Multi-temporal remote sensing endmember extraction method based on multidimensional purity evaluation

Abstract

The present invention discloses a multi-temporal remote sensing end-member extraction method based on multi-dimensional purity evaluation, comprising: first, evaluating the temporal stability of the multi-temporal remote sensing image, as well as the spatial heterogeneity and texture characteristics of the single-temporal remote sensing image included in the multi-temporal remote sensing image, to obtain the temporal stability index, spatial heterogeneity index and texture characteristic index of the pixel. Then, introducing a weight coefficient adaptive adjustment mechanism to adapt the weight coefficients of the spatial heterogeneity index, texture characteristic index and temporal stability index under different ground object scenes. Finally, based on the weight coefficients of each index, the spatial heterogeneity index, texture characteristic index and temporal stability index are weightedly fused to obtain a comprehensive purity score for each pixel, and combined with the spatial uniformity constraint to automatically screen uniformly distributed high-purity end-member points, thereby achieving high-purity and high-representative automatic extraction of end-member points, significantly improving the accuracy of end-member extraction and scene adaptability.

Description

Technical Field

[0001] The present invention relates to the technical field of image analysis, and in particular to a multi-temporal remote sensing endmember extraction method based on multidimensional purity evaluation. Background Art

[0002] Remote sensing technology, as a core means of Earth observation, provides critical data support for natural resource surveys, ecological monitoring, urban planning, and other fields through the acquisition and analysis of spectral information. In remote sensing data processing, the construction of endmember libraries is fundamental to spectral unmixing, object identification, and classification, and their quality directly impacts the accuracy of subsequent applications. As the fundamental units of pure object spectra, endmembers must accurately reflect the differences in spectral characteristics of objects across time and space, but traditional construction methods face multiple technical bottlenecks.

[0003] Traditional endmember extraction methods primarily rely on spectral information, ignoring the spatial distribution characteristics (such as heterogeneity), texture structure, and temporal variation of ground objects. For example, in vegetated areas or complex urban scenes, relying solely on spectral curves makes it difficult to effectively distinguish pure endmembers from mixed pixels, resulting in an underrepresentational endmember library. Furthermore, ground object spectra are significantly affected by seasonal variations, growth cycles, and environmental conditions. Traditional methods lack quantitative analysis of the temporal stability of multi-temporal remote sensing imagery, making it difficult to capture the temporal evolution of ground object spectra. This leads to mismatching of endmember libraries in cross-temporal applications.

[0004] Due to the lack of an effective fusion model for spatial heterogeneity, texture features, and temporal stability, there is a lack of adaptive weight optimization strategies for different landform scenarios. For example, in farmland monitoring, temporal features are crucial for distinguishing crop types; while in urban remote sensing, spatial texture features have a greater impact on building material identification. Traditional methods are unable to dynamically adjust feature weights based on scene differences, resulting in limited endmember extraction accuracy. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this paper proposes a multi-temporal remote sensing endmember extraction method based on multidimensional purity evaluation, which can improve the accuracy and scene adaptability of endmember extraction. The specific technical solution is as follows:

[0006] A multi-temporal remote sensing endmember extraction method based on multidimensional purity evaluation is provided. In a first achievable method, the method includes:

[0007] Evaluate the temporal stability of the acquired multi-temporal remote sensing images, as well as the spatial heterogeneity and texture characteristics of each single-temporal remote sensing image in the multi-temporal remote sensing image, and obtain a temporal stability index, a spatial heterogeneity index, and a texture characteristic index corresponding to each pixel;

[0008] Identify the object categories in multi-temporal remote sensing images through the spatial heterogeneity index, texture feature index and temporal stability index of the pixels, and adjust the weight coefficients corresponding to the spatial heterogeneity index, texture feature index and temporal stability index according to the discrimination degree of different object categories;

[0009] According to the spatial heterogeneity index, texture feature index and temporal stability index and their corresponding weight coefficients, a comprehensive purity score for each pixel is generated by fusion and weighting, and the uniformly distributed high-purity end-member points are automatically screened in combination with the spatial uniformity constraint.

[0010] In combination with the first feasible method, in the second feasible method, the acquired multi-temporal remote sensing images are evaluated, including:

[0011] Each single-temporal remote sensing image is cropped using the obtained ROI vector, and the near-infrared band of the cropped single-temporal remote sensing image is normalized.

