A method for extracting spectral index from hyperspectral data based on information gain

Abstract

The present invention relates to a method for extracting spectral index of hyperspectral data based on information gain, comprising the steps of: (1) acquiring and preprocessing sample data; (2) constructing multiple similar spectral index forms of different band combinations with the same formula by using band replacement, and calculating spectral indexes of different classes under each spectral index form; (3) calculating the information gain of the spectral index in its class: firstly dividing nodes according to the continuity of the spectral index values ​​in each class, calculating the information entropy of each divided node, calculating the information gain according to the information entropy, selecting the division point with the largest information gain as the optimal division point, and the information gain of the division point as the information gain of the spectral index form a; (4) calculating the information gain of each spectral index, and taking the spectral index corresponding to the maximum value as the optimal classification spectral index. The method can effectively expand the form and combination method of the spectral index, and can quickly screen out the optimal classification index from a large number of spectral indexes and band combinations.

Description

Technical Field

[0001] The present invention relates to a spectral index extraction method, in particular to a spectral index extraction method for hyperspectral data based on information gain, and belongs to the technical field of remote sensing feature extraction. Background Art

[0002] Spectral index, as a key parameter in remote sensing classification, significantly impacts the accuracy of remote sensing classification results. Traditional spectral indices are derived by studying the relationship between the spectral curves of a large amount of measured data and the research target. Furthermore, some spectral indices are further modified or band-replaced using existing spectral indices. These methods require extensive data research, resulting in inefficient spectral index screening and difficulty in quickly obtaining the appropriate spectral index. Furthermore, with the application of hyperspectral data, the number of bands has increased, and the number of band combinations for spectral indices has increased significantly. This makes the selection of spectral indices suitable for hyperspectral data a challenge in rapidly processing hyperspectral data. Summary of the Invention

[0003] The purpose of the present invention is to overcome the above-mentioned shortcomings and provide a spectral index extraction method for hyperspectral data based on information gain, which can effectively expand the form and combination method of spectral index and can quickly screen out the optimal classification index from a large number of spectral indices and band combinations.

[0004] The technical solution adopted by the present invention is:

[0005] A method for extracting spectral index from hyperspectral data based on information gain includes the following steps:

[0006] (1) Preparation and processing of sample data

[0007] Obtain hyperspectral measured remote sensing reflectance data or hyperspectral remote sensing image remote sensing reflectance data, group the obtained hyperspectral data according to the category of the measured target, each category of hyperspectral data is a group, and each group of hyperspectral data corresponds to a category label L1...L n There are n groups in total. Each group of hyperspectral data corresponds to N bands. The hyperspectral data categories are renamed, and the category labels to be distinguished are set to 1, and the rest are set to 2.

[0008] (2) Using band replacement, we construct multiple similar spectral index forms with different band combinations using the same formula, and calculate the spectral index of different classes (or groups) under each spectral index form:

[0009] The spectral index is calculated in different forms according to different calculation methods. The main calculation methods include single band value (Formula 1), band difference (Formula 2), band sum (Formula 3), single band ratio (Formula 4), ratio of single band to band difference (Formula 5), ​​and ratio of band difference to sum (Formula 6). The calculation methods are as follows:

[0010] SI=b i Formula (1)

[0011] SI=b i -b m Formula (2)

[0012] SI=b i +b m Formula (3)

[0013]

[0014] Where: SI is the spectral index, b is the band, i, m, n are the i-th, m-th, and n-th bands of the hyperspectral data (i, m, n are natural numbers, i, m, n ≤ N);

[0015] According to the band setting of hyperspectral data, the bands in the hyperspectral data are used for replacement to construct multiple similar spectral index forms with the same formula but different band combinations of existing spectral indices. The spectral index of each spectral index form in different categories (or groups) is calculated using the acquired hyperspectral data and formula.

