Multispectral remote sensing image de-mixing method combined with superpixels
By combining the time series change detection model of superpixel and multispectral remote sensing images, pure cells are extracted and demixed using vegetation index for post-processing, the variability problem of traditional spectral mixed analysis model in terminal element selection is solved, and high Accurate image demixing and abundance map separation.
Patent Information
- Application Number
- CN202411410996.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-10
- Publication Date
- 2025-01-03
AI Technical Summary
The traditional spectral mixing analysis model has problems such as large intra-class variability and small inter-class variability in the selection of end-member elements, which leads to large errors in the selection error of pure cell and the area proportional error of image demix results, especially soil covers. The degree and impermeable surface cover are easily confused.
Using a multi-spectral remote sensing image demixing method combining superpixels, a time series change detection model is constructed through long-term series low-resolution multi-spectral images and high-resolution land use classification diagrams, superpixel partitioning and pure cell extraction are performed, and kmeans unsupervised clustering and vegetation index NDVI, drought bare soil index DBSI, and building index NISI are demixed to reduce missed scores and misses.
Automatic end-member extraction is realized, artificial participation is reduced, the number of pure pixels is increased, the understanding of mixing is improved, the intra-class variability and inter-class variability errors are reduced, and the vegetation, impermeable surfaces and soil are accurately separated. Abundance map.
Smart Images
Figure CN119251702A_ABST
Abstract
Description
Technical Field
[0001] The invention provides a multi-spectral remote sensing image unmixing method combined with superpixels, belonging to the field of remote sensing science and technology. Background Art
[0002] Researchers have developed a series of spectral mixture analysis models to solve the mixed pixel problem, which can be divided into linear and nonlinear models. The choice of mixture analysis model mainly depends on the degree of nonlinearity of the relationship between the spectral reflectance and the cover proportion of each land cover class within the pixel. Both linear and nonlinear methods require end members (represented by the pure spectrum of each land cover class) to construct the corresponding mathematical model.
[0003] The traditional spectral mixture analysis model is mainly divided into three steps: ① determine the number and type of endmembers of the target image; ② determine the spectral characteristic curve of the endmember; ③ spectrally decompose the target image. However, the traditional method has high technical requirements for endmember selection, and there are few pure pixels in multispectral images, which makes it difficult to overcome the problem of large intra-class variability and small inter-class variability.
[0004] At present, due to the error in the spectral information of the selected pure pixels, the "pure pixels" selected in the endmember selection are not completely pure single target land use types, but are more or less mixed with other objects; in addition, there are external factors of the image that affect the spectral characteristics of the endmembers, such as environmental factors (atmosphere, topography, climate) and the spectral influence of surrounding objects. Due to the above reasons, the area ratio (abundance) of each surface cover type after unmixing the mixed pixels using spectral unmixing has errors that cannot be ignored. In the unmixing results of spectral unmixing of mixed pixels in remote sensing images, soil cover and impervious surface cover are often confused: on the one hand, a large amount of vegetation and soil in the high reflectivity and low reflectivity impervious surface cover components are mistakenly divided into impervious surfaces; on the other hand, the vegetation and soil cover components also contain impervious surface cover. Summary of the invention
[0005] In order to solve the problem that the traditional spectral mixture analysis model has large intra-class variability and small inter-class variability in endmember selection, the present invention proposes a multispectral remote sensing image unmixing method combined with superpixels.
[0006] The technical solution adopted by the present invention is: a multispectral remote sensing image unmixing method combined with superpixels, comprising the following steps: ① Collect long-term low-resolution multispectral remote sensing images, existing high-resolution land use classification maps within the long-term series, and high-resolution multispectral remote sensing images at the corresponding time of the high-resolution land use classification maps; ② Assume that the long time series range is T0-T nAt each moment, a time series change detection model of the surface reflectance of each pixel in the selected study area is constructed based on the continuous change detection and classification method; ③Select T k2 The low-resolution multispectral remote sensing image of the study area at the moment is partitioned into superpixels, and each superpixel partition is called a superpixel block, where T k2 ∈[T0,T n] ; ④Extract pure pixels; ⑤ Perform kmeans unsupervised clustering on the extracted pure pixels, and assign pure pixel land type labels to the unsupervised clustering results combined with manual interaction, dividing them into five categories: vegetation, soil, shadow, impervious surface, and water body, which constitute T k2 Endmember spectral library of low-resolution multispectral images at all times; ⑥ Assume that the acquisition time of high-resolution multispectral remote sensing images and high-resolution land use classification maps is T k1 , where T k1 <T k2 , according to the change detection model constructed in step ②, we get T k1 To T k2 The stable pixel distribution map at time T and the land class labels and spectral reflectance values of the sub-pixels in the stable pixels constitute k2 Endmember spectral library at sub-pixel level at every moment; ⑦ Combined with the endmember spectral library in steps ⑤ and ⑥, according to the spatial synthetic spectral reflectance method, a candidate endmember spectral set for each pixel is constructed for T k2 Fully constrained linear spectral decomposition of low-resolution multispectral remote sensing images pixel by pixel in the study area at the moment; ⑧ Combined with high-resolution remote sensing images, unmixing and post-processing analysis is carried out based on the vegetation index NDVI, drought bare soil index DBSI, and building index NISI to obtain the final abundance map of vegetation, impervious surface, and soil.
