A method for estimating the number of non-target categories for remote sensing classification

Abstract

The application discloses a non-target category number estimation method for remote sensing classification and belongs to the technical field of remote sensing classification and identification. The method comprises the following steps: S1, acquiring a remote sensing image, wherein the remote sensing image contains a plurality of target ground classes with known category numbers and a plurality of non-target ground classes with unknown category numbers, and image samples of each target ground class and non-target ground class are prepared; S2, based on the image samples of each target ground class and non-target ground class in step S1, an optimization index for non-target category number estimation is constructed; and S3, based on the optimization index in step S2, a judgment standard for the best non-target category number is constructed, and the value of the best non-target ground class category number is determined. The optimization index for non-target category number estimation and the judgment standard can be constructed to quickly and accurately determine the non-target category number, and the problems of the existing non-target category number being general and rough, the non-target ground class labeling and category division being time-consuming and labor-consuming, and the inconsistency are solved.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing classification and recognition technology, specifically relating to a method for estimating the number of non-target categories for remote sensing classification. Background Technology

[0002] In recent years, with advancements in satellite spatial resolution, temporal resolution, and spectral resolution, remote sensing technology has become an effective means of providing massive amounts of Earth observation data. The dramatic increase in the amount of data from high-resolution remote sensing images provides more detailed data sources for remote sensing image target detection systems. However, this increase also raises the computational costs of remote sensing image land cover classification and target detection. Improving the accuracy of remote sensing classification has become a pressing problem. In traditional remote sensing classification, due to the difficulty or lack of interest in interpreting non-target land cover types, multiple non-target land cover types are often grouped into a single category, such as "others" or "unknown," and samples from multiple non-target land cover types are input into the same classification model along with samples from target land cover types for training and classification. This can interfere with the classification of target land cover types, leading to a decrease in recognition accuracy. In addition, the labeling and classification of non-target land cover types have traditionally been done manually. Due to the large volume of remote sensing image data, manual processing is not only time-consuming and labor-intensive, but may also be affected by subjective human factors, leading to inconsistencies between the labeling and classification of non-target land cover types and increasing the uncertainty of the final target land cover type extraction. Summary of the Invention

[0003] To overcome the aforementioned technical problems, this invention provides a method for estimating the number of non-target categories for remote sensing classification. Using image samples as the object, by constructing an optimization index for estimating the number of non-target categories and determining the criteria for the optimal number of categories, the number of non-target categories can be determined quickly and accurately. This solves the problems of existing methods for estimating the number of non-target categories being vague and coarse, as well as the time-consuming, labor-intensive, and inconsistent labeling and classification of non-target land cover types.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] A method for estimating the number of non-target categories for remote sensing classification includes the following steps:

[0006] Step S1: Acquire remote sensing images containing multiple target land cover types with known numbers and multiple non-target land cover types with unknown numbers, and create image samples for each target land cover type and non-target land cover type;

[0007] Step S2: Based on the image samples of each target land class and non-target land class in Step S1, construct an optimized index for estimating the number of non-target categories;

[0008] Step S3: Based on the optimization index in step S2, construct a criterion for determining the number of non-target optimal categories and determine the value of the number of non-target optimal categories.

[0009] Further, step S2 includes the following steps:

[0010] Step S21: Classify the non-target land types from Step S1 using clustering methods;

[0011] Step S22: Calculate the distance from each non-target land type to the target land type;

[0012] Step S23: Construct an optimization index for estimating the number of non-target categories based on the calculation results of step S22.

[0013] Further, the classification method in step S21 is as follows: Assuming the current remote sensing image contains K target land classes with known class numbers, the K-means clustering method is used, and based on n-dimensional features, non-target land classes with unknown class numbers are divided into U classes. Then, the Euclidean distance from target land class k to non-target land class u can be denoted as: The calculation formula is as follows:

[0014]

[0015] Furthermore, in step S21, the n-dimensional features include RGB features, spectral index features, and texture features.

[0016] Further, in step S22, the method for calculating the distance from the non-target land class to the target land class is as follows: First, calculate the Euclidean distances from the current non-target land class u to all target land classes K, and then take the minimum value as the distance d from the current non-target land class to the target land class. u The calculation formula is as follows:

[0017]

[0018] Furthermore, in step S23, the optimization index is constructed as follows: First, the distance d from all non-target land classes to the target land class is calculated. u Then, the minimum value is taken as the optimization index I for estimating the number of non-target categories. The calculation formula is as follows:

[0019]

[0020] Furthermore, step S3 includes the following steps:

[0021] Step S31: Using the dataset of image samples obtained in step S1 as the object, the number U of non-target land types is iteratively clustered using a clustering method in an order that gradually increases from 1, 2, 3, ..., m.

