
[0041] This application screens feature variables by assessing the correlation between vegetation indices and leaf LCC, their expressive power for various aspects of chlorophyll content, and the information redundancy between indices. Ultimately, the following combination of feature variables with high sensitivity to changes in leaf LCC is selected: MERIS Terrestrial Chlorophyll Index (MTCI, formula MTCI = (B6-B5) / (B5-B4)), Modified Simple Ratio Index (mSR705, formula mSR705 = (B6-B1) / (B5-B1)), Modified Red Edge Normalized Difference Vegetation Index (MNDVIred, formula MNDVIred = (B6-B5) / (B6+B5-B1)), and Red Edge Chlorophyll Index (CIred-edge, formula CIred-edge = B6 / B5-1). The specific screening process is as follows: S231: Candidate index grouping and initial screening based on spectral mechanisms This application first groups the 16 candidate vegetation indices based on the band combination method and their characterization mechanism of chlorophyll absorption characteristics, dividing them into four functional groups: (1) Red edge position sensitive group, including MTCI, NDRE, and RERVI, which mainly uses the change in reflectance of the red edge region (705–783 nm) to characterize the red edge displacement characteristics caused by chlorophyll; (2) Red edge amplitude sensitive group, including mSR705, MNDVIred, CIred-edge, and SR, which characterizes the change in red edge reflectance amplitude through the ratio of red edge to reference band or normalization operation; (3) Broadband integrated group, including NDVI, EVI, and NGRDI, which mainly uses the visible light and near-infrared broadband information; (4) Scattering / reciprocal correction group, including IR700, SaRR, PSND650, DRP550, NTBI, and MDRI, which eliminates the influence of canopy scattering through reciprocal transformation or multi-band correction.
[0042] S232: Correlation-based screening and elimination of intragroup redundancy and intergroup complementarity analysis In the red-edge position sensitive group, MTCI uses three consecutive red-edge bands (B4, B5, B6) to construct the difference ratio. Its response to chlorophyll-induced red-edge shift shows an approximately linear relationship, and it is not easily saturated under high chlorophyll concentration conditions. In contrast, NDRE only uses the normalized difference between two bands, B5 and B8A. Correlation matrix analysis revealed that the correlation coefficient between NDRE and MTCI is as high as 0.93, indicating severe information redundancy. The correlation coefficient between RERVI and CIred-edge is 0.97, also forming a highly redundant pair. Therefore, only MTCI is retained in the red-edge position sensitive group.
[0043] In the red-edge amplitude-sensitive group, mSR705 corrects for leaf scattering effects by introducing the blue light band B1 (443 nm). The formula mSR705 = (B6-B1) / (B5-B1) effectively eliminates the systematic bias in the red-edge ratio caused by specular reflection from the leaf wax layer and epidermal trichomes. This correction mechanism is particularly crucial in complex forest stands of mixed subtropical evergreen broad-leaved and coniferous forests—the study area in this application includes various tree species with significant differences in leaf structure, such as Chinese fir, bamboo, oak, and sweetgum. The uncorrected simple ratio index SR exhibits a large systematic bias (coefficient of variation reaching 23.6%) among different tree species, while mSR705 reduces this bias to 11.2%. MNDVIred was normalized to (B6-B5) / (B6+B5-B1). This normalization effectively suppresses the influence of irradiance variations caused by differences in solar altitude angle and observation geometry, which is crucial for the cross-temporal consistency of multi-temporal images (covering all 12 months of the year) in this study area. CIred-edge = (B6 / B5)⁻¹, its mathematical form ensures an approximately linear proportional relationship with chlorophyll content in the medium-to-high concentration range (35–60 μg / cm²), while SR shows a clear saturation trend in this range. Therefore, this group retains mSR705, MNDVIred, and CIred-edge, while SR is discarded.
