A method for inverting stellar spectral parameters based on a feature fusion random forest model

Abstract

The present application relates to a kind of star spectrum parameter inversion methods based on feature fusion random forest model, first to star spectrum catalogue and spectrum file are matched and screened, according to public wavelength intersection completes sample alignment, then obtains stable spectrum characterization using Gaussian smoothing self-normalization;Subsequently, standard deviation, skewness, kurtosis and equivalent width features are extracted from global statistics, segmented local statistics and sensitive spectral line window, and are spliced into fusion features with self-normalized spectrum sequence;Finally, effective temperature Teff, surface gravity logg and metal abundance [Fe / H] are jointly estimated using random forest regressor.Compared with prior art, the present application has the advantages of considering original sequence information and physical prior feature information, improving star parameter inversion accuracy, enhancing model robustness and interpretability, etc.

Description

Technical Field

[0001] This invention relates to the technical field of parameter inversion, and in particular to a method for stellar spectral parameter inversion based on a feature fusion random forest model. Background Technology

[0002] Low-resolution stellar spectra contain rich physical information such as stellar temperature, surface gravity, and metallicity, making them an important data source for rapid inversion of stellar physical parameters in large-scale sky survey data processing. Traditional parameter estimation methods based on empirical templates, spectral line indices, or single statistical features often struggle to comprehensively utilize both full-spectrum morphological information and local information from sensitive spectral lines.

[0003] When stellar spectra are directly used for machine learning modeling, problems often arise such as continuous spectrum fluctuations, incomplete consistency in sample wavelength coverage, local anomalous pixels, and insufficient utilization of spectral window information. If only the entire spectral sequence is used for modeling, although fine-grained information is preserved, the model has a weak explicit perception of physically sensitive features; if only a small number of manually statistical features are used, it is easy to cause excessive information compression, resulting in limited parameter inversion capabilities.

[0004] Random forest models possess advantages such as strong ability to model nonlinear relationships, robustness to noise, high training efficiency, and a certain degree of feature interpretation capability, making them suitable for stellar spectral parameter inversion tasks. However, existing random forest schemes often lack specialized preprocessing and feature fusion mechanisms for stellar spectral characteristics, and cannot fully leverage the synergistic effect of full-spectrum sequence information and physically inspired features. Summary of the Invention

[0005] The purpose of this invention is to provide a method for stellar spectral parameter inversion based on a feature fusion random forest model. Addressing the problems of inconsistent wavelength coverage, continuous spectrum fluctuations, and the difficulty of a single representation method in simultaneously considering global and local spectral information in stellar spectral parameter inversion, this invention proposes a feature fusion random forest model for stellar spectral parameter inversion to overcome these issues.

[0006] The objective of this invention can be achieved through the following technical solutions: A method for inverting stellar spectral parameters based on a feature fusion random forest model, comprising the following steps: S1, matching and quality-screening the stellar spectral catalog data and spectral files to construct a sample set that can be used for parameter inversion; S2, scan the common wavelength intersection range of the sample set, and perform finite value filtering and length alignment on the spectral sequence within the common wavelength intersection to obtain the aligned spectrum; S3 performs Gaussian smoothing and self-normalization on the aligned spectra and constructs the normalized spectral sequence channels within the sample. S4. Based on the Gaussian smoothed and self-normalized spectrum, structured statistical features are extracted. The structured statistical features include global statistical features, local statistical features, and sensitive spectral line features. S5, the spectral sequence channels are spliced ​​and fused with structured statistical features to obtain a feature fusion input vector; S6. Establish a random forest regression model, input the feature fusion input vector into the random forest regression model, and perform joint inversion of the star's effective temperature Teff, surface gravity logg, and metallicity [Fe / H] to output the estimated results of the star's spectral parameters.

[0007] Furthermore, the quality screening includes signal-to-noise ratio screening and parameter physical range screening.

[0008] Furthermore, the common wavelength intersection interval is obtained by taking the maximum value of the minimum wavelength and the minimum value of the maximum wavelength in the sample set; When performing finite value filtering and length alignment on a spectral sequence within a common wavelength intersection, the spectrum is first truncated based on the wavelength mask, then finite value filtering is performed on the spectral flux and wavelength in combination, and the shortest effective length within the intersection is used as the uniform length for forward truncation alignment.

[0009] Furthermore, the Gaussian smoothing self-normalization process is as follows: the smoothing kernel width is calculated based on the wavelength sampling interval, and the original flux is divided by the Gaussian smoothing result to obtain the Gaussian smoothed self-normalized spectrum.