[0022] Step S32: After each iteration of clustering is completed, calculate the optimization index I of that iteration, and obtain the sequence of optimization indices I1, I2, ..., I1. m ;

[0023] Step S33: Construct the judgment criteria as (I i-1 -I i )>0.5*(I max -I min In the formula, I i-1 I represents the value of I in the (i-1)th iteration. i I represents the value of I in the i-th iteration. max The maximum value of the exponential I sequence, I min It is the minimum value of the exponential I sequence;

[0024] Step S34: Determine I in the i-th iteration i If the value meets the criteria, then i is the optimal number of non-target land categories.

[0025] Furthermore, in step S31, the clustering method is the K-means clustering method.

[0026] By adopting the above technical solution, the present invention has the following beneficial effects:

[0027] 1. This invention takes image samples as the object and, by constructing an optimization index for estimating the number of non-target categories and determining the criteria for the optimal number of categories, can quickly and accurately determine the number of non-target categories. It can effectively solve the problems of existing methods for determining the number of non-target categories being general and coarse, as well as the time-consuming, labor-intensive, and inconsistent labeling and classification of non-target land cover types.

[0028] 2. This invention not only improves work efficiency and reduces computing costs, but also avoids subjective judgment errors caused by human differences. It can improve the overall accuracy of automatically extracting target land types from large and complex actual remote sensing images, thereby better serving research in various fields.

[0029] 3. The non-target category quantity estimation method proposed in this invention provides a good foundation for the training of subsequent image target extraction models, and can improve the overall effect of automatic extraction of target land categories in actual remote sensing images, thus having high practical application value. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the method implementation of Embodiment 1 of the present invention;

[0031] Figure 2 This is a schematic diagram illustrating the calculation of the optimization index for estimating the number of non-target categories in Embodiment 1 of the present invention;

[0032] Figure 3 This is a schematic diagram of the "inflection point" of the non-target category number estimation optimization index in Embodiment 1 of the present invention. Figure 1 ;

[0033] Figure 4 This is a schematic diagram of the "inflection point" of the non-target category number estimation optimization index in Embodiment 1 of the present invention. Figure 2 ;

[0034] Figure 5 This is a schematic diagram of the "inflection point" of the non-target category number estimation optimization index in Embodiment 1 of the present invention. Figure 3 ;

[0035] Figure 6 This is a schematic diagram of the non-target land type before optimal classification in Embodiment 1 of the present invention;

[0036] Figure 7 This is a schematic diagram of the optimal classification of non-target land types in Embodiment 1 of the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The accompanying drawings are for illustrative purposes only and represent schematic diagrams, not actual pictures, and should not be construed as limiting this patent. In order to better illustrate the specific implementation of the present invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product size. It is understandable for those skilled in the art that some well-known structures, components, and their descriptions may be omitted in the drawings. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set" and "connection" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0039] Example 1

[0040] Please see Figure 1 A method for estimating the number of non-target categories for remote sensing classification includes steps S1, S2, and S3.

[0041] Step S1: Acquire remote sensing images containing multiple target land cover types with known numbers and multiple non-target land cover types with unknown numbers, and create image samples for each target land cover type and non-target land cover type.

[0042] Step S2: Based on the image samples of each target land class and non-target land class in Step S1, construct an optimized index for estimating the number of non-target categories.

[0043] Specifically, step S2 includes steps S21, S22 and S23.

[0044] Step S21: Classify the non-target land cover types from Step S1 using a clustering method. The classification method is as follows: Assuming the current remote sensing image contains K target land cover types with known class numbers, use K-means clustering and n-dimensional features to classify the non-target land cover types with unknown class numbers into U categories. The Euclidean distance from target land cover type k to non-target land cover type u can be denoted as: The calculation formula is as follows:

[0045]

[0046] The n-dimensional features include RGB features, spectral index features, and texture features.

[0047] Step S22: Calculate the distance from each non-target land class to the target land class. The method for calculating the distance from a non-target land class to the target land class is as follows: First, calculate the Euclidean distances from the current non-target land class u to all target land classes K, then take the minimum value as the distance d from the current non-target land class to the target land class. u The calculation formula is as follows:

[0048]

[0049] Step S23: Construct an optimization index for estimating the number of non-target land categories based on the calculation results of step S22. The optimization index is constructed as follows: First, calculate the distance d from all non-target land categories to the target land category. u Then, the minimum value is taken as the optimization index I for estimating the number of non-target categories. The calculation formula is as follows:

[0050] I = min(d) 1 ,d 2 ,...,d U ).

