Remote sensing image color conversion method based on rough classification

Abstract

The invention discloses a remote sensing image color conversion method based on coarse classification, which relates to the field of remote sensing image processing and comprises the following steps: acquiring an original remote sensing image and a reference remote sensing image; performing rough classification processing on the original remote sensing image; performing communication and merging processing on the original rough classification result to obtain a merged classification result; extracting the texture features of each ground feature category in the merged classification result, and calculating the texture similarity of the ground feature categories between the original remote sensing image and the reference remote sensing image; based on the texture similarity, matching ground feature categories in the original remote sensing image and the reference remote sensing image; adjusting the number of pixels of the reference remote sensing image according to the pixel proportion of the ground object category; and performing color conversion based on the adjusted pixel proportion. According to the method, a technical route of combining coarse classification prior, texture matching and pixel proportion regulation and control is innovatively provided, and a new thought is provided for solving the problem of color distortion under the condition of unbalanced remote sensing image ground object categories.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing, and in particular relates to a method for color conversion of remote sensing images based on coarse classification. Background Technology

[0002] Remote sensing technology plays an important role in the extraction and application of ground feature information, and can provide support for fields such as agriculture, forestry, environmental monitoring and urban planning. For example, Li Xue et al. proposed a recurrent consistent generative adversarial network (ACGAN) that introduces a U-shaped attention mechanism. Through unsupervised learning, it realizes color conversion of multi-temporal remote sensing images and effectively improves the generalization ability of building semantic segmentation models on images of different temporal phases.

[0003] However, factors such as different sensors, atmospheric conditions, and lighting environments can lead to color variations in remote sensing images, thus affecting the accuracy of ground object identification and classification results. Xie Bin et al. achieved high-quality visual effects in emotion color transfer by combining cartoon texture decomposition, self-sparse fuzzy clustering, and adaptive brightness correction strategies, which better preserved the image's hierarchical structure and detail information. However, this method relies on manually selecting emotion color combinations from a color library, and the number of main colors, N, is fixed, failing to achieve fully adaptive transfer. While existing color conversion methods can alleviate this problem to some extent, they still suffer from shortcomings such as global shift, local distortion, and high computational complexity under conditions of complex categories and multi-source images.

[0004] Specifically, during global color conversion, it is often necessary to consider the overall color distribution between the original image and the reference image for subsequent color feature information matching. If the original image or the reference image contains different categories of color regions or the number of corresponding category pixels between the images is unbalanced, the color distribution matching and mapping between the original image and the reference image will be unsatisfactory. This will lead to the color conversion being unable to distinguish the statistical information of different ground features, and will ultimately confuse color regions. For example, if the reference image contains large areas of green space, but the original image mainly contains large areas of buildings, problems such as overall color shift, local ground feature color distortion, hue changes, and unnatural image blurring will occur during the color conversion process. Summary of the Invention

[0005] The purpose of this invention is to address the problem of inaccurate color conversion in remote sensing images due to the imbalance of corresponding category pixels between the original image and the reference image, and to propose a color conversion method for remote sensing images based on coarse classification.

[0006] This invention is achieved through the following technical solution: A method for color conversion of remote sensing images based on coarse classification includes the following steps: S1. Acquire raw remote sensing images and reference remote sensing images; S2. Perform coarse classification processing on the original remote sensing image to obtain the original coarse classification result; S3. Perform connectivity merging on the original coarse classification results to obtain the merged classification results; S4. Extract the texture features of each land cover category in the merged classification results, and calculate the texture similarity of land cover categories between the original remote sensing image and the reference remote sensing image. S5. Based on the texture similarity, match the land cover categories in the original remote sensing image and the reference remote sensing image; S6. Adjust the number of pixels in the reference remote sensing image according to the pixel ratio of land cover categories in the original remote sensing image and the reference remote sensing image. S7. Based on the adjusted pixel ratio, perform color conversion on the original remote sensing image to obtain the color-converted image.

[0007] A storage device that stores instructions and data for implementing a color conversion method for remote sensing images based on coarse classification.

[0008] A color conversion device for remote sensing images based on coarse classification includes: a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement a color conversion method for remote sensing images based on coarse classification.