[0010] Furthermore, the global statistical features include the standard deviation, skewness, and kurtosis of the Gaussian-smoothed and self-normalized spectrum; the local statistical features include dividing the Gaussian-smoothed and self-normalized spectrum into 8 sub-segments along the wavelength direction and calculating the standard deviation, skewness, and kurtosis of each sub-segment.

[0011] Furthermore, the sensitive spectral line characteristics are obtained by calculating the standard deviation and equivalent width within the preset spectral line window.

[0012] Furthermore, the equivalent width is calculated using the percentile continuous spectrum estimation method, specifically: within each spectral window, the percentile value of the set of edge pixels at the blue and red ends is selected as the endpoint of the continuous spectrum, and the line connecting the endpoints is used as the baseline of the continuous spectrum. The equivalent width is obtained by wavelength integration of 1 minus the normalized linear curve.

[0013] Furthermore, the feature fusion input vector is composed of a concatenation of spectral sequence channels and structured statistical feature vectors.

[0014] Furthermore, the random forest regression model employs a multi-output regression structure.

[0015] Furthermore, the parameter inversion expression for the random forest regression model is: ŷ_RF=f_RF([vec(Z_self(x,λ)),φ(Z_self(x,λ),λ)]) Where x represents the input stellar spectral flux, λ represents the corresponding wavelength sequence, Z_self represents the Gaussian smoothing self-normalization operator, vec(·) represents the sequence flattening operator, and φ(·) represents the structured statistical feature extraction operator.

[0016] Compared with the prior art, the present invention has the following beneficial effects: First, by using common wavelength intersection and self-normalization, the adverse effects of differences in wavelength ranges among different samples and continuous spectrum variations on modeling are reduced. Second, by fusing sequence information with physically inspired statistical features, the model's comprehensive perception of global spectral morphology, local spectral perturbations, and sensitive spectral line information is improved. Third, the random forest model is stable to train, simple to implement, and fast to infer, making it suitable for large-scale, low-resolution sky survey spectral parameter inversion scenarios. Attached Figure Description

[0017] Figure 1 This is a flowchart of a feature fusion random forest model for stellar spectral parameter inversion in an embodiment of the present invention; Figure 2 This is a summary graph of the performance of the feature fusion random forest model in the embodiments of the present invention on the validation set. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are implemented based on the technical solution of the present invention, providing detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0019] This invention proposes a feature-fusion random forest model for stellar spectral parameter inversion, belonging to the field of stellar spectral analysis and astrophysical parameter estimation technology. Addressing the problems of continuous spectral fluctuations, insufficient utilization of spectral sensitive regions, and limited single-feature characterization capabilities in the joint inversion of temperature, surface gravity, and metallicity in low-resolution stellar spectra, this invention proposes a random forest parameter inversion scheme based on the fusion of spectral sequences and structured statistical features. The scheme first matches and filters stellar spectral catalogs and spectral files, aligning samples based on the intersection of common wavelengths, and then uses Gaussian smoothing and self-normalization to obtain stable spectral representations. Subsequently, standard deviation, skewness, kurtosis, and equivalent width features are extracted from global statistics, piecewise local statistics, and sensitive spectral window, and concatenated with the self-normalized spectral sequence to form fused features. Finally, a random forest regressor is used to jointly estimate the effective temperature (Teff), surface gravity (logg), and metallicity [Fe / H]. This invention takes into account both original sequence information and prior physical feature information, improving the accuracy of stellar parameter inversion, enhancing model robustness and interpretability, and is suitable for rapid estimation of spectral parameters in large-scale surveys.

[0020] The present invention adopts the following technical solution: First, the star catalog and spectral data are matched and quality controlled to select samples that meet the constraints of signal-to-noise ratio and physical parameter range; then, the common wavelength intersection of all samples is scanned, and effective wavelength interception, finite value filtering and length alignment are uniformly completed within the intersection; then, Gaussian smoothing and self-normalization are applied to the aligned spectrum to obtain a stable morphological characterization; based on the self-normalized spectrum, global statistical features, 8 local statistical features and sensitive spectral window features are extracted simultaneously, and the obtained structured features are spliced ​​and fused with the flattened vector of the self-normalized spectral sequence; finally, the data are input into a random forest regression model to output the estimated values ​​of Teff, logg and [Fe / H].

[0021] In addition to fluctuation statistics within the window, the sensitive spectral line window feature also uses the percentile continuous spectrum method to calculate the equivalent width of the spectral line, thereby enhancing the model's ability to characterize absorption line depth, line shape, and metal-sensitive regions.

