Method and system for machine learning inversion of sunlight-induced chlorophyll fluorescence and medium

Abstract

The invention discloses a method, system and medium for machine learning inversion of sunlight-induced chlorophyll fluorescence, and relates to the technical field of vegetation sunlight-induced chlorophyll fluorescence remote sensing inversion, and the method comprises the steps: obtaining an actual measurement data set, and generating a synthetic data set; introducing principal component analysis to carry out dimension reduction preprocessing on features in the synthetic data set to obtain a preprocessed synthetic data set, and taking the preprocessed synthetic data set as a training set; introducing noise of different levels into the training set to train each machine learning model, and obtaining each trained machine learning model; pre-processing the incident irradiance and the reflected radiance in the measured data set, taking the pre-processed incident irradiance and reflected radiance as input of each trained machine learning model, and comparing performance indexes of output results of each model to obtain an optimal machine learning model; executing an SIF inversion task based on the optimal machine learning model; the method is used for high-resolution and rapid inversion of the SIF.

Description

Technical Field

[0001] This invention relates to the technical field of remote sensing inversion of sunlight-induced chlorophyll fluorescence in vegetation, and particularly to methods, systems, and media for machine learning inversion of sunlight-induced chlorophyll fluorescence. Background Technology

[0002] Sun-induced chlorophyll fluorescence (SIF), as a non-invasive method for monitoring photosynthesis, has attracted widespread attention in fields such as ecology, remote sensing, and global carbon cycle research. SIF signals are mainly located in the red and near-infrared bands and are composed of contributions from photosystem II and photosystem I. Therefore, simultaneously monitoring red and far-infrared fluorescence can directly reflect the dynamic processes of photosynthesis and light energy utilization efficiency. Existing studies have shown that SIF responds significantly to environmental conditions (such as temperature, light, water, and nutrients) and non-photochemical quenching, and that canopy structure also influences SIF signals by altering light absorption, scattering, and emission processes.

[0003] To interpret complex spectral signals and retrieve fluorescence, researchers have developed various inversion methods. The spectral fitting method (SFM), based on physical mechanisms, can well describe the shape of the fluorescence spectrum and is currently the standard processing method for ground-based systems such as FloX. However, it is computationally expensive and its accuracy is limited in the presence of atmospheric reabsorption. The singular value decomposition (SVD) method has successfully retrieved far-red fluorescence using solar Fraunhofer lines, but this method heavily relies on high spectral resolution and has limitations in residual modeling. In recent years, integrated radiative transfer models such as the SCOPE (Soil Canopy Observation of Photosynthesis and Energy fluxes) model have provided references for validating SIF inversion, but the model inversion process is complex and highly dependent on input parameters. Overall, existing methods still have shortcomings in terms of computational efficiency, noise robustness, and adaptability to complex canopy conditions. Summary of the Invention

[0004] Based on the technical problems existing in the background technology, this invention proposes a method, system and medium for machine learning to invert sunlight-induced chlorophyll fluorescence, which can be used for high-resolution and rapid inversion of SIF.

[0005] The method for machine learning to invert sunlight-induced chlorophyll fluorescence proposed in this invention includes: Obtain the actual test dataset and generate the synthetic dataset; Principal component analysis is introduced to perform dimensionality reduction preprocessing on the features in the synthetic dataset to obtain a preprocessed synthetic dataset, which is then used as the training set. Different levels of noise are introduced into the training set to train each machine learning model, resulting in trained machine learning models. After preprocessing the incident irradiance and reflected irradiance in the measured dataset, they are used as inputs to each trained machine learning model. The optimal machine learning model is obtained by comparing the performance metrics of the output results of each model. Perform the SIF inversion task based on the optimal machine learning model.

[0006] Furthermore, the generation of the synthetic dataset specifically involves: The LESS model, which includes a leaf optical properties module and a chlorophyll fluorescence simulation module, is used. The LESS model is a large-scale remote sensing data and image simulation framework model. The blade optical properties module serves as the calculation engine for the blade's optical properties. It uses a layered flat plate model to quantify the absorption process of photosynthetically active radiation and generate the blade's optical properties. The chlorophyll fluorescence simulation module is extended into a photosynthesis-fluorescence coupling system, and the distribution path of absorbed energy is deconstructed through a quenching competition model to calculate the fluorescence emission spectrum; The leaf optical properties module and the chlorophyll fluorescence simulation module achieve cross-scale coupling through bidirectional path tracing, output reflectance, and then convert the reflectance into canopy radiance. Grassland structural units were constructed using real measurement parameters, and the grassland structural units were run in the LESS model to simulate canopy sunlight-induced chlorophyll fluorescence, thereby creating a synthetic radiation spectrum dataset. The spectral dataset is randomly shuffled and oversampled by a set factor to generate a synthetic dataset.

[0007] Furthermore, the core equation of the blade optical characteristic module is expressed as: ; in, The absorption coefficient of the leaf. For wavelength, These are the blade structure parameters. These are chlorophyll content, equivalent water thickness, and dry matter content, respectively. For chlorophyll at wavelength The specific absorption coefficient at that location, For water at wavelength The specific absorption coefficient at that location, For dry matter at wavelength The specific absorption coefficient at that location.