[0009] The present invention has the following beneficial effects: This invention proposes a color conversion method for remote sensing images based on coarse classification to address the color distortion problem caused by uneven pixel distribution across land cover categories. Systematic qualitative and quantitative validation was performed on the LoveDA dataset and self-made high-resolution remote sensing images. The results are as follows: (1) The images obtained by color conversion using the method of the present invention are superior to traditional global methods in both visual effect and objective indicators. Qualitative experiments show that the method can effectively alleviate the color shift caused by the imbalance of the proportion of vegetation and buildings, and achieve natural color migration. In quantitative experiments, the FSIM and SSIM indices reached 0.951 and 0.822, respectively, which are 3.7% and 23.2% higher than the suboptimal method, fully verifying its significant advantages in maintaining image structure and perceptual features.

[0010] (2) The method of this invention introduces a pixel ratio constraint mechanism to explicitly consider the difference in the number of pixels of land cover categories during the color matching process. Analysis shows that the method achieves a better balance between ΔE (12.437) and SSIM (0.822), overcoming the problem of the contrast method excessively pursuing color consistency at the expense of image structure quality.

[0011] (3) Based on an in-depth analysis of existing color conversion methods, this invention innovatively proposes a technical route that combines coarse classification prior, texture matching and pixel ratio adjustment, providing a new approach to overcome the color distortion problem under the imbalance of land cover categories in remote sensing images, and also providing an application reference for color consistency processing of multi-source remote sensing images. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of the process of a remote sensing image color conversion method based on coarse classification according to the present invention; Figure 2 This is a schematic diagram of the RGB histogram of the experimental image; Figure 3 This is a schematic diagram illustrating the qualitative experimental analysis results of the method of the present invention; Figure 4 This is a schematic diagram illustrating the experimental results of the self-made dataset of this invention; Figure 5 This is a schematic diagram of the hardware device of the present invention. Detailed Implementation

[0014] 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. 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.

[0015] Example 1 Please see Figure 1 As shown, this invention is a method for color conversion of remote sensing images based on coarse classification, comprising the following steps: S1. Acquire raw remote sensing images and reference remote sensing images; S2. Perform coarse classification processing on the original remote sensing image to obtain the original coarse classification result; It should be noted that step S2 is as follows: S21. Calculate multiple thresholds based on the pixel distribution information of the original remote sensing image; In this invention, the original remote sensing image and the reference remote sensing image are input into the traditional Otsu algorithm to obtain multiple thresholds.

[0016] S22. For each threshold, calculate the ratio of pixel cumulative probability to overall cumulative probability, and iteratively adjust the threshold by moving the step size so that the ratio of pixel cumulative probability to overall cumulative probability for each threshold reaches a predetermined ratio, thus completing the threshold adjustment. It should be noted that in traditional Otsu's inter-class variance image segmentation, when the ratio of pixel cumulative probability to overall cumulative probability reaches a certain threshold, the rate of change of the gray-level histogram will increase significantly. This phenomenon usually indicates the potential optimal threshold region.

[0017] Based on this, the present invention proposes a threshold adjustment strategy that combines categories with smaller intra-class variance. By initially classifying land features in remote sensing images, reliable pixel statistical information is obtained, thereby providing accurate category ratio support for subsequent adaptive color conversion, as shown in the following equation (1): (1) Among them, for the first threshold in the multi-threshold set Threshold Regarding its upper and lower limits and The scope is adjusted accordingly. Indicates the threshold index. This indicates the number of thresholds in the multi-threshold set. Indicates in The average gray value within the interval, For a moving value used for estimation The size of the threshold. Based on the pixel accumulation ratio, within a specified range of multiple threshold sets, the thresholds are iteratively adjusted by moving the step size to ensure that the accumulation ratio of each threshold is... Approaching the predetermined rate This optimizes the segmentation threshold.

[0018] S23. After threshold adjustment, the difference between the actual category and the predicted category of each pixel is measured by the misclassification error, and the optimal threshold is determined from the adjusted thresholds based on the principle of minimum error.

