Remote sensing image geological disaster monitoring method based on supervised classification and space-time fusion

Abstract

The invention discloses a remote sensing image geological disaster monitoring method based on supervised classification and space-time fusion, and relates to the field of remote sensing image processing and geological disaster monitoring, and the method comprises the steps: generating high-resolution multispectral image data through registration and fusion of multisource remote sensing images; supervised classification is carried out by using a support vector machine, and the spatial distribution proportion of each ground feature category is obtained; based on a spectral linear decomposition theory, time variation of unchanged pixels is extracted; determining a residual calculation formula based on the time sequence change, and optimizing residual distribution under the spatial dependency by adopting a thin-plate spline function and minimizing the difference between the residual and a predicted value; introducing a homogenization index and constructing a dynamic weighting function; similar pixels are screened based on an STARFM method, and a high-temporal-spatial-resolution fusion image is generated through weighted fusion of distance and spectral weight. The remote sensing fusion influence has high temporal-spatial resolution and precision, and efficient and reliable support is provided for geological disaster monitoring and early warning.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing and geological disaster monitoring, and more particularly to a remote sensing image geological disaster monitoring method based on supervised classification and spatiotemporal fusion. Background Art

[0002] Effective monitoring and early warning of geological disasters such as collapse and landslides are of great significance to protecting people's lives and property.

[0003] However, the flexible spatiotemporal data fusion method (FSDAF), currently widely used in monitoring geological hazards such as collapses and landslides, mainly relies on unsupervised classification (such as the ISODATA algorithm) to process high-resolution imagery. However, its classification results are significantly affected by initial parameters and the iterative process, resulting in high randomness. This makes it difficult to stably capture detailed information in high-dimensional feature spaces, especially in scenarios with limited samples or dynamic changes in feature types. Furthermore, traditional methods rely on a fixed weight strategy for residual allocation. When feature types suddenly change, they are unable to dynamically distinguish between intra-class and inter-class changes, resulting in blurred or distorted fused images. This makes it difficult to balance spatiotemporal resolution and detail preservation, severely restricting the accuracy and timeliness of geological hazard monitoring. While existing improved technologies attempt to optimize classification efficiency or adjust weight allocation, the inherent flaws of unsupervised classification remain unresolved. Practical applications still face bottlenecks such as poor classification stability and insufficient reliability of fusion results.

[0004] Therefore, how to solve the problems of unstable classification and large residual allocation errors in existing remote sensing image spatiotemporal fusion technology is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0005] In view of this, the present invention provides a remote sensing image geological disaster monitoring method based on supervised classification and spatiotemporal fusion, aiming to solve the problems of unstable classification and large residual allocation error in the existing remote sensing image spatiotemporal fusion technology. The SVM-FSDAF method based on supervised classification support vector machine (SVM) and dynamic residual optimization is used to improve the spatiotemporal resolution and classification accuracy of remote sensing images in geological disaster monitoring.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] The present invention discloses a remote sensing image geological disaster monitoring method based on supervised classification and spatiotemporal fusion, the specific steps of which are as follows:

[0008] S1: Generate high-resolution multispectral image data by registering and fusing multi-source remote sensing images;

[0009] S2: Using a support vector machine to perform supervised classification on the high-resolution multispectral image data to obtain the spatial distribution ratio of each object category;

[0010] S3: Based on the theory of spectral linear decomposition and the spatial distribution ratio of each feature category, the temporal variation of unchanged pixels is extracted to predict the temporal variation of high-resolution images.

[0011] S4: Determine the residual calculation formula based on the time series changes, use the thin plate spline function and optimize the residual distribution under the spatial dependence relationship by minimizing the difference between the residual and the predicted value; introduce the homogeneity index and construct a dynamic weight function;

[0012] S5: Based on the STARFM method, similar pixels are screened and fused by distance and spectral weight to generate a high-temporal-resolution fused image; based on the high-temporal-resolution fused image, potential geological disaster areas are identified by analyzing changes in land feature types and terrain characteristics.