[0016] (3) Calculate the information gain of the spectral index in its class

[0017] The basic idea behind information gain is to discretize the continuous spectral index values. Using a dichotomy method, a split point is used to divide the dataset into two categories. The information entropy of each category is then calculated based on the weights. Finally, the information gain at that split point is calculated using the overall information entropy. The calculated spectral index values ​​can be considered continuous attributes, so they need to be discretized using a dichotomy method. The basic idea behind this method is that given a sample set D and a continuous attribute α, the dichotomy method attempts to find a split point t that divides the sample set D into categories ≤ t and > t on attribute α.

[0018] In order to further quantify the classification effect of each spectral index, the following information gain calculation is performed:

[0019] Arrange all the values ​​of different classes corresponding to the spectral index form a to be calculated in ascending order to obtain the sample set D, the continuous attribute α, and the mean of all adjacent attributes as candidate partition points. There are n-1, where n is the number of all values ​​of α.

[0020] According to the renaming in step (1) and the proportion of the categories to be distinguished in all categories, the proportion of the k-th category samples in the sample set D is p k (k=1, 2), then the information entropy of D is defined as:

[0021]

[0022] The smaller the value of Ent(D), the higher the purity of D and the greater the possibility of belonging to the same class;

[0023]

[0024] After dividing the set D with each partition point, the first branch node contains all the nodes in D whose values ​​on the spectral index form a are less than or equal to T a The sample is recorded as D V (v=1), the value is greater than T a The sample is recorded as D V (v=2); calculate D according to formula 7 V The information entropy of , and considering that different branch nodes contain different numbers of samples, the branch nodes are given weights |D V | / |D|, that is, the more samples a branch node has, the greater its influence. The difference between the information entropy of the spectral index form a when the sample set D is unclassified and the information entropy of each partition point after partitioning is calculated, which is the obtained "information gain". The partition point with the maximum information gain is selected as the optimal partition point, and the information gain of this partition point is used as the information gain of the spectral index form a.

[0025]

[0026] Where a is the name of the spectral index form, a i is a value of the spectral index form a in D, and t is a specific division threshold, which is calculated as T a , v is the classification category (v = 1, 2);

[0027] (4) Determine the optimal classification spectral index

[0028] The greater the information gain, the greater the "purity improvement" obtained by using spectral index a to divide the data set. The information gain of all spectral indices is calculated. When the information gain takes the maximum value, the spectral index corresponding to the maximum information gain is the optimal classification spectral index. The method is as shown in formula (10):

[0029] SI y =max(Gain(D, SI)) Formula (10)

[0030] Where: SI yThe spectral index represented is a specific optimal spectral index classification form, and Gain(D, SI) is the information gain of all spectral indices.

[0031] Based on hyperspectral data, this method utilizes band integration, replacing each band with hyperspectral data one by one to construct multiple similar spectral indices of the same form as the existing spectral index. The corresponding spectral index values ​​are then calculated, and the optimal classification spectral index is selected by calculating the information gain of the spectral index. This method can quickly select the optimal classification index from a large number of spectral index and band combinations and can be widely applied to hyperspectral data classification problems in different scenarios and categories. It has high portability and improves the efficiency of hyperspectral data classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0033] In order to better illustrate the method for extracting spectral index from hyperspectral data based on information gain according to the present invention, the technical solutions in the embodiments of the present invention are described in detail, clearly and completely.

[0034] Example 1

[0035] Hyperspectral data spectral index extraction method based on information gain:

[0036] 1. Data Preparation and Processing

[0037] Some data are selected to illustrate the method. Three categories (1, 2, 3) of hyperspectral data (S1, S2, S3) are obtained. Each category has one group of hyperspectral data, for a total of three groups. Each group of hyperspectral data has four bands (bands 1, 2, 3, 4), as shown in Table 1.