[0008] Furthermore, the core formula of the chlorophyll fluorescence simulation module is as follows: ; in, To be at wavelength The fluorescence intensity emitted at that location, The fluorescence emission rate constant is . For heat dissipation rate constants other than those for photochemical quenching, It is a non-photochemical quenching rate constant. Photosynthetically active radiation flux Let f be the fluorescence escape probability function. For the blade at wavelength The absorption coefficient at that location.

[0009] Furthermore, the cross-scale coupling formula is as follows: ; in, The radiance of the canopy received by the sensor is the vector sum of the reflected radiance and the fluorescence radiance. For wavelength, The reflected radiance produced by canopy reflection, The fluorescence radiance produced by the canopy fluorescence component.

[0010] Furthermore, the process of generating the preprocessed synthetic dataset is as follows: During the operation of the machine learning model, the SHAP method is used to interpret the feature importance of the machine learning model and obtain the correspondence between SIF changes and band reflection characteristics. The spectra in the synthetic data are subjected to first-order derivative transformation, followed by normalization. Principal component analysis is then introduced to perform dimensionality reduction preprocessing on the normalized spectra, resulting in the preprocessed synthetic dataset.

[0011] Furthermore, the machine learning models are: random forest, extreme gradient boosting, convolutional neural network, and dual-input spectral neural network.

[0012] Furthermore, the machine learning model uses the metric R during training. 2 The root mean square error is used to evaluate the model's fit and prediction accuracy.

[0013] A computer system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0014] A computer-readable storage medium storing a plurality of classification programs, the plurality of classification programs being invoked by a processor to execute the method described above.

[0015] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as ROM, RAM, magnetic disk, or optical disk.

[0016] The advantages of the machine learning-based method, system, and medium for inverting sunlight-induced chlorophyll fluorescence (SIF) provided by this invention are as follows: By combining principal component analysis (PCA) with a machine learning framework, the computational efficiency and stability of SIF inversion are significantly improved. Traditional spectral fitting (SFM)-based inversion methods are often computationally expensive and sensitive to the selection of spectral bands; while principal component analysis (PCA) dimensionality reduction reduces the computational burden while ensuring prediction accuracy, enabling the model to quickly process large-scale observational data. This framework achieves stable advantages in key indicators such as RMSE and MAE, and keeps error fluctuations within a low range, thus making high-resolution and rapid SIF inversion possible. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 The diagram illustrates the performance metrics of the XGBoost-based regression model after dimensionality reduction using PCA (Principal Component Analysis). (a) shows the Root Mean Square Error (RMSE), and (b) shows the Coefficient of Determination (R²). 2 Indicator diagram; Figure 3 This diagram illustrates the performance of each machine learning model before and after PCA dimensionality reduction, where (a) represents the coefficient of determination R. 2 The diagram shows the following indicators: (b) is the root mean square error (RMSE) indicator, (c) is the mean absolute error (MAE) indicator, and (d) is the training time indicator. Figure 4 This is a schematic diagram illustrating the results of training machine learning models using a synthetic dataset, where (a) represents the determination coefficients R of each model. 2 The diagram shows (b) the root mean square error (RMSE) of each model and (c) the mean absolute error (MAE) of each model. Figure 5 This is a comparison chart and error chart showing the SIF values ​​of each model and those calculated by traditional SFM on the actual test dataset. Detailed Implementation

[0018] The technical solution of the present invention will now be described in detail through specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0019] like Figures 1 to 5 As shown, the machine learning method for retrieving sunlight-induced chlorophyll fluorescence proposed in this invention includes: Step 1: Obtain the actual test dataset and generate the synthetic dataset; Step 2: Principal component analysis is introduced to perform dimensionality reduction preprocessing on the features in the synthetic dataset to obtain the preprocessed synthetic dataset, which is then used as the training set. Step 3: Introduce different levels of noise into the training set to train each machine learning model, and obtain the trained machine learning models. Step 4: After performing the same preprocessing as in Step 2 on the incident irradiance and reflected irradiance in the measured dataset, use them as input to each machine learning model. By comparing the performance metrics of the output results of each model, the optimal machine learning model is obtained. Step 5: Perform the SIF inversion task based on the optimal machine learning model.

[0020] Building upon the background technologies, machine learning methods have gradually become an important approach for ground-based SIF inversion. Machine learning can capture the complex mapping relationship between high-dimensional inputs and nonlinear outputs, avoiding reliance on excessive prior assumptions. Machine learning techniques can automatically learn nonlinear mapping relationships from high-dimensional spectra without strong prior assumptions, demonstrating potential for inversion accuracy and generalization. However, hyperspectral data itself suffers from dimensional redundancy and noise amplification effects, and directly inputting it into machine learning models can easily lead to overfitting or instability. Principal Component Analysis (PCA), as a classic dimensionality reduction method, can effectively compress the spectral feature space, reduce collinearity and noise interference, thereby providing more representative inputs for machine learning models.

