A characteristic spectrum extraction method for target object classification and recognition

Abstract

The present invention discloses a characteristic spectrum extraction method for target object classification and identification, including: obtaining multiple object spectra of the target object and constructing a spectrum matrix, obtaining the spectrum domain space of the target object spectrum according to the spectrum matrix, constructing multiple interpolation matrices corresponding to the multiple target object spectra with integer reflectance as row number and wavelength as column number, calculating the normalized interpolation factor of each matrix to be interpolated according to the reflectance corresponding to the wavelength and the distance in the spectrum domain space, interpolating each matrix to be interpolated by using the inverse distance interpolation method, obtaining multiple spectrum domain space matrices, accumulating multiple spectrum domain space matrices, obtaining the cumulative spectrum domain space matrix of the object, and calculating the characteristic spectrum of the target object. The characteristic spectrum constructed by this method can be used to improve the characteristic spectrum extraction process of the target object in the classification scene, improve the target recognition accuracy; optimize the construction of characteristic parameters in the inversion scene, and improve the characterization ability of the characteristic parameters.

Description

Technical Field

[0001] The present invention relates to the technical field of spectral domain interpolation, and more specifically to a characteristic spectrum extraction method for target object classification and identification. Background Art

[0002] Hyperspectral remote sensing plays an important role in national economy and national defense construction. It has been widely used in the fields of surface classification, target detection, agricultural monitoring, mineral mapping, environmental management and national defense construction. Compared with traditional optical images, hyperspectral images add one-dimensional spectral data on the basis of ordinary two-dimensional spatial images. The entire data forms a spectral image cube. The spectral data of each pixel is expanded to correspond to a spectral curve, forming a data feature of unified graph and spectrum. For hyperspectral images, whether for classification or parameter inversion, it is necessary to construct the characteristic spectral curve of the target object. Affected by the characteristics of the target object itself and environmental factors, there are certain fluctuations and oscillations between multiple spectra of the same object. Therefore, it is crucial to obtain the characteristic spectrum of the object with sufficient characterization ability for the analysis and classification of hyperspectral data. The traditional characteristic spectrum extraction method mainly takes the arithmetic mean of multiple target object spectral curves as the characteristic spectrum of the target object. Some scholars also use the fluctuation range of multiple target object spectra to form a band spectrum as the characteristic spectrum of the target object.

[0003] Traditional feature spectrum extraction methods have been widely used in hyperspectral image classification, analysis and quantitative inversion, and have achieved fruitful results. However, some common problems have also been exposed in the application process.

[0004] The arithmetic mean is a representative value with strong generalization ability, characterizing the general level of the population, and easy to calculate and analyze, but it is highly sensitive and poorly resistant, and is greatly affected by a few extreme values. This is reflected in the characteristic spectrum based on the arithmetic mean, which will weaken some characteristic spectra and produce a "peak elimination and valley flattening" effect similar to the "atmospheric effect".

[0005] Based on the band-like characteristic spectrum of the fluctuation range, several target object spectra are used to obtain the maximum and minimum values ​​of the unified band to construct a band-like characteristic spectrum. This method is also easily affected by extreme values. At the same time, compared with a single linear characteristic spectrum, its uneven spectral band morphology is not conducive to the extraction of some features and subsequent calculation and analysis. Summary of the invention

[0006] The embodiment of the present invention provides a characteristic spectrum extraction method for target object classification and identification, comprising:

[0007] Acquire multiple object spectra of the target object, and construct a spectrum matrix based on the multiple object spectra;

[0008] The spectrum matrix is ​​used to calculate the maximum and minimum values ​​of each column of the spectrum matrix to obtain the spectrum domain space of the target object;

[0009] Construct multiple interpolation matrices corresponding to multiple target object spectra with integer reflectance as row number and wavelength as column number, where the element of 1 in each column of each interpolation matrix represents that there is a corresponding reflectance at the wavelength in the column;

[0010] According to the reflectivity corresponding to the wavelength and the distance in the spectral domain space, the normalized interpolation factor of each column of each matrix to be interpolated is calculated;

[0011] Each column of each matrix to be interpolated is interpolated by using an inverse distance interpolation method to obtain multiple spectral domain space matrices;

[0012] Multiple spectral domain space matrices are accumulated to obtain the cumulative spectral domain space matrix of the object, and the maximum value of each column is calculated. According to the maximum value of each column, the characteristic spectrum of the target object is constructed.

[0013] Preferably, obtaining the spectral domain space of the target object spectrum includes:

[0014] Obtain the reflectivity corresponding to different wavelengths in multiple object spectra of the target object;

[0015] Multiply the reflectivity by 10000 to convert it into an integer;

[0016] Construct an N×M matrix representing N spectra and M wavelengths;

[0017] Calculate the maximum value max and minimum value min of each column of the N×M matrix as the boundary of the spectral domain space.

[0018] Preferably, constructing a matrix to be interpolated includes:

[0019] Construct a 10000×M matrix Ai of N ground feature spectra, i∈[1, 2, …, N], where N represents the number of spectra. An element of 1 in each column of the 10000×M matrix represents that there is a corresponding reflectivity at the wavelength in the column, and 0 represents that there is no corresponding reflectivity at the wavelength in the column.

[0020] Preferably, the normalized interpolation factor of each column of each matrix to be interpolated is calculated according to the reflectivity corresponding to the wavelength and the distance in the spectral domain space, including:

[0021] Obtain the reflectivity Rm corresponding to the wavelength in each column of the matrix to be interpolated;

[0022] Calculate the distances |max-Rm| and |min-Rm| between the maximum value max and the minimum value min corresponding to the wavelength in the spectral domain space and the reflectivity Rm;

[0023] Use 1 / |max-Rm| and 1 / |min-Rm| as normalized interpolation factors.

[0024] Preferably, a plurality of spectral domain space matrices are obtained, including:

[0025] In each column of the matrix to be interpolated, the element with value 1 is used as the boundary, and the interpolation is performed row by row until the value is not greater than 0. The normalized interpolation factor for upward decrease is 1 / |max-Rm|, and the normalized interpolation factor for downward decrease is 1 / |min-Rm|, until the value is not greater than 0;

[0026] Update N 10000×M matrices to obtain multiple spectral domain space matrices.

[0027] Preferably, accumulating a plurality of spectral domain space matrices to obtain a cumulative spectral domain space matrix of a ground object and constructing a characteristic spectrum of a target ground object comprises:

[0028] Add N 10000×M spectral domain space matrices to obtain a cumulative spectral domain space matrix;

[0029] Retrieve the maximum value of each column of the cumulative spectrum domain space matrix and record its row number to obtain a 1×M matrix;

[0030] Divide the 1×M matrix by 10000 to obtain M 0-1 values ​​and obtain the characteristic spectrum of the target object.

[0031] Preferably, the method further includes verifying the accuracy of identifying different types of ground objects, which includes:

[0032] Use aerial hyperspectral set annotation data to generate validation pixels in proportion and in layers;

[0033] The confusion matrix is ​​used to verify the recognition accuracy of different types of ground objects.

[0034] Preferably, the method further includes evaluating the accuracy of identifying different types of ground objects, which includes:

[0035] The overall accuracy OA, Kappa coefficient, mapping accuracy PA and user accuracy UA are used as accuracy indicators to evaluate the recognition accuracy of different types of ground objects.