[0051] Step S3: Based on the optimization index in step S2, construct a criterion for determining the number of non-target optimal categories and determine the value of the number of non-target optimal categories.

[0052] Specifically, step S3 includes steps S31, S32, S33 and S34.

[0053] Step S31: Using the dataset of image samples obtained in Step S1 as the object, iteratively cluster the number U of non-target land classes in an increasing order from 1, 2, 3, ..., m using a clustering method. The clustering method is K-means clustering.

[0054] Step S32: After each iteration of clustering is completed, calculate the optimization index I of that iteration, and obtain the sequence of optimization indices I1, I2, ..., I1. m .

[0055] Step S33: Construct the judgment criteria as (I i-1 -I i )>0.5*(I max -I min In the formula, I i-1 I represents the value of I in the (i-1)th iteration. i I represents the value of I in the i-th iteration. max The maximum value of the exponential I sequence, I min It is the minimum value of the exponential I sequence.

[0056] Step S34: Determine I in the i-th iteration i If the value meets the criteria, then i is the optimal number of non-target land categories.

[0057] When the I of the i-th iteration i When the value meets the judgment criteria, it indicates that the optimal estimate i of the number of non-target land types has been found. That is, the "inflection point" of the curve of the optimization index changing with the number of iterations corresponds to the current optimal estimate of the number of non-target land types.

[0058] Figure 2 This diagram illustrates the optimization index calculation for estimating the number of non-target land cover categories. First, it assumes that the number of target land cover categories in the current remote sensing image sample is 2. K-means clustering is used to divide the non-target land cover categories with unknown numbers into 2 categories based on n-dimensional features (including RGB, spectral index, and texture). Second, the distance d from each non-target land cover category to the target land cover category is calculated. u Then, the minimum distance from the current non-target land class to all target land classes is taken as the distance from the current non-target land class to the target land class; finally, the distance d from all non-target land classes to the target land class is calculated in turn. u Then, the minimum value among these distances is taken as the optimization index I for estimating the number of non-target categories.

[0059] Figure 3-5This diagram illustrates the "inflection point" of the optimization index for estimating the number of non-target categories. The method of this invention was applied to remote sensing image samples containing multiple known target land types and multiple known non-target land types to calculate the optimization index for estimating the number of non-target categories. Figure 3 To verify Experiment 1, the remote sensing image samples included two target land types (wasteland and orchards) and three non-target land types (desert, water bodies, and silkworms). K-means clustering was used to classify the non-target land types in the current remote sensing image samples multiple times, and the corresponding non-target category number estimation optimization index was calculated. Figure 3 It can be observed that the optimization index drops sharply when non-target land cover is divided into three categories, i.e., an "inflection point" occurs, which is consistent with the number of known non-target land cover categories in the current remote sensing image sample. Furthermore, Figure 4 Verification Experiment 2 and Figure 5 The "inflection point" of the optimization index for estimating the number of non-target land cover categories in Experiment 3 also coincides with the known number of non-target land cover categories in the corresponding remote sensing image sample. All the above experiments confirm the reliability of determining the optimal number of non-target land cover categories in the current remote sensing image by finding the "inflection point" of the optimization index for estimating the number of non-target land cover categories.

[0060] Using an image sample dataset with known category names and quantities, samples are artificially divided into two groups with unequal numbers of categories: target land category samples and non-target land category samples (their quantities are known and used as the true values). The practicality and reliability of the proposed non-target category quantity estimation method are verified by comparing the non-target land category quantities calculated using the method of this invention with the true quantities.

[0061] During the experiment, K-means clustering was used to automatically divide non-target land features into different numbers of categories multiple times using n-dimensional features. An optimization index for estimating the number of non-target categories in the current remote sensing image sample was calculated based on the number of categories in each division. It can be clearly observed that the optimization index for estimating the number of non-target categories in the current remote sensing image sample changes with the K-means clustering classification. Furthermore, there is an "inflection point," where the optimization index suddenly drops sharply from its previous gradual change at a certain classification position. After this "inflection point," the optimization index tends to change gradually with the classification.

[0062] By comparing the calculated optimization index for the number of non-target land cover categories with the known number of non-target land cover categories in the current remote sensing image sample, it was found that the "inflection point" precisely corresponds to the known number of non-target land cover categories in the current remote sensing image sample. This clarifies the criterion for finding the optimal number of non-target land cover categories in remote sensing images based on the optimization index proposed in this invention, i.e., finding the "inflection point" of the optimization index for estimating the number of non-target land cover categories in the current remote sensing image. Furthermore, it confirms the applicability and accuracy of the optimization index for estimating the number of non-target land cover categories proposed in this invention. The non-target category quantity estimation method for remote sensing classification proposed in this invention provides a good foundation for the training of subsequent image target extraction models, and can improve the overall effect of automatic target land cover extraction in actual remote sensing images, thus having high practical application value.

