CN106778738A - Ground feature extraction method based on decision theory rough set - Google Patents

Abstract

The present invention relates to a feature extraction method based on rough set of decision theory, comprising the steps of: (1) acquiring multi-source remote sensing data, and preprocessing the remote sensing multi-band data to obtain band data vectors; (2) collecting expert knowledge base And adding the expert knowledge vector into the band data vector as a condition set, the condition set is open; (3) normalize the condition set; (4) form a decision set by the real ground class, And transform the decision set into a vector and align and combine it with the condition set as a remote sensing decision table; (5) Use the forward greedy search algorithm to process the decision table to obtain the feature ranking, get the core features of the ground objects and distinguish different land types According to the decision-making rules, the features of ground objects are extracted under the decision-making rules. The present invention is based on the feature extraction of the decision set, eliminates the noise in the data, and improves the classification accuracy.

Description

A Method of Feature Extraction Based on Rough Sets of Decision Theory

technical field

The invention relates to image processing, in particular to a feature extraction method of ground objects based on decision theory rough sets.

Background technique

The feature extraction of remote sensing features refers to the extraction of key information on different features of urban and rural industrial and mining residential land (urban construction land), water body, agricultural land, forest land, and bare land, such as obtaining specific feature information from hyperspectral data. At present, manual extraction and computer-aided semi-automatic methods are mainly used to extract ground feature information from remote sensing images; these methods require a lot of manual operations and are inefficient, while hyperspectral and multispectral information mining is less.

At present, the eCognition system in Germany is better for feature extraction of remote sensing features. It is based on multi-scale segmentation, and a classification protocol is required for classification. The classification protocol for each classification needs to be modified when it is used for other classifications. The classification protocol here is similar. Based on knowledge support, it is characterized by classification and refinement, but it relies heavily on classification protocols (protocol library support is better), and it is not a system specifically for feature extraction of ground features, and its application efficiency is not high.

Since ground features have the characteristics of reflecting or radiating electromagnetic waves, that is, the spectral response of ground features, or expert knowledge formed by summarizing previous research results, such as vegetation index VI (vegetation indices), is composed of two or more band responses in a certain form. The resulting parameters have certain indicative significance for vegetation growth, biomass, etc. A certain combination of bands can eliminate the interference of the atmosphere and soil background in the vegetation canopy spectrum to a certain extent, and has a close functional relationship with some important biophysical parameters (such as LAI, dry matter, etc.). Commonly used vegetation indexes include normalized difference vegetation index (NDVI), ratio vegetation index (SR), environmental vegetation index (EVI), green vegetation index (GVI) in tasseled cap transformation, and orthogonal vegetation index (PVI), etc. ( Statement Peng et al., 1990). In addition to vegetation, Qian Lexiang et al. found that water bodies have (TM2+TM3)>(TM4+TM5) interspectral spectral features in TM. Duggin et al. found that TM4 and TM6 can be used to distinguish forest land from other vegetation. Interpretation experience shows that TM1/TM3, TM5/TM7, TM5+TM7, (TM4-TM3)/(TM4+TM3), etc. will have certain features of the ground. The data addition, subtraction, multiplication, division and linear nonlinear combination of multiple spectral bands constitute indicative values, which is the inherent knowledge hidden in the data set, while the predecessors only performed random combination experiments based on literature and experience to achieve practical significance.

Since the key features of ground features involve remote sensing big data, the performance of ground features is complex, which increases the difficulty of information extraction and affects the extraction effect. Aiming at the shortcomings of the existing method of ground object extraction using remote sensing images, such as serious interference information, low extraction efficiency, and poor extraction methods, the present invention proposes an efficient Feature extraction method of remote sensing images.

Contents of the invention

In order to solve the above technical problems, the object of the present invention is to provide a feature extraction method based on decision theory rough sets with small error and high classification accuracy.

