A spatio-temporal fusion method and device for remote sensing images based on adaptive texture constraints
The spatiotemporal fusion method of remote sensing images with adaptive texture constraints solves the problems of similar pixel identification and window search in complex surface environments of the ESTARFM algorithm, improves the spatial detail fidelity and computational efficiency of the fused images, and is suitable for a variety of remote sensing application scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INSPUR OPTOELECTRONICS SATELLITE TECHNOLOGY (SHANDONG) CO LTD
- Filing Date
- 2026-04-24
- Publication Date
- 2026-07-21
AI Technical Summary
The existing ESTARFM algorithm suffers from limitations in handling complex surface environments, including a single similar pixel identification criterion, a lack of adaptability in window search, insufficient ability to preserve spatial structural details, and limitations in linear assumptions. This results in spectral overflow, ground feature confusion, and low computational efficiency in fused images at texture edges, making it difficult to meet the requirements for accurate sub-pixel-level inversion.
An adaptive texture constraint is introduced, and contrast texture features are extracted through the gray-level co-occurrence matrix. An adaptive search window is constructed, and similar pixels are screened by combining spectral and contrast texture features. Dynamic weights are calculated, and nonlinear transformation coefficients are optimized through texture residuals. Fusion prediction is then performed by combining time weights.
It significantly improves the accuracy of similar pixel screening in complex and heterogeneous regions, enhances the ability of fusion results to preserve the spatial features of fragmented landmass edges and linear feature outlines, and generates high spatiotemporal resolution remote sensing products with low noise and clear texture details, which are suitable for monitoring urban dynamic expansion, precision agriculture assessment, and analysis of complex landscape and geomorphological evolution.
Smart Images
Figure CN122434751A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method and device for spatiotemporal fusion of remote sensing images based on adaptive texture constraints, belonging to the field of remote sensing image processing technology. Background Technology
[0002] With the rapid development of remote sensing technology, acquiring high spatial resolution (HSR) and high temporal resolution (HTR) remote sensing data is of great significance for vegetation phenology monitoring, environmental change analysis, and land surface abrupt change detection. However, limited by the physical imaging mechanism of satellite sensors, a single satellite payload often cannot simultaneously achieve high spatial resolution and high revisit frequency. For example, Landsat satellite has high spatial resolution but a long revisit period, while MODIS data has a high revisit frequency but blurred spatial details.
[0003] To address this contradiction, spatiotemporal fusion technology has emerged. Among numerous algorithms, the Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM) is widely used due to its ability to handle complex and heterogeneous regions and to make predictions using two pairs of high- and low-resolution images at known times. By introducing a conversion coefficient and a similar pixel weighting mechanism, ESTARFM demonstrates superior accuracy compared to the traditional STARFM algorithm in regions with strong surface heterogeneity.
[0004] However, in practical applications, the existing ESTARFM algorithm still has the following significant drawbacks: (1) The similar pixel identification criteria are singular: Traditional ESTARFM mainly relies on spectral distance and spatial distance to screen similar pixels. However, in areas with complex surface cover, there is often a phenomenon of "same spectrum but different land cover" (i.e., similar spectrum but different land cover types, such as woodland and green space). It is difficult to accurately distinguish them by spectral features alone, resulting in spectral overflow or land cover confusion at the texture edge of the fused image; (2) Lack of adaptability of search window: Existing algorithms usually use a fixed-size sliding window to search for similar pixels. In homogeneous areas such as plains, a fixed large window will lead to computational redundancy; while in cities or mountainous areas with high fragmentation, a fixed small window may not be able to search for enough similar pixels, resulting in poor model robustness and low computational efficiency; (3) Insufficient ability to preserve spatial structural details: ESTARFM does not fully consider the inherent texture features and structural information of high spatial resolution images when calculating weights. When processing ground feature boundaries and small patches, the generated predicted images often have blurred edges and lost details, making it difficult to meet the requirements of accurate sub-pixel level inversion; (4) Limitations of the linear assumption: The core of the algorithm is based on the assumption that high and low resolution pixels change linearly over time. When faced with periods of intense vegetation growth or sudden changes in land cover (such as fires or urban construction), the prediction accuracy will drop significantly.