[0063] Figure 6-7 This diagram illustrates the differences between non-target land types before and after optimal classification. In sample images where non-target land types were never classified, it's evident that different categories, such as deserts, concrete areas, mulberry fields, and water bodies, are grouped into a single category. This coarse classification ignores the distinctive features between non-target land types, failing to differentiate them and easily causing confusion between non-target and target land types, severely impacting the accuracy of target land type extraction. Furthermore, previous labeling and classification of non-target land types were primarily done manually, which is inefficient and susceptible to human error, leading to inconsistencies between labeling and classification, and increasing the uncertainty of the final target land type extraction. By using the non-target category number estimation optimization index proposed in this invention to find the number of non-target land cover categories, the non-target land cover categories in the current remote sensing image samples are divided into 4 categories. It can be found that the non-target land cover categories have similar characteristics and the differences between categories are obvious. This classification method not only clarifies the categories of non-target land cover categories, making it easy to distinguish between non-target land cover categories and target land cover categories, thus improving the extraction accuracy of target land cover categories, but also avoids the inefficiency and result errors caused by human judgment of non-target land cover categories.

[0064] This invention takes image samples as the object and, by constructing an optimization index for estimating the number of non-target categories and determining the criteria for the optimal number of categories, can quickly and accurately determine the number of non-target categories. It can effectively solve the problems of existing methods for determining the number of non-target categories being general and coarse, as well as the time-consuming, labor-intensive, and inconsistent labeling and classification of non-target land cover types.

[0065] This invention not only improves work efficiency and reduces computing costs, but also avoids subjective judgment errors caused by human differences. It can improve the overall accuracy of automatically extracting target land types from large and complex real remote sensing images, thereby better serving research in various fields.

[0066] The above description is a detailed description of the preferred embodiments of the present invention. However, the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modifications made under the technical spirit of the present invention should fall within the patent scope covered by the present invention.

Claims

1. A non-target class number estimation method for remote sensing classification, characterized in that, The method comprises the following steps: Step S1: obtaining a remote sensing image, the remote sensing image containing a plurality of target land classes with known class numbers and a plurality of non-target land classes with unknown class numbers, and making image samples of each target land class and non-target land class; Step S2: constructing an optimization index for non-target class number estimation based on the image samples of each target land class and non-target land class in step S1; Step S2 comprises the following steps: Step S21: performing class division on the non-target land classes in step S1 by a clustering method; The classification method in step S21 is as follows: assuming the current remote sensing image contains... Given a known number of target land categories, using Mean clustering method and based on Dimensional features divide non-target land classes with unknown number of categories into If there are several categories, then the target location category is... To non-target location The Euclidean distance can be denoted as: The calculation formula is as follows: ； In step S21, the The features include RGB features, spectral index features, and texture features. Step S22: respectively calculating the distance from each non-target land class to the target land classes; In step S22, the distance calculation method from the non-target land class to the target land class is as follows: first, the Euclidean distance from the current non-target land class to all target land classes is counted, and then the minimum value is taken as the distance from the current non-target land class to the target land class , and the calculation formula is as follows: ​​ ； Step S23: constructing an optimization index for non-target class number estimation according to the calculation results of step S22; In step S23, the construction method of the optimization index is as follows: first, the distance of all non-target land types to the target land type is counted , and then the minimum value is taken as the optimization index of the non-target type number estimation , and the calculation formula is as follows: ; step S3: based on the optimization index in step S2, the determination standard of the best number of non-target land types is constructed, and the value of the best number of non-target land types is determined; Step S3 comprises the following steps: Step S31: Taking the data set of the image sample obtained in step S1 as the object, the number of non-target land classes is determined In the order of gradually increasing from 1, 2, 3, …, m, iterative clustering is performed by using a clustering method; Step S32: After each round of iterative clustering is completed, the optimization index of the round of iterative clustering is calculated I , the optimization index is obtained I of the sequence I 1 ，I 2 ，…，I m ; Step S33: constructing the decision criterion as wherein represents the value of the first i -1 iteration, represents the value of the first iteration, represents the value of the first I iteration, is the maximum value of the exponential sequence, is the minimum value of the exponential sequence; Step S34: Determine the first i The next iteration Does the value meet the conditions of the judgment criterion? If it does, then... i This represents the optimal number of categories for non-target land types.

2. The non-target class number estimation method for remote sensing classification according to claim 1, wherein: In step S31, the clustering method is a K-means clustering method.

Images