[0044] In the broadband composite group, NDVI exhibited a severe saturation effect under high vegetation cover conditions in this study area (mean NDVI = 0.82), with a correlation coefficient of only 0.41 with LCC, significantly lower than that of the red-edge indexes (MTCI and LCC had a correlation coefficient of 0.78). Although EVI partially mitigated the influence of atmospheric scattering by introducing the blue light band, its initial design purpose was to characterize canopy structure parameters (leaf area index LAI), and its sensitivity to changes in chlorophyll content at the leaf scale was limited, with a correlation coefficient of 0.35 with LCC. NGRDI only utilizes the green and red bands of visible light and is easily affected by atmospheric aerosol and water vapor absorption, with a correlation coefficient of 0.22 with LCC. Therefore, all three indices in the broadband composite group were excluded.
[0045] In the scattering / reciprocal correction group, although the six indices IR700, SaRR, PSND650, DRP550, NTBI, and MDRI could theoretically improve their sensitivity to chlorophyll through reciprocal transformation or multi-band correction, actual data verification revealed that: (1) the correlation coefficients between IR700 and DRP550 and LCC were 0.52 and 0.48, respectively, lower than those of the previously selected MTCI (0.78) and mSR705 (0.74); (2) the correlation coefficients between NTBI and SaRR and mSR705 reached 0.91 and 0.88, respectively, indicating significant information redundancy; (3) PSND650 was mainly sensitive to chlorophyll a and had a weak response to chlorophyll b, and its stability was insufficient under the condition of large variations in the chlorophyll a / b ratio in the mixed coniferous and broad-leaved forests of this study area; (4) the correlation coefficient between MDRI and CIred-edge was 0.94, which is a redundant feature. Therefore, all six indices in this group were removed.
[0046] Spectral complementarity and synergistic effect analysis of the final feature combination After the above screening, the four vegetation indices MTCI, mSR705, MNDVIred, and CIred-edge were determined to constitute the optimal combination of characteristic variables. This combination has the following spectral complementarity characteristics: (a) MTCI uses the difference ratio of the three red edge bands to mainly capture the red edge position shift information caused by chlorophyll changes, which belongs to the red edge "displacement type" feature; CIred-edge uses the simple ratio of B6 / B5 to mainly reflect the amplitude change of the red edge reflectance steep slope segment, which belongs to the red edge "amplitude type" feature; the two represent two independent dimensions of red edge spectral change (displacement direction and amplitude direction), and their information cross-correlation coefficient is 0.61, which is lower than the intragroup redundancy threshold of 0.85, confirming that they are complementary rather than redundant.
[0047] (b) mSR705 provides red edge signal extraction capability independent of leaf optical properties by correcting leaf scattering in the blue light band, and has cross-species stability advantage for mixed forests with significant differences in leaf structure in the study area.
[0048] (c) The normalized operation form of MNDVIred gives it robustness to irradiance changes and can effectively suppress spurious signals caused by seasonal changes in solar altitude angle in monthly multi-time inversion scenarios, thus complementing the other three indices in the time dimension.
[0049] The unintended technical effects of the above feature-based dimensionality reduction In traditional remote sensing inversion research, it is generally believed that increasing the dimensionality of input features can improve the fitting ability of machine learning models. However, this application, through systematic experiments, discovered a technical effect that contradicts this conventional understanding: after reducing the input features from all 16 vegetation indices to the four indices mentioned above, the coefficient of determination (R²) of the extreme random tree model on the test set increased from 0.4326 to 0.5198, and the normalized root mean square error (nRMSE) decreased from 0.1823 to 0.1585. This result indicates that, under conditions of limited sample size (approximately 160 samples in this application) and complex collinearity among features, high-dimensional redundant features can actually lead to noise splitting during node partitioning in the machine learning model, increasing the risk of overfitting. The low-dimensional complementary feature space formed by the four indices allows the extreme random tree model to select feature subsets with independent information contributions from the LCC with a higher probability during random feature subset sampling, thereby effectively reducing the correlation of prediction errors between base learners and achieving a substantial reduction in the variance of the ensemble model.