[0005] Therefore, how to optimize ESTARFM's similar pixel extraction, window search, and weight allocation mechanisms by introducing richer feature constraints for complex surface environments, thereby improving the spatial detail fidelity and computational efficiency of fused images, has become a pressing technical challenge in the field of spatiotemporal fusion. Summary of the Invention
[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a remote sensing image spatiotemporal fusion method based on adaptive texture constraints. This method not only effectively improves the physical accuracy of similar pixel screening in complex and heterogeneous regions, but also significantly enhances the ability of the fusion results to preserve the spatial features of fragmented landmass edges and linear feature outlines.
[0007] The technical solution adopted in this invention is as follows: The spatiotemporal fusion method for remote sensing images based on adaptive texture constraints includes the following steps: S1. Obtain high spatial resolution image datasets and low spatial resolution image datasets for the area to be studied, and preprocess them; S2. Use the gray-level co-occurrence matrix to extract high spatial resolution images for the two reference times before and after the prediction date. Contrast texture feature maps, utilizing low spatial resolution images in The information entropy E is calculated within a local region centered on the Earth to characterize surface heterogeneity. S3. Construct an adaptive search window based on local heterogeneity; S4. Combine spectral and contrast texture features to filter similar pixels. Within the search window, filter out candidate pixels that are closest to the attributes of the pixel to be predicted. S5. Calculate the dynamic weights of similar pixels to the center pixel using structural similarity, spectral difference values, and spatial distance coupling; S6. Optimize the calculation of nonlinear transformation coefficients based on texture residuals; S7. Utilize the calculated dynamic weights and nonlinear conversion coefficients to perform reflectivity prediction based on a single reference time. S8. Calculate the time weight based on the overall spectral difference between the low spatial resolution image and the reference time. Use the high spatial resolution images before and after the prediction date as reference images to calculate the prediction values respectively. Then, perform a linear weighted sum of the two prediction values to obtain the high spatial resolution image of the final prediction time.
[0008] In the above method, the preprocessing in step S1 includes mask removal and denoising, spatial reference alignment and coordinate system transformation, and resampling.
[0009] The formula for calculating the contrast texture feature map in step S2 is as follows: , in Representing coordinates Texture feature values at the location; Image grayscale level; It represents the probability of gray-level pairs appearing in the gray-level co-occurrence matrix, while the contrast texture feature map is used for secondary screening of similar pixels.
[0010] The formula for calculating information entropy E in step S2 is as follows: , in For the measurement of pixel heterogeneity; It is the first one within this local window The probability of grayscale values appearing.
[0011] The calculation formula for the search window in step S3 is as follows: , in Representing coordinates The adaptive search window radius corresponding to the pixel; This represents the preset minimum search radius. This represents the adaptive adjustment coefficient, used to control the step size of window growth; This represents the maximum and minimum values of the local information entropy of the entire image.
[0012] In step S4, similar pixel filtering involves selecting candidate pixels within a defined window W that are closest in attributes to the pixel to be predicted. The calculation formula is as follows: , in Indicates the relationship between the center cell to be predicted and the first... The joint distance of the candidate pixels; These represent the reflectance of the center pixel and the candidate pixel in the b-th band of the reference image, respectively. These represent the contrast texture feature vectors of the center pixel and the candidate pixel, respectively. This represents the weighting coefficient, balancing the contributions of spectral similarity and texture similarity, and is selected through screening. The smallest n pixels are considered as similar pixels.
