CN103927538B - Threshold selection method for improving spectral angle mapping precision - Google Patents

Abstract

The invention belongs to the field of remote sensing information science and technology, and specifically discloses a threshold selection method for improving the accuracy of spectral angle mapping: obtaining pure endmember spectral curves of target images; classifying endmember spectral curves, and constructing spectral curve sets for each type of ground object; Calculate the spectral angle arc similarity value between different types of ground objects; obtain the maximum value of the spectral angle arc similarity value set between two ground objects; according to the maximum value of the spectral angle arc similarity value set, obtain the upper threshold value of the ground object extraction Limit value; Calculate the arc similarity value of the spectral angle of the same type of feature; According to the similar value of the spectral angle arc of the same type of feature, obtain the minimum value of the arc similar value set of each type of feature; according to the minimum value of the arc similar value set of each type of feature , to obtain the lower limit value of the extraction threshold of the ground feature; compare the upper limit value and the lower limit value, select the extraction threshold value, and ensure the accuracy and completeness of the ground feature information extraction. This method can ensure the accuracy and completeness of object information extraction from imagery.

Description

A Threshold Selection Method for Improving the Accuracy of Spectral Angle Mapping

technical field

The invention belongs to the field of remote sensing information science and technology. In particular, the invention relates to a threshold value selection method of a spectral angle mapping method.

Background technique

The spectral angle mapping method is one of the main methods to identify ground objects in imaging spectral image processing technology. This algorithm is an interactive classification method that directly matches the image spectrum with the reference spectrum, and is an automatic classification method that compares the image spectrum with the ground object spectrum or the ground object spectrum in the spectrum library. This method makes full use of the information of the spectral dimension, emphasizes the shape characteristics of the spectrum, and greatly reduces the characteristic information.

At present, the spectral angle mapping method has been successfully applied to ground object recognition in many fields, mainly because this method mainly focuses on the shape of the spectrum and reduces the influence of spectral feature gain and drift. In the spectral angle mapping technique, the selection of the threshold is very important. If the threshold is small, the information extraction results will be missing, and if the threshold is large, the accuracy of extraction will decrease. Therefore, a new method is needed to select the threshold to improve the accuracy and completeness of the spectral angle mapping results.

Contents of the invention

The purpose of the present invention is to provide a threshold selection method that improves the accuracy of spectral angle mapping for the defect of threshold value setting of spectral angle.

In order to achieve the above object, the technical solution of the present invention is as follows: a threshold selection method for improving spectral angle mapping accuracy, the method comprises the following steps:

Step 1, obtain the pure endmember spectral curve of the target image object;

Step 2. According to the shape and characteristic position of the spectral curve, classify the pure endmember spectral curve of the target image object obtained in the above step (1), determine the number of object types in the target image, and construct each type of target image feature. Spectral curve collection of objects;

Step 3, use the spectral analysis tool to calculate the similarity value rad of the spectral angle radian among different types of target image objects A, B, C, among them, rad(A,B), rad(A,C), rad(B , C) Respectively represent the sets of spectral angle radian similarity values between the three types of target images, A and B, A and C, B and C, and two types of ground objects;

Step 4, obtain the spectral angle radian similarity value sets rad(A,B), rad(A,C), The maximum value of rad(B,C) δ _AB , δ _AC , δ _BC ;

Step 5, according to the maximum value δ _AB , δ _AC , δ _BC of the spectral angle radian similar value set obtained in the above step 4, obtain the extraction thresholds A _threshold , B _threshold , C _threshold of the target image features A, B, C Upper limit A _thresholdmax , B _thresholdmax , C _thresholdmax ;

Step 6, use the spectral analysis tool to calculate the radian similarity values rad(A,A), rad(B,B), rad(C,C) of the spectral angle of the object image of the same category of objects;

Step 7: According to the radian similarity values rad(A,A), rad(B,B) and rad(C,C) of the spectral angles of the target images of the same category obtained in the above step 6, obtain each type of target image features The minimum value β _AA , β _BB , β _CC of the set of spectral angle radian similarity values;