[0007] In step ③, superpixel partitioning of the study area is achieved based on a simple linear iterative clustering method.
[0008] In step ④, pure pixels are extracted by performing spectral purity analysis and similarity analysis on each superpixel block.
[0009] In step ⑤, the distance measurement function of kmeans unsupervised clustering adopts the spectral angle distance function.
[0010] In step ⑧, the DBSI, NISI, and NDVI thresholds are finally determined through comparison of pixel values and multiple threshold selection experiments.
[0011] In step ⑧, the characteristics of different indices are combined to use NDVI and NISI in combination to eliminate the confusion between vegetation and impervious surfaces, and NISI, DBSI, and NDVI in combination to eliminate the confusion between vegetation and soil within the impervious surface; NISI and DBSI in combination to eliminate the confusion between soil and impervious surfaces.
[0012] The thresholds of DBSI, NISI, and NDVI are 0.1, 0.4, and 0.4, respectively.
[0013] The beneficial effects of the present invention compared with the prior art are as follows: (1) In the process of end member extraction in the prior art, the degree of manual involvement is high. The present invention combines the Simple Linear Iterative Clustering (SLIC) method to adjust the global search range of clustering to the local area, thereby realizing the automatic extraction of end members with low manual involvement and achieving good accuracy.
[0014] (2) There are relatively few pure pixels in low-resolution multispectral images that can be used as representative end members. The present invention combines long-term series low-resolution multispectral image change detection analysis to obtain stable pixels, and obtains pure sub-pixels based on high-resolution multispectral images and corresponding land use classification maps, thereby increasing the number of pixels that can be used as representative end members.
[0015] (3) In order to solve the problems of missed classification and misclassification, the present invention designs a post-processing analysis based on the vegetation index NDVI, the soil index NDSI and the building index NISI. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] The present invention will be further described below in conjunction with the accompanying drawings: Figure 1 The present invention is a flow chart of the method. DETAILED DESCRIPTION
[0017] like Figure 1 As shown, the present invention provides a multispectral remote sensing image unmixing method combined with superpixels, which can overcome the defects of large intra-class variability and small inter-class variability, and at the same time increase the number of pure pixels in the image. The multispectral unmixing method comprehensively considers various factors, reasonably uses multi-source long-time series multispectral remote sensing image data, combines superpixel segmentation to achieve homogeneous region clustering, realizes intelligent end member extraction, and then performs spectral unmixing and unmixing post-processing operations. The method specifically includes the following steps: A. Use multispectral remote sensing images to perform pure pixel analysis, including the following analysis contents: a) Image segmentation and clustering standard analysis: ① Collect long-term low-resolution multispectral remote sensing images, existing high-resolution land use classification maps within the long-term series, and high-resolution multispectral remote sensing images at the corresponding time.
[0018] b) Analysis of change detection methods: ② Assume that the long time series range is T0-T n At each moment, a time series change detection model (Continuous change detection, CCD) of the surface reflectance of each pixel in the selected study area is constructed based on the continuous change detection and classification method.
[0019] c) Analysis of the number of superpixel partitions: ③Select T k2 The low-resolution multispectral remote sensing image of the study area is used at all times, and the superpixel partition of the study area is realized according to the simple linear iterative clustering (SLIC) method. Each superpixel partition is called a superpixel block, and it is assumed that the superpixel block is a homogeneous area, and the global search range of clustering is adjusted to the local area.
[0020] d) Analysis of pure pixel extraction methods: ④ Perform spectral purity analysis and similarity analysis on each superpixel block to extract pure pixels.
[0021] B. Use multispectral time-series remote sensing images to analyze stable pixels within two specified time intervals, including the following analysis contents: a) Analysis of sub-pixel labels within stable pixels: ⑤ Perform kmeans unsupervised clustering on the extracted pure pixels, and use the spectral angle distance function as the distance measurement function; combine manual interaction to assign pure pixel land type labels to the unsupervised clustering results, and divide them into five categories: vegetation, soil, shadow, impervious surface, and water body, which constitute T k2 Moment-by-moment low-resolution multispectral image endmember spectral library.
[0022] ⑥ Assume that the acquisition time of high-resolution multispectral remote sensing images and high-resolution land use classification maps is T k1 (T k1 <T k2 ), according to the change detection model constructed in step ②, we get T k1 To T k2 The stable pixel distribution map at time T and the land class labels and spectral reflectance values of the sub-pixels in the stable pixels constitute k2 Endmember spectral library at sub-pixel level at each moment.
[0023] ⑦ Combine the endmember spectral library of steps ⑤ and ⑥ to minimize intra-class changes and solve the problem of few pure pixels in low-resolution remote sensing images. According to the spatial synthetic spectral reflectance method, a candidate endmember spectral set for each pixel is constructed for T k2 Fully constrained linear spectral decomposition of low-resolution multispectral remote sensing images of the study area on a pixel-by-pixel basis.