[0013] The method for calculating the dynamic weights in step S5 is as follows: (1) Calculate structural similarity: , Where μ represents the mean of the local window of the central pixel or the local window of similar pixels, and σ represents the standard deviation of the local window of the central pixel or the local window of similar pixels; Let be the covariance of the two local windows; It is a constant; (2) Calculate spectral differences: Spectral difference measures how similar pixels within the search window are close to the target pixel in terms of "reflectance value," and the calculation formula is shown below: , in Indicates the target center pixel and the first The combined spectral difference value between similar pixels; The reference time phase is represented by the formula, which calculates the sum of the spectral differences between the two times; B represents the total number of bands in the image. Indicates the reference time Below, the high spatial resolution image is located at the center of the target. The reflectance value of the b-th band at that location; Indicates the reference time Below, the i-th similar pixel is at position The first Band reflectivity value; (3) Calculate spatial distance: , in This represents the relative spatial distance between the center pixel and the i-th similar pixel. Represents the coordinates of the center pixel of the target to be predicted; Represents the coordinates of the i-th similar pixel within the search window; A represents the distance adjustment parameter; (4) Coupled calculation of dynamic weights, the formula is as follows: , in For the first Weights of similar pixels; Spatial distance; This is due to spectral differences.
[0014] In step S6, the conversion coefficient is dynamically amplified or reduced by introducing texture variation as a correction factor, as shown in the following formula: , , in This represents the initial transformation coefficients obtained by regressing high-resolution images at the reference time. This represents the corrected nonlinear conversion coefficients; Indicates the amount of texture change at the reference time; This represents the residual compensation factor.
[0015] The reflectivity prediction process based on a single reference time in step S7 is as follows: , in Based on Predicted in time Time-dependent reflectivity; This represents the difference between the predicted and baseline times for the low spatial resolution image.
[0016] The formula for calculating the time weight in step S8 is as follows: , In the formula t k k=m,n are the base times before and after the predicted date, t p Indicates the predicted date, x i y j For the position of a pixel, A k These are the calculated time weights; After obtaining the time weights, the predicted values generated at the two baseline times are linearly weighted and summed to obtain the final prediction result: , in For the final fusion result, Time weighting.
[0017] An electronic device for spatiotemporal fusion of remote sensing images based on adaptive texture constraints includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the spatiotemporal fusion method for remote sensing images based on adaptive texture constraints as described above.
[0018] The beneficial effects of this invention are: This invention introduces local texture feature extraction, an adaptive window driven by surface heterogeneity, and a weighted model that couples structural similarity. This not only effectively improves the physical accuracy of similar pixel selection in complex heterogeneous regions and significantly enhances the ability of the fusion results to preserve the spatial features of fragmented landmass edges and linear feature outlines, but also generates high spatiotemporal resolution remote sensing products with lower noise, clearer texture details, and more realistic spatial reconstruction. It is applicable to various scenarios such as urban dynamic expansion monitoring, refined agricultural assessment, complex landscape and geomorphological evolution analysis, and high-fragmentation vegetation cover surveys. Attached Figure Description
[0019] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0020] The present invention will be further described below with reference to specific embodiments.
[0021] Example 1: A spatiotemporal fusion method for remote sensing images based on adaptive texture constraints, comprising the following steps: S1. Obtain high spatial resolution image datasets and low spatial resolution image datasets for the area to be studied, and preprocess them.
[0022] Using the Google Earth Engine (GEE) distributed computing platform, a spatiotemporal matching image sequence of the target area was acquired and constructed, providing a highly consistent data foundation for subsequent texture feature extraction. Specific steps include: (1) Definition and selection of heterogeneous remote sensing datasets: Using the GEE API interface, surface reflectance datasets that have undergone atmospheric and orthorectification corrections were called. The high spatial resolution dataset selected was the COPERNICUS / S2_SR_HARMONIZED dataset (Level-2A product) with a spatial resolution of 10 meters; the low spatial resolution dataset selected was the MODIS / 061 / MOD09GQ dataset (250-meter resolution). The vector boundary (ROI) of the study area was set, and the prediction time was selected. MODIS images, and distance The two most recent reference times Paired Sentinel-2 and MODIS images.
[0023] (2) Demasking and denoising based on QA band: Pixel-level denoising is performed using the quality assessment band (QA Band) built into the sensor to ensure that the fusion model is not affected by clouds, shadows and aerosols.