Step 8, according to the minimum value β _AA , β _BB , β _CC of the radian similarity value set of each type of target image features obtained in the above step 7, obtain the extraction thresholds A _threshold , B _threshold of the target image features A, B, C , the lower limit of C _threshold A _thresholdmin , B _thresholdmin , C _thresholdmin ;

Step 9: Compare the extraction threshold upper limit values A _thresholdmax , B _thresholdmax , C _thresholdmax obtained in the above step 5 with the lower limit values A _thresholdmin , B _thresholdmin , and C _thresholdmin obtained in the above step 8, so as to select the extraction threshold and ensure Accuracy and completeness of target image feature information extraction.

In the step 1, the hourglass method or the continuous maximum angle convex cone method is used to obtain the pure endmember spectral curve of the target image object.

The specific steps of the above step 2: there are three types of image target features, namely A, B, and C, and the spectral curve sets of each type of target image features A, B, and C are expressed as SA={a ₁ , a ₂ , a ₃ ...a _i }, SB={b ₁ , b ₂ , b ₃ ...b _j }, SC={c ₁ , c ₂ , c ₃ ...c _k }.

The specific formulas of rad(A, B), rad(A, C), and rad(B, C) in the step 3 are as follows:

The specific formulas of δ _AB , δ _AC , and δ _BC in the step 4 are as follows:

δ _AB ＝max(rad(A,B))＝max{rad(a ₁ ,b ₁ ),rad(a ₁ ,b ₂ ),rad(a ₁ ,b ₃ )...rad(a _i ,b _j )}

δ _AC ＝max(rad(A,C))＝max{rad(a ₁ ,c ₁ ),rad(a ₁ ,c ₂ ),rad(a ₁ ,c ₃ )...rad(a _i ,c _k )}

δ _BC ＝max(rad(B,C))＝max{rad(b ₁ ,c ₁ ),rad(b ₁ ,c ₂ ),rad(b ₁ ,c ₃ )...rad(b _j ,c _k )}.

The concrete formulas of A _thresholdmax , B _thresholdmax , C _thresholdmax in the described step 5 are as follows respectively:

A _thresholdmax ＝1-Max(δ _AB ,δ _AC )

B _thresholdmax ＝1-Max(δ _AB ,δ _BC )

C _thresholdmax = 1-Max(δ _AC , δ _BC ).

The concrete formulas of rad(A, A), rad(B, B), rad(C, C) in the described step 6 are as follows respectively:

rad(A,A)={rad(a _n ,a _m ), where n=1,2,…i-1, m=n+1,n+2,…i and n≠m}

rad(B,B)={rad(b _n ,b _m ), where n=1,2,...j-1, m=n+1,n+2,...j and n≠m}

rad(C,C)={rad(c _n ,c _m ), where n=1,2,...k-1, m=n+1,n+2,...k and n≠m}

The specific formulas of β _AA , β _BB , and β _CC in the step 7 are as follows:

β _AA =min(rad(A,A))=min{rad(a _n , _am ), where n=1,2,...i-1, m=n+1,n+2,...i and n≠m}

β _BB =min(rad(B,B))=min{rad(b _n ,b _m ), where n=1,2,...j-1, m=n+1,n+2,...j and n≠m}

β _CC =min(rad(C,C))=min{rad(c _n ,c _m ), where n=1,2,...k-1, m=n+1,n+2,...k and n≠m}.

The specific formulas of A _thresholdmin , B _thresholdmin , C _thresholdmin in the described step (8) are as follows respectively:

A _thresholdmin = 1-β _AA

B _thresholdmin = 1-β _BB

C _thresholdmin = 1 - β _CC .

The step 9 specifically includes the following two situations:

(9.1) When the upper limit value of the extraction threshold of a certain type of target image feature is less than the lower limit value, the accuracy and completeness of the extraction of this type of target image feature information cannot be guaranteed at the same time;

(9.2) When the upper limit value of a certain type of target image feature is greater than or equal to the lower limit value, the threshold value of the spectral angle mapping of this type of target image feature is equal to these two values or takes the interval between the upper limit value and the lower limit value Any value in , can guarantee the completeness and accuracy of this kind of target image feature information extraction at the same time.