[0012] In combination with the first feasible method, in the third feasible method, the temporal stability of multi-temporal remote sensing images is evaluated, including:

[0013] Calculate the mean and standard deviation of the spectral reflectance of pixels in multi-temporal remote sensing images, and calculate the coefficient of variation of the pixels based on the mean and standard deviation;

[0014] Calculating the Euclidean distance between the timing curve of a single pixel and the timing curves of other pixels, and determining the timing similarity index corresponding to the pixel based on the Euclidean distance between the single pixel and other pixels;

[0015] The temporal stability index of the pixel is obtained by weighted fusion of the coefficient of variation score and the temporal similarity index obtained by the coefficient of variation.

[0016] In combination with the first implementable manner, in a fourth implementable manner, evaluating the spatial heterogeneity of the single-temporal remote sensing image includes:

[0017] Based on the local Moran's I index, the spatial autocorrelation between the pixel in the single-temporal remote sensing image and its neighboring pixels is calculated through a sliding window;

[0018] According to different ground object scales, a multi-scale Gaussian weighted smoothing method is used to perform multi-scale fusion on the spatial autocorrelations corresponding to the pixels.

[0019] The multi-scale fusion results of pixels are normalized to obtain the spatial heterogeneity index of pixels.

[0020] In combination with the fourth implementable manner, in a fifth implementable manner, mirror filling is used to process the boundary generated by the sliding window when processing a single-phase remote sensing image, and / or a Gaussian weight kernel is adaptively generated.

[0021] In combination with the first possible implementation, in a sixth possible implementation, evaluating the texture features of the single-temporal remote sensing image includes:

[0022] Dividing the single-phase remote sensing image into blocks to obtain a plurality of divided images;

[0023] The texture features of each block image are extracted in parallel using a gray level co-occurrence matrix method to obtain a texture feature index of each pixel in the single-temporal remote sensing image.

[0024] In combination with the first possible implementation, in a seventh possible implementation, evaluating the texture features of the single-temporal remote sensing image includes:

[0025] An edge detection algorithm is used to detect edge regions in the single-temporal remote sensing image, and penalty processing is performed on texture features of pixels located in the edge regions.

[0026] In combination with the first feasible method, in the eighth feasible method, the ground object categories in the multi-temporal remote sensing image are identified, including:

[0027] The spatial heterogeneity index, texture characteristic index and temporal stability index of all pixels are standardized;

[0028] Based on the standardized spatial heterogeneity index, texture feature index and temporal stability index, a K-means clustering algorithm is used to classify and identify all pixels to determine the category of the ground objects in the multi-temporal remote sensing image.

[0029] In combination with the first implementable manner, in a ninth implementable manner, adjusting the weight coefficients of the spatial heterogeneity index, the texture feature index, and the temporal stability index includes:

[0030] The standard deviations of the spatial heterogeneity index, texture feature index and temporal stability index in different land feature categories are respectively calculated as discrimination;

[0031] Normalizing the standard deviations of the spatial heterogeneity index, texture feature index, and temporal stability index in different land feature categories;

[0032] The weight coefficients corresponding to the spatial heterogeneity index, texture feature index and temporal stability index are determined according to the corresponding normalization results.

[0033] In combination with the first feasible method, in a tenth feasible method, screening for uniformly distributed high-purity end member points includes:

[0034] Determining an adaptive fractional threshold corresponding to each pixel according to the temporal stability index;

[0035] The adaptive fractional thresholds of all pixels are combined to generate corresponding stability masks, and the high-purity end-member points are screened through the stability masks.

[0036] Beneficial effects: The multi-temporal remote sensing endmember extraction method based on multidimensional purity evaluation of the present invention is adopted, by fusing the three-dimensional purity indicators of spatial heterogeneity, texture characteristics and temporal stability, and introducing a scene adaptive weight optimization mechanism, to achieve high-purity and high-representative automatic extraction of endmember points, which significantly improves the accuracy of endmember extraction and scene adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the specific embodiments of the present invention, the following briefly introduces the drawings required for use in the specific embodiments. In all the drawings, each element or part is not necessarily drawn according to the actual scale.

[0038] Figure 1 A flowchart of a multi-temporal remote sensing endmember extraction method based on multidimensional purity evaluation provided by one embodiment of the present invention;

[0039] Figure 2 It is the spatial heterogeneity distribution map corresponding to different single-temporal remote sensing images in the multi-temporal remote sensing images of a certain area;

[0040] Figure 3 It is the texture feature distribution map corresponding to different single-temporal remote sensing images in the multi-temporal remote sensing images of a certain area;

[0041] Figure 4 The distribution map of spatial heterogeneity indicators corresponding to multi-temporal remote sensing images of a certain area;

[0042] Figure 5 The distribution map of texture feature indicators corresponding to multi-temporal remote sensing images of a certain area;

[0043] Figure 6 The temporal stability index distribution map corresponding to the multi-temporal remote sensing images of a certain area;