The present invention's method for feature extraction based on decision theory rough sets comprises the steps of:

(1) Obtain multi-source remote sensing data, and preprocess the remote sensing multi-band data to obtain band data vectors;

(2) Collect expert knowledge base and add expert knowledge vector in described band data vector as condition set, and described condition set is open type;

(3) performing normalization processing on the condition set;

(4) The decision set is formed by the real ground class, and the decision set is converted into a vector and aligned with the condition set and merged as a remote sensing decision table;

(5) Use the forward greedy search algorithm to process the decision table to obtain the feature ranking, get the core features of the ground objects and the decision rules for distinguishing different land types from the primary and secondary features, and extract the ground feature features under the decision rules.

Further, the specific steps of processing the decision table using the forward greedy search algorithm are as follows:

(11) Set up the core feature set R and assign it to an empty set;

(12) For each feature a _k belonging to the condition set CR, calculate k=1,2...,C={a ₁ ,a ₂ ,...a _k ,...}, for the probability that _X belongs to D, is the description of x, X is a decision class, and satisfies criterion Ⅰ: ≥α, α is the confidence level;

(13) Calculate the cost function ω _i is the i-th category, i=1, 2, ..., c, X is the decision-making category, and the largest P represents the largest possible category in the boundary domain;

(14) Calculate the cost function for each unit i in the remote sensing data, and accumulate it into the total cost function i=1,2,...,n;

(15) calculation |U| is the total number of units, γ is an indicator that integrates confidence level, cost or risk measurement, and universality;

(16) First calculate a γ value for each feature, and select the largest one, which meets the following two conditions:

Criterion II: The total cost does not increase between cycles,

Criterion Ⅲ: maximum universality;

(17) The initial value of γ ₀ is assigned 0;

(18) If the absolute value of the difference between γ and γ ₀ is greater than the importance threshold Δ, the feature a _m is considered important, and added to the core feature set R, and enters the cycle of step (12) until the absolute value of the difference between γ and γ ₀ When it is less than or equal to the importance threshold Δ, it is considered that no important features can be extracted, the loop ends, and R is returned;

(19) The order of features in the core feature set R corresponds to the decision rule.

Further, the preprocessing of the remote sensing multi-band data in the step (1) includes vector encoding and coordinate registration processing.

Further, the expert knowledge in the step (2) includes the vegetation index, the parameters indicative of vegetation growth, biomass, etc., which are combined in a certain form from the responses of two or more bands.

Further, the decision set composed of real ground types in the step (4) is based on one or more materials in topographic maps, thematic maps, and planning maps combined with manual field surveys, and assigns different values to each pixel unit. The value of the category is obtained from the thematic category of the corresponding area.

By means of the above scheme, the present invention uses composite standards to minimize errors, and the proposed γ measure is an index that integrates confidence level, cost or risk measure, and universality, and has the following three advantages:

1. It is easy to process discrete and numerical data;

2. The confidence guarantees its strong noise resistance;

3. The cost function reflects the real misclassification;

The invention controls the real error, obtains the core feature set, and improves the classification accuracy.

The above description is only an overview of the technical solutions of the present invention. In order to understand the technical means of the present invention more clearly and implement them according to the contents of the description, the preferred embodiments of the present invention and accompanying drawings are described in detail below.

Description of drawings

Fig. 1 is a flow chart of the present invention;

Fig. 2 is a comparison table of accuracy performance of the feature extraction method of the present invention in four typical classification algorithms.

detailed description

The specific implementation manners of the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

A feature extraction method based on rough sets of decision-making theory described in a preferred embodiment of the present invention, through on-site investigation to understand the characteristics of various features and the surrounding environment in the area, and then proceed from the three aspects of object space, image space and people Analysis, use the rough set of decision theory to extract the feature sequence and association rules of the ground object, under the formal background of a set of conditional feature (attribute) set and decision attribute set, establish the remote sensing decision table, design a heuristic search algorithm to process the decision table to get Feature sorting, the decision rules are obtained from the primary and secondary features. Under the decision-making rules, ground feature extraction is carried out, as shown in Figure 1, first collect remote sensing data and terrain (shape) maps; then select the most accurate reference map, and other data are corrected to the coordinate system of the reference map; and then divide into two steps , to construct the condition set and decision set respectively. The condition set is composed of hyperspectral band, band algebra, ratio and other expert knowledge, and it is open, allowing new features to be added; the decision set is composed of real ground classes; thus, the remote sensing decision table is passed The condition set and the decision set are connected according to the order of the image units; the number of features in the decision table may be many, which is an NP-hard problem. In order to obtain a suitable solution, the present invention implements a forward greedy search algorithm, and the steps are as follows:

Step 1: set up the core feature set R and assign it to an empty set;

Step 2: Calculate for each feature a _k that belongs to the condition set CR

Here, k=1, 2..., C={a ₁ , a ₂ ,...a _k ,...}, for Probability of (description of x) belonging to D _X (decision class X), satisfying criterion I: α is the confidence level;

Step 3: Calculate the cost function, namely Here ω _i is the i (i=1,2,...,c) class; X is the decision class; the largest P represents the largest possible class in the boundary domain, and at the same time, if The equivalence class is an increase in the probability of small classes in the boundary domain, which is more likely to cause misclassification, so the cumulative cost is 1;

Step 4: Calculate the cost function for each unit i (i=1,2,…,n) in the remote sensing data, and accumulate it into the total cost function:

Step 5: Calculate |U| is the total number of units; the γ measure is an indicator that integrates confidence level, cost or risk measure, and universality;

Step 6: First calculate a γ value for each feature (a total of k), and select the largest one, which meets the following two conditions:

Criterion II: The total cost does not increase between cycles;

Criterion Ⅲ: maximum universality;

Step 7: assign 0 to the initial value of γ ₀ ;

Step 8: If the absolute value of the difference between γ and γ ₀ is greater than the importance threshold Δ, the feature a _m is considered important, and added to the core feature set R, and enters the cycle of step 2 until the absolute value of the difference between γ and γ ₀ is not greater than When the importance threshold is Δ, it is considered that no important features can be proposed, the loop ends, and R is returned;

Step 9: The order of features in the core feature set R corresponds to the decision rule.

Through the above steps, the core features of ground objects and the decision rules for distinguishing different land types are obtained.

The specific implementation method will be described below with an implementation example, that is, feature extraction is performed on many bands of remote sensing images to obtain thematic information vector data (.shp files) that meet the requirements of general GIS software. Its implementation is as follows:

1. Remote sensing multi-band data preprocessing, specifically including vector encoding, registration, etc.

The remote sensing spectrum vector encoding adopts discretization encoding, specifically: according to the remote sensing image data, each band is regarded as one dimension, the number of bands determines the dimension of the vector, and the data other than the thermal infrared band are selected to form a vector, and the vector arrangement is strictly in accordance with Bit-aligned to form a vector of band data.

2. Add expert knowledge, that is, expert knowledge formed by summarizing previous research results, such as vegetation index VI (vegetation indices), which is composed of two or more band responses in a certain form and has a certain effect on vegetation growth and biomass. Parameters that indicate meaning. A certain combination of bands can eliminate the interference of the atmosphere and soil background in the vegetation canopy spectrum to a certain extent, and has a close functional relationship with some important biophysical parameters (such as LAI, dry matter, etc.). Commonly used vegetation indices include normalized difference vegetation index (NDVI), ratio vegetation index (SR), environmental vegetation index (EVI), green vegetation index (GVI) in tasseled cap transformation, and orthogonal vegetation index (PVI), etc. ( Statement Peng et al., 1990). In addition to vegetation, Qian Lexiang et al. found that water bodies have (TM2+TM3)>(TM4+TM5) spectral features in TM. Duggin et al. found that TM4 and TM6 can be used to distinguish forest land from other vegetation. Interpretation experience shows that TM1/TM3, TM5/TM7, TM5+TM7, (TM4-TM3)/(TM4+TM3), etc. will all propose certain ground features.

Add expert knowledge TM2+TM3, TM4+TM5, TM2-TM4, TM3-TM7, TM1/TM3, TM5/TM7 to the band data vector, and align by bit, as a condition set.

3. Normalize the condition set. Different indicators often have different dimensions and dimensional units. This situation will affect the results of data analysis. In order to eliminate the dimensional influence between indicators, normalization is required Processing, that is, y=(x-MinValue)/(MaxValue-MinValue), where: x, y are the values before and after conversion respectively, and MaxValue and MinValue are the maximum value and minimum value of the feature respectively.