The beneficial effects of the present invention are as follows: the method of the present invention can ensure that the extracted target features are not In the case of other ground objects mixed in, the accuracy and completeness of the spectral angle mapping results of the target image and ground objects can be improved, and the phenomenon of omission and wrong extraction due to inappropriate selection of the threshold can be compensated, thereby improving the information of the target image and ground features. Extraction accuracy.

detailed description

The present invention will be further described below in conjunction with the accompanying drawings and examples of implementation.

A threshold selection method suitable for improving the accuracy of spectral angle mapping provided by the present invention comprises the following steps:

Step 1. Obtain the pure endmember spectral curve of the target image

The purpose of obtaining the pure endmember spectral curve of the target image is to minimize the amount of computation in the later stage without reducing the classification accuracy. Usually the number of pixels in an image is thousands or more, and the number of pure endmember spectra is generally less than 100.

The hourglass method or the continuous maximum angle convex cone method can be used to obtain the pure endmember spectral curve of the target object in the actual image.

Step 2. According to the shape and characteristic position of the spectral curve, classify the endmember spectral curves of the target image features obtained in the above step (1), determine the number of target image types, and construct each type of target image feature collection of spectral curves.

Assuming that the number of pure spectrum endmembers obtained in step (1) is 50, and there are three types of actual image target objects, A, B, and C, the spectral curve sets of each type of image target objects A, B, and C represent respectively SA={a ₁ , a ₂ , a ₃ ...a _i }, SB={b ₁ , b ₂ , b ₃ ...b _j }, SC={c ₁ , c ₂ , c ₃ ...c _k }. Among them, a, b, and c are the two-dimensional spectral scatter curves of the target image objects A, B, and C respectively, a ₁ represents the first curve of the object A, and so on, and i, j, and k are the target image ground objects respectively. The number of wave curves of objects A, B, and C. 0<i<50, 0<j<50, 0<k<50, and the sum of the three is equal to 50.

Step 3, use spectral analysis tools to calculate the similarity values of spectral angles between different types of target image features A, B, and C (Radians Similarity is abbreviated as rad)

The value range of rad is (0,1), with two decimal places reserved. The larger the value is, the higher the similarity between the two spectral curves is, and the radian similarity value of two completely identical spectral curves is 1.

rad(A,B), rad(A,C), and rad(B,C) respectively represent the sets of spectral angle radian similarity values between the three types of target images of A and B, A and C, B and C, and two types of ground objects. The specific formulas of rad(A,B), rad(A,C), and rad(B,C) are as follows:

rad(a ₁ ,b ₁ ) represents the similarity value of the spectral angle radian between the first spectral curves in the two types of surface features A and B, and rad(a ₁ ,b ₁ )=rad(b ₁ ,a ₁ ). The meanings of i, j, and k in the above formula are the same as those in step (2).

The smaller the similarity value of the spectral angle radian of the two types of target object spectral curves, the greater the separability of the two types of target image objects, that is, the greater the difference in the shape of the spectral curves of the two types of object objects; the larger the similarity value of the spectral angle radian , which proves that the smaller the separability of the two types of target images, that is, the smaller the difference in the shape of the spectral curves of the two types of objects.

Using the ENVI spectral analysis tool to directly calculate the similarity value of the spectral angle radian of the two spectral curves.