[0013] Furthermore, the S2 specifically includes:

[0014] S21: Construct the optimal classification hyperplane of the support vector machine and use the minimized weight norm ||ω|| and slack variable ζ i The weighted sum of maximizing the classification interval of samples;

[0015] S22: For the nonlinearly separable original features in the support vector machine, introducing a Gaussian kernel function to map the original features to a high-dimensional space;

[0016] S23: Based on the classification result of the support vector machine, output the proportion of high-resolution pixels of each object category within the low-resolution pixels.

[0017] Furthermore, the S3 specifically includes:

[0018] S31: Filter pixels whose ground feature types have not changed between the reference time and the prediction time, and calculate the temporal variation of low-resolution pixels using a linear model;

[0019] S32: Superimposing the time variation of the low-resolution pixel with the reference moment image to generate a time variation prediction value, and predicting the temporal variation of the high-resolution image.

[0020] Furthermore, the S4 specifically includes:

[0021] S41: defining a residual as the difference between a true value and a predicted value of the time series change, and reflecting a local error in spatiotemporal fusion through the residual;

[0022] S42: A spatial prediction model is constructed using thin plate spline function, and the residual distribution under spatial dependence is optimized by minimizing the difference between the residual and the predicted value;

[0023] S43: Introducing a homogeneity index to quantify the uniformity of the landform types within the pixel neighborhood; constructing a dynamic weight function to adaptively adjust the residual allocation weight according to the homogeneity index value.

[0024] Furthermore, the S5 specifically includes:

[0025] S51: normalizing the dynamic weight function to obtain a new weight function, performing normalized distribution on the residuals; and reconstructing the change value of the high-resolution pixel using the normalized distribution residuals combined with the category change amount;

[0026] S52: Using the STARFM method to screen pixels with similar spectra and space, calculating spatial distance weights, and calculating normalized weights of reference pixels based on the spatial distance weights; using the normalized weights to perform weighted fusion on the change values ​​of the high-resolution pixels to obtain weighted residuals; superimposing the initial moment image with the weighted residuals to output a high spatiotemporal resolution fused image;

[0027] S53: Based on the high temporal and spatial resolution fused image, potential geological disaster areas are identified by analyzing changes in land feature types and terrain characteristics.

[0028] It can be seen from the above technical solution that compared with the existing technology, the present invention discloses a remote sensing image geological disaster monitoring method based on supervised classification and spatiotemporal fusion. By replacing unsupervised classification with supervised classification, dynamic residual allocation mechanism and collaborative fusion of multi-source data, while maintaining high spatial resolution, the temporal resolution is also greatly improved, and the fusion accuracy is significantly better than the traditional FSDAF method, providing efficient and reliable technical support for the monitoring and early warning of geological disasters such as collapse and landslides. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0030] Figure 1 Schematic diagram of the overall process of an embodiment of the present invention.

[0031] Figure 2 Schematic diagram of the study area of ​​the embodiment of the present invention.

[0032] FIG3( a ) is a real image of the GF-1 according to an embodiment of the present invention.

[0033] FIG3( b ) is a real image of the Gaofen-6 satellite according to an embodiment of the present invention.

[0034] FIG3( c ) is a real image of Sentinel-2 according to an embodiment of the present invention.

[0035] FIG3( d ) is a schematic diagram of the high-resolution image fusion result of an embodiment of the present invention.

[0036] FIG3( e ) is a schematic diagram of a sentinel prediction image according to an embodiment of the present invention.

[0037] FIG4( a ) is an enlarged view of a local portion of a real GF-1 image according to an embodiment of the present invention.

[0038] FIG4( b ) is an enlarged view of a partial portion of the GF-6 real image according to an embodiment of the present invention.

[0039] FIG4( c ) is an enlarged view of a partial portion of a real image of Sentinel-2 according to an embodiment of the present invention.

[0040] FIG4( d ) is a schematic diagram showing a partial enlargement of the high-resolution image fusion result according to an embodiment of the present invention.