4. Based on other materials (topographic maps, thematic maps, planning maps) and field surveys, the thematic category of the region is obtained, and each pixel unit is given a different category value to form a decision set, and the decision set is converted into a vector and Align with condition set and merge into remote sensing decision table.

5. Use the forward greedy search algorithm to process the decision table, get the feature ranking, and get the decision rules of the local features based on the primary and secondary features, which can achieve better classification accuracy than that without feature extraction.

As shown in Figure 2, the five data sets in the first column are: TM band, 12-dimensional data, RI, RII, and RIII, among which RI, RII, and RIII are feature sets, which are the core features proposed by this method; The second column is the feature number of the dataset. The same four classification algorithms are used for these data sets, namely, parallelepiped method (Parallelepiped), Mahalanobis distance method (MahaDist), maximum likelihood method (MaxLike), and support vector machine (SVM). Accuracy OA and kappa coefficient express. It can be seen that the feature reduction (that is, feature extraction) of the present invention eliminates the noise in the data, so that the classification accuracy is improved.

The above is only a preferred embodiment of the present invention, and is not intended to limit the present invention. It should be pointed out that for those of ordinary skill in the art, some improvements can be made without departing from the technical principle of the present invention. and modifications, these improvements and modifications should also be considered as the protection scope of the present invention.

Claims (5)

1. A feature extraction method based on decision theory rough set, is characterized in that, comprises steps: (1) Obtain multi-source remote sensing data, and preprocess the remote sensing multi-band data to obtain band data vectors; (2) Collect expert knowledge base and add expert knowledge vector in described band data vector as condition set, and described condition set is open type; (3) performing normalization processing on the condition set; (4) The decision set is formed by the real ground class, and the decision set is converted into a vector and aligned with the condition set and merged as a remote sensing decision table; (5) Use the forward greedy search algorithm to process the decision table to obtain the feature ranking, get the core features of the ground objects and the decision rules for distinguishing different land types from the primary and secondary features, and extract the ground feature features under the decision rules. 2. the feature extraction method based on decision theory rough set according to claim 1, is characterized in that: utilize forward greedy search algorithm to the concrete step that decision table is processed as (11) Set up the core feature set R and assign it to an empty set; (12) For each feature a _k belonging to the condition set CR, calculate k=1,2...,C={a ₁ ,a ₂ ,...a _k ,...}, for the probability that _X belongs to D, is the description of x, X is a decision class, and satisfies criterion Ⅰ: α is the confidence level; (13) Calculate the cost function ω _i is the i-th category, i=1, 2, ..., c, X is the decision-making category, and the largest P represents the largest possible category in the boundary domain; (14) Calculate the cost function for each unit i in the remote sensing data, and accumulate it into the total cost function i=1,2,...,n; (15) calculation |U| is the total number of units, γ is an indicator that integrates confidence level, cost or risk measurement, and universality; (16) First calculate a γ value for each feature, and select the largest one, which meets the following two conditions: Criterion II: The total cost does not increase between cycles, Criterion Ⅲ: maximum universality; (17) The initial value of γ ₀ is assigned 0; (18) If the absolute value of the difference between γ and γ ₀ is greater than the importance threshold Δ, the feature a _m is considered important, and added to the core feature set R, and enters the cycle of step (12) until the absolute value of the difference between γ and γ ₀ When it is less than or equal to the importance threshold Δ, it is considered that no important features can be extracted, the loop ends, and R is returned; (19) The order of features in the core feature set R corresponds to the decision rule. 3. The feature extraction method based on decision theory rough sets according to claim 1, characterized in that: the preprocessing of the remote sensing multi-band data in the step (1) includes vector encoding and coordinate registration processing. 4. the feature extraction method based on decision theory rough set according to claim 1, is characterized in that: in described step (2), expert knowledge comprises vegetation index, is responded to by two or more wave bands in a certain form Combined parameters that are indicative of vegetation growth, biomass, etc. 5. the feature extraction method based on decision theory rough set according to claim 1, is characterized in that: in the described step (4), the decision set that is made up of real ground class is based on topographic map, thematic map, planning One or more types of data in the picture are combined with manual field surveys to assign different category values to each pixel unit to obtain thematic category of the corresponding area.