[0050] Furthermore, cross-validation of the aforementioned four index combinations with nine subsequent machine learning models revealed that the performance improvement (R² improvement of 0.087) of this feature combination on the Extremely Random Tree model was significantly higher than that on the Random Forest (R² improvement of 0.031) and Gradient Boosting Tree (R² improvement of 0.022). This differentiated synergistic effect stems from the completely random splitting mechanism unique to Extremely Random Trees: unlike Random Forests which select the optimal value among candidate split points, Extremely Random Trees randomly generate splitting thresholds for each candidate feature, thus being more sensitive to the quality of the input feature space—when redundant dimensions in the input features are reduced, the randomly generated splitting thresholds are more likely to fall within the discriminative feature value range, thereby enabling the model to achieve a more significant performance improvement. This synergistic effect between feature selection and model architecture has not been revealed in existing technologies and is one of the core technical contributions of this application.
[0051] Based on the correlation analysis results between LCC and vegetation indices and the correlation structure between indices, and taking into account both the intensity of the index response to LCC and the independence of features, and combined with the verification results of multiple model training and testing experiments on the inversion accuracy, the four vegetation index combinations of MTCI, mSR705, MNDVIred, and Cired_edge were finally selected as the input features for subsequent model training and inversion.
[0052] S3: Construction of a vegetation chlorophyll content inversion model based on extreme random trees Model Principles Extremely randomized tree (ERT) is an ensemble learning method based on decision trees. It constructs multiple decision trees and integrates their results to achieve regression prediction. During model construction, each decision tree is generated independently based on training samples. At each node split, a split threshold is further randomly generated from a randomly selected subset of features, rather than using the optimal splitting criterion. This enhances the overall randomness of the model. Finally, the inversion result is obtained by averaging the predictions from all decision trees. This method further introduces a random mechanism in node splitting and subsequent feature selection, effectively reducing model variance, mitigating overfitting risk, and improving the model's stability and generalization ability in modeling complex nonlinear relationships.
[0053] Parameter settings This application automatically optimizes the key parameters of the extreme random tree model. The number of decision trees (n_estimators) is set to 500 to improve model stability; the maximum tree depth (max_depth) is set to 20 to enhance the model's ability to characterize complex feature relationships; the minimum number of split samples per node (min_samples_split) is set to 5 to avoid overfitting caused by excessively small sample splits; and the minimum number of leaf samples (min_samples_leaf) is set to 1 to ensure the model's flexibility in characterizing details.
[0054] Model advantages Extremely randomized tree models introduce strong randomness during node partitioning, significantly increasing the differences between decision trees. This effectively reduces the overall model variance, mitigates overfitting risk, and enhances the stability of modeling complex nonlinear relationships. Furthermore, this model does not require strict distribution assumptions about input features, enabling it to directly handle multi-dimensional and correlated vegetation index data. It also weakens the influence of outliers and noise through multi-tree ensemble averaging, thereby enhancing the model's stability and generalization ability.
[0055] Given the nonlinear relationship between vegetation chlorophyll content and multispectral features, the limited number of samples, and the certain correlation between features in this application, the extreme random tree model can achieve stable prediction while ensuring computational efficiency, and is suitable for remote sensing inversion scenarios of vegetation chlorophyll content.
[0056] S31: Training Sample Construction This application uses the combination of highly sensitive vegetation index feature variables obtained through screening as the model input and the ground-measured leaf LCC of the corresponding sample points as the model output to construct LCC inversion training samples.
[0057] S32: Model Training and Accuracy Evaluation Based on the constructed training samples, an inversion model for vegetation chlorophyll content is constructed using the Extra Trees method.
[0058] The sample data was divided into training and test sets in a 7:3 ratio, and stratified sampling was used. During model training, the key parameters of the model were automatically optimized based on the characteristics of the training data to obtain the optimal parameter configuration of the model, and the model training was completed under the optimal parameter conditions.
[0059] The accuracy of the extreme random tree model was validated using a test set, and the coefficient of determination (R²) was selected. 2 R0 and normalized root mean square error (nRMSE) are used as evaluation metrics, where R0 2 nRMSE is used to characterize the model's ability to explain the measured LCC variation. It measures the relative magnitude of the model's prediction error, and the formula is: In the formula, Indicates the first i The measured LCC of a sample This represents the corresponding model prediction value. This is the average value of the measured LCC. n This represents the number of samples.