[0041] FIG4(e) is a schematic diagram showing an enlarged portion of a sentinel prediction image according to an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0043] The embodiment of the present invention discloses a remote sensing image geological disaster monitoring method based on supervised classification and spatiotemporal fusion, such as Figure 1 The specific steps are as follows:

[0044] S1: Generate high-resolution multispectral imagery by registering and fusing multi-source remote sensing images. This integrates remote sensing images from different satellites or sensors, resolving potential geometric misalignments and integrating their respective strengths to generate a baseline imagery dataset with high spatial resolution and rich spectral information. This provides the necessary input for subsequent high-precision object classification and change detection.

[0045] S2: Supervised classification of high-resolution multispectral imagery data using support vector machines (SVMs) was used to determine the spatial distribution of feature categories. Using supervised classification methods, manually labeled training samples were used to allow the machine to learn feature characteristics and construct a highly accurate classification model. This model outputs the various high-resolution feature categories within each low-resolution pixel. Supervised classification methods are more stable and accurate than unsupervised classification, significantly improving the reliability of the surface composition information that is relied upon for subsequent change prediction and residual allocation.

[0046] S3: Based on the theory of spectral linear decomposition and the spatial distribution ratios of local feature categories, the temporal variation of unchanged pixels is extracted to predict the temporal variation of the high-resolution imagery. The fine feature ratio information obtained in S2 is combined with the principles of spectral mixture models to predict the temporal variation of the high-resolution imagery. Its core logic is to screen out pixels whose feature types have not changed during the prediction period. It is assumed that the spectral variations of these pixels primarily reflect changes in environmental factors or the feature's state, rather than category conversion. By analyzing the overall variation of these unchanged pixels in the low-resolution imagery and combining it with the high-resolution feature ratio information at the baseline time, it is possible to predict the expected change trend of each high-resolution pixel at the prediction time. This provides a preliminary, global estimate of variation for subsequent fine-grained fusion.

[0047] S4: Based on the time series variation, the residual calculation formula is determined. A thin plate spline function is used to optimize the residual distribution under spatial dependence by minimizing the difference between the residual and the predicted value. A homogeneity index is introduced and a dynamic weight function is constructed. This effectively addresses the fuzzy fusion distortion caused by inaccurate residual distribution in traditional methods.

[0048] S5: Using the STARFM method to screen similar pixels, a high-resolution fused image is generated through distance and spectral weighted fusion. Based on this high-resolution fused image, potential geological disaster areas are identified by analyzing changes in ground feature types and terrain characteristics. This ultimately generates a high-resolution fused image for geological disaster monitoring.

[0049] Specifically, S1 includes: acquiring multiple satellite remote sensing image data, fusing them with NNDiffuse PanSharpening (pan-sharpening based on nearest neighbor diffusion) to generate 2 high-resolution multispectral image data, and then marking samples such as bare rocks, vegetation, and water bodies.

[0050] In a specific embodiment, S2 specifically includes:

[0051] S21: Construct the optimal classification hyperplane of the support vector machine, the specific formula is formula (1); and use the minimized weight norm ||ω|| and the slack variable ζ iThe weighted sum of is used to maximize the classification interval of the sample. The specific expression is shown in formula (2), and the constraint conditions are shown in formula (3);

[0052] f(xw)=w T x+b (1);

[0053]

[0054] y i (w T x i +b)≥1-ζ i (3);

[0055] Among them, f(xw) represents the classification hyperplane, x is the sample, w is the weight vector, b is the bias term; C represents; y i is the true label of the i-th sample (value is ±1), x i is the i-th sample; in formula (2), ||ω|| is used to control the complexity of the model to avoid overfitting. The smaller the weight norm, the flatter the hyperplane, the larger the classification interval, and the stronger the generalization ability; the slack variable ζ i Allowing some samples not to meet strict classification conditions improves the model's robustness to noise and outliers.

[0056] S22: For the nonlinearly separable original features in the support vector machine, a Gaussian kernel function is introduