CN-122023955-A - Method for extracting and inverting hyperspectral characteristic wave bands

Abstract

The invention discloses a hyperspectral characteristic wave band extraction and inversion method for the remote sensing field, which comprises ADSPA characteristic wave band extraction and B-XGBoost inversion. The characteristic wave band extraction comprises data collection and preprocessing to obtain target variable and hyperspectral image data, defining a set of unselected wave bands, randomly selecting initial iteration wave bands, setting random seeds and setting iteration times N to construct ADSPA on the basis of SPA, and extracting characteristic wave bands after the wave band reflectivity data are iterated on ADSPA. The inversion model adopts a Bayesian optimization strategy to tune a plurality of parameters of the model, sets the optimal parameter result as the model parameter of B-XGBoost, inverts the characteristic wave band output by ADSPA, and outputs the R2, RMSE and MAE results to evaluate the accuracy. Finally, the SHAP value may also be calculated, and the degree of contribution of the characteristic bands screened by ADSPA to the inversion may be again assessed using the SHAP value.

Inventors

• ZHAO BO
• ZHANG ANBING
• LIU XINXIA
• Hou Yikai
• JIA PENGFEI
• WANG ZEPENG
• ZHANG XUDONG

Assignees

• 河北工程大学
• 邯郸职业技术学院

Dates

Publication Date

20260512

Application Date

20250814

Claims (3)

1. 1. The hyperspectral characteristic wave band extraction and inversion method is characterized by comprising ADSPA characteristic wave band extraction and B-XGBoost inversion, wherein the step of ADSPA characteristic wave band extraction comprises the following steps: 1) Data collection and preprocessing: collecting hyperspectral image data by an unmanned aerial vehicle and a hyperspectral sensor, and synchronously carrying out sample collection, wherein an actual monitoring index of a sample is used as a target variable, and hyperspectral image data is preprocessed to obtain a hyperspectral reflectivity matrix, wherein the preprocessing comprises SG smoothing processing; 2) Construction ADSPA: define a set Z of unselected bands: ; Selecting an initialization wave band, namely randomly selecting an index k (0) column from the set Z as an initial wave band; Setting a random seed, wherein the value range of the random seed is set to be 0-100, and the first iteration is set to be 0; setting the range of iteration times N to be more than or equal to 5 and less than or equal to 50; For each band index j ∉ Z of the unselected bands in the set Z, calculate the orthogonal projection of the jth column data vector Spec j in the hyperspectral matrix in the selected band subspace as follows: ; Calculating the L2 norm is calculated with the orthogonal projection as P Specj ∥ 2 : ; And selecting a characteristic wave band index, namely finding the maximum norm from the L2 norms calculated in the previous step, and selecting the characteristic wave band index k (n) corresponding to the maximum norm from the set Z: ; Adding k (n) to the characteristic band index set, removing k (n) from the set Z, replacing n with n+1 for the number of iterations, and setting k (n) as an initial band X k （n） of the next iteration; Outputting a final characteristic band index set, namely returning to continue iteration if N is less than N, and outputting the final characteristic band index set if n=N and the iteration is finished: ; The final characteristic wave band index set W is evaluated, wherein the set W is divided into 5 parts, a random forest regressor is trained through 5 times of cross validation, and the decision coefficient R2 predicted by the random forest regressor is validated each time; 3) Extracting characteristic wave bands: And 3) respectively combining the target variable in the step 1) with the hyperspectral reflectance matrix of each wave band to form an input matrix ADSPA, and outputting a final characteristic wave band index set after the calculation in the step 2) to finish the extraction of characteristic wave bands.
2. 2. The method of hyperspectral signature band extraction and inversion as claimed in claim 1 wherein said B-XGBoost inversion includes the steps of: 1) B-XGBoost model construction, wherein the XGBoost model is a B-XGBoost model by adopting Bayesian optimization, and a global search key parameter n_ estimators, max _ depth, learning _rate is obtained; 2) And B-XGBoost inversion, wherein n_ estimators, max _ depth, learning _rate is set as a B-XGBoost model parameter, a characteristic wave band dataset output by ADSPA is read for inversion, the inversion comprises model training and parameter optimization, 5-fold cross validation and target optimization, and R2, RMSE and MAE are output after inversion so as to evaluate the accuracy of the characteristic wave band.
3. 3. The method for extracting and inverting hyperspectral characteristic bands as claimed in claim 2, wherein the B-XGBoost is inverted and then subjected to SHAP value analysis, and the SHAP value is calculated according to the following formula: 。