[0060] S4: Comparison of inversion performance between the Extremely Random Tree Model and other major models Based on the training samples constructed using S31, leaf LCC inversion models were constructed using the remaining 8 machine learning algorithms. The 8 machine learning methods include: Random Forest (RF), Gradient Boosting Decision Tree (GBDT), Category Feature Gradient Boosting (CatBoost), Support Vector Regression (SVR), Artificial Neural Network (MLP), K Nearest Neighbor Regression (KNN), Adaptive Boosting (AdaBoost), and Bagging.
[0061] Overall, different machine learning models exhibit significant differences in their ability to retrieve vegetation LCC. Most models achieve high R-values on the training set. 2 The low nRMSE and low accuracy of the ensemble learning models indicate good fitting ability, but their accuracy on the test set generally decreases, reflecting differences in generalization performance and varying degrees of overfitting in some models. In contrast, the ensemble learning models show more stable overall performance on the test set, balancing model complexity and generalization ability to a certain extent, and demonstrating stronger resistance to overfitting.
[0062] Among all models, the Extra Trees model has the best overall performance, with R0 values for both the training and test sets. 2The R² values were 0.7418 and 0.5198 respectively, and the nRMSE values were 0.0975 and 0.1585 respectively, both of which were the best among all models. In terms of accuracy, Extra Trees not only had the highest fitting ability during training but also maintained the highest R² value during testing. 2 The lowest error indicates that it effectively characterizes the nonlinear relationship between LCC and vegetation indices without significant overfitting. Figure 5 (i) It can be seen that the overall distribution of the predicted values and the measured values of the Extra Trees model is closer to the 1:1 reference line, the point dispersion is smaller, and the error distribution is more uniform. This indicates that the model has good predictive ability for different LCC intervals, and its stability and generalization performance are the most outstanding.
[0063] In comparison, Random Forest (RF) and Bagging models also achieved high accuracy on the training set (R²). 2 The scores were 0.6981 and 0.7087 respectively, but the test set R... 2 The nRMSE values decreased to 0.4807 and 0.4771 respectively, and were also slightly higher than Extra Trees, indicating that they tend to overfit to some extent and their generalization ability is slightly inferior. Although GBDT and CatBoost models have strong nonlinear fitting ability, the test set R... 2 The values were only 0.3542 and 0.3973, and the error was relatively large, indicating that it failed to fully leverage its advantages under the current sample size and feature conditions, and the model stability was relatively insufficient.
[0064] AdaBoost model test set R 2 The value is 0.4495, which is at a moderate level overall, but it is quite sensitive to outliers, leading to some fluctuations in the prediction results. The SVR model training set is R. 2 With a value of only 0.3035, it exhibits significant underfitting, indicating its limited ability to characterize complex nonlinear relationships; the KNN and MLP models tested on the R... 2 The values were 0.4458 and 0.4655 respectively. Although these values were an improvement over SVR, the distribution of predicted points was relatively discrete, and the ability to characterize high or low value ranges was insufficient. The stability still needs to be improved.
[0065] A comprehensive comparison of the accuracy and inversion performance of all models on the test set revealed that the Extra Trees model outperformed the other eight models in multiple aspects, including training accuracy, testing accuracy, and error control. It effectively improved generalization performance while maintaining high fitting ability, and its predictions more closely approximate the actual distribution characteristics. Therefore, the Extra Trees model was ultimately selected as the optimal model.
[0066] S5: Chlorophyll Content Retrieval and Spatial Distribution Analysis S51: Remote sensing inversion of chlorophyll content This application utilizes the optimal machine learning model (extreme random tree model) and takes the vegetation index feature variables calculated from Sentinel-2 multispectral remote sensing images as input. It performs pixel-by-pixel calculations on remote sensing images of the study area for each month to obtain the leaf LCC prediction values of the corresponding pixels and generate multi-temporal LCC inversion result data.
[0067] S52: Spatial Distribution and Variation Characteristics Analysis Based on LCC inversion results, this application generates monthly LCC spatial distribution maps of the study area and analyzes their spatial distribution patterns, identifying the spatial differentiation characteristics between high-value and low-value areas. Statistical analysis of multi-temporal inversion results is performed to extract the temporal variation characteristics of LCC, revealing its seasonal variation patterns. The application also analyzes the magnitude of LCC variation and spatial heterogeneity at different growth stages, thereby achieving a quantitative characterization of the vegetation growth status and dynamic changes in the study area.
