A high-dimensional image unsupervised classification method

Abstract

This invention relates to the field of high-resolution image technology, specifically disclosing an unsupervised classification method for high-resolution images, comprising the following steps: S01: Input: The input file is a high-resolution remote sensing image, and the mean of the sum of the bands is calculated to obtain a single-band high-resolution remote sensing image; S02: Output: The output file is a classification result image with the same number of rows and columns as the high-resolution image; S03: Create a classification result image: Create a single-channel image with the same number of rows and columns as the input high-resolution remote sensing image, denoted as C, and record the classification result; S04: Consistency judgment of objects within the window: Read the input single-band high-resolution image G, traverse the high-resolution image G and the classification result image C, calculate the kurtosis coefficient of the pixel values ​​of G within the window, and then move the window to the image area that has not been traversed. This solves the problems of traditional supervised classification requiring a large amount of prior knowledge and consuming manpower and resources to create labels, and the problem of fragmented classification results in traditional unsupervised classification methods.

Description

Technical Field

[0001] This application relates to the field of high-resolution image technology, and specifically discloses an unsupervised classification method for high-resolution images. Background Technology

[0002] Currently, there are two main methods for classifying high-resolution images: one is supervised classification based on machine learning, which uses convolutional neural networks to classify high-resolution images; the other is unsupervised classification, such as K-Means, K-mediods, and Fuzzy C-Means.

[0003] Machine learning-based methods typically involve creating labeled data from field sampling, existing land use classification data, and remote sensing images. The labeled data and the original remote sensing images are then cropped to specific sizes as needed and divided into training and datasets according to a set ratio for model training. Commonly used models include SVM, CNN, and DNN.

[0004] Unsupervised classification methods refer to classification processes where humans do not impose any prior knowledge on the classification process, but only classify objects based on the similarity of their color features. These methods are based on clustering theory, are computationally simple, and possess rotation and translation invariance. Examples include K-Means, which groups similar pixels into one class.

[0005] Disadvantages of existing technology:

[0006] The two commonly used high-resolution image classification methods have the following main problems.

[0007] The main problems with machine learning-based methods are: dealing with massive amounts of remote sensing data requires a large amount of labeled data, which is cumbersome to produce and consumes a lot of manpower and resources. As the amount of data increases, the cost of training the model also increases, the hardware requirements become more demanding, and it becomes more time-consuming.

[0008] The main problems with unsupervised classification methods are: while unsupervised classification involves less human constraint before machine classification, requiring only the definition of land cover categories, remote sensing imagery contains instances of the same object with different spectra and the same spectra with different objects. Unsupervised classification only considers pixel similarity, leading to fragmented and unreasonable classification results. Furthermore, existing unsupervised classification methods do not consider the first law of geography, meaning the same land cover may be classified into different categories. Existing unsupervised classification methods also require pre-determining the number of categories, typically using scree maps. Similarity coefficients, such as Euclidean distance and cosine similarity, are easily affected by extreme values, and some only consider directional similarity, not magnitude. Kurtosis, on the other hand, can determine whether pixels belong to the same class based on the concentration of data distribution, resulting in more reasonable classification results. Summary of the Invention

[0009] The purpose of this invention is to solve the problem that traditional supervised classification relies on a large amount of prior knowledge and requires manpower and resources to produce labels.

[0010] To achieve the above objectives, the present invention provides the following basic solution:

[0011] A high-resolution image unsupervised classification method includes the following steps:

[0012] S01: Input: The input file is a high-resolution remote sensing image, and the mean of the sum of the bands is calculated to obtain a single-band high-resolution remote sensing image;

[0013] S02: Output: The output file is a classification result image with the same number of rows and columns as the high-resolution image; the number of categories does not need to be determined.

[0014] S03: Create a classification result image: Create a single-channel image with the same number of rows and columns as the input high-resolution remote sensing image, with the initial pixel value of 0, denoted as C, and record the classification result;

[0015] S04: Consistency judgment of objects in the window: Read the input single-band high-resolution image G, traverse the high-resolution image G and the classification result image C, scan the entire image, calculate the kurtosis coefficient of the pixel values ​​of G in the window, and then move the window to the image area that has not been traversed, and continue to make judgments until the entire image has been traversed.