[0021] In the PCA feature space, different types of machine learning models can further explore potential nonlinear relationships. For example, Random Forest (RF) has good feature selection and noise resistance; Gradient Boosting Tree (XGBoost) can efficiently fit complex nonlinear features; Multilayer Perceptron (MLP) and Backpropagation Neural Network (BP) can establish nonlinear regression relationships; while Two-Input Spectral Neural Network (DSNN) and Convolutional Neural Network (CNN) have stronger deep feature learning and automatic pattern extraction capabilities. The combination of these methods with PCA reduces data dimensionality and noise interference, while fully leveraging the advantages of different models in nonlinear fitting and feature representation, providing new ideas for achieving high-precision and robust SIF inversion.

[0022] In one embodiment, step one, obtaining the measured dataset and generating the synthetic dataset, specifically involves: (a1) Using the LESS model, which includes a leaf optical properties module and a chlorophyll fluorescence simulation module, the LESS model is a large-scale remote sensing data and image simulation framework model.

[0023] The LESS (Large-scale Remote Sensing data and image Simulation framework) model improves upon the radiative transfer model (RTM), overcoming the technical limitations of traditional methods in simulating complex three-dimensional vegetation scenes. Unlike traditional one-dimensional radiative transfer models based on the assumption of a uniform canopy, LESS employs advanced ray tracing algorithms to accurately simulate the interactions between photons and vegetation components in heterogeneous environments.

[0024] In the LESS model, photosynthesis and canopy reflectance are simulated through two core modules: the leaf optical properties module (PROSPECT, reference: JACQUEMOUD S, BARET F. PROSPECT: A model of leaf optical properties spectra[J]. Remote Sensing of Environment, 1990, 34(2):75-91. DOI:10.1016 / 0034-4257(90)90100-Z.) and the chlorophyll fluorescence simulation module (FLUSPECT, reference: VILFAN N, VANDERTOL C, MULLER O, et al. Fluspect-B: A model for leaf fluorescence, reflectance and transmittance spectra[J]. Remote Sensing of Environment, 2016, 186: 596-615.). (DOI:10.1016 / j.rse.2016.09.017) to achieve high-precision simulation driven by physical mechanisms.

[0025] The blade optical properties module, serving as the computational engine for blade optical properties, employs a layered slab model to quantify the absorption process of photosynthetically active radiation (PAR). Its core equation is expressed as: ; in, The absorption coefficient is... For wavelength, These are the blade structure parameters. These are chlorophyll content, equivalent water thickness, and dry matter content, respectively. For chlorophyll at wavelength The specific absorption coefficient at that location, For water at wavelength The specific absorption coefficient at that location, For dry matter at wavelength The specific absorption coefficient at that location.

[0026] The blade optical properties module treats the blade as an optical medium with a rough surface, accurately simulating the reflection / transmission of photons at the cell wall interface. The absorption and internal scattering processes of biochemical components were studied to establish a quantitative mapping from leaf biochemical parameters to light energy capture. For example, when... Increase from 10 to 40 µmg·cm 2 At that time, the wavelength of red light (650 nm) increased by 62%.

[0027] The chlorophyll fluorescence simulation module is extended from this physics to a photosynthesis-fluorescence coupled system, and the distribution path of absorbed energy is deconstructed through a quenching competition model: ,(1); (2); in, To be at wavelength The fluorescence intensity emitted at that location, The fluorescence emission rate constant is . For heat dissipation rate constants other than those for photochemical quenching, It is a non-photochemical quenching rate constant. Photosynthetically active radiation flux The fluorescence escape probability function is determined by the blade tilt angle. Determined by the geometry of the canopy, For the blade at wavelength The absorption coefficient at that point, i.e., the two-way fluorescence spectral morphology (red light 685nm / far-infrared 740nm). Given the fluorescence quantum efficiency, its dynamic response to environmental stress is as follows: Based on fluorescence emission (determined by the PSII structure). It is related to photochemical quenching (regulated by the electron transport rate ETR). This is non-photochemical quenching (such as the lutein cycle caused by drought).

[0028] This chlorophyll fluorescence simulation module can quantify the response of photosynthetic mechanisms to adverse conditions. For example, under high-temperature stress, An increase of 40% led to Reduced by 30%.

[0029] In the three-dimensional radiative transport framework of the LESS model, the two modules achieve cross-scale coupling through bidirectional path tracing: (3); in, The canopy radiance received by the sensor is the vector sum of reflected radiance and fluorescence radiance, representing the canopy at wavelength. The total energy radiated towards the sensor. For wavelength, The reflected radiance produced by canopy reflection, The fluorescence radiance produced by the canopy fluorescence component.

[0030] For the cross-scale coupling of formula (3), which is the top-level output expression of the LESS model, it integrates the outputs of the leaf optical properties module and the chlorophyll fluorescence simulation module through bidirectional path tracing, realizing scale coupling from leaf to canopy.

[0031] in, The portion of solar radiation (or artificial light source) that reaches the sensor after being scattered (reflected and transmitted) by various components of the canopy (leaves, stems, etc.) and the soil background, excluding fluorescence emission. The optical properties of the leaves (from...) The derived function (BSDF) is used to define the two-way scattering distribution function, which describes the scattering behavior of photons interacting with leaves. Multiple scattering within the canopy and diffuse reflection from the soil background both depend on the optical properties of the leaves; therefore... It is a calculation One of the basic inputs.

[0032] final, It is generated by the modulation of canopy structure (such as leaf tilt distribution and canopy self-shading), the optical properties of the leaves (described by BSDF), and the multiple scattering effect.