[0068] The following explanation, based on experimental results, further clarifies this application: Based on the optimal model and combined with Sentinel-2 multispectral imagery, the monthly LCC of vegetation in the study area in 2024 was retrieved. The results show that the method proposed in this application can achieve the acquisition of the spatially continuous distribution of vegetation LCC. Spatially, the vegetation LCC in the study area exhibits obvious spatial heterogeneity. High chlorophyll values are mainly concentrated in areas with relatively continuous vegetation cover and good growth, while low value areas are mostly found in sparse vegetation or areas with strong human disturbance, indicating that the retrieval results can well reflect the spatial differences in vegetation growth status in the study area.
[0069] Based on the monthly chlorophyll content results, a statistical analysis was conducted on the temporal variation characteristics of vegetation LCC in the study area in 2024 (excluding April). (See attached data.) Figure 6 In spring (January to March), the overall LCC (Limited Photosynthesis Coefficient) is relatively low, with the median and mean values for each month being close, indicating weaker vegetation physiological activity from winter to early spring. After entering the growing season, the LCC increases significantly from May to July, with both the monthly mean and median values rising synchronously. June and July reach the highest level of the year, reflecting the strongest photosynthetic capacity of vegetation in the study area during the peak summer season. It gradually declines after August, returning to the level at the beginning of the year by November and December. The dispersion of LCC distribution varies across different months, with a wider distribution during the peak growing season (May to July) and a relatively concentrated distribution outside the growing season.
[0070] In summary, this application takes Changle Forest Farm in Hangzhou, Zhejiang Province as the study area, integrates ground-measured LCC data and Sentinel-2 multispectral remote sensing imagery to construct a vegetation index feature system including red-edge band information. By comparing and analyzing various machine learning models, the extreme random tree model was determined as the optimal inversion model, thus achieving the inversion modeling of vegetation LCC. The inversion results obtained based on this model exhibit good spatial continuity and consistency, effectively characterizing the spatial differences in vegetation growth status in different areas of the study area. Temporally, it reflects the dynamic process of vegetation LCC changes with the seasons, showing a gradual increase from the early growth stage, reaching a high level during the vigorous growth period, and then gradually decreasing in the later stages. Therefore, the method in this application can achieve stable inversion of vegetation LCC, possessing good model stability and applicability, and can be applied to regional-scale vegetation growth monitoring and ecological environment assessment.
Claims
1. A remote sensing inversion method for vegetation chlorophyll content based on red-edge vegetation index and extreme random tree, characterized in that... Includes the following steps: S1: Data Acquisition and Sample Construction in the Study Area We acquired measured chlorophyll content data of vegetation leaves and Sentinel-2 multispectral remote sensing image data in the study area and established the spatial matching relationship between the two, thereby constructing a sample dataset containing ground measured chlorophyll content and remote sensing spectral information. S2: Vegetation Index Feature Construction and Feature Analysis Multispectral reflectance data related to vegetation chlorophyll content were obtained from Sentinel-2 multispectral remote sensing images. Through calculation and analysis, the two independent dimensions that can express red edge spectral changes, as well as the combination of vegetation indices that can extract red edge signals without relying on leaf optical properties and irradiance changes were selected. S3: Construction of a vegetation chlorophyll content inversion model based on extreme random trees S31: Using the selected vegetation index combination as the input variable and the measured chlorophyll content as the output variable, construct training samples; S32: An extreme random tree method was used to construct a vegetation chlorophyll content inversion model, and the model was validated and its accuracy evaluated. S4: Comparison of the inversion results of the vegetation chlorophyll content inversion model with other major models; S5: Inversion and spatial distribution analysis of vegetation chlorophyll content.