[0016] S05: Output the classification results.

[0017] The principle and effect of this basic scheme are as follows:

[0018] 1. Compared with existing technologies, this invention reduces the prior knowledge required for supervised classification and the significant manpower and resources spent on creating a large number of labels. Unlike traditional unsupervised classification, it considers the first law of geography during classification, avoiding the fragmentation of classification results by classifying the same feature into different categories. It utilizes kurtosis coefficients to determine whether images within a window belong to the same feature, fully leveraging the effective information of pixels within the window. Traversing the entire high-resolution remote sensing imagery by window not only reduces computational load but also accelerates calculation speed. The adaptive size of the traversal window and the use of kurtosis coefficients to determine whether pixels belong to the same class maximize the classification of the same feature into one category, resulting in more reasonable classification results.

[0019] 2. Compared with existing technologies, ① supervised classification requires the creation of a large number of data labels, consuming significant manpower and resources, while unsupervised classification requires no prior knowledge and is faster; ② incorporating proximity relationships of ground features into unsupervised classification improves classification effectiveness and avoids overly fragmented results; ③ this unsupervised classification method does not require pre-defining the number of categories; ④ using kurtosis coefficients to determine whether pixels within a window are more concentrated and belong to the same category makes the clustering results more reasonable.

[0020] Further, in step S04, the specific steps are as follows: Read the input single-band high-resolution image G, traverse the high-resolution image G and the classification result image C, that is, starting from the upper left corner of the image, from left to right and from top to bottom, scan the entire image using a window (the calculation area). Calculate the kurtosis coefficient of the pixel values ​​of G within the window.

[0021] Further, determine whether the kurtosis coefficient is greater than 0. If it is greater than 0, expand the window by one pixel along the window outline. If it is less than 0, discard the one pixel expansion along the window outline in the previous step.

[0022] Furthermore, in step S04, the window size for the first calculation of the kurtosis coefficient (initialization window) is M×M, where M is the width and height of the window.

[0023] Furthermore, if the kurtosis coefficient is less than or equal to 0 and the window size is greater than 3×3, then remove the one-pixel width that was expanded along the window outline in the previous step, and assign the pixel value of C within the window to i.

[0024] Furthermore, if the window contains pixels with C not equal to 0, then remove the G pixels at the positions corresponding to the pixels with C not equal to 0 in the window, and then continue to calculate the kurtosis coefficient of the G pixels in the window.

[0025] Furthermore, in step S04, the entire image is traversed, and it is determined whether the image has been traversed. If the image has been traversed, the classification result C is output. If the image has not been traversed, the process returns to the window where the kurtosis coefficient was calculated for the first time and continues to calculate the kurtosis coefficient. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 A schematic diagram illustrating the principle of a high-resolution image unsupervised classification method proposed in an embodiment of this application is shown.

[0028] Figure 2The code diagram shows the classification result of a high-resolution image unsupervised classification method proposed in an embodiment of this application. Detailed Implementation

[0029] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0030] The reference numerals in the accompanying drawings are: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12.

[0031] Implementation, for example Figure 1 As shown:

[0032] A high-resolution image unsupervised classification method includes input description, output description, consistency judgment of objects within the window, and adaptive window size.

[0033] For a sample with n values, the sample kurtosis is:

[0034]

[0035] Where m4 is the fourth-order sample central moment, m2 is the second-order central moment (i.e., sample variance), n is the number of samples, and x i It is the i-th value. This is the sample mean. Note that the divisor used when calculating the variance here is N, not (N-1) used when calculating the sample variance separately.

[0036] (1) Input instructions

[0037] The input file is a high-resolution remote sensing image. The average value of the sum of the bands is calculated to obtain a single-band high-resolution remote sensing image. Specifically, the pixel values ​​of the corresponding pixels in each band are added together, and then divided by the number of bands to obtain the average value of each pixel. After calculating all pixels, the average value of the sum of the bands is obtained.