[0044] Figure 7 A high-purity endmember point distribution map of a certain area extracted using the multi-temporal remote sensing endmember extraction method provided by the present invention;

[0045] Figure 8 The spatial heterogeneity distribution map corresponding to different single-temporal remote sensing images in the multi-temporal remote sensing image of another region;

[0046] Figure 9The texture feature distribution map corresponding to different single-temporal remote sensing images in the multi-temporal remote sensing image of another region;

[0047] Figure 10 The spatial heterogeneity index distribution map corresponding to the multi-temporal remote sensing image of another region;

[0048] Figure 11 A distribution map of texture feature indicators corresponding to multi-temporal remote sensing images of another region;

[0049] Figure 12 The temporal stability index distribution map corresponding to the multi-temporal remote sensing images of another region;

[0050] Figure 13 This is a high-purity endmember point distribution map of another region extracted using the multi-temporal remote sensing endmember extraction method provided by the present invention. DETAILED DESCRIPTION

[0051] The following embodiments of the technical solution of the present invention will be described in detail with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore only examples and are not intended to limit the scope of protection of the present invention.

[0052] like Figure 1 The flowchart of the multi-temporal remote sensing endmember extraction method based on multidimensional purity evaluation is shown, and the extraction method includes:

[0053] Step 1: Evaluate the temporal stability of the acquired multi-temporal remote sensing image, as well as the spatial heterogeneity and texture characteristics of each single-temporal remote sensing image in the multi-temporal remote sensing image, and obtain a temporal stability index, a spatial heterogeneity index, and a texture characteristic index corresponding to each pixel;

[0054] Step 2: Identify the object categories in the multi-temporal remote sensing imagery through the spatial heterogeneity index, texture feature index, and temporal stability index of each pixel, and adjust the weight coefficients corresponding to the spatial heterogeneity index, texture feature index, and temporal stability index according to the discrimination degree of different object categories;

[0055] Step 3: Based on the spatial heterogeneity index, texture feature index and temporal stability index and their corresponding weight coefficients, a weighted fusion is performed to generate a comprehensive purity score for each pixel, and uniformly distributed high-purity end-member points are automatically screened in combination with the spatial uniformity constraint.

[0056] Specifically, the temporal stability of acquired multi-temporal remote sensing imagery can be evaluated, yielding a temporal stability index for each pixel within the image. Simultaneously, the spatial heterogeneity and texture characteristics of each single-temporal remote sensing image within the multi-temporal remote sensing image can be evaluated, yielding a spatial heterogeneity index and a texture characteristic index for each pixel. Spatial heterogeneity, texture characteristic, and temporal stability indices each have independent evaluation principles, parameter systems, and engineering implementations. These three complementary approaches comprehensively reflect the purity and representativeness of the ground feature end-members.

[0057] Then, by introducing an adaptive weight coefficient adjustment mechanism, we can automatically adapt to different ground object scenarios, improving the versatility and scientificity of endmember extraction. Specifically, we can use existing classification and recognition algorithms to identify ground objects in multi-temporal remote sensing imagery based on the spatial heterogeneity index, texture feature index, and temporal stability index of each pixel, and determine the corresponding status category of each ground object. We then calculate the degree to which the spatial heterogeneity index, texture feature index, and temporal stability index distinguish different ground object categories, and use this to dynamically assign weight coefficients for these indicators.

[0058] Finally, based on the weight coefficients of the spatial heterogeneity index, texture feature index, and temporal stability index, the spatial heterogeneity index, texture feature index, and temporal stability index corresponding to each pixel can be weighted and fused to obtain a comprehensive purity score for each pixel. This achieves a three-dimensional comprehensive evaluation of pixel purity, improves the representativeness and discrimination of pixels, and further enables the automatic extraction of high-purity and highly representative endmember points, significantly improving the accuracy of endmember extraction and scene adaptability.

[0059] In this embodiment, optionally, evaluating the acquired multi-temporal remote sensing images includes:

[0060] Each single-temporal remote sensing image is cropped using the obtained ROI vector, and the near-infrared band of the cropped single-temporal remote sensing image is normalized.

[0061] Specifically, when acquiring multi-temporal remote sensing images, the corresponding ROI vector can be obtained. Based on the ROI vector, existing high-precision masking algorithms are used to crop each single-temporal image within the multi-temporal remote sensing image. The near-infrared band of the cropped single-temporal image is then normalized to eliminate differences in imaging conditions and improve multi-temporal comparability. After processing is complete, the results are automatically saved for subsequent retrieval.