[0036] The embodiment of the present invention provides a characteristic spectrum extraction method for target object classification and identification. Compared with the prior art, the method has the following beneficial effects:

[0037] Based on the law of geographic similarity and the idea of ​​spatial interpolation, this paper proposes a characteristic spectrum construction algorithm based on spectral domain interpolation. This method couples the idea of ​​spectral bands with spatial interpolation. While retaining the easy calculation and analysis characteristics of the mean characteristic spectrum, it effectively makes up for the shortcomings of the poor resistance of the mean characteristic spectrum, improves the resistance of the characteristic spectrum to abnormal extreme values, and strengthens the characterization ability of the characteristic spectrum. The experimental results show that compared with MCS, ICS has an overall accuracy of 4.24% in SAM that characterizes the characteristic spectrum morphology; in the importance evaluation of amplitude characteristic parameters and LDA that characterize the details of the characteristic spectrum, the importance score of the amplitude parameter is increased by an average of 0.35, the discrimination accuracy of each tree species is increased by 2.51%, and the overall accuracy is increased by 2.5%. The study shows that ICS is superior to traditional MCS in both the characterization of the overall morphology of spectral features and the reproduction of detailed features. It can be used to improve the characteristic spectrum extraction process of target objects in classification scenes and improve the separability between classes; optimize the construction of characteristic parameters in inversion scenes and improve the characterization ability of spectra. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 An ICS calculation flow chart of a characteristic spectrum extraction method for target object classification and identification provided by an embodiment of the present invention;

[0039] Figure 2 Mean (a) and interpolation (b) characteristic spectrum SAM results of a characteristic spectrum extraction method for target object classification and identification provided by an embodiment of the present invention;

[0040] Figure 3 A comparison diagram of characteristic spectra of misclassified tree species in a characteristic spectrum extraction method for target object classification and identification provided by an embodiment of the present invention;

[0041] Figure 4 A characteristic parameter FS of a characteristic spectrum extraction method for target object classification and identification provided by an embodiment of the present invention;

[0042] Figure 5 A discriminant analysis accuracy diagram of a characteristic spectrum extraction method for target object classification and identification provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] See also Figures 1 to 5, an embodiment of the present invention provides a characteristic spectrum extraction method for target object classification and identification, the method comprising: the first step is to calculate the spectral domain. Assuming that N ground object spectra with M bands are used to calculate the characteristic spectrum, the reflectivity is first multiplied by 10000 to be converted into an integer for ease of calculation, and then the N ground object spectra are constructed into an N×M matrix, and the value of the nth row and mth column of the matrix is ​​the reflectivity of the mth band of the nth spectrum. Calculate the maximum value max and the minimum value min of each column of the matrix to construct the spectral domain space of the target object, that is Figure 1 The blue line in the middle represents the spectral domain boundary.

[0045] The second step is to construct a matrix. Construct N 10000×M matrices Ai, i∈[1, 2, …, N], corresponding to N ground feature spectra. In each column of the matrix, only one value is 1, and the rest are 0. The row number of each column where the value 1 is located is the reflectivity of that band. For example, if the m-band reflectivity of the i-th ground feature spectrum is Rm, the corresponding matrix Ai has the m-th column Rm row value of 1, and the other values ​​of the column are 0.

[0046] The third step is spectral domain interpolation. Calculate the distance between the reflectance of spectrum i and the boundary of the spectral domain column by column, that is, |max-Rm|, |min-Rm|. Then, in the corresponding matrix Ai, use the Rm row as the center to interpolate row by row until the value is no greater than 0. The upward decreasing value is 1 / |max-Rm|, and the downward decreasing value is 1 / |min-Rm|. After completing the interpolation update of the matrix Ai, N 10000×M matrices are obtained, which are called the interpolation spectral domain space.

[0047] The fourth step is ICS extraction. N matrices, namely the interpolated spectral domain space, are accumulated to generate a 10000×M matrix, namely the cumulative spectral domain space. Then the maximum value of the cumulative spectral domain space is calculated column by column, and the row number of the maximum value is recorded. The resulting 1×M one-dimensional matrix is ​​the ICS.

[0048] In order to verify the effectiveness and superiority of the spectral domain interpolation feature extraction algorithm in the construction of characteristic spectral morphology and amplitude, the mean characteristic spectrum (MCS) and spectral domain interpolation characteristic spectrum (ICS) were obtained respectively using aerial hyperspectral remote sensing images and ASD measured tree species spectra as data sources. Based on the two characteristic spectra, the spectral angle mapping (SAM) of aerial hyperspectral data and the characteristic parameter extraction, importance evaluation and linear discriminant analysis (LDA) of ASD measured data were performed to explore the ability of ICS to characterize the overall shape and reproduce the detailed features, and verify its effectiveness and superiority.