[0024] (3) Spatial reference alignment and coordinate system transformation: In order to eliminate the geometric offset caused by different satellite orbits, a unified spatial projection transformation is performed in the GEE cloud to reproject the MODIS image to the UTM (Universal Transverse Mercator) projection coordinate system consistent with Sentinel-2, and the same WGS84 geodetic coordinate reference is used.
[0025] (4) Resampling: Due to the significant difference in spatial scale between Sentinel-2 (10m) and MODIS (250m), the low-resolution image needs to be upsampled to a high-resolution grid. A bicubic convolution interpolation algorithm is used to reconstruct the spatial resolution of the MODIS image to 10m. The formula is as follows: , in The MODIS reflectance at a grid depth of 10m after resampling; These are the original low-resolution pixel values; The weighting function is a cubic convolution. Compared to bilinear interpolation, this method better preserves surface texture edges, which is beneficial for subsequent matching with structural features of high-resolution images.
[0026] S2. Use the gray-level co-occurrence matrix to extract high spatial resolution images for the two reference times before and after the prediction date. Contrast texture feature maps, utilizing low spatial resolution image M in Calculate the information entropy E within a local region centered on the Earth to characterize surface heterogeneity: Input two pairs of reference times High-resolution (H) and low-resolution (M) imagery, and predicted time. Low-resolution images. Before extracting features, it is necessary to ensure that data from different sensors and at different times are within the same quantization range to eliminate the interference of system noise on texture calculation.
[0027] (1) Linear stretching and quantization: The preprocessed 10-meter resolution Sentinel-2 reflectivity image (0-1 floating-point type) is linearly mapped to the integer range of (0-255). This step is a necessary prerequisite for constructing the Gray-Level Co-occurrence Matrix (GLCM). The formula is as follows: , in This represents the original surface reflectance. This represents the quantized grayscale level.
[0028] (2) Spectral decorrelation: Principal component analysis (PCA) is performed on the multi-band image to extract the first principal component (PC1) as the reference band for texture extraction, so as to preserve spatial structure information to the maximum extent and reduce redundancy.
[0029] (3) Texture feature calculation: Reflects the image sharpness. The more obvious the boundary, the higher the contrast. High-resolution variable image at the reference time is extracted using the gray-level co-occurrence matrix (GLCM). The contrast texture feature map. The formula is as follows: , in Representing coordinates Texture feature values at the location; Image grayscale level; This represents the probability of gray-level pairs appearing in the gray-level co-occurrence matrix. A texture feature map that perfectly matches the Sentinel-2 spatial location is generated for subsequent secondary selection of similar pixels.
[0030] (4) Calculation of local information entropy: Using the low-variance image M in the local information entropy calculation. The information entropy E is calculated within a local region centered on the Earth's surface, used to characterize surface heterogeneity. The formula is as follows: , in For the measurement of pixel heterogeneity; It is the first one within this local window The probability of grayscale values appearing.
[0031] S3. Construct an adaptive search window based on local heterogeneity: Instead of using a fixed-size window, the search range of similar pixels for each pixel is dynamically determined based on the degree of surface fragmentation. This step ensures that more samples are obtained by expanding the window in fragmented areas and reducing the window in uniform areas to eliminate redundant calculations. The search window calculation formula is as follows: , in Representing coordinates The adaptive search window radius corresponding to the pixel; This indicates the preset minimum search radius (e.g., 25 pixels). This represents the adaptive adjustment coefficient, used to control the step size of window growth; This represents the maximum and minimum values of the local information entropy of the entire image.
[0032] S4. Combine spectral and contrast texture features to perform similar pixel filtering. Within the search window, select candidate pixels whose attributes are closest to those of the pixel to be predicted. Within the window W defined in the above steps, candidate pixels that are closest in attributes to the pixel to be predicted are selected. The formula is as follows: , in Indicates the relationship between the center cell to be predicted and the first... The joint distance of the candidate pixels; This represents the reflectance of the center pixel and candidate pixels in the b-band of the reference image; The contrast texture feature vector representing the center pixel and the candidate pixels; This represents the weighting coefficient, balancing the contributions of spectral similarity and texture similarity, and is selected through screening. The smallest n pixels are considered as similar pixels.