[0062] In this embodiment, optionally, evaluating the temporal stability of multi-temporal remote sensing images includes:

[0063] Calculate the mean and standard deviation of the spectral reflectance of pixels in multi-temporal remote sensing images, and calculate the coefficient of variation of the pixels based on the mean and standard deviation;

[0064] Calculating the Euclidean distance between the timing curve of a single pixel and the timing curves of other pixels, and determining the timing similarity index corresponding to the pixel based on the Euclidean distance between the single pixel and other pixels;

[0065] The temporal stability index of the pixel is obtained by weighted fusion of the coefficient of variation score and the temporal similarity index obtained by the coefficient of variation.

[0066] Specifically, after all single-temporal remote sensing images are cropped and normalized, statistical analysis can be performed on all single-temporal remote sensing images in the multi-temporal remote sensing image to determine the mean and standard deviation of the spectral reflectance corresponding to each pixel, and the coefficient of variation of the pixel can be calculated based on the mean and standard deviation. Then, based on the spectral reflectance of the pixel in each single-temporal remote sensing image, a time series curve corresponding to each pixel can be generated, and the Euclidean distance between the time series curve corresponding to each pixel and the time series curve of other pixels can be calculated. The Euclidean distance is used as the time series similarity of the pixel. The specific calculation formula is as follows:

[0067] ;

[0068] ;

[0069] .

[0070] in, For pixels The corresponding pixel timing curve is: For other pixels The timing curve of . Represents the pixel All corresponding similarities Taking the average value is the temporal similarity of the pixels.

[0071] In this embodiment, the coefficient of variation of the pixel can be calculated based on the mean and standard deviation of the spectral reflectance of the pixel. The specific calculation formula is as follows:

[0072] ;

[0073] in, is the standard deviation, is the mean.

[0074] Converting the coefficient of variation into a coefficient of variation score can better reflect the stability of the time series and can be weighted combined with other stability indicators such as similarity scores. The specific conversion formula is as follows:

[0075] .

[0076] Finally, the temporal similarity and variation coefficient scores can be weighted and fused to obtain the temporal stability index of the pixel. The specific calculation formula is as follows:

[0077] .

[0078] In this embodiment, optionally, evaluating the spatial heterogeneity of the single-temporal remote sensing image includes:

[0079] Based on the local Moran's I index, the spatial autocorrelation between the pixel in the single-temporal remote sensing image and its neighboring pixels is calculated through a sliding window;

[0080] According to different ground object scales, a multi-scale Gaussian weighted smoothing method is used to perform multi-scale fusion on the spatial autocorrelations corresponding to the pixels.

[0081] The multi-scale fusion results of pixels are normalized to obtain the spatial heterogeneity index of pixels.

[0082] Specifically, first, the neighborhood pixels adjacent to the pixel to be evaluated can be determined through a sliding window, and the local Moran's I index between the pixel to be evaluated and the neighborhood pixels can be calculated. The spatial autocorrelation between the pixel and its neighborhood pixels can be evaluated by the local Moran's I index.

[0083] Then, for different feature scales, a multiscale Gaussian weighted smoothing method can be used to perform multi-scale fusion of the local Moran's I index corresponding to the target pixel to be evaluated, enhancing its adaptability to different spatial structures. Finally, the multi-scale fusion results for the target pixel can be normalized and the corresponding normalized results in all single-temporal remote sensing images can be superimposed and averaged to obtain the spatial heterogeneity index for the target pixel. This spatial heterogeneity index can effectively identify pure endmembers with complex spatial structures and concentrated distributions.

[0084] In this embodiment, optionally, mirror filling is used to process the boundary generated by the sliding window when processing a single-phase remote sensing image, and / or a Gaussian weight kernel is adaptively generated.

[0085] Specifically, the boundary is the portion of the window that exceeds the image range when the sliding window moves along the edge of the image. For this portion, the mirror filling method can be used to process it. Specifically, for pixels that exceed the image boundary, the mirror value of the pixel at the image boundary is used to fill it. The adaptive generation of the Gaussian weight kernel includes:

[0086] First, the kernel size is automatically determined based on the spatial scale of the target object. For example, a small kernel of 3×3 is used in urban areas, and a large kernel of 7×7 is used in farmland areas. The kernel size calculation formula is:

[0087] ;

[0088] in, is the standard deviation of the Gaussian distribution and can be dynamically adjusted according to the characteristics of the terrain.

[0089] Then, create a 2D coordinate grid and calculate the distance from each point to the center point , using the Gaussian function Calculate the weight value.

[0090] in, The value can be dynamically adjusted according to the local spatial structure characteristics. For areas with complex textures, a smaller sigma value (such as 0.5) is used, and for areas with smooth textures, a larger sigma value (such as 1.5) is used.