[0022] Random forests employ a multi-output regression structure, utilize random sampling and multi-tree ensemble mechanisms to suppress overfitting, and achieve nonlinear mapping of stellar parameters through learning fused features.

[0023] In this embodiment, the stellar parameter catalog file and the corresponding FITS spectral file are first read, and a one-to-one correspondence between the catalog data and the spectral data is established based on the observation number OBSID. For each spectrum, the flux sequence and wavelength sequence are read, and multi-band signal-to-noise ratio information is extracted. If a sample has empty spectra, invalid wavelengths, a g-band signal-to-noise ratio below a threshold, or Teff, logg, and [Fe / H] exceeding a preset physical range, then the sample is discarded.

[0024] For samples that meet the criteria, a common wavelength intersection interval [w_min, w_max] is first determined based on the effective wavelength range of each sample. Then, the spectrum is truncated within this intersection interval. To reduce the impact of NaN and infinite values ​​on modeling, joint finite value filtering is performed on flux and wavelength within the intersection. If the number of effective pixels after filtering is insufficient or the variance is zero, the sample is discarded. For the retained samples, the shortest effective length within the intersection is used as the uniform length for forward truncation, thereby ensuring that the input length of all samples is consistent.

[0025] For each aligned spectrum, two types of sequence channels are constructed. The first is a historically compatible intra-sample z-score channel, which is normalized using the median as the center and the standard deviation as the scale. The second is a Gaussian smoothed self-normalized channel. Specifically, a Gaussian kernel equivalent to a 50 Å smoothing scale is constructed based on the wavelength sampling interval, and convolution smoothing is performed on the original spectrum. Then, the ratio of the original spectrum to the smoothed result is used as the self-normalized output, and further intra-sample standardization is performed. The subsequent fusion model preferentially uses the second type of channel as the primary sequence representation.

[0026] In the structured feature extraction stage, the standard deviation, skewness, and kurtosis of the entire self-normalized spectrum are first calculated as global statistical features. Then, the spectrum is divided into 8 local segments, and the standard deviation, skewness, and kurtosis are calculated for each segment, resulting in a total of 24 local statistical features. Then, the standard deviation and equivalent width features of the window are extracted in the predefined sensitive spectral window H_alpha, H_beta, H_gamma, Fe5270, Fe5015, Mgb_5175, NaD, and CaII_8542, respectively, to finally form a 43-dimensional structured statistical feature.

[0027] The equivalent width feature is estimated using the percentile continuous spectrum method. For a given spectral window, a certain proportion of edge pixels are selected at the blue and red ends of the line region, and their 90th percentile values ​​are calculated as the continuous spectrum levels at the blue and red ends. Then, the continuous spectrum within the window is obtained through linear interpolation. Finally, the equivalent spectral width is obtained by calculating the integral of 1 minus the linearly normalized spectrum within the window. This design improves the robustness of continuous spectrum estimation even in the presence of local noise.

[0028] The self-normalized spectral sequence is flattened and concatenated with the 43-dimensional structured statistical features to form a feature fusion input vector. A random forest regressor is then built, preferably with 200 trees, a maximum feature subset of sqrt, and a random seed of 42. The training and validation sets are divided in an 8:2 ratio for model training and performance evaluation.

[0029] In a specific experiment, the feature fusion random forest model was used to jointly invert three types of stellar parameters on the validation set. The Teff model achieved a determination coefficient R² of 0.9668, a mean absolute error (MAE) of 87.8642K, and a root mean square error (RMSE) of 131.9420K. For logg, the R² was 0.7053, MAE was 0.2288, and RMSE was 0.3359. For [Fe / H], the R² was 0.4661, MAE was 0.1766, and RMSE was 0.2400. The corresponding mean residuals were approximately -5.7526, -0.0098, and 0.0079, respectively, indicating that the overall model bias was small.

[0030] like Figure 1 As shown, the process of this embodiment includes inputting stellar spectra, extracting global statistical features, local statistical features, and sensitive spectral line features, performing normalization processing, training using a random forest algorithm, and outputting parameter prediction results. Compared with schemes that rely solely on global statistical features, adding local statistical features and sensitive spectral line features can effectively alleviate the underfitting problem and improve the accuracy of stellar parameter inversion.

[0031] like Figure 2 As shown, the scatter plot of the validation set predicted values ​​versus the actual values ​​and the residual distribution plot reveal that the Teff inversion results are more tightly distributed along the diagonal. While logg and [Fe / H] are more challenging than with Teff, the model still maintains good stability after introducing structured features. These results demonstrate that this invention can improve the stellar parameter inversion effect while maintaining the simplicity of model implementation.