[0033] against On the one hand, the leaf absorption coefficient The absorption coefficient determines the leaf's ability to absorb photosynthetically active radiation flux (PAR), thus affecting the energy source for fluorescence excitation. In the chlorophyll fluorescence simulation module, the absorption coefficient... (usually with) (Related) directly participates in fluorescence intensity calculations. Therefore, the leaf optical properties module provides key absorption parameters for fluorescence simulation. On the other hand, The fluorescence originates from chlorophyll re-emitted after leaves absorb photosynthetically active radiation (PAR), and reaches the sensor after being transported within the canopy (including fluorescence escape, multiple scattering, and self-masking processes). Its intensity depends on the PAR absorbed by the leaves, the fluorescence quantum efficiency, the canopy escape probability, and spectral characteristics. In cross-scale coupling, the transport process of these fluorescent photons within the canopy (including the fluorescence escape probability function) is simulated through bidirectional path tracing. (Due to factors such as expansion, multiple scattering, and canopy self-masking effect), the final fluorescence radiance at the top of the canopy is obtained. Therefore, It is a calculation The starting point.

[0034] Finally, the leaf optical properties module first generates the optical properties of the leaf, and the chlorophyll fluorescence simulation module calculates the fluorescence emission spectrum. Subsequently, the ray tracing algorithm simulates the following in a 3D scene: collisions between photons and vegetation components (leaf scattering described by the two-way scattering distribution function BSDF); diffuse reflection from the soil background; canopy self-shading effect; and modulation of the SIF signal by multiple scattering. The final output is the canopy radiance. Includes reflection component and fluorescence intensity The vector superposition, where Incident irradiance is provided by the integrated atmospheric module 6S (aerosol optical thickness 0.05-0.4, water vapor content: 1-4 g / cm²). The bidirectional reflectivity factor represents the reflectivity at wavelengths of ... The reflectivity of the entire canopy system (including all components such as vegetation leaves, stems, soil background, etc., and taking into account its three-dimensional structure, shadows and multiple scattering interactions) to incident light is not the optical reflectivity of the leaves, but the comprehensive reflectivity of the entire pixel (canopy) as observed from the perspective of the remote sensing sensor.

[0035] This mechanistic modeling capability enables the LESS model to decode the "biochemical-physiological-structural" cascade effect: when the canopy leaf tilt angle increases from 35° to 55°... A 22% reduction led to an observed SIF 760 Reduced by 15%. Under drought stress, through the leaf optical characteristic module, It decreased by 30%, reducing near-infrared absorption. Meanwhile, in the chlorophyll fluorescence simulation module... The increase made It decreased by 25%. Ultimately, it exhibited SIF at the canopy scale. 680 / SIF 760 The ratio characteristic increases. This collaborative system provides a virtual laboratory for monitoring vegetation photosynthesis, in which SIF 760 It is the SIF in the 760 nm band. 680 It is the SIF in the 680 nm band.

[0036] (a2) Canopy data simulation; To recreate a more realistic 3D grassland scene in the LESS model, real-world measurement parameters, such as canopy height and grass height, were used when constructing the grassland structural units. Multiple different scenes were constructed, taking into account the differences in height, spacing, and other factors between grasses in real-world scenes, thus creating a more realistic 3D structural scene. The spectral characteristics of the grass structural units included those of the leaves and soil.

[0037] The canopy height was set to 0.05 ± 0.01 m, the leaf tilt angle was spherically distributed (μ = 35°, σ = 15°), and the scene size was set to 5 m × 5 m. To ensure the simulation resolution matched the actual measurement and to obtain the most realistic simulation results, the solar zenith angle and azimuth angle were set according to the information provided in the image file, and the skylight ratio was calculated using the 6s atmospheric model. The built-in leaf optical characteristics module was used to simulate the leaf spectral information. After completing the parameter settings, the LESS model simulation was run to simulate the canopy SIF. Specific parameter settings are shown in Table 1. Table 1. Parameter settings for the LESS model

[0038] A synthetic upflow radiation spectrum dataset was created using LESS model simulations, which includes downflow irradiance: (W·m -2 ·nm -1 ), reflectivity and (W·m -2 ·sr -1 ·nm -1 These spectral data were randomly shuffled and oversampled by a set factor (e.g., three-fold oversampling) to obtain a synthetic dataset of 15,210 spectra at TOC.

[0039] (a3) Acquisition of measured data; The experiment used a GratingWorks high-resolution spectrometer (Ningbo Fulu Co., Ltd., LIBS series), with a spectral coverage of 660–780 nm, a resolution of 0.01 nm, a sampling interval of 0.007 nm, and a signal-to-noise ratio (SNR) > 500, meeting the spectral accuracy requirements for SIF inversion. The sensor was mounted 2.0 m above the ground, and a bare fiber optic probe with a 25° field of view (FOV) was used to vertically observe the grassland canopy. To simultaneously acquire incident irradiance, a path switching device (PLS) combined with a cosine corrector with a 180° FOV was used to measure the canopy incident irradiance, effectively eliminating radiation bias caused by zenith angle variations.

[0040] Data collection was conducted from 7:30 to 17:30 (UTC+8), with spectral data recorded every 3 minutes to ensure coverage of the peak noon solar altitude angle period and clear sky conditions during the observation period, with the solar altitude angle ranging from 25° to 65°.