[0038] (2) Output Description

[0039] The output file contains classification result images with the same number of rows and columns as the high-resolution image. The number of categories cannot be determined in advance. Regarding the classification result images: a two-dimensional array with an initial value of 0 is created using NumPy (a third-party library) to store the classification results, marking whether pixels in the single-band high-resolution image have been traversed. For example, see attached... Figure 2 The code below creates a 1000×1000 two-dimensional array with an initial value of 0.

[0040] (3) Create a classification result image

[0041] Create a single-channel image, denoted as C, with the same number of rows and columns as the input high-resolution remote sensing image, and record the classification results.

[0042] (4) Consistency judgment of objects in the window

[0043] Read the input single-band high-resolution image G, and iterate through the high-resolution image G and the classification result image C, starting from the top left corner of the image, from left to right and from top to bottom, using a window (the calculation area) to scan the entire image. Calculate the kurtosis coefficient of the pixel values ​​of G within the window. The initial window size for the first calculation of the kurtosis coefficient (initial window) is M×M (M is the width and height of the window, which can be set according to the size of the terrain features, with a minimum of 3×3. If the terrain features are large, the initial window size can be set larger to speed up the traversal). If the kurtosis coefficient is less than or equal to 0 and the window size is greater than 3×3, then remove the one pixel width expanded along the window outline in the previous step, and assign the pixel value of C within the window to i (the minimum window size after reduction is 3×3; if the window size is further reduced, the classification result will lose its meaning, for example, classifying a single pixel into one category). The window size M×M is increased or decreased by 1, such as a 5×5 window. If the kurtosis coefficient is small, add 1 to make it 6×6, subtract 1 to make it 4×4); if the kurtosis coefficient is less than or equal to 0 and the window size is equal to 3×3, then directly assign the pixel value of C in the window to i; if the kurtosis coefficient is greater than 0, then expand the window by one pixel along the window outline, and then determine whether the window contains pixels with C not equal to 0. If not, continue to calculate the kurtosis coefficient until the kurtosis coefficient is less than or equal to 0, then remove the one pixel width expanded along the window outline in the previous step, and assign the pixel value of C in the window to i; if the window contains pixels with C not equal to 0, then remove the G pixel at the position of the pixel with C not equal to 0 in the window, and then continue to calculate the kurtosis coefficient of the G pixel in the window. Then the window moves to the untraversed image area and continues to make judgments until the entire image is traversed. The specific process is as follows:

[0044] (4-1) Enter the system, load the input file: high-resolution image, and then proceed to 4-2);

[0045] (4-2) Calculate the average value of the corresponding pixels in each band of the high-resolution remote sensing image to obtain the high-resolution remote sensing of a single band, denoted as G, and then proceed to 4-3);

[0046] (4-3) Create a single-channel image with the same number of rows and columns as the input high-resolution remote sensing image, initialize the pixel values ​​to 0, and denote the image as C. Then proceed to 4-4.

[0047] (4-4) The window size is initialized to M×M (because each new judgment is made by gradually increasing the size of the small window until it contains all pixels of that class), and then proceed to 4-5).

[0048] (4-5) Calculate the kurtosis coefficient of the pixels of G within the window, and then proceed to 4-6);

[0049] (4-6) Determine if the kurtosis coefficient is greater than 0. If it is greater than 0, proceed to 5-1); if it is less than or equal to 0, proceed to 5-4.

[0050] (5) Window size auto-adjustment

[0051] If the kurtosis coefficient within the window is greater than 0, then the feature belongs to a class. However, it is impossible to determine whether the current window covers all pixels of that class of feature. Therefore, the window is expanded by one pixel along the window outline, and the window cannot contain any non-zero pixels in C. Then, the kurtosis coefficient is recalculated until the kurtosis coefficient of all pixel values ​​contained in this window is less than 0. The one-pixel expansion along the window outline in the previous step is removed, and the 0 pixels in C within the window are assigned the value i.