[0032] In summary, this invention provides a stable, interpretable, and easily deployable technical solution for the rapid estimation of physical parameters of large-scale survey stellar spectra by fusing Gaussian self-normalized spectral sequences with multi-level statistical features and combining them with a random forest regressor to achieve joint inversion of stellar parameters.

[0033] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for inverting stellar spectral parameters based on a feature fusion random forest model, characterized in that, The method includes the following steps: S1, matching and quality-screening the stellar spectral catalog data and spectral files to construct a sample set that can be used for parameter inversion; S2, scan the common wavelength intersection range of the sample set, and perform finite value filtering and length alignment on the spectral sequence within the common wavelength intersection to obtain the aligned spectrum; S3 performs Gaussian smoothing and self-normalization on the aligned spectra and constructs the normalized spectral sequence channels within the sample. S4. Based on the Gaussian smoothed and self-normalized spectrum, structured statistical features are extracted. The structured statistical features include global statistical features, local statistical features, and sensitive spectral line features. S5, the spectral sequence channels are spliced ​​and fused with structured statistical features to obtain a feature fusion input vector; S6. Establish a random forest regression model, input the feature fusion input vector into the random forest regression model, and perform joint inversion of the star's effective temperature Teff, surface gravity logg, and metallicity [Fe / H] to output the estimated results of the star's spectral parameters.

2. The method for stellar spectral parameter inversion based on a feature fusion random forest model according to claim 1, characterized in that, The quality screening includes signal-to-noise ratio screening and parameter physical range screening.

3. The method for stellar spectral parameter inversion based on a feature fusion random forest model according to claim 1, characterized in that, The common wavelength intersection interval is obtained by taking the maximum value of the minimum wavelength and the minimum value of the maximum wavelength in the sample set; When performing finite value filtering and length alignment on a spectral sequence within a common wavelength intersection, the spectrum is first truncated based on the wavelength mask, then finite value filtering is performed on the spectral flux and wavelength in combination, and the shortest effective length within the intersection is used as the uniform length for forward truncation alignment.

4. The method for stellar spectral parameter inversion based on a feature fusion random forest model according to claim 1, characterized in that, The Gaussian smoothing and self-normalization process is as follows: calculate the smoothing kernel width based on the wavelength sampling interval, and divide the original flux by the Gaussian smoothing result to obtain the Gaussian smoothed and self-normalized spectrum.

5. The method for stellar spectral parameter inversion based on a feature fusion random forest model according to claim 4, characterized in that, The global statistical features include the standard deviation, skewness, and kurtosis of the Gaussian-smoothed and self-normalized spectrum; the local statistical features include dividing the Gaussian-smoothed and self-normalized spectrum into 8 sub-segments along the wavelength direction and calculating the standard deviation, skewness, and kurtosis of each sub-segment.

6. The method for stellar spectral parameter inversion based on a feature fusion random forest model according to claim 1, characterized in that, The sensitive spectral line characteristics are obtained by calculating the standard deviation and equivalent width within the preset spectral line window.

7. The method for stellar spectral parameter inversion based on a feature fusion random forest model according to claim 1, characterized in that, The equivalent width is calculated using the percentile continuous spectrum estimation method. Specifically, the percentile value of the set of edge pixels at the blue and red ends within each spectral window is selected as the endpoint of the continuous spectrum. The line connecting the endpoints is then used as the baseline of the continuous spectrum. The equivalent width is obtained by wavelength integration of 1 minus the normalized linear curve.

8. The method for stellar spectral parameter inversion based on a feature fusion random forest model according to claim 1, characterized in that, Feature fusion The input vector is composed of a concatenation of spectral sequence channels and structured statistical feature vectors.

9. The method for stellar spectral parameter inversion based on a feature fusion random forest model according to claim 1, characterized in that, The random forest regression model employs a multi-output regression structure.

10. The method for stellar spectral parameter inversion based on a feature fusion random forest model according to claim 1, characterized in that, The parameter inversion expression for the random forest regression model is: ŷ_RF=f_RF([vec(Z_self(x,λ)),φ(Z_self(x,λ),λ)]) Where x represents the input stellar spectral flux, λ represents the corresponding wavelength sequence, Z_self represents the Gaussian smoothing self-normalization operator, vec(·) represents the sequence flattening operator, and φ(·) represents the structured statistical feature extraction operator.

Images