Step 4, obtain the spectral angle radian similarity value sets rad(A,B), rad(A,C ), the maximum value of rad(B,C) δ _AB , δ _AC , δ _BC , the specific formulas are as follows:

δ _AB ＝max(rad(A,B))＝max{rad(a ₁ ,b ₁ ),rad(a ₁ ,b ₂ ),rad(a ₁ ,b ₃ )...rad(a _i ,b _j )}

δ _AC ＝max(rad(A,C))＝max{rad(a ₁ ,c ₁ ),rad(a ₁ ,c ₂ ),rad(a ₁ ,c ₃ )...rad(a _i ,c _k )}

δ _BC ＝max(rad(B,C))＝max{rad(b ₁ ,c ₁ ),rad(b ₁ ,c ₂ ),rad(b ₁ ,c ₃ )...rad(b _j ,c _k )}

Among them, δ _AB , δ _AC , and δ _BC represent the spectral arc similarity between each two types of target image features, respectively. For the extraction threshold of type A ground object, when δ _AB > δ _AC , it means that the separation of type A ground object and C type ground object is greater, and the threshold for extracting type A ground object should not be greater than 1-δ _AB , otherwise the extraction The results will be mixed with some type B features, and the extraction accuracy will be reduced; similarly, when δ _AB < δ _AC , it means that the separability between type A and type B features is greater, and the threshold for extracting type A features should not be If it is greater than 1-δ _AC , otherwise the extraction result will be mixed with some C-type objects, and the extraction accuracy will decrease. The accuracy of other feature extraction can be deduced by analogy, and the meanings of i, j, and k in the above formula are the same as those in step (2).

Step 5, according to the maximum value δ _AB , δ _AC , δ _BC of the spectral angle radian similarity value set obtained in the above step (4), obtain the extraction thresholds A _threshold , B _threshold , C of the target image features A, B, C The upper limit of the _threshold A _thresholdmax , B _thresholdmax , C _thresholdmax , the specific formulas of A _thresholdmax , B _thresholdmax , and C _thresholdmax are as follows:

A _thresholdmax ＝1-Max(δ _AB ,δ _AC )

B _thresholdmax ＝1-Max(δ _AB ,δ _BC )

C _thresholdmax ＝1-Max(δ _AC ,δ _BC )

Assuming that δ _AB , δ _AC , and δ _BC in the above formula are 0.83, 0.91, and 0.88 respectively, A _thresholdmax =0.09, B _thresholdmax =0.12, and C _thresholdmax =0.09 are obtained through calculation. If the upper limit value of the extraction threshold is not calculated, and the default threshold (Maximum radians) of 0.1 in the ENVI software is used directly to extract the target features, the pixels whose arc similarity with the spectral curve of the target features is greater than or equal to 0.9 will be taken as The target features are extracted.

When extracting type A ground objects, the pixels of type C with a similarity of 0.91 to A are extracted as type A ground objects. When extracting type B ground objects, there is no case where other types of ground objects are extracted as type B ground objects; when extracting type C ground objects, pixels of type A with a similarity of 0.91 to C are taken Class C features are extracted.

Step 6. Use the spectral analysis tool to calculate the arc similarity values of the spectral angles of the object images of the same category rad(A,A), rad(B,B), rad(C,C), rad(A,A), rad( B, B), rad(C, C) represent the similar value sets of spectral angle radians of the spectral curves among the three types of target imagery objects of A, B, and C, respectively, and the specific formulas are as follows:

rad(A,A)={rad(a _n ,a _m ), where n=1,2,…i-1, m=n+1,n+2,…i and n≠m}

rad(B,B)={rad(b _n ,b _m ), where n=1,2,...j-1, m=n+1,n+2,...j and n≠m}

rad(C,C)={rad(c _n ,c _m ), where n=1,2,...k-1, m=n+1,n+2,...k and n≠m}

The meanings of i, j, and k in the above formula are the same as those in step (2), and m, n are integer variables with different value ranges for different types of ground objects.

The radian similarity values of spectral angles rad(A,A), rad(B,B), and rad(C,C) indicate the degree of variation of the spectral curves of ground objects between classes, and the greater the radian similarity value of spectral angles, it indicates The smaller the variation of the spectral curve of the object is, the greater it is vice versa.

Step 7: According to the radian similarity values rad(A,A), rad(B,B), and rad(C,C) of the same category of target images obtained in the above step (6), obtain each type of target image The minimum values β _AA , β _BB , and β _CC of the radian similarity value set of ground features are as follows:

β _AA =min(rad(A,A))=min{rad(a _n , _am ), where n=1,2,...i-1, m=n+1,n+2,...i and n≠m}

β _BB =min(rad(B,B))=min{rad(b _n ,b _m ), where n=1,2,...j-1, m=n+1,n+2,...j and n≠m}

β _CC =min(rad(C,C))=min{rad(c _n ,c _m ), where n=1,2,...k-1, m=n+1,n+2,...k and n≠m}

The meanings of i, j, k, m and n in the formula are the same as those in step (6).

The above three values β _AA , β _BB , and β _CC respectively represent the approximation of the spectral radian in each type of target image object. The smaller the value, the greater the spectral variation of this type of object. The threshold value setting for extracting each type of object is The fixed value should be greater than or equal to 1-β to ensure that there is no missing phenomenon in the extraction results of ground objects.

Step 8: _Obtain the extraction _{thresholds A threshold ,} The lower limit values of B _threshold and C _threshold are A _thresholdmin , B _thresholdmin and C _thresholdmin , and the specific formulas of A _thresholdmin , B _thresholdmin and C _thresholdmin are as follows:

A _thresholdmin = 1-β _AA

B _thresholdmin = 1-β _BB

C _thresholdmin = 1-β _CC

Assuming that β _AA , β _BB , and β _CC in the above formula are 0.93, 0.88, and 0.89 respectively, A _thresholdmin =0.07, B _thresholdmin =0.12, and C _thresholdmin =0.11 are obtained through calculation. If the lower limit value of the extraction threshold is not calculated, directly set the thresholds satisfying the conditions arbitrarily according to the upper limit values A _thresholdmax = 0.09, B _thresholdmax = 0.12, C _thresholdmax = 0.09 in step (5), that is, A _threshold = 0.05, B _threshold = 0.05, C _threshold =0.05, these three thresholds can guarantee the accuracy of each type of extraction, but cannot guarantee the integrity of the extraction results.

When extracting type A ground objects, the target pixels with a similarity greater than or equal to 0.95 to the type A standard curve are extracted, while the target pixels with a similarity between 0.93 and 0.95 are not extracted; when extracting type B ground objects, The target pixels whose similarity with the Class B standard curve is greater than or equal to 0.95 are extracted, while the target pixels with a similarity between 0.88 and 0.95 are not extracted; when extracting Class C ground objects, the similarity with the Class B standard curve The target pixels greater than or equal to 0.95 are extracted, but the target pixels with a similarity between 0.89 and 0.95 are not extracted; therefore, simply relying on the upper threshold value can only improve the accuracy of spectral angle classification, but cannot guarantee the spectral angle. completeness of classification.

Step 9: Compare the extraction threshold upper limit values A _thresholdmax , B _thresholdmax , C _thresholdmax obtained in the above step 5 with the lower limit values A _thresholdmin , B _thresholdmin , and C _thresholdmin obtained in the above step 8, so as to select the extraction threshold and ensure Accuracy and completeness of target image feature information extraction.

(9.1) When the upper limit value of the extraction threshold of a certain type of target image feature is less than the lower limit value, the accuracy and completeness of the extraction of this type of target image feature information cannot be guaranteed at the same time.

For example, the above-mentioned situation exists in the extraction threshold of C-type ground features in step (5) and step (8), that is, C _thresholdmin =0.11, C _thresholdmax =0.09, C _thresholdmin >C _thresholdmax , and only C-type ground features can be satisfied during the extraction process. Integrity of object information extraction, that is, C _threshold =0.11; or meet the accuracy of class C object information extraction, that is, C _threshold =0.09.

(9.2) When the upper limit value of a certain type of target image feature is greater than or equal to the lower limit value, the completeness and accuracy of the information extraction of this type of target image feature information can be guaranteed at the same time;

When the upper limit value of a certain type of target image feature is equal to the lower limit value, the threshold value of the spectral angle mapping of this type of target image feature is equal to these two values. For example, the above-mentioned situation exists in the extraction threshold of B-type feature in step (5) and step (8), that is, B _thresholdmin =0.12, B _thresholdmax =0.12, B _thresholdmin =B _thresholdmax , and the threshold value B _threshold is usually set in the extraction process = 0.12, it can ensure that there are no omissions and false mentions in the extraction results of Class B features, that is, it can ensure the integrity of the information extraction of Class B features; at the same time, it can ensure the accuracy of the information extraction of Class B features.