[0091] Finally, the weights are normalized to ensure that the sum of all weights is 1, and multiple weight kernels of different scales are generated for weighted averaging to obtain a comprehensive spatial heterogeneity index.

[0092] In this embodiment, optionally, evaluating the texture features of the single-temporal remote sensing image includes:

[0093] Dividing the single-phase remote sensing image into blocks to obtain a plurality of divided images;

[0094] The texture features of each block image are extracted in parallel using a gray level co-occurrence matrix method to obtain a texture feature index of each pixel in the single-temporal remote sensing image.

[0095] Specifically, a single-temporal remote sensing image can be divided into multiple image blocks. Then, the gray-level co-occurrence matrix method is used to extract texture features from each block in parallel. The texture features of each pixel across all single-temporal remote sensing images are then averaged to obtain a texture feature index for each pixel. This texture feature index effectively distinguishes texture-rich areas from average areas, improving the discrimination of endmember points. The use of block-by-block parallel processing can effectively improve the efficiency of endmember extraction.

[0096] When constructing the gray-level co-occurrence matrix, first, normalize the block image to the range of 0-255 and convert it to uint8 type. Then, based on the normalized block image, distance, angle and other parameters, call the skimage featuregraycomatrix function to generate the gray-level co-occurrence matrix glcm. The specific code is:

[0097] glcm=graycomatrix(chunk_uint8,

[0098] distances=distances,

[0099] angles=angles,

[0100] symmetric=True,

[0101] normed=True).

[0102] Among them, distances is the distance between pixels, angles is the angle between pixels, symmetric=True means generating the matrix, and normed=True means normalizing the matrix.

[0103] The constructed gray-level co-occurrence matrix is ​​a four-dimensional numpy array with a shape of (levels, levels, len(distances), len(angles)). Levels is the number of grayscale levels. len(distances) and len(angles) correspond to different distance and angle parameters, respectively. Each element represents the co-occurrence probability of the corresponding grayscale pair at a specified distance and angle. For example, if there is a small 4x4 image with grayscale levels 0 to 3, a distance of 1, and an angle of 0 (i.e., horizontally), then GLCM[0, 1, 0, 0] represents the normalized horizontal frequency of the grayscale 1 pixel immediately to the right of the grayscale 0 pixel.

[0104] It should be understood that this embodiment only uses the gray level co-occurrence matrix to extract texture features as an example, but the present invention is not limited to this. Other texture feature extraction methods can also be used to extract texture features, such as gray level run length matrix, wavelet transform or LBP local binary pattern. Since the gray level co-occurrence matrix method can quantify the contrast, correlation, homogeneity, energy, entropy and other texture features of the ground objects from multiple angles and distances, and has a strong ability to distinguish ground objects in remote sensing images and efficient engineering implementation, the present application uses the gray level co-occurrence matrix method to extract texture features.

[0105] In this embodiment, optionally, evaluating the texture features of the single-temporal remote sensing image includes:

[0106] An edge detection algorithm is used to detect edge regions in the single-temporal remote sensing image, and penalty processing is performed on texture features of pixels located in the edge regions.

[0107] Specifically, after extracting the texture feature indicators of all pixels in the single-phase remote sensing image, existing edge detection algorithms, such as the Canny operator edge detection algorithm, the Sobel operator edge detection algorithm, etc., can also be used to detect the edge area in the single-phase remote sensing image, and the texture feature indicators corresponding to the edge pixels located in the edge area are penalized, and the texture feature index values ​​corresponding to the edge pixels are reduced to half of the original values, thereby reducing the interference of edge noise on end-member extraction and improving the accuracy and noise resistance of texture feature evaluation.

[0108] The Canny operator edge detection algorithm includes:

[0109] First, Gaussian filtering is performed on the single-phase remote sensing image. The specific calculation formula is as follows:

[0110] ;

[0111] ;

[0112] in, is the Gaussian kernel standard deviation, Number the columns, Number the rows;

[0113] Then calculate the gradient amplitude of the filtered single-phase remote sensing image and gradient direction , the specific calculation formula is as follows:

[0114] ;

[0115] ;

[0116] ;

[0117] .

[0118] in, For pixels in The gradient in direction, For pixels in Gradient in direction.

[0119] Afterwards, non-maximum suppression is performed on the gradient image corresponding to the calculated single-phase remote sensing image to eliminate redundant responses caused by edge detection. This process can make the edges more refined.

[0120] Then, double threshold detection is used to determine strong edges, weak edges, and non-edges. If the gradient value of the pixel point is greater than the high threshold , it is considered a strong edge. If the gradient value of the pixel point is between the high threshold and low threshold If the gradient value of the pixel point is less than the low threshold, it is considered a weak edge. , it is considered as noise. Finally, the strong edges and weak edges are connected to form a complete edge line.