[0019] It should be noted that, in order to accurately distinguish pixel categories in image segmentation, misclassification error (ME) is also used to measure the difference between the actual category and the predicted category of each pixel, so as to determine the optimal threshold misclassification error based on the principle of minimum error, as shown in the following formula (2): [Changtan 1] (2) For each threshold in the multi-threshold set In the corresponding positive class and negative classes The discussion will take place within the scope of the composition. Representing negative class The middle class was incorrectly classified as positive. These misclassified data points actually belong to the negative class, but are misclassified due to their eigenvalues. Greater than the threshold Therefore, it was incorrectly classified. Represents the positive class The middle class was incorrectly classified as negative. The data points. These misclassified points actually belong to the positive class, but are misclassified due to their eigenvalues. Less than or equal to the threshold Therefore, it was incorrectly classified. Indicates threshold The total number of all data points within the corresponding range.

[0020] By analyzing each threshold in the multi-threshold set Lower positive and negative data points To assess the impact of the threshold on classification accuracy.

[0021] S3. Perform connectivity merging on the original coarse classification results to obtain the merged classification results; It should be noted that because small fragments or noise may exist in the coarse classification results, which may interfere with the acquisition of image pixel statistics, scattered regions are merged by setting an area threshold.

[0022] Step S3 is as follows: S31. Statistically analyze the connectivity information of various pixel blocks and set area thresholds; S32. Traverse the pixel blocks of each type of land feature. If the area of ​​a pixel block is smaller than the area threshold, merge it with the nearest pixel block.

[0023] Specifically, by statistically obtaining the connectivity information of various pixel blocks, an area threshold is set, and pixel blocks of different types of land features are traversed and merged, as shown in the following formula (3): (3) For pixel blocks Each pixel in ,implement Pixel category replacement and merging operation, It is the pixel value at the corresponding position in the nearest neighbor pixel block.

[0024] S4. Extract the texture features of each land cover category in the merged classification results, and calculate the texture similarity of land cover categories between the original remote sensing image and the reference remote sensing image. It should be noted that, based on the obtained pixel block merging results, key texture features are used to perform quantitative analysis on the segmented regions in order to accurately characterize the structure and texture characteristics of the regions, thereby extracting the corresponding category of ground feature pixel blocks.

[0025] Specifically, the differences between texture features are calculated by calculating four texture feature values ​​(including contrast, homogeneity, energy, and correlation) for each land feature. The texture similarity is obtained by taking the square root of the sum of the squares of the differences in texture features, as shown in equation (4): (4) in, This represents the difference between the texture features corresponding to different land features. This represents the final similarity estimate. Indicates contrast, ranging from A contrast interval, or grayscale interval, represents the difference between the highest and lowest values ​​of adjacent pixel sets in an image, based on the spatial relationship of grayscale levels. This difference is positively correlated with the contrast value. Contrast also reflects the degree of local variation in an image; for example, smooth areas with little variation will produce lower contrast values, while low-contrast images typically display high density near the main diagonal of the matrix. Indicates homogeneity, ranging from The range not only reflects the consistency between pixels in an image, but also has a negative correlation with contrast in the gray-level co-occurrence matrix. This means that when contrast increases, i.e., when the gray-level differences in the image increase, homogeneity usually decreases; conversely, when contrast decreases, homogeneity usually increases. Representing energy value, it measures the uniformity of gray-level distribution and texture roughness in an image, ranging from... The energy value is close to 1 when the gray level distribution in the image is very uniform; conversely, if the gray level distribution in the image is very non-uniform, the energy value is close to 0. Indicates correlation, ranging from An interval measures the correlation between the gray values ​​of a pixel in an image and its neighboring pixels, reflecting the spatial distribution characteristics of gray level pairings in the image, especially their linear dependence.

[0026] S5. Based on the texture similarity, match the land cover categories in the original remote sensing image and the reference remote sensing image; It should be noted that the matching in step S5 is achieved by minimizing the difference in texture similarity between land cover categories in the original remote sensing image and the reference remote sensing image.

[0027] Specifically, based on the calculated texture values, the similarity between each land parcel category in the original image and the reference image is evaluated to obtain the best matching correspondence for each corresponding land cover category for subsequent operations, as shown in equation (5): (5) In order to find a reference image Each region in the original image The best matching region in the middle, through Describe the similarity evaluation function and minimize each reference image. Neutralize the original image The similarity difference between them is used to define a matching function. Used to store these matching results.