[0024] C. Use vegetation index NDVI, dry bare soil index (DBSI), and building index NISI to achieve unmixing post-processing analysis, including the following analysis contents: a) Analysis of threshold intervals of each index: ⑧ Since spectral unmixing will cause errors in the area ratio (abundance) of each surface cover type after unmixing mixed pixels, in order to reduce this error and solve the problem of misclassification and omission, high-resolution remote sensing images are combined with vegetation index NDVI, drought bare soil index DBSI (reference Rasul, A., Balzter, H., Ibrahim, GRF, et al. (2018) Applying Built-Up and Bare-Soil Indices from Landsat 8 to Cities in Dry Climates. Land, 7, Article No. 81. https: / / doi.org / 10.3390 / land7030081) and building index NISI to realize post-unmixing analysis. Through comparison of pixel values and multiple threshold selection experiments, the DBSI, NISI and NDVI thresholds are finally determined to be 0.1, 0.4 and 0.4.
[0025] b) Design index portfolio analysis: ⑨ Combining the characteristics of different indices, the combination of NDVI and NISI is used to eliminate the confusion between vegetation and impervious surfaces; the combination of NISI, DBSI, and NDVI is used to eliminate the confusion between vegetation and soil within the impervious surface; and the combination of NISI and DBSI is used to eliminate the confusion between soil and impervious surfaces.
[0026] ⑩ After the above technical analysis steps, the final abundance map of vegetation, impervious surface and soil is obtained.
[0027] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. 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 replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A multispectral remote sensing image unmixing method combined with superpixels, characterized by: The following steps are involved: ① Collect long-term low-resolution multispectral remote sensing images, existing high-resolution land use classification maps within the long-term series, and high-resolution multispectral remote sensing images at the corresponding time of the high-resolution land use classification maps; ② Assume that the long time series range is T0-T n At each moment, a time series change detection model of the surface reflectance of each pixel in the selected study area is constructed based on the continuous change detection and classification method; ③Select T k2 The low-resolution multispectral remote sensing image of the study area at the moment is partitioned into superpixels, and each superpixel partition is called a superpixel block, where T k2 ∈[T0,T n] ; ④Extract pure pixels; ⑤ Perform kmeans unsupervised clustering on the extracted pure pixels, and assign pure pixel land type labels to the unsupervised clustering results combined with manual interaction, dividing them into five categories: vegetation, soil, shadow, impervious surface, and water body, which constitute T k2 Endmember spectral library of low-resolution multispectral images at all times; ⑥ Assume that the acquisition time of high-resolution multispectral remote sensing images and high-resolution land use classification maps is T k1 , where T k1 <T k2 , according to the change detection model constructed in step ②, we get T k1 To T k2 The stable pixel distribution map at time T and the land class labels and spectral reflectance values of the sub-pixels in the stable pixels constitute k2 Endmember spectral library at sub-pixel level at every moment; ⑦ Combined with the endmember spectral library in steps ⑤ and ⑥, according to the spatial synthetic spectral reflectance method, a candidate endmember spectral set for each pixel is constructed for T k2 Fully constrained linear spectral decomposition of low-resolution multispectral remote sensing images pixel by pixel in the study area at the moment; ⑧ Combined with high-resolution remote sensing images, unmixing and post-processing analysis is carried out based on the vegetation index NDVI, drought bare soil index DBSI, and building index NISI to obtain the final abundance map of vegetation, impervious surface, and soil.
2. The method for unmixing multispectral remote sensing images combined with superpixels according to claim 1, characterized in that: In step ③, superpixel partitioning of the study area is achieved based on a simple linear iterative clustering method.
3. The method for unmixing multispectral remote sensing images combined with superpixels according to claim 1, characterized in that: In step ④, pure pixels are extracted by performing spectral purity analysis and similarity analysis on each superpixel block.
4. The method for unmixing multispectral remote sensing images combined with superpixels according to claim 1, characterized in that: In step ⑤, the distance measurement function of kmeans unsupervised clustering adopts the spectral angle distance function.
5. The method for unmixing multispectral remote sensing images combined with superpixels according to claim 1, characterized in that: In step ⑧, the DBSI, NISI, and NDVI thresholds are finally determined through comparison of pixel values and multiple threshold selection experiments.
6. The method for unmixing multispectral remote sensing images combined with superpixels according to claim 5, characterized in that: In step ⑧, the characteristics of different indices are combined to use NDVI and NISI in combination to eliminate the confusion between vegetation and impervious surfaces, and NISI, DBSI, and NDVI in combination to eliminate the confusion between vegetation and soil within the impervious surface; NISI and DBSI in combination to eliminate the confusion between soil and impervious surfaces.
7. The method for unmixing multispectral remote sensing images combined with superpixels according to claim 6, characterized in that: The thresholds of DBSI, NISI, and NDVI are 0.1, 0.4, and 0.4, respectively.