[0121] The texture feature index scores corresponding to the edge pixels in the edge area are penalized. Specifically, the detected edge pixels are penalized (multiplied by 0.5) to reduce the texture score of the edge area and highlight the texture features of the non-edge area. The specific code is:

[0122] edges=canny(image,sigma=1.2);

[0123] texture_score[edges]*=0.5.

[0124] Among them, edges represents edge pixels.

[0125] In this embodiment, optionally, identifying the ground object category in the multi-temporal remote sensing image includes:

[0126] The spatial heterogeneity index, texture characteristic index and temporal stability index of all pixels are standardized;

[0127] Based on the standardized spatial heterogeneity index, texture feature index and temporal stability index, a K-means clustering algorithm is used to classify and identify all pixels to determine the category of the ground objects in the multi-temporal remote sensing image.

[0128] Specifically, after obtaining the spatial heterogeneity index, texture feature index, and temporal stability index corresponding to each pixel in the multi-temporal remote sensing image, the objects in the multi-temporal remote sensing image can be classified based on the spatial heterogeneity index, texture feature index, and temporal stability index, and the object category corresponding to each object in the image can be determined. The weight coefficients of the spatial heterogeneity index, texture feature index, and temporal stability index can be adaptively adjusted according to the object scene corresponding to the multi-temporal remote sensing image, thereby improving the versatility and scientificity of the end-member library. For example, the spatial heterogeneity weight can be increased in urban areas, the temporal stability weight can be increased in farmland areas, and the texture feature weight can be increased in wetland areas.

[0129] Specifically, the spatial heterogeneity index, texture feature index, and temporal stability index of each pixel can be normalized first. Then, based on the normalized spatial heterogeneity index, texture feature index, and temporal stability index, the existing K-means clustering algorithm can be used to classify and identify all pixels, thereby determining the status categories corresponding to different land features in multi-temporal remote sensing images.

[0130] It should be understood that this embodiment only uses the K-means clustering algorithm as an example, but the present invention is not limited to this. Other classification and recognition algorithms can also be used to determine the status categories in multi-temporal remote sensing images, such as Gaussian mixture models, hierarchical clustering algorithms, etc.

[0131] In this embodiment, optionally, adjusting the weight coefficients of the spatial heterogeneity index, the texture feature index, and the temporal stability index includes:

[0132] The standard deviations of the spatial heterogeneity index, texture feature index and temporal stability index in different land feature categories are respectively calculated as discrimination;

[0133] Normalizing the standard deviations of the spatial heterogeneity index, texture feature index, and temporal stability index in different land feature categories;

[0134] The weight coefficients corresponding to the spatial heterogeneity index, texture feature index and temporal stability index are determined according to the corresponding normalization results.

[0135] Specifically, after determining the feature type in multi-temporal remote sensing imagery, the standard deviations of the spatial heterogeneity, texture, and temporal stability indicators for each feature category are first calculated. The standard deviation is used as the discrimination factor, with larger standard deviations indicating a greater ability of the feature to distinguish between categories. Next, the standard deviations of the spatial heterogeneity, texture, and temporal stability indicators for each feature category are normalized, and the normalized results are used as weight coefficients for the spatial heterogeneity, texture, and temporal stability indicators. This approach automatically adjusts the weight coefficients based on the discriminatory power of the feature indicators, enabling automatic adaptation to different feature scenarios and improving the scientific nature and adaptability of endmember extraction.

[0136] Specifically, first, the spatial heterogeneity index, texture feature index, and temporal stability index of each pixel are spliced ​​into a feature vector to form a feature matrix. Then, each column of the feature matrix is ​​standardized so that its mean is 0 and its standard deviation is 1. The specific calculation formula is as follows:

[0137] ;

[0138] in, represents the original eigenvalue, is the mean value of the characteristic index, is the standard deviation of the characteristic index.

[0139] Then, K-means clustering is performed on the standardized feature indicators to obtain the category label of each pixel. For each feature indicator, the standard deviation is calculated within each category and the standard deviations calculated for all categories are summed. The specific calculation formula is as follows:

[0140] .

[0141] in, is the number of category labels, For the No. All pixel values ​​of the characteristic indicators.

[0142] Finally, the cumulative standard deviation of all characteristic indicators is normalized to obtain the weight coefficients corresponding to the spatial heterogeneity index, texture characteristic index, and temporal stability index. The specific calculation formula is as follows:

[0143] ;

[0144] For the The weight of the characteristic index, is the number of characteristic indicators.