When the upper limit value of a certain type of target image features is greater than the lower limit value, the threshold for spectral angle mapping is any value between the upper limit value and the lower limit value. For example, the above-mentioned situation exists in the extraction threshold of A-type feature in step (5) and step (8), that is, A _thresholdmin = 0.07, A _thresholdmax = 0.09, A _thresholdmin < A _thresholdmax , if the threshold A _threshold is set during the extraction process =0.08, it can ensure that there is no omission and misrepresentation in the extraction results, that is, it can ensure the integrity of the information extraction of Class A ground features; at the same time, it can ensure the completeness and accuracy of the extraction of Class A ground features.

The present invention has been described in detail above in conjunction with the implementation examples, but the present invention is not limited to the above-mentioned embodiments, and various changes can also be made without departing from the gist of the present invention within the scope of knowledge possessed by those of ordinary skill in the art . The content that is not described in detail in the present invention can adopt the prior art.

Claims (8)

1. A method for selecting a threshold to improve spectral angle mapping accuracy, characterized in that: the method may further comprise the steps: Step 1, obtain the pure endmember spectral curve of the target image object; Step 2. According to the shape and characteristic position of the spectral curve, classify the pure endmember spectral curve of the target image object obtained in the above step 1, determine the number of object types in the target image, and construct the collection of spectral curves; Step 3, use the spectral analysis tool to calculate the similarity value rad of the spectral angle radian among different types of target image objects A, B, C, among them, rad(A,B), rad(A,C), rad(B , C) Respectively represent the sets of spectral angle radian similarity values between the three types of target images, A and B, A and C, B and C, and two types of ground objects; Step 4, obtain the spectral angle radian similarity value sets rad(A,B), rad(A,C), The maximum value of rad(B,C) δ _AB , δ _AC , δ _BC ; Step 5, according to the maximum value δ _AB , δ _AC , δ _BC of the spectral angle radian similar value set obtained in the above step 4, obtain the extraction thresholds A _threshold , B _threshold , C _threshold of the target image features A, B, C Upper limit A _thresholdmax , B _thresholdmax , C _thresholdmax ; The concrete formulas of A _thresholdmax , B _thresholdmax , C _thresholdmax in the described step 5 are as follows respectively: A _thresholdmax ＝1-Max(δ _AB ,δ _AC ) B _thresholdmax ＝1-Max(δ _AB ,δ _BC ) C _thresholdmax = 1-Max(δ _AC ,δ _BC ); Step 6, use the spectral analysis tool to calculate the radian similarity values rad(A,A), rad(B,B), rad(C,C) of the spectral angle of the object image of the same category of objects; Step 7: According to the radian similarity values rad(A,A), rad(B,B) and rad(C,C) of the target images of the same category obtained in the above step 6, obtain the spectrum of each type of target image and feature The minimum value β _AA , β _BB , β _CC of the similar value set of angle radian; Step 8, according to the minimum values β _AA , β _BB , and β _CC of the similar value sets of spectral angles and radians of each type of target image features obtained in the above step 7, obtain the extraction thresholds A _threshold , B _threshold , the lower limit value of C _threshold A _thresholdmin , B _thresholdmin , C _thresholdmin ; The concrete formulas of A _thresholdmin , B _thresholdmin , C _thresholdmin in the described step 8 are as follows respectively: A _thresholdmin = 1-β _AA B _thresholdmin = 1-β _BB C _thresholdmin = 1-β _CC ; Step 9: Compare the extraction threshold upper limit values A _thresholdmax , B _thresholdmax , C _thresholdmax obtained in the above step 5 with the lower limit values A _thresholdmin , B _thresholdmin , and C _thresholdmin obtained in the above step 8, so as to select the extraction threshold and ensure Accuracy and completeness of target image feature information extraction. 