[0028] S6. Adjust the number of pixels in the reference remote sensing image according to the pixel ratio of land cover categories in the original remote sensing image and the reference remote sensing image. It should be noted that after obtaining the matching results for various land features, the pixel ratio of land feature categories is taken into consideration, and the mapping in image color conversion is adjusted according to the pixel ratio of land feature categories.

[0029] Step S6 is as follows: S61. Determine the land cover category with the fewest pixels in the reference remote sensing image; S62. Based on the minimum number of pixels, adjust the number of pixels for other land cover categories so that the adjusted number of pixels is proportional to the number of pixels for the corresponding land cover category in the original remote sensing image.

[0030] The relevant mathematical formulas are as follows: (6) (7) for Land features, the number of pixels for each land feature type are as follows: First, the category with the fewest pixels is determined using formula (6). This is to ensure that, when dealing with multiple categories, the number of pixels in each category after rounding is not less than the category with the fewest pixels in the reference image. Next, we use formula (7) for each category... Adjust the number of pixels involved in the color mapping process respectively.

[0031] S7. Based on the adjusted pixel ratio, perform color conversion on the original remote sensing image to obtain the color-converted image.

[0032] It should be noted that step S7 is as follows: S71. Transform the original remote sensing image to the color gamut of the reference remote sensing image using a transformation matrix; S72. Shift the transformed image back to the original center of the reference remote sensing image to obtain the color-converted image.

[0033] Specifically, after adjusting the pixels of the image involved in color gamut alignment, the original image is transformed by a transformation matrix and shifted back to the original center of the reference color gamut to obtain the result image, as shown in equation (8): (8) in, This represents the final image result of the color conversion. Represents the original image, through Perform transformation operations on the original image. This represents the original center of the color gamut of the reference image.

[0034] The present invention will introduce some specific implementation examples of the method below. Before introducing the examples, the dataset used in this application will be introduced first.

[0035] This invention mainly includes two research areas: the LoveDA dataset and a self-made dataset.

[0036] The LoveDA dataset covers the cities of Nanjing, Changzhou, and Wuhan in East and Central China, with a combined latitude and longitude range of 118°21'~114°17'E, 30°22'~32°37'N, and a total coverage area of ​​536.15 km². The study area exhibits a complex and diverse surface landscape, clearly divided into two typical domains—urban and rural—based on land cover characteristics and socioeconomic structure. Urban areas feature densely built, regularly laid-out structures and wide, continuous roads; rural areas have sparsely built, loosely shaped structures, narrow and winding roads, interspersed with ponds, ditches, and patches of farmland. Land cover types primarily include seven categories: background, buildings, roads, water bodies, bare land, forests, and farmland. This dataset contains 5987 high spatial resolution (HSR) images and 166,768 annotated objects, demonstrating significant diversity and representativeness. Covering real urban and rural areas, this dataset provides a more comprehensive reflection of the actual distribution characteristics of land cover categories. Its multi-scale characteristics allow objects of the same category to exhibit significant scale differences across different geographical landscapes, providing a more challenging test scenario for evaluating the performance of segmentation algorithms. Furthermore, the LoveDA dataset offers rich and complex background samples, including diverse high-resolution images and complex scenes, increasing intra-class variability and detail representation. Therefore, the LoveDA dataset not only serves as an important benchmark for evaluating image segmentation methods but also provides ideal data support for subsequent color conversion experiments due to the color style differences between urban and rural images.

[0037] The images in this self-made dataset are sourced from Google Earth. The study area is located in Wuhan, Hubei Province, China, with latitude and longitude ranging from 114°53′ to 114°23′E and 30°22′ to 30°46′N. The dataset contains 30 high spatial resolution remote sensing images, each with a size of 800 × 800 pixels. Urban areas are characterized by dense man-made buildings and a well-organized road network, while rural areas are characterized by extensive vegetation cover, scattered settlements, and ponds.