[0049] 1 Overall morphological analysis of characteristic spectra based on SAM

[0050] SAM measures the similarity between spectra by calculating the angle between pixel spectral vectors, focusing on the similarity of spectral morphology. Therefore, the accuracy of SAM can measure the representation of the overall morphology of the characteristic spectrum for the target land type. Based on the MCS and ICS extracted by the process described in the previous section, SAM is used to classify and map the aerial hyperspectral set. The classification results are as follows: Figure 2 The validation pixels were generated in proportion to the labeled data of the aerial hyperspectral set. The accuracy of the classification results was verified using the confusion matrix. The overall accuracy (OA), Kappa coefficient, producer's accuracy (PA) and user's accuracy (UA) were used as accuracy indicators for evaluation. The results are shown in Table 1.

[0051] The following conclusions can be drawn from the chart. First, overall, compared with the traditional mean characteristic spectrum based on arithmetic mean, the characteristic spectrum extracted based on spectral domain interpolation has an overall accuracy of 4.24% higher when performing spectral angle mapping, and the Kappa coefficient is also 0.042 ahead, which shows that ICS is superior to traditional MCS in terms of accuracy and stability when combined with the spectral angle mapping algorithm for classification. Secondly, by comparing the recognition accuracy of each target land type, it can be seen that among the 32 single accuracy indicators (PA, UA) of 16 land types, 23 are better performed by ICS. At the same time, it can be seen that the indicators with better performance of MCS are mainly concentrated on land features such as water areas, bare land, houses, and rice fields with significant differences in spectral characteristics and lower recognition difficulty. For land features with similar spectral characteristics, ICS shows a more obvious recognition advantage, which also proves to a certain extent that ICS can better retain the detailed characteristics of the target land features and enhance the ability of characteristic spectra to identify similar land features. Finally, a comparative analysis of the misclassification phenomenon caused by the algorithm shows that the spectral angle mapping algorithm has achieved good accuracy in the recognition of water areas, bare land, houses, grasslands, and rice fields. Mistakes and omissions mainly occur in vegetation such as poplar, peach, and koelreuteria paniculata. Figure 3 As shown in the figure, although the inter-class distinguishability of ICS in spectral amplitude and detail features at 535-571nm and 715-751nm is significantly better than that of MCS, the misclassified vegetation types are very similar in overall morphology and cannot be effectively distinguished only by the spectral vector angle. This phenomenon exists in both MCS and ICS.

[0052] Table 1 Classification result accuracy verification table

[0053]

[0054] 2 Characteristic spectrum amplitude parameter analysis based on FS

[0055] In order to better explore the ability of different characteristic spectrum extraction algorithms to reproduce the detailed features and amplitude of the target objects, the measured spectra of ASD tree species with similar spectral morphology were used as the data source. Referring to previous studies, the parameters that can represent the detailed features and amplitude of the spectrum were extracted (see Table 2) to analyze the importance of FS features. First, the characteristic spectra of tree species were extracted based on the ICS and MCS algorithms, and the characteristic parameters were extracted according to the definition. Then, the characteristic parameters were normalized by category to eliminate the influence of different dimensions. Finally, the normalized characteristic parameters were used to calculate the FS score, and ICS and MCS were evaluated based on the characteristic parameter FS that can characterize the spectral amplitude.