2. A threshold selection method for improving spectral angle mapping accuracy according to claim 1, characterized in that: in the step 1, the hourglass method or the continuous maximum angle convex cone method are used to obtain the pure end of the target image feature Metaspectral curve. 3. a kind of threshold selection method that improves spectral angle mapping precision according to claim 2, it is characterized in that: the concrete step of described step 2: target image feature type has three kinds and is respectively A, B, C, Then the set of spectral curves of each type of target image object A, B, C is expressed as SA={a ₁ , a ₂ , a ₃ ...a _i }, SB={b ₁ , b ₂ , b ₃ ...b _j } ,SC={c ₁ , c ₂ , c ₃ ... c _k }; Among them, i, j, k are the number of spectral curves of target image features A, B, C, respectively. 4. a kind of threshold selection method that improves spectral angle mapping accuracy according to claim 3, is characterized in that: rad (A, B), rad (A, C), rad (B in described step 3 , and the specific formulas of C) are as follows: Among them, i, j, k are the number of spectral curves of target image features A, B, C, respectively. 5. a kind of threshold selection method that improves spectral angle mapping precision according to claim 4, is characterized in that: the concrete formula of δ _AB , δ _AC , δ _BC in the described step 4 is respectively as follows: δ _AB ＝max(rad(A,B))＝max{rad(a ₁ ,b ₁ ),rad(a ₁ ,b ₂ ),rad(a ₁ ,b ₃ )...rad(a _i ,b _j )} δ _AC ＝max(rad(A,C))＝max{rad(a ₁ ,c ₁ ),rad(a ₁ ,c ₂ ),rad(a ₁ ,c ₃ )...rad(a _i ,c _k )} δ _BC ＝max(rad(B,C))＝max{rad(b ₁ ,c ₁ ),rad(b ₁ ,c ₂ ),rad(b ₁ ,c ₃ )...rad(b _j ,c _k )}; Among them, i, j, k are the number of spectral curves of target image features A, B, C, respectively. 6. a kind of threshold selection method that improves spectral angle mapping accuracy according to claim 5, is characterized in that: rad (A, A), rad (B, B), rad (C in described step 6 , and the specific formulas of C) are as follows: rad(A,A)={rad(a _n ,a _m ), where n=1,2,…i-1, m=n+1,n+2,…i and n≠m} rad(B,B)={rad(b _n ,b _m ), where n=1,2,...j-1, m=n+1,n+2,...j and n≠m} rad(C,C)={rad(c _n ,c _m ), where n=1,2,...k-1, m=n+1,n+2,...k and n≠m} Among them, i, j, k are the number of spectral curves of target image features A, B, C, respectively. 7. a kind of threshold selection method that improves spectral angle mapping accuracy according to claim 6, is characterized in that: the concrete formula of β _AA , β _BB , β _CC in described step 7 is respectively as follows: β _AA =min(rad(A,A))=min{rad(a _n , _am ), where n=1,2,...i-1, m=n+1,n+2,...i and n≠m} β _BB =min(rad(B,B))=min{rad(b _n ,b _m ), where n=1,2,...j-1, m=n+1,n+2,...j and n≠m} β _CC =min(rad(C,C))=min{rad(c _n ,c _m ), where n=1,2,...k-1, m=n+1,n+2,...k and n≠m}; Among them, i, j, k are the number of spectral curves of target image features A, B, C, respectively. 8. a kind of threshold selection method that improves spectral angle mapping accuracy according to claim 7, is characterized in that: specifically comprise following two situations in described step 9: (9.1) When the upper limit value of the extraction threshold of a certain type of target image feature is less than the lower limit value, the accuracy and completeness of the extraction of this type of target image feature information cannot be guaranteed at the same time; (9.2) When the upper limit value of a certain type of target image feature is greater than or equal to the lower limit value, the threshold value of the spectral angle mapping of this type of target image feature is equal to these two values or takes the interval between the upper limit value and the lower limit value Any value in , can guarantee the completeness and accuracy of this kind of target image feature information extraction at the same time.