[0145] After adaptively assigning weight coefficients for spatial heterogeneity, texture characteristics, and temporal stability, the corresponding spatial heterogeneity, texture characteristics, and temporal stability indicators for each pixel are weighted and fused based on the corresponding weight coefficients to obtain a comprehensive purity score for each pixel. All pixels are sorted according to the comprehensive purity score. Combined with set spatial uniformity constraints, such as a minimum distance constraint, high-purity endmember points can be automatically selected from all pixels, generating a corresponding endmember spectral library and spatial distribution products, such as endmember point spatial distribution maps, single-phase and average spectra. Endmember point coordinates can also be exported as CSV and Shapefiles for GIS integration and subsequent analysis.

[0146] Applying a minimum distance constraint ensures that the extracted endmember points are spatially evenly distributed, avoiding oversampling in local areas. When a desired number of endmember points is specified, if the current minimum distance cannot extract a sufficient number of endmember points, the system automatically adjusts the minimum distance (by 5 pixel units at a time) until the specified number of endmember points is extracted, but never falls below 1 pixel unit. This parameter can be adjusted based on the characteristics and needs of the study area to balance the spatial uniformity and quantity of endmember points.

[0147] In this embodiment, optionally, screening for uniformly distributed high-purity end member points includes:

[0148] Determining an adaptive fractional threshold corresponding to each pixel according to the temporal stability index;

[0149] The adaptive fractional thresholds of all pixels are combined to generate corresponding stability masks, and the high-purity end-member points are screened through the stability masks.

[0150] Specifically, a stability mask is generated by weighted fusion of the coefficient of variation and the temporal similarity score, and then filtering using an adaptive quantile threshold, marking stable pixel locations as 1 and unstable ones as 0. The stability mask helps identify and select stable pixel locations suitable as candidate points for endmember extraction, while filtering out unstable areas affected by factors such as noise and cloud cover, thereby improving the accuracy and reliability of endmember extraction.

[0151] In order to verify the technical effect of the extraction method provided in this application, multi-temporal remote sensing images of two real areas were collected for endmember extraction. One of the areas is a hilly agricultural area with undulating terrain, complex distribution of cultivated land, forest land and water bodies, rich types of land features and significant temporal changes. The spatial heterogeneity and texture feature distribution corresponding to different temporal phases are shown in Figure 2. Figure 2 、 Figure 3 As shown in the figure, the corresponding distributions of spatial heterogeneity index, texture feature index and temporal stability index are as follows: Figure 4 、 Figure 5 、 Figure 6 The distribution of end-member points extracted by the extraction method provided in this application is shown in Figure 7 shown.

[0152] Depend on Figure 2-7 It can be seen that the extraction method provided in this application shows a high ability to distinguish farmland, woodland and other land features. The spatial heterogeneity index effectively identifies the boundaries of fields and woodland patches, the texture characteristics prominently reflect the farming structure of farmland, and the temporal stability accurately selects the end-member points with stable growth cycles. The scene adaptive weight optimization mechanism automatically improves the temporal stability weight, so that the final end-member points are mainly distributed in the main area of ​​farmland and the edge of woodland, with significantly improved representativeness and purity, which overall meets the needs of agricultural remote sensing interpretation and mixed pixel decomposition.

[0153] The other area is the urban expansion area, which contains a large number of buildings, roads, bare land and a small amount of green land. The spatial structure of the land features is complex, heterogeneous and has rich texture changes. The spatial heterogeneity and texture feature distribution corresponding to different phases are as follows: Figure 8 、 Figure 9 As shown in the figure, the corresponding distributions of spatial heterogeneity index, texture feature index and temporal stability index are as follows: Figure 10 、 Figure 11 、 Figure 12 The distribution of end-member points extracted by the extraction method provided in this application is shown in Figure 13 shown.

[0154] Depend on Figure 8-13As can be seen, spatial heterogeneity and texture feature indicators perform particularly well in urban areas, effectively distinguishing between building, road, and green space types. Areas with high spatial heterogeneity are often densely built-up areas, while areas with high texture features correspond to roads and bare land. The scenario-adaptive weight optimization mechanism automatically increases the weights of spatial heterogeneity and texture features. Endpoints are primarily distributed in urban core areas, road intersections, and typical green spaces, demonstrating a uniform spatial distribution and strong representativeness. This provides high-quality basic data for urban feature classification and change detection.

[0155] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and description of the present invention.