[0056] Table 2 Characteristic parameter definition table

[0057]

[0058] FS results are as follows Figure 4 As shown in the figure, the digital label above the histogram is the difference between the FS of the ICS parameter and the FS of the MCS parameter. The following conclusions can be drawn from the figure. First, the order of importance of the ICS and MCS amplitude characteristic parameters is generally consistent, and the order from high to low is Sb, Sr, Sy, Db, Rr, Dy, Dr, Rg, indicating that the importance and characterization ability of the amplitude feature are mainly affected by the physical and chemical properties of the target itself. At the same time, the FS difference of the same characteristic parameters in different characteristic spectra is no more than 1.62. This is because the ASD measured spectral data are standardized repeated observations of the same sample points, and the differences between multiple spectra of the same sample point and the spectra of the same sample points are relatively small. Secondly, the importance of ICS amplitude parameters is generally better than that of the same type of MCS amplitude parameters, and the average FS increases by 0.35. Among the 8 selected amplitude parameters, the FS of ICS amplitude parameters is better than that of MCS in 5 parameters such as Sb, Sy, and Sr, and the 3 amplitude parameters with the strongest characterization ability and the highest score are all dominated by ICS. It shows that the spectral domain interpolation feature spectrum extraction algorithm can better retain the characteristic information of the target and has better amplitude detail reproduction ability than the traditional mean feature spectrum. Finally, by comparing the differences in FS of the same type of amplitude features, it can be seen that for tree species identification, the more effective feature parameters are the characteristic area information, such as Sb, Sy and Sr, and ICS has the most obvious optimization effect on this type of parameters, indicating that ICS has more potential and advantages in feature construction and classification scenarios.

[0059] 3 Characteristic spectrum detail parameter analysis based on LDA

[0060] As a single-feature evaluation method, FS only considers the relationship between a single parameter and the response factor. Although it helps to understand the data, it does not consider the interaction between multiple parameters. In order to further comprehensively evaluate the different characteristic spectral detail parameters, the amplitude parameters Sb, Sr, Sy, Db and Rr with higher scores in the previous section, plus the normalized red edge position, blue edge position, and yellow edge position (wavelengths corresponding to Dr, Db, and Dy) parameters were used for LDA classification, and 120 samples of 6 categories in the validation set were used for accuracy evaluation. At the same time, the original mean characteristic spectrum full band was used for LDA classification and accuracy verification to verify the tree species characterization ability and recognition effectiveness of the characteristic detail parameters through comparison.

[0061] The discrimination accuracy of each tree species is plotted as Figure 5 , the discrimination accuracy in the figure is the average of UA and PA of the tree species, the discrimination accuracy difference is the difference between ICS discrimination accuracy and MCS discrimination accuracy, and the values ​​in brackets in the legend are the average discrimination accuracy difference and OA of each LDA classification. The following conclusions can be drawn from the figure. First, compared with the full-band LAD, the characteristic parameter LAD has improved the recognition accuracy of each tree species, and the overall accuracy has increased by 13.93%. It shows that the characteristic spectrum detail parameters can well characterize the physical and chemical properties of different tree species, effectively reduce information redundancy, and improve the accuracy of tree species identification. It also proves that this paper uses detail parameters to characterize the characteristic spectrum for analysis and evaluation with certain scientific basis. Secondly, by comparing the ICS and MCS characteristic parameter LAD classifications, it can be seen that except for the ICS characteristic parameter discrimination accuracy being lower than MCS on fir, the discrimination accuracy of other tree species is significantly better than MCS. Compared with MCS, the average discrimination accuracy of LDA classification using ICS detail features is improved by 2.51%, and the overall accuracy is improved by 2.5%, indicating that the characteristic spectrum extracted by the spectral domain interpolation algorithm can better retain the target detail characteristics and has better characterization ability.

[0062] Disclosed above are only a few specific embodiments of the present invention. Those skilled in the art may make various changes and modifications to the embodiments of the present invention without departing from the spirit and scope of the present invention. However, the embodiments of the present invention are not limited thereto. Any changes that can be conceived by those skilled in the art should fall within the scope of protection of the present invention.