2. The remote sensing inversion method for vegetation chlorophyll content according to claim 1, characterized in that... Step S2 includes: S21: Vegetation Index Calculation Based on Sentinel-2 multispectral band reflectance data, vegetation indices related to chlorophyll content were calculated, including the simple ratio index, normalized difference vegetation index, MERIS terrestrial chlorophyll index, corrected red-edge normalized difference vegetation index, red band chlorophyll vegetation index, corrected difference ratio index, reciprocal reflectance index, scattering corrected reflectance ratio index, pigment-specific normalized difference index, reciprocal reflectance difference index, new three-band index, corrected simple ratio index, normalized green-red difference index, normalized red-edge vegetation index, red-edge ratio index, and enhanced vegetation index. S22: Construction of Feature Variable Set and Correlation Analysis The calculated vegetation indices are combined with the ground-measured leaf chlorophyll content data of the corresponding sample points to construct a set of characteristic variables. The correlation coefficient between each vegetation index and leaf chlorophyll content is calculated by statistical analysis methods, and a correlation matrix between vegetation indices is established. S23: Feature Selection and Sensitivity Index Determination Feature variables are screened by judging the correlation between vegetation indices and leaf chlorophyll content, their expressive power over various aspects of chlorophyll content, and the information redundancy between indices.
3. The remote sensing inversion method for vegetation chlorophyll content according to claim 1, characterized in that... Step S23 includes: S231: Candidate index grouping and initial screening based on spectral mechanisms Based on the band combination of each vegetation index and its characterization mechanism of chlorophyll absorption characteristics, the groups are grouped to cover all relevant dimensions of chlorophyll content in vegetation leaves as much as possible. S232: Correlation-based screening and elimination of intragroup redundancy and intergroup complementarity analysis Ultimately, the optimal combination of feature variables was determined to be four vegetation indices: the MERIS terrestrial chlorophyll index, the modified simple ratio index, the modified red-edge normalized difference vegetation index, and the red-edge chlorophyll index.
4. The remote sensing inversion method for vegetation chlorophyll content according to claim 1, characterized in that... Step S1 includes: S11: Select the study area, set up survey plots in the study area and collect vegetation leaf samples, determine the chlorophyll content of the leaves by laboratory chemical analysis methods, and obtain the ground-measured leaf chlorophyll content data of the study area. S12: Acquire contemporaneous Sentinel-2 multispectral remote sensing images of the study area on the Google Earth Engine platform, and perform preprocessing such as cloud and shadow filtering and removal, and image cropping; S13: Based on the spatial coordinates of the sample points, the ground-measured samples are spatially matched with the corresponding remote sensing image pixels to extract the multispectral reflectance data of the sample point locations, thereby constructing a sample dataset containing ground-measured chlorophyll content and remote sensing spectral information.
5. The remote sensing inversion method for vegetation chlorophyll content according to claim 4, characterized in that: The Sentinel-2 multispectral remote sensing image mentioned in step S12 is a Sentinel-2 Level-2A surface reflectance product, which includes visible light band, near-infrared band and red edge band, wherein the red edge band includes 705 nm, 740 nm and 783 nm bands.
6. The remote sensing inversion method for vegetation chlorophyll content according to claim 1, characterized in that: The parameters of the extreme random tree include: 500 decision trees, a maximum tree depth of 20, a minimum number of samples per node partition of 5, and a minimum number of samples per leaf node of 1.
7. The remote sensing inversion method for vegetation chlorophyll content according to claim 1, characterized in that... Step S5 includes: S51: Remote sensing inversion of vegetation chlorophyll content The vegetation chlorophyll content inversion model takes the vegetation index characteristic variables calculated from Sentinel-2 multispectral remote sensing images as input, performs pixel-by-pixel calculations on remote sensing images of the study area for each month, obtains the predicted leaf chlorophyll content of the corresponding pixel, and generates multi-temporal chlorophyll content inversion result data. S52: Analysis of Spatial Distribution and Change Characteristics of Vegetation Based on the multi-temporal chlorophyll content inversion results, a monthly spatial distribution map of chlorophyll content in the study area was generated, and its spatial distribution pattern was analyzed. The spatial differentiation characteristics of high-value areas and low-value areas were identified, the temporal variation characteristics of chlorophyll content were extracted, its seasonal variation pattern was revealed, and the variation range and spatial heterogeneity of chlorophyll content at different growth stages were analyzed, so as to achieve a quantitative characterization of the vegetation growth status and dynamic change process in the study area.