[0038] Example 2 Regarding the analysis of experimental results for color conversion methods, this invention takes both objective and subjective perspectives. The objective analysis includes qualitative and quantitative analysis, analyzing the experimental results of color conversion effects on remote sensing images. Original and reference images are selected for color conversion, and the experimental results are compared with those of Reinhard, TMR (Transportation Map Regularization), and L2_RCT (L2 Robust Color Transfer). Specific experimental results are analyzed below: Since this paper aims to address issues such as overall color shift and land cover category imbalance caused by color conversion, and the essence of land cover category imbalance is the difference in color information distribution in the image, there are obvious differences in color features and pixel distribution between the original image and the reference image selected in this experiment. The RGB histogram of the experimental image is shown below. Figure 2 As shown: The differences described above can be observed through the RGB distribution histogram. As shown in the image above, the reference image and its corresponding RGB distribution histogram reveal that the orange-red building area is relatively small, while the green area is large. Therefore, the number of pixels in the R channel is generally low across all pixel levels. Furthermore, the image above clearly shows that the original image's histogram generally covers a larger area and a wider pixel range, while the reference image covers a narrower color region. This can lead to inaccurate mapping of the color distribution information from the original image to the color space of the reference image, easily resulting in color shifts and distortions of ground features. This places higher demands on the color conversion method presented in this paper. The following section provides an analysis of the relevant color conversion experiments, such as... Figure 3 As shown: The results show that adjusting the number of pixels of different types of features in the reference image to match the original image effectively avoids color distortion caused by feature imbalance and achieves overall color style transfer. However, some problems exist: First, sacrificing the number of pixels of features involved in color conversion in the reference image balances pixel loss and coloring effect, but this also makes the coloring effect of features during conversion closer to a "conservative" average level. Second, since the color conversion method is based on global considerations and color perception has a certain degree of subjectivity, the results focus more on the overall color style of the image and lack an understanding of local image features.

[0039] Example 3 This invention provides a quantitative evaluation of color conversion-related experimental methods using indicators such as color deviation, structural similarity (SSIM), and feature similarity (FSIM). On one hand, it primarily assesses the accuracy and quality of color conversion by evaluating the color difference between the reference image and the resulting image, thereby quantifying the actual effect of the color conversion method. On the other hand, SSIM and FSIM mainly focus on evaluating the image quality of the color conversion result image, considering factors such as gradient information and brightness information to assess the structural similarity between the converted image and the original image. In summary, by comprehensively considering these two aspects of indicators, each method is objectively analyzed, as shown in Table 1. Table 1. Quantitative Evaluation Indicators for Different Color Conversion Methods

[0040] From a quantitative perspective, TMR performs best in color consistency (ΔE = 9.526), ​​but its structural and feature similarity are relatively low (SSIM = 0.667, FSIM = 0.917), indicating an overemphasis on color matching at the expense of structure and detail. In contrast, our proposed method achieves the highest values ​​for both FSIM (0.951) and SSIM (0.822), improving upon TMR by 0.034 and 0.155 respectively, significantly enhancing its ability to preserve structural and perceptual features. Although our proposed method's ΔE (12.437) is slightly higher than TMR's, it achieves a better overall balance between color reproduction and structural fidelity.

[0041] In summary, the TMR and Reinhard methods still suffer from an overemphasis on color transformation relationships in images, while neglecting the structure and quality of the images themselves. Although the L2_RCT method achieves a balance between color saturation and visual appeal, and focuses more on good visual effects, it does not strictly match the color information of the reference image. For example, the water body does not match the color of the reference image, so the overall performance is still poor.

[0042] Example 4 This invention discusses in detail the performance of the proposed adaptive color conversion method on other widely used datasets. To clearly assess the method's effectiveness in different application scenarios, a self-made Google remote sensing image dataset is still selected for experiments. This experimental analysis is based on the preliminary work on coarse classification of remote sensing images, and is used to evaluate the conversion effect of the color conversion method on different remote sensing images, such as... Figure 4 As shown.

[0043] When there is an imbalance between different land cover categories in the original image and the reference image, by adjusting the pixel ratio between different land cover categories participating in the global color conversion, the problems of unnaturalness and color distortion introduced by the color conversion are solved, and the color characteristics of the original image are maintained during the final conversion. This effectively adjusts the color matching process between the original image and the reference image.

[0044] Example 5 Please see Figure 5 , Figure 5 This is a schematic diagram of the hardware device in operation according to an embodiment of the present invention. The hardware device specifically includes: a remote sensing image color conversion device 401 based on coarse classification, a processor 402, and a storage device 403.

[0045] A remote sensing image color conversion device 401 based on coarse classification: The remote sensing image color conversion device 401 based on coarse classification implements the remote sensing image color conversion method based on coarse classification.

[0046] Processor 402: The processor 402 loads and executes the instructions and data in the storage device 403 to implement the remote sensing image color conversion method based on coarse classification.

[0047] Storage device 403: The storage device 403 stores instructions and data; the storage device 403 is used to implement the remote sensing image color conversion method based on coarse classification.

[0048] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0049] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for color conversion of remote sensing images based on coarse classification, characterized in that: Includes the following steps: S1. Acquire raw remote sensing images and reference remote sensing images; S2. Perform coarse classification processing on the original remote sensing image to obtain the original coarse classification result; S3. Perform connectivity merging on the original coarse classification results to obtain the merged classification results; S4. Extract the texture features of each land cover category in the merged classification results, and calculate the texture similarity of land cover categories between the original remote sensing image and the reference remote sensing image. S5. Based on the texture similarity, match the land cover categories in the original remote sensing image and the reference remote sensing image; S6. Adjust the number of pixels in the reference remote sensing image according to the pixel ratio of land cover categories in the original remote sensing image and the reference remote sensing image. S7. Based on the adjusted pixel ratio, perform color conversion on the original remote sensing image to obtain the color-converted image.

2. The method for color conversion of remote sensing images based on coarse classification according to claim 1, characterized in that, Step S2 is as follows: S21. Calculate multiple thresholds based on the pixel distribution information of the original remote sensing image; S22. For each threshold, calculate the ratio of pixel cumulative probability to overall cumulative probability, and iteratively adjust the threshold by moving the step size so that the ratio of pixel cumulative probability to overall cumulative probability for each threshold reaches a predetermined ratio, thus completing the threshold adjustment. S23. After threshold adjustment, the difference between the actual category and the predicted category of each pixel is measured by the misclassification error, and the optimal threshold is determined from the adjusted thresholds based on the principle of minimum error.

3. The method for color conversion of remote sensing images based on coarse classification according to claim 1, characterized in that, Step S3 is as follows: S31. Statistically analyze the connectivity information of various pixel blocks and set area thresholds; S32. Traverse the pixel blocks of each type of land feature. If the area of ​​a pixel block is smaller than the area threshold, merge it with the nearest pixel block.

4. The method for color conversion of remote sensing images based on coarse classification according to claim 1, characterized in that, In step S4, the texture features include contrast, homogeneity, energy, and correlation, and the texture similarity is obtained by taking the square root of the sum of squared differences in the texture features.

5. The method for color conversion of remote sensing images based on coarse classification according to claim 1, characterized in that, In step S5, matching is achieved by minimizing the difference in texture similarity between land cover categories in the original remote sensing image and the reference remote sensing image.

6. The remote sensing image color conversion method based on coarse classification as described in claim 1, characterized in that: Step S6 is as follows: S61. Determine the land cover category with the fewest pixels in the reference remote sensing image; S62. Based on the minimum number of pixels, adjust the number of pixels for other land cover categories so that the adjusted number of pixels is proportional to the number of pixels for the corresponding land cover category in the original remote sensing image.

7. The method for color conversion of remote sensing images based on coarse classification according to claim 1, characterized in that, Step S7 is as follows: S71. Transform the original remote sensing image to the color gamut of the reference remote sensing image using a transformation matrix; S72. Shift the transformed image back to the original center of the reference remote sensing image to obtain the color-converted image.

8. A storage device, characterized in that: The storage device stores instructions and data for implementing the remote sensing image color conversion method based on coarse classification as described in any one of claims 1 to 7.

9. A color conversion device for remote sensing images based on coarse classification, characterized in that: include: A processor and a storage device; the processor loads and executes instructions and data in the storage device to implement the remote sensing image color conversion method based on coarse classification as described in any one of claims 1 to 7.

Images