Claims

1. A characteristic spectrum extraction method for target object classification and identification, characterized in that: include: Obtain multiple object spectra of the target object and construct a spectrum matrix; Calculate the maximum and minimum values ​​of each column of the spectrum matrix to obtain the spectrum domain space of the target object; The method of obtaining the spectral domain space of the target object spectrum includes: obtaining reflectances corresponding to different wavelengths in multiple object spectra of the target object; multiplying the reflectance by 10000 to convert it into an integer reflectance; constructing an N×M matrix representing N spectra and M wavelengths; calculating the maximum value max and the minimum value min of each column of the N×M matrix as the boundaries of the spectral domain space; Construct multiple interpolation matrices corresponding to multiple ground feature spectra with integer reflectance as row number and wavelength as column number, wherein the element of 1 in each column of each interpolation matrix represents that the wavelength in the column has a corresponding reflectance; In each matrix to be interpolated, the reflectivity corresponding to the wavelength in each column and the distance in the spectral domain space are calculated to obtain the normalized interpolation factor; Each column of each matrix to be interpolated is interpolated by using an inverse distance interpolation method to obtain multiple spectral domain space matrices; Accumulate multiple spectral domain space matrices to obtain the cumulative spectral domain space matrix of the object, and extract the characteristic spectrum of the target object; The characteristic spectrum of the target object is combined with the spectral angle mapping algorithm to classify and map the collected aerial hyperspectral set to identify different types of ground objects.

2. A characteristic spectrum extraction method for target object classification and identification as claimed in claim 1, characterized in that: The method of constructing a plurality of matrices to be interpolated corresponding to a plurality of ground object spectra with integer reflectance as row number and wavelength as column number includes: Construct a 10000×M matrix Ai of N ground feature spectra, i∈[1, 2, …, N], where N represents the number of spectra. An element of 1 in each column of the 10000×M matrix represents that there is a corresponding reflectivity at the wavelength in the column, and 0 represents that there is no corresponding reflectivity at the wavelength in the column.

3. A characteristic spectrum extraction method for target object classification and identification as claimed in claim 2, characterized in that: In each matrix to be interpolated, the reflectivity corresponding to the wavelength in each column and the distance in the spectral domain space are calculated to obtain a normalized interpolation factor, including: Obtain the reflectivity Rm corresponding to the wavelength in each column of the matrix to be interpolated; Calculate the distances |max-Rm| and |min-Rm| between the maximum value max and the minimum value min corresponding to the wavelength in the spectral domain space and the reflectivity Rm; Use 1 / |max-Rm| and 1 / |min-Rm| as normalized interpolation factors.

4. A characteristic spectrum extraction method for target object classification and identification as claimed in claim 3, characterized in that: The obtaining of a plurality of spectral domain space matrices comprises: In each column of the matrix to be interpolated, the element with value 1 is used as the boundary, and the interpolation is performed row by row until the value is not greater than 0. The normalized interpolation factor for upward decrease is 1 / |max-Rm|, and the normalized interpolation factor for downward decrease is 1 / |min-Rm|, until the value is not greater than 0; Update N 10000×M matrices to obtain multiple spectral domain space matrices.

5. A characteristic spectrum extraction method for target object classification and identification as claimed in claim 4, characterized in that: Accumulate multiple spectral domain space matrices to obtain the cumulative spectral domain space matrix of the object and construct the characteristic spectrum of the target object, including: Add N 10000×M spectral domain space matrices to obtain a cumulative spectral domain space matrix; Retrieve the maximum value of each column of the cumulative spectrum domain space matrix and record its row number to obtain a 1×M matrix; Divide the 1×M matrix by 10000 to obtain M 0-1 values ​​and obtain the characteristic spectrum of the target object.

6. The characteristic spectrum extraction method for target object classification and identification according to claim 1, characterized in that: It also includes accuracy verification for identifying different types of ground objects, including: Use aerial hyperspectral set annotation data to generate validation pixels in proportion and in layers; The confusion matrix is ​​used to verify the recognition accuracy of different types of ground objects.

7. The characteristic spectrum extraction method for target object classification and identification according to claim 1, characterized in that: It also includes the accuracy evaluation of identifying different types of ground objects, including: The overall accuracy OA, Kappa coefficient, mapping accuracy PA and user accuracy UA are used as accuracy indicators to evaluate the recognition accuracy of different types of ground objects.

Images