[0033] S5. Calculate the dynamic weights of similar pixels to the center pixel using structural similarity, spectral difference values, and spatial distance coupling: Introducing dynamic weight calculation of structural similarity (SSIM): To ensure the edge sharpness of the fused image, the structural similarity index from computer vision is introduced when calculating the weights.
[0034] (1) Structural similarity calculation: , Where μ represents the mean of the local window of the central pixel or the local window of similar pixels, and σ represents the standard deviation of the local window of the central pixel or the local window of similar pixels; Let be the covariance of the two local windows; It is a constant; (2) Calculation of spectral differences: Spectral difference measures how similar pixels within the search window are close to the target pixel in terms of "reflectance value". The smaller the value, the more consistent the similar pixel is with the target pixel in terms of spectral characteristics. The calculation formula is as follows: , in Indicates the target center pixel and the first The combined spectral difference value between similar pixels; To indicate the reference time phase, there are usually two reference times in ESTARFM ( and This formula calculates the sum of spectral differences between two moments; B represents the total number of bands in the image (such as red band, near-infrared band). Indicates the reference time Below, the high spatial resolution image is located at the center of the target. The reflectance value of the b-th band at that location; Indicates the reference time Below, the i-th similar pixel is at position The reflectance value of the b-th band at that location.
[0035] (3) Spatial distance calculation: Pixels that are spatially closer are more likely to share similar terrain attributes and trends with the target pixel. The formula for calculating spatial distance is shown below: , in This represents the relative spatial distance between the center pixel and the i-th similar pixel. Represents the coordinates of the center pixel of the target to be predicted; This represents the coordinates of the i-th similar pixel within the search window; A represents the distance adjustment parameter (usually half the radius of the search window), used to normalize the physical distance and balance the sensitivity of the distance to the weights. The "1+" before the formula serves as a smoothing factor to prevent the error of dividing by zero when the similar pixel is the center pixel itself (distance is 0), and also to avoid the weights being infinitely amplified due to excessively small physical distances.
[0036] The two traditional factors mentioned above are coupled with structural similarity (SSIM), and the weight calculation formula is as follows:
[0037] in For the first Weights of similar pixels; Spatial distance; This is due to spectral differences. The more similar a pixel is to the central pixel texture structure (SSIM), the greater its contribution to the prediction result.
[0038] S6. Optimization calculation of nonlinear transformation coefficients based on texture residuals: The transformation coefficient V is obtained by performing linear regression on the nth similar pixel selected within the search window. In ESTARFM, to improve robustness, two reference times are used ( and The transformation coefficient V is calculated by using similar pixel sample pairs as the solution. The formula for calculating the transformation coefficient V is as follows: , in The initial transformation coefficient corresponding to the center pixel (reflecting the linear scaling relationship between high- and low-resolution images); n represents the total number of similar pixels selected within the search window; j represents the index of the similar pixels (from 1 to n). These represent the j-th similar pixel at the reference time. and High-resolution image reflectance values; These represent the j-th similar pixel at the reference time. and Low-resolution image reflectance values (after resampling); This represents the arithmetic mean of the reflectance of all similar pixels on the high-resolution image at the corresponding time. This represents the arithmetic mean of the reflectance of all similar pixels in the low-resolution image at the corresponding time.
[0039] The traditional V model is based on the linear assumption, which holds that between two time points, the land cover only changes in intensity (e.g., crop height), without any class jumps or structural abrupt changes. Therefore, by introducing texture variation as a correction factor, the conversion coefficient can be dynamically amplified or reduced, thus forcing the model to pay attention to the nonlinear reflectance fluctuations caused by structural changes. The formula is as follows: , , in This represents the initial conversion coefficients obtained by regressing high-resolution images at the reference time (reflecting the linear ratio between high and low pixels). This represents the corrected nonlinear conversion coefficients; Indicates the amount of texture change at the reference time; This represents the residual compensation factor. Drastic changes in texture indicate a potential change in land cover type, which can be addressed through… The conversion coefficient is increased to enhance the ability to capture mutations.