Description

Method for extracting and inverting hyperspectral characteristic wave bands Technical Field The invention is applied to the field of remote sensing monitoring, relates to a hyperspectral technology, and particularly relates to a method for extracting and inverting hyperspectral characteristic wave bands by combining a self-adaptive continuous projection algorithm (ADSPA) and an extreme gradient lifting (XGBoost) model. Background The remote sensing technology, in particular to the unmanned aerial vehicle hyperspectral technology, can realize environmental monitoring with large range, high frequency and high precision. However, hyperspectral data typically contains a large amount of redundant information, and efficient dimension reduction methods are required to improve inversion accuracy, and typically involve feature band extraction and data compression. The characteristic wave band is that a kind of target only has certain sensibility to the incident light of a specific wave band, and the wave band can better reflect the information of the property, the characteristics and the like of the target ground object, namely the characteristic wave band. In the field of remote sensing water quality monitoring, the method for monitoring the water quality by adopting the hyperspectral technology has more patent applications, and the method has the advantages of low requirement on a training sample, simple measurement and more accurate result compared with the prior art because the training sample is used for training to obtain the DBN dimension reduction model and the ELASTICNET inversion model, for example, CN110567905B is used for training. While DBNs are capable of extracting deep features, the hierarchical structure of DBNs may limit their flexibility in feature combination and abstraction. CN117092047A punctures satellite hyperspectral remote sensing data to a determined band range based on a satellite hyperspectral image, determines a high-correlation two-band data set with a determination coefficient larger than a determination coefficient threshold value, traverses and solves all comprehensive remote sensing reflectances meeting requirements according to the principle that the high-correlation two-band does not appear in a characteristic remote sensing band of the same water quality index, traverses and establishes an inversion model between a target water quality index and the comprehensive remote sensing reflectances by adopting an artificial neural network method, and estimates the water quality index of a target area by adopting the established inversion model. The method can be used for quickly and accurately determining the characteristic remote sensing wave bands of different water quality indexes, and has more accurate, more precise and more reliable characteristic extraction and water quality index estimation capability. However, the method only uses a linear interpolation method to reduce the dimension, nonlinear characteristics in the data may not be fully extracted, and an artificial neural network model may have an overfitting problem, so that the generalization capability of the model is not strong. The CN119337251A integrated hyperspectral water quality analysis method adopts three dimension reduction methods to carry out dimension reduction treatment, namely a principal component analysis method, t-distributed random neighborhood embedding, uniform manifold approximation and projection, and fusion dimension reduction through parameter balance selection, and adopts three machine learning algorithms of a support vector machine, a random forest and a decision tree to carry out training and testing of a hyperspectral water quality inversion model on the spectrum data after dimension reduction, and finally, a decision coefficient R2 and a root mean square error RMSE are selected as test indexes of fitting precision of the hyperspectral water quality inversion model. Although a plurality of dimension reduction methods are integrated, the distribution of weights in the process of merging dimension reduction may be subjective, a method for automatically optimizing the weights is lacked, and optimization of the number of characteristic wave bands is lacked, so that unstable performance of a model may be caused. The continuous projection algorithm (SPA) is a feature selection method for minimizing the co-linearity of vector space through forward circulation, and utilizes the projection analysis of vectors to select the feature wave bands with the lowest redundancy and the smallest co-linearity, so that the number of parameters required by modeling is reduced, the modeling efficiency is improved, and the covariance matrix of the whole data set is not required to be calculated, which is particularly useful in processing large-scale data sets. Furthermore, the SPA can be adapted gradually to the nature of the data, which makes it potentially more promising than traditional approaches in some cases. Journ