[0041] Based on incident irradiance and canopy radiance data acquired using a high-precision spectral measurement system, a 760 nm band SIF (Spectral Fitting Method, SFM) was achieved. 760The inversion calculation of ) is performed. As a physical model-driven spectral analysis method, the SFM algorithm is based on the principle of establishing a mathematical model of the radiative transfer process, decomposing the observed spectrum into a linear combination of different component signals, thereby achieving effective extraction of weak fluorescence signals.

[0042] Compared to the traditional Fraunhofer line depth method, the SFM method exhibits better anti-interference capability and stability at a spectral resolution of (0.1-1 nm), making it particularly suitable for high spatiotemporal resolution SIF monitoring on ground platforms. Therefore, the SFM algorithm was chosen for comparative validation with machine learning.

[0043] In this embodiment, step two, introducing principal component analysis to perform dimensionality reduction preprocessing on the features in the synthetic dataset to obtain a preprocessed synthetic dataset, which is then used as the training set, specifically involves: During the operation of the machine learning model, the SHAP method is used to interpret the feature importance of the machine learning model and obtain the correspondence between SIF changes and band reflection characteristics. SHAP (SHapley Additive ex Planations) is an interpretable machine learning method based on game theory, used to quantify the contribution of features to the model output.

[0044] The synthetic data undergoes a first-order derivative transformation, followed by normalization. Principal component analysis (PCA) is then introduced to perform dimensionality reduction preprocessing on the normalized spectra, resulting in a preprocessed synthetic dataset.

[0045] Specifically, the choice of input and output for the machine learning framework is driven by the physical principles of daylight-induced chlorophyll fluorescence (SIF) inversion and computational efficiency requirements. Incident irradiance and reflected radiance are used as inputs to the machine learning model, rather than reflectance, because these quantities directly represent the radiative components in the fundamental equations: ; in, The canopy radiance received by the sensor. For parameters, This is the bidirectional reflectivity factor. This form of the above formula preserves... Physical relationships as an additional term of the reflection component.

[0046] In this embodiment, the incident irradiance and reflected irradiance across the entire wavelength range are used as input features, and solar-induced chlorophyll fluorescence (SIF) is used as the input feature. 760 Using as the target variable, a regression model based on XGBoost was constructed. To explore the contribution of different spectral bands to the model's prediction results, the SHAP (SHapley Additive exPlanations) method was used to interpret the feature importance of the model.

[0047] The model's main contribution is concentrated in the reflected radiance around 760 nm, with a significantly higher SHAP value than other bands. This indicates that SIF 760 The changes in spectral density are closely related to the reflection characteristics of this band, and the model exhibits the highest sensitivity in this band. In contrast, the contributions from other bands are relatively small, indicating that redundant spectral information may introduce noise interference into model training.

[0048] Based on this finding, this embodiment further introduces Principal Component Analysis (PCA) to perform dimensionality reduction preprocessing on the original high-dimensional spectral features, in order to retain the main information components, reduce feature redundancy, and improve the robustness of the model. The features reduced in dimensionality by PCA can better characterize spectral information and SIF. 760 The relationship between the features is analyzed to reduce the risk of overfitting from high-dimensional data. To enhance features and suppress noise, the input spectrum is subjected to a first-order derivative transformation to highlight spectral morphology features, followed by normalization to eliminate the influence of dimensions, and finally PCA is used to reduce the dimensionality to 200 feature vectors. Figure 2 To minimize the impact of PCA dimensionality reduction on the fitting results, only traditional XGBoost was used for fitting. Figure 2 It can be seen that when PCA is not used for dimensionality reduction (PCA=0), the coefficient of determination R is... 2 The RMSE (Root Mean Square Error) results were poor. When the dimensions were reduced to 200, the best fitting results were achieved with an RMSE of 0.0531 and R0. 2 =0.9879.

[0049] from Figure 3 The results show that PCA dimensionality reduction significantly improves the performance of all models. In terms of accuracy, the predictive ability of all models is significantly enhanced after using PCA, for example, the R value of CNN is significantly improved. 2 The value increased from 0.855 without PCA to 0.975, while the MAE decreased from 0.112 to 0.041; the MLP value increased without PCA. 2 The accuracy jumped from 0.769 to 0.952 after using PCA, an improvement of nearly 24%. In terms of efficiency, PCA significantly reduced training time, decreasing CNN training time from 5126.9 seconds to 112.8 seconds (approximately 45 times faster), and RF training time from 3664.0 seconds to 123.6 seconds (approximately 27 times faster). Even for models less sensitive to high-dimensional data (such as XGBoost), PCA enabled them to maintain high accuracy (R²). 2 While achieving a precision of 0.988, the training time was kept to an efficient level of 3.0 seconds. In summary, whether from the perspective of improving accuracy or optimizing efficiency, PCA demonstrates clear and significant value in this scenario, and is an effective means to enhance the overall performance of the model.

[0050] The preprocessed synthetic dataset was used as the training set to train various machine learning models, establishing a model from spectral derivative to SIF. 760 The nonlinear mapping relationship is established. This machine learning model, while maintaining the physical mechanism of the LESS model, improves the inversion speed by several orders of magnitude, achieving rapid SIF estimation capabilities.