[0038] Table 1: Raw data

[0039] Data Name Band 1 (B1) Band 2 (B2) Band 3 (B3) Band 4 (B4) category S1 32 100 75 8 1 S2 27 97 73 202 2 S3 56 241 13 844 3

[0040] Select the spectral index to distinguish category 2 (e.g., S2 data) from the remaining two categories (category 1 and category 3). Set the label of the category to be distinguished to 1 and the remaining categories to 2, and rename the hyperspectral data categories as shown in Table 2. At this point, the hyperspectral data contains only the category to be distinguished, with category label 1, and the remaining categories, with category label 2.

[0041] Table 2: Category renaming data

[0042] Data Name Band 1 Band 2 Band 3 Band 4 New Category S1 32 100 75 8 2 S2 27 97 73 202 1 S3 56 241 13 844 2

[0043] 2. Calculate the spectral index of different groups (classes) of each spectral index form:

[0044] Based on the band settings of the hyperspectral data, the bands in the hyperspectral data are used for substitution, and the spectral index forms (partial examples) in Table 3 are obtained through formulas 1, 2, 3, 4, 5, and 6. Then, the corresponding bands of the three sets of data (S1, S2, and S3) are substituted to obtain the spectral index values ​​shown in Table 4.

[0045] Table 3: Spectral index form

[0046]

[0047]

[0048] Table 4

[0049]

[0050] 3. Calculate the information gain of the spectral index in its class

[0051] The proportion of samples with label 1 in the current sample set is The proportion of samples with label 2 is Then the information entropy of D is defined as:

[0052] Ent(D)=-p0log2p0-(1-p0)log2(1-p0)=0.917

[0053] In order to further quantify the classification effect of each spectral index, we further optimize it based on the information gain and select the information gain as the evaluation index. The calculation process of information gain is as follows:

[0054] Taking the spectral index B2 as an example, the information gain is calculated and compared:

[0055] First, all values ​​of the spectral index B2 are arranged in ascending order, and the mean of all adjacent attributes is used as a candidate partition point. There are two candidate partition points, namely T1(97+100) / 2=98.5 and T2(100+241) / 2=170.5.

[0056] After the first partition point (T1) divides the set D (composed of S1, S2, and S3) into two parts, the first branch node contains all the nodes in D whose B2 values ​​are less than T I (98.5) is recorded as D1. The information entropy of D1 is calculated according to the formula, and considering that the number of samples contained in different branch nodes is different, the branch nodes are given a weight of |D1| / |D|=1 / 3; the second branch node contains all the samples in D whose B2 value is greater than T IThe sample with a value of (98.5) is denoted as D2. The information entropy of D2 is calculated according to the formula. Taking into account the different numbers of samples contained in different branch nodes, the branch nodes are assigned a weight of |D2| / |D|=2 / 3. The information gain of the T1 partition point is calculated using the formula.

[0057]

[0058] Similarly, at the T2 division point, the information entropy less than T2 (170.5) is

[0059]

[0060] Comparing Gain(D, B2, T1) with Gain(D, B2, T2), the information gain of the dividing point (T1) with the largest information gain is selected as the information gain of the spectral index B2, and the information gain value is 0.917.

[0061] 4. Determine the optimal classification spectral index

[0062] Similarly, calculate the information gain for all spectral indices according to step 3. Select the maximum information gain for each spectral index as the information gain value for that spectral index. Then, select the spectral index with the highest information gain as the optimal classification. Multiple optimal spectral index classifications may exist for the same data to be classified (S1, S2, S3). In this calculation, using B2 as the spectral index and 98.5 as the classification basis is one of the optimal solutions for extracting S2.

[0063] The above description of the present invention is intended to enable any person skilled in the art to make or use the present invention. Various modifications of the present invention will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the examples and designs described herein, but is intended to be embodied in accordance with the widest scope of the principles and novel features disclosed herein.