[0040] In areas with simple land cover, the linear assumption is usually accurate, and little correction is needed. Smaller texture constraints are needed in areas with fragmented landscapes and dramatic changes in landforms. (Large). The formula is as follows: , in This represents the basic compensation strength, which is usually set based on experiments (generally between 0.05 and 0.2). This represents the local information entropy obtained in the above steps; This represents the maximum information entropy in the entire image.
[0041] S7. Reflectance prediction based on a single reference time using calculated dynamic weights and nonlinear conversion coefficients: (1) Reflectance prediction based on a single reference time: , in Based on Predicted in time Time-dependent reflectivity; This represents the difference between the low-variance image at the prediction time and the reference time.
[0042] S8. Calculate the time weight based on the overall spectral difference between the low spatial resolution image and the reference time. Use the high spatial resolution images before and after the prediction date as reference images to calculate the prediction values respectively. Perform a linear weighted sum of the two prediction values to obtain the high spatial resolution image at the final prediction time. In the ESTARFM algorithm, time weights are typically determined based on the overall spectral difference between the low-resolution image and the reference time. The smaller the difference, the higher the weight. The formula for calculating the time weight is shown below: , In the formula t k， k=m,n are the base times before and after the predicted date, respectively, and t p Indicates the predicted date, x i y j For the position of a pixel, A k It is the calculated time weight.
[0043] After obtaining the time weights, the predicted values generated at the two reference times are... Perform a linear weighted summation to obtain the final prediction result: , in The final fusion result; Time weights (usually determined by the prediction time) The distance from the reference time determines the time.
[0044] Example 2: An electronic device for spatiotemporal fusion of remote sensing images based on adaptive texture constraints, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the spatiotemporal fusion method for remote sensing images based on adaptive texture constraints as described in Example 1 above.
[0045] A computer-readable storage device having a computer program stored thereon, characterized in that, when executed by a processor, the program implements the steps of the remote sensing image spatiotemporal fusion method based on adaptive texture constraints as described in Example 1.
[0046] The above is a further description of the present invention in conjunction with specific embodiments, and the scope of protection of the present invention is not limited thereto.
Claims
1. A spatiotemporal fusion method for remote sensing images based on adaptive texture constraints, characterized by: The steps include the following: S1. Obtain high spatial resolution image datasets and low spatial resolution image datasets for the area to be studied, and preprocess them; S2. Use the gray-level co-occurrence matrix to extract high spatial resolution images for the two reference times before and after the prediction date. Contrast texture feature maps, utilizing low spatial resolution images in The information entropy E is calculated within a local region centered on the Earth to characterize surface heterogeneity. S3. Construct an adaptive search window based on local heterogeneity; S4. Combine spectral and contrast texture features to filter similar pixels. Within the search window, filter out candidate pixels that are closest to the attributes of the pixel to be predicted. S5. Calculate the dynamic weights of similar pixels to the center pixel using structural similarity, spectral difference values, and spatial distance coupling; S6. Optimize the calculation of nonlinear transformation coefficients based on texture residuals; S7. Utilize the calculated dynamic weights and nonlinear conversion coefficients to perform reflectivity prediction based on a single reference time. S8. Calculate the time weight based on the overall spectral difference between the low spatial resolution image and the reference time. Use the high spatial resolution images before and after the prediction date as reference images to calculate the prediction values respectively. Perform a linear weighted sum of the two prediction values to obtain the high spatial resolution image of the final prediction time.
2. The remote sensing image spatiotemporal fusion method based on adaptive texture constraints according to claim 1, characterized in that, The formula for calculating the contrast texture feature map in step S2 is as follows: ， in Representing coordinates Texture feature values at the location; Image grayscale level; It represents the probability of gray-level pairs appearing in the gray-level co-occurrence matrix, while the contrast texture feature map is used for secondary screening of similar pixels.