[0051] In one embodiment, step three, introducing different levels of noise into the training set to train each machine learning model, yields the trained machine learning models, specifically as follows: To enhance the robustness of the model, different levels of noise were introduced into the training set, and a systematic comparison was made between various machine learning models (RF, XGBoost, DSNN, CNN, MLP, BP). RF stands for Random Forest, XGBoost is Extreme Gradient Boosting, DSNN is a two-input spectral neural network, CNN is a convolutional neural network, MLP is a multilayer perceptron, and BP is a backpropagation neural network.

[0052] Figure 4 The performance of each machine learning model on the independent test set is summarized. Overall, all machine learning models have high prediction accuracy (R²>0.95) on the synthetic dataset (i.e., simulated data), and can effectively learn the derivative from the synthetic dataset to SIF. 760 The nonlinear mapping relationships were analyzed. XGBoost performed best, achieving the highest accuracy (R²=0.9987, RMSE=0.0163) under noise-free conditions, while maintaining strong stability even after noise was introduced. Random Forest (RF) also performed well under noise-free conditions (R²=0.9964, RMSE=0.0269), but its robustness under noise enhancement conditions was slightly lower than XGBoost. Deep neural network methods (DSNN and CNN) showed good overall performance (R² around 0.99), with DSNN maintaining high accuracy even after noise enhancement, demonstrating good generalization ability. MLP and BP performed relatively worse (R²≈0.92–0.97), being more sensitive to noise.

[0053] In one embodiment, step four involves preprocessing the incident irradiance and reflected irradiance from the measured dataset and using them as input to each trained machine learning model. The optimal machine learning model is obtained by comparing the performance metrics of the outputs of each model. Specifically: The preprocessing process for incident irradiance and reflected irradiance in the measured dataset is the same as that for the synthetic dataset. The preprocessing process is as follows: principal component analysis is introduced to perform dimensionality reduction preprocessing on the features in the measured dataset to obtain the preprocessed measured dataset.

[0054] The preprocessed incident irradiance and reflected irradiance are input into trained XGBoost, RF, DSNN, CNN, BP, and MLP models to predict SIF. 760 The SIF value calculated using the SFM algorithm was used as a reference for verification.

[0055] Figure 5 This demonstrates how the SFM model compares with several other prediction models (XGBoost, DSNN, RF, CNN, BP, and MLP) in SIF. 760 The performance of the models in the prediction task is compared to the SIF reference value calculated by SFM. However, different models exhibit different errors, with models having smaller residuals indicating higher prediction accuracy. In contrast, the DSNN model performs better in prediction and has smaller residuals.

[0056] For performance metrics of XGBoost, RF, DSNN, CNN, BP, and MLP, please refer to Table 2. Table 2 Performance Indicators

[0057] Based on the performance metrics comparison in Table 2, DSNN significantly outperforms other models in terms of model accuracy, achieving the best performance (R²). 2 =0.9314, RMSE=0.0429, MAE=0.0322). Following closely behind is XGBoost (R... 2 =0.9193, RMSE=0.0465) and Random Forest (R 2 =0.9189, RMSE=0.0466), both showed similar performance on the actual test dataset, demonstrating strong stability.

[0058] In contrast, CNN(R) 2 =0.9099, RMSE=0.0492, MAE=0.0388) and MLP(R 2 The performance of the BP neural network (RMSE=0.0516, MAE=0.0406) was slightly inferior, indicating that it still has some shortcomings in complex feature extraction. It is worth noting that the BP neural network performed the worst (R... 2 =0.6873, RMSE=0.0915, MAE=0.0669), although its relative error was low (13.9%), its overall fitting ability and generalization performance were significantly inferior to other models.

[0059] Overall, the neural network-based DSNN demonstrated the strongest adaptability and stability on the experimental dataset. For example, its RMSE was reduced by approximately 8.0% compared to Random Forest (0.0429 vs 0.0466), and its MAE was reduced by approximately 6.4% compared to XGBoost (0.0322 vs 0.0344), further validating the advantages of DSNN's dual-channel structure in spectral feature modeling. Considering all four core metrics (R², RMSE, MAE, Relative Error), the DSNN algorithm performed best in external data prediction, indicating higher modeling accuracy and robustness in the SIF inversion task.

[0060] In this embodiment, daily variation data is used to verify that the DSNN algorithm is the optimal machine learning model, specifically as follows: The SIF emission intensity of plants varies with daytime photosynthetically active radiation flux (PAR), chlorophyll content, and photosynthetic intensity. To better evaluate the method of this embodiment, predictions for three days, from May 1st to May 3rd, 2025, were calculated to measure the model's validity. Data within a single day are displayed chronologically to observe the dynamic changes in SIF values ​​due to diurnal variations. The R-values ​​of different machine learning models were compared. 2 The metrics revealed that the DSNN model achieved the highest R-value on all three datasets. 2 The R-values ​​(5.1: 0.979, 5.2: 0.961, 5.3: 0.955) significantly outperform other comparative models (including BP neural network, convolutional neural network CNN, multilayer perceptron MLP, random forest RF, and XGBoost). This indicates that the DSNN model can more accurately capture the diurnal variation pattern of SIF and has excellent predictive performance and stability. Furthermore, the DSNN model maintains high R-values ​​on datasets from three different dates. 2 The values ​​further validated its good generalization ability. Therefore, the DSNN model provides an effective tool for accurate prediction of SIF, which helps to further understand the dynamics of plant photosynthesis.