[0052] (5-1) Expand the window by one pixel along the window outline, then proceed to 5-2);

[0053] (5-2) Determine if the window contains any non-zero pixels in C. If it does not contain any non-zero pixels in C, proceed to 4-5); if it does contain any non-zero pixels in C, proceed to 5-3.

[0054] (5-3) Remove the G pixel corresponding to the pixel in the window where C is not 0, and then proceed to 4-5);

[0055] (5-4) Determine if the window size M×M is greater than 3×3. If it is greater than 3×3, proceed to 5-5); if it is less than or equal to 3×3, proceed to 5-6.

[0056] (5-5) Remove the one-pixel width that was expanded along the window outline in the previous step, assign the value i to the pixel point C inside the window, and then proceed to 5-7);

[0057] (5-6) Assign the value i to the 0 pixel point of C in the window, and then proceed to 5-7);

[0058] (5-7)i = i + 1 (If the current i is 5, it means that 5 categories have been divided, so the next category should be the 5+1, that is, the 6th category), and then proceed to 5-8);

[0059] (5-8) Determine if the image has been traversed. If it has been traversed, proceed to 5-9); if it has not been traversed, proceed to 5-10.

[0060] (5-9) The remaining 0 pixels in C are assigned the value i, and then proceed to 6-1);

[0061] (5-10) Traverse the untraversed areas, that is, the image areas that have been classified before do not need to be classified again. Only classify the remaining image areas, and then proceed to 4-4).

[0062] (6) Output classification results

[0063] (6-1) Output the classification result C, which is raster data. Each class of pixel value will be assigned a number, with the first class being 1, the second class being 2, and so on up to n.

[0064] The purpose of this invention is to solve the problems of traditional supervised classification, which relies on a large amount of prior knowledge and requires manpower and resources to produce labels, as well as the problem of fragmented classification results in traditional sub-supervised classification.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A high-resolution image unsupervised classification method, characterized in that: Includes the following steps: S01: Input: The input file is a high-resolution remote sensing image, and the mean of the sum of the bands is calculated to obtain a single-band high-resolution remote sensing image; S02: Output: The output file is a classification result map with the same number of rows and columns as the high-resolution remote sensing image; the number of categories does not need to be determined. S03: Create a classification result image: Create a single-channel image with the same number of rows and columns as the input high-resolution remote sensing image, with an initial pixel value of 0, denoted as C, and record the classification result; S04: Consistency judgment of objects in the window: Read the input single-band high-resolution remote sensing image G, traverse the single-band high-resolution remote sensing image G and the classification result image C, scan the entire image, calculate the kurtosis coefficient of the pixel values ​​of G in the window, and then move the window to the image area that has not been traversed, and continue to make judgments until the entire image has been traversed. S05: Output classification results; In step S04, the consistency judgment of the features within the window is as follows: Read the input single-band high-resolution remote sensing image G, traverse the high-resolution remote sensing image G and the classification result image C, that is, start from the upper left corner of the image, from left to right, from top to bottom, use the window, i.e. the calculated area, to scan the entire image, and calculate the kurtosis coefficient of the pixel values ​​of G within the window. The first window used to calculate the kurtosis coefficient is the initialization window, which has a size of M×M, where M is the width and height of the window; Determine if the kurtosis coefficient is greater than 0. If it is greater than 0, expand the window by one pixel along the window outline. If it is less than 0, discard the one pixel expansion along the window outline in the previous step. If the kurtosis coefficient is less than or equal to 0 and the window size is greater than 3×3, then remove the one-pixel width that was expanded along the window outline in the previous step, assign the pixel value of C in the window to i, i = i + 1, and determine whether the image has been traversed. If it has been traversed, output the classification result C. If it has not been traversed, return to the window where the kurtosis coefficient was calculated for the first time and continue to calculate the kurtosis coefficient.

2. The unsupervised classification method for high-resolution images according to claim 1, characterized in that, If the window contains pixels with C not equal to 0, then remove the G pixels at the positions corresponding to the pixels with C not equal to 0 in the window, and then continue to calculate the kurtosis coefficient of the G pixels in the window.

Images