Claims

1. A multi-temporal remote sensing endmember extraction method based on multidimensional purity evaluation, characterized in that: include: Evaluate the temporal stability of the acquired multi-temporal remote sensing images, as well as the spatial heterogeneity and texture characteristics of each single-temporal remote sensing image in the multi-temporal remote sensing image, and obtain the temporal stability index, spatial heterogeneity index and texture characteristic index corresponding to each pixel; The spatial heterogeneity index, texture feature index and temporal stability index of each pixel are used to identify the ground object categories in multi-temporal remote sensing images, and the weight coefficients corresponding to the spatial heterogeneity index, texture feature index and temporal stability index are adjusted according to the discrimination degree of different ground object categories. According to the spatial heterogeneity index, texture feature index and temporal stability index and their corresponding weight coefficients, a comprehensive purity score for each pixel is generated by fusion and weighting, and uniformly distributed high-purity end-member points are automatically screened in combination with spatial uniformity constraints; Adjusting the weight coefficients of the spatial heterogeneity index, texture feature index, and temporal stability index includes: The standard deviations of the spatial heterogeneity index, texture feature index and temporal stability index in different land feature categories are respectively calculated as discrimination; Normalizing the standard deviations of the spatial heterogeneity index, texture feature index, and temporal stability index in different land feature categories; The weight coefficients corresponding to the spatial heterogeneity index, texture feature index and temporal stability index are determined according to the corresponding normalization results.

2. The multi-temporal remote sensing endmember extraction method based on multidimensional purity evaluation according to claim 1, characterized in that: Evaluate the acquired multi-temporal remote sensing images, including: Each single-temporal remote sensing image is cropped using the obtained ROI vector, and the near-infrared band of the cropped single-temporal remote sensing image is normalized.

3. The multi-temporal remote sensing endmember extraction method based on multidimensional purity evaluation according to claim 1, characterized in that: Evaluate the temporal stability of multi-temporal remote sensing images, including: Calculate the mean and standard deviation of the spectral reflectance of pixels in multi-temporal remote sensing images, and calculate the coefficient of variation of the pixels based on the mean and standard deviation; Calculating the Euclidean distance between the timing curve of a single pixel and the timing curves of other pixels, and determining the timing similarity index corresponding to the pixel based on the Euclidean distance between the single pixel and other pixels; The temporal stability index of the pixel is obtained by weighted fusion of the coefficient of variation score and the temporal similarity index obtained by the coefficient of variation.

4. The multi-temporal remote sensing endmember extraction method based on multidimensional purity evaluation according to claim 1, characterized in that: Evaluating the spatial heterogeneity of the single-temporal remote sensing image, including: Based on the local Moran's I index, the spatial autocorrelation between the pixel in the single-temporal remote sensing image and its neighboring pixels is calculated through a sliding window; According to different ground object scales, a multi-scale Gaussian weighted smoothing method is used to perform multi-scale fusion on the spatial autocorrelations corresponding to the pixels. The multi-scale fusion results of pixels are normalized to obtain the spatial heterogeneity index of pixels.

5. The multi-temporal remote sensing endmember extraction method based on multidimensional purity evaluation according to claim 4 is characterized in that: Mirror filling is used to process the boundary generated by the sliding window when processing a single-phase remote sensing image, and / or a Gaussian weight kernel is adaptively generated.

6. The multi-temporal remote sensing endmember extraction method based on multidimensional purity evaluation according to claim 1, characterized in that: Evaluating the texture features of the single-temporal remote sensing image, including: Dividing the single-phase remote sensing image into blocks to obtain a plurality of divided images; The texture features of each block image are extracted in parallel using a gray level co-occurrence matrix method to obtain a texture feature index of each pixel in the single-temporal remote sensing image.

7. The multi-temporal remote sensing endmember extraction method based on multidimensional purity evaluation according to claim 1, characterized in that: Evaluating the texture features of the single-temporal remote sensing image, including: An edge detection algorithm is used to detect edge regions in the single-temporal remote sensing image, and penalty processing is performed on texture features of pixels located in the edge regions.

8. The multi-temporal remote sensing endmember extraction method based on multidimensional purity evaluation according to claim 1, characterized in that: Identify object categories in multi-temporal remote sensing images, including: The spatial heterogeneity index, texture characteristic index and temporal stability index of all pixels are standardized; Based on the standardized spatial heterogeneity index, texture feature index and temporal stability index, a K-means clustering algorithm is used to classify and identify all pixels to determine the category of the ground objects in the multi-temporal remote sensing image.

9. The multi-temporal remote sensing endmember extraction method based on multidimensional purity evaluation according to claim 1, characterized in that: Screening for uniformly distributed, high-purity end-members, including: Determining an adaptive fractional threshold corresponding to each pixel according to the temporal stability index; The adaptive fractional thresholds of all pixels are combined to generate corresponding stability masks, and the high-purity end-member points are screened through the stability masks.

Images