[0061] This embodiment employs a machine learning model and a PCA framework. The performance difference mechanism of the machine learning model in SIF inversion is analyzed as follows: This embodiment evaluates the performance differences of various machine learning models in the sunlight-induced chlorophyll fluorescence retrieval task. During the validation process from synthetic datasets (i.e., simulated data) to measured datasets (i.e., actual measured data), different machine learning models exhibited significant performance divergence. This phenomenon can be analyzed in depth from two dimensions: the structural features of the machine learning models and the characteristics of the spectral data.

[0062] Traditional tree models (RF, XGBoost) exhibit excellent nonlinear feature capture capabilities in synthetic datasets, primarily due to their decision tree-based ensemble learning mechanism. Random Forest constructs multiple decision trees through bootstrap sampling and random feature selection, effectively reducing model variance; while XGBoost achieves high-precision fitting of complex nonlinear relationships through a gradient boosting framework combined with regularization constraints. However, in real-world observation scenarios, the spatial heterogeneity of natural canopy structures, the temporal variability of atmospheric conditions, and the randomness of instrument noise collectively constitute a more complex spectral response environment. Spectral features in such real-world observation environments often exhibit higher-dimensional nonlinear coupling relationships, exceeding the feature representation range of traditional tree models, thus leading to performance degradation on real-world datasets.

[0063] In contrast, deep neural network architectures (especially DSNN and CNN) exhibit unique advantages in complex feature extraction. The dual-branch structure used in DSNN models has a clear physical meaning: one branch is dedicated to handling incident irradiance. The spectral characteristics, another branch processes canopy radiance. The spectral characteristics, this separation feature learning mechanism and the physical process of SIF formation ( This approach is highly compatible. Convolutional neural networks, through local receptivity and weight sharing mechanisms, construct a hierarchical feature extraction network in the spectral dimension, effectively identifying local gradient features in the fluorescence-sensitive band around 760nm. This structural characteristic allows it to maintain robust feature recognition capabilities even when faced with complex spectral interference in real-world observation environments.

[0064] In addition, the role of PCA dimensionality reduction in spectral feature optimization is as follows: The successful application of PCA in this embodiment reveals an important principle in high-dimensional spectral data processing. The original hyperspectral data (650-780 nm, 0.035 nm resolution) in both the synthetic and experimental datasets contain approximately 3648 feature dimensions, including information closely related to SIF emission, as well as a large amount of redundant information and random noise. PCA projects the original spectrum into an orthogonal feature space through eigenvalue decomposition, achieving effective separation of signal and noise. From an information theory perspective, the essence of PCA dimensionality reduction is to optimally compress the spectral feature space while maximizing information entropy. In this embodiment, 200 principal components are selected. This parameter is established based on a systematic analysis of the eigenvalue decay curve: the first 200 principal components have a cumulative contribution rate exceeding 99.5%, while the eigenvalues ​​of subsequent components rapidly decay to noise levels. This feature selection strategy significantly reduces the risk of overfitting while ensuring information integrity. Particularly noteworthy is that the feature space processed by PCA not only has reduced dimensionality, but more importantly, the orthogonality between its dimensions eliminates the multicollinearity problem among the original spectral features. This characteristic is crucial for improving the training stability and generalization ability of machine learning models. For example, the training time of a CNN model after PCA processing was drastically reduced from 5126 seconds to 112 seconds, while the prediction accuracy (R²) also improved. 2 The improvement from 0.855 to 0.975 fully demonstrates the dual benefits of feature optimization for improving the performance of machine learning models.

[0065] Therefore, based on this embodiment, the following conclusions can be drawn: Machine learning models exhibit varying performance characteristics across different validation scenarios. Traditional tree models (such as RF and XGBoost) effectively capture nonlinear feature interactions and achieve high accuracy on simulated data. However, their performance on real observational data is often affected by the heterogeneity of natural canopy structures, leading to decreased accuracy. This limitation highlights the inadequacy of relying solely on raw hyperspectral data for modeling. In contrast, machine learning methods incorporating PCA effectively mitigate the negative impacts of noise interference and redundant features by compressing hyperspectral data into a low-dimensional latent space. For example, the DSNN model, after incorporating PCA features, achieves optimal performance on experimental datasets (R0). 2 =0.9314, MSE=0.0429, MAE=0.0322), significantly outperforming most models in the case of no dimensionality reduction. This advantage stems from PCA's feature extraction capabilities, which not only preserve the main spectral variation information but also enhance the generalization ability of different models. Deep learning models such as CNN and MLP also perform robustly on inputs after PCA dimensionality reduction (R... 2The R² > 0.90 demonstrates the dual value of this method in reducing model complexity and improving computational efficiency. In contrast, the BP network still performs poorly even with PCA feature support (R² > 0.90). 2 =0.6873), indicating that it has inherent limitations in capturing complex spectral-structure relationships.