Claims

1. A method for extracting spectral index from hyperspectral data based on information gain, characterized in that: The following steps are involved: (1) Acquisition and preprocessing of sample data: Acquire hyperspectral measured remote sensing reflectance data or hyperspectral remote sensing image remote sensing reflectance data, group the acquired hyperspectral data by category, divide each group of data into bands, and rename the hyperspectral data categories as needed; (2) Using band replacement to construct multiple similar spectral index forms with different band combinations using the same formula, and calculate the spectral index of different classes under each spectral index form; (3) Calculate the information gain of the spectral index in its class: first divide the nodes according to the continuity of the spectral index in each class, calculate the information entropy of each divided node, calculate the information gain based on the information entropy, select the division point with the largest information gain as the optimal division point, and the information gain of this division point is used as the information gain of the corresponding spectral index form; calculate the difference between the information entropy of the sample set when it is not classified and the information entropy of each division point after the division, which is the information gain obtained; (4) Calculate the information gain of each spectral index, and take the spectral index corresponding to the maximum information gain as the optimal classification spectral index.

2. The method for extracting spectral index from hyperspectral data based on information gain according to claim 1, characterized in that: The hyperspectral data described in step (1) are grouped according to the category of the measured target, each category of hyperspectral data is a group, and the category label L1...L corresponding to each group of hyperspectral data n There are n groups in total. Each group of hyperspectral data corresponds to N bands. The hyperspectral data categories are renamed, and the category labels to be distinguished are set to 1, and the rest are set to 2.

3. The method for extracting spectral index from hyperspectral data based on information gain according to claim 1, characterized in that: The spectral index form described in step (2) is constructed according to the spectral index obtained by the single band value as formula (1), the band difference as formula (2), the band sum value as formula (3), the single band ratio as formula (4), the ratio of the single band to the band difference as formula (5), and the ratio of the band difference to the sum as formula (6), and the calculation methods are as follows: SI=b i Formula (I) SI=b i -b m Formula (II) SI=b i +b m Formula (III) Where: SI is the spectral index, b is the band, i, m, n are the i-th, m-th, n-th bands of the hyperspectral data, i, m, n are natural numbers, i, m, n ≤ N.

4. The method for extracting spectral index from hyperspectral data based on information gain according to claim 1, characterized in that: The information entropy in step (3) is calculated as follows: Arrange all the values ​​of different classes corresponding to the spectral index form a to be calculated in ascending order to obtain the sample set D, the continuous attribute α, and the mean of all adjacent attributes as candidate partition points. There are n-1, where n is the number of all values ​​of α; According to the renaming in step (1) and the proportion of the categories to be distinguished in all categories, the proportion of the k-th category samples in the sample set D is p k , k=1,2, then the information entropy of D is defined as:

5. The method for extracting spectral index from hyperspectral data based on information gain according to claim 4, characterized in that: The information gain calculation method described in step (3) is as follows: after dividing the set D by each partition point, the first branch node includes all the nodes in D whose values ​​in the spectral index form a are less than or equal to T a The sample is denoted as D V , v=1, the value is greater than T a The sample is denoted as D V , v = 2; according to formula (VII) calculate D V The information entropy of , and considering the different number of samples contained in different branch nodes, the branch nodes are given weights |D V | / |D|, calculate the difference between the information entropy of the spectral index form a for the sample set D when it is not classified and the information entropy of each partition point after the partition, which is the obtained "information gain", and select the partition point with the maximum information gain as the optimal partition point, and the information gain of this partition point is used as the information gain of the spectral index form a; Where a is the name of the spectral index form, a i is a value of the spectral index form a in D, and t is a specific division threshold, which is calculated as T a , v is the classification category, v=1,2.

6. The method for extracting spectral index from hyperspectral data based on information gain according to claim 1, characterized in that: The optimal classification spectrum index calculation method described in step (4) is as follows: SI y =max(Gain(D, SI)) Formula (10) Where: SI y The spectral index represented is a specific optimal spectral index classification form, and Gain(D, SI) is the information gain of all spectral indices.

Images