3. The remote sensing image spatiotemporal fusion method based on adaptive texture constraints according to claim 1, characterized in that, The calculation formula for the search window in step S3 is as follows: ， in Representing coordinates The adaptive search window radius corresponding to the pixel; This represents the preset minimum search radius. This represents the adaptive adjustment coefficient, used to control the step size of window growth; This represents the maximum and minimum values of the local information entropy of the entire image.
4. The remote sensing image spatiotemporal fusion method based on adaptive texture constraints according to claim 1, characterized in that, In step S4, similar pixel filtering involves selecting candidate pixels within a defined window W that are closest in attributes to the pixel to be predicted. The calculation formula is as follows: ， in Indicates the relationship between the center cell to be predicted and the first... The joint distance of the candidate pixels; These represent the reflectance of the center pixel and the candidate pixel in the b-th band of the reference image, respectively. These represent the contrast texture feature vectors of the center pixel and the candidate pixel, respectively. This represents the weighting coefficient, balancing the contributions of spectral similarity and texture similarity, and is selected through screening. The smallest n pixels are considered as similar pixels.
5. The remote sensing image spatiotemporal fusion method based on adaptive texture constraints according to claim 1, characterized in that, The method for calculating the dynamic weights in step S5 is as follows: (1) Calculate structural similarity: ， Where μ represents the mean of the local window of the central pixel or the local window of similar pixels, and σ represents the standard deviation of the local window of the central pixel or the local window of similar pixels; Let be the covariance of the two local windows; It is a constant; (2) Calculate spectral differences: Spectral difference measures how similar pixels within the search window are close to the target pixel in terms of "reflectance value". The calculation formula is shown below: ， in Indicates the target center pixel and the first The combined spectral difference value between similar pixels; The reference time phase is represented by the formula, which calculates the sum of the spectral differences between the two times; B represents the total number of bands in the image. Indicates the reference time Below, the high spatial resolution image is located at the center of the target. The reflectance value of the b-th band at that location; Indicates the reference time Below, the i-th similar pixel is at position The reflectance value of the b-th band at that location; (3) Calculate spatial distance: ， in This represents the relative spatial distance between the center pixel and the i-th similar pixel. Represents the coordinates of the center pixel of the target to be predicted; Represents the coordinates of the i-th similar pixel within the search window; A represents the distance adjustment parameter; (4) Coupled calculation of dynamic weights, the formula is as follows: ， in For the first Weights of similar pixels; Spatial distance; This is due to spectral differences.
6. The remote sensing image spatiotemporal fusion method based on adaptive texture constraints according to claim 1, characterized in that, In step S6, the conversion coefficient is dynamically amplified or reduced by introducing texture variation as a correction factor, as shown in the following formula: ， ， in This represents the initial transformation coefficients obtained by regressing high-resolution images at the reference time. This represents the corrected nonlinear conversion coefficients; Indicates the amount of texture change at the reference time; This represents the residual compensation factor.
7. The spatiotemporal fusion method for remote sensing images based on adaptive texture constraints according to claim 1, characterized in that, The reflectivity prediction process based on a single reference time in step S7 is as follows: ， in Based on Predicted in time Time-dependent reflectivity; This represents the difference between the predicted and baseline times for the low spatial resolution image.
8. The spatiotemporal fusion method for remote sensing images based on adaptive texture constraints according to claim 1, characterized in that, The formula for calculating the time weight in step S8 is as follows: ， In the formula t k k=m,n are the base times before and after the predicted date, t p Indicates the predicted date, x i y j For the position of a pixel, A k These are the calculated time weights; After obtaining the time weights, the predicted values generated at the two baseline times are linearly weighted and summed to obtain the final prediction result: ， in For the final fusion result, Time weighting.
9. An electronic device for spatiotemporal fusion of remote sensing images based on adaptive texture constraints, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the remote sensing image spatiotemporal fusion method based on adaptive texture constraints as described in any one of claims 1-8.
10. A computer-readable storage device having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the remote sensing image spatiotemporal fusion method based on adaptive texture constraints as described in any one of claims 1-8.