[0066] This embodiment significantly improves the computational efficiency and stability of SIF inversion by combining principal component analysis (PCA) and a machine learning framework. Traditional spectral fitting (SFM)-based inversion methods are often computationally expensive and sensitive to scene complexity; while PCA dimensionality reduction reduces computational burden while maintaining prediction accuracy, enabling the model to quickly process large-scale observational data. This framework achieves stable advantages in key metrics such as RMSE and MAE, and keeps error fluctuations within a low range, thus enabling high-resolution and rapid SIF inversion.

[0067] Therefore, the PCA+machine learning model framework set up in this embodiment significantly improves the computational efficiency and stability of SIF inversion. It provides a reliable technical option for SIF inversion on ground-based and future satellite platforms.

[0068] Based on the above description of the embodiments, those skilled in the art will understand that the method, system, and medium for machine learning to invert sunlight-induced chlorophyll fluorescence described in this embodiment can be implemented purely in software or deployed and run on a general-purpose or dedicated computing hardware platform. Based on this essence, the technical solution of this embodiment can be specifically implemented in the form of a software product containing program instructions. This software product can be stored on various non-volatile storage media or directly deployed as a local or cloud service. The program instructions are used to cause computer devices with processing capabilities—including but not limited to personal computers, server clusters, mobile terminals, or other network devices—to execute the steps described in this embodiment.

[0069] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for machine learning to invert sunlight-induced chlorophyll fluorescence, characterized in that, include: Obtain the actual test dataset and generate the synthetic dataset; Principal component analysis is introduced to perform dimensionality reduction preprocessing on the spectral data in the synthetic dataset, resulting in a preprocessed synthetic dataset, which is then used as the training set. Different levels of noise are introduced into the training set to train each machine learning model, resulting in trained machine learning models. After preprocessing the incident irradiance and reflected irradiance in the measured dataset, they are used as inputs to each trained machine learning model. The optimal machine learning model is obtained by comparing the performance metrics of the output results of each machine learning model. The task of retrieving sunlight-induced chlorophyll fluorescence was performed based on the optimal machine learning model.

2. The method according to claim 1, characterized in that, The generation of the synthetic dataset specifically involves: The LESS model, which includes a leaf optical properties module and a chlorophyll fluorescence simulation module, is used. The LESS model is a large-scale remote sensing data and image simulation framework model. The blade optical properties module serves as the calculation engine for the blade's optical properties. It uses a layered flat plate model to quantify the absorption process of photosynthetically active radiation and generate the blade's optical properties. The chlorophyll fluorescence simulation module is extended into a photosynthesis-fluorescence coupling system, and the distribution path of absorbed energy is deconstructed through a quenching competition model to calculate the fluorescence emission spectrum; The leaf optical properties module and the chlorophyll fluorescence simulation module achieve cross-scale coupling through bidirectional path tracing, outputting canopy reflectivity, which is then converted into canopy radiance. Grassland structural units were constructed using real measurement parameters, and the grassland structural units were run in the LESS model to simulate canopy sunlight-induced chlorophyll fluorescence, thereby creating a synthetic upflow radiation spectrum dataset. The spectral dataset is randomly shuffled and oversampled by a set factor to generate a synthetic dataset.

3. The method according to claim 2, characterized in that, The core equation of the blade optical property module is expressed as follows: ； in, The absorption coefficient of the leaf. For wavelength, These are the blade structure parameters. These are chlorophyll content, equivalent water thickness, and dry matter content, respectively. For chlorophyll at wavelength The specific absorption coefficient at that location, For water at wavelength The specific absorption coefficient at that location, For dry matter at wavelength The specific absorption coefficient at that location.

4. The method according to claim 2, characterized in that, The core formula of the chlorophyll fluorescence simulation module is as follows: ； in, To be at wavelength The fluorescence intensity emitted at that location, The fluorescence emission rate constant is . For heat dissipation rate constants other than those for photochemical quenching, It is a non-photochemical quenching rate constant. Photosynthetically active radiation flux Let f be the fluorescence escape probability function. For the blade at wavelength The absorption coefficient at that location.

5. The method according to claim 2, characterized in that, The cross-scale coupling formula is as follows: ； in, The radiance of the canopy received by the sensor is the vector sum of the reflected radiance and the fluorescence radiance. For wavelength, The reflected radiance produced by canopy reflection, The fluorescence radiance produced by the canopy fluorescence component.

6. The method according to claim 1, characterized in that, The process of generating the preprocessed synthetic dataset is as follows: During the operation of the machine learning model, the SHAP method was used to interpret the feature importance of the machine learning model and obtain the correspondence between sunlight-induced chlorophyll fluorescence changes and band reflectance characteristics. The spectra in the synthetic data are subjected to first-order derivative transformation, followed by normalization. Principal component analysis is then introduced to perform dimensionality reduction preprocessing on the normalized spectra, resulting in the preprocessed synthetic dataset.

7. The method according to claim 1, characterized in that, The machine learning models mentioned are: random forest, extreme gradient boosting, convolutional neural network, and dual-input spectral neural network.

8. The method according to claim 1, characterized in that, Machine learning models use the metric R during training. 2 The root mean square error is used to evaluate the model's fit and prediction accuracy.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the method according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of classification programs, which are used by a processor to execute the method as described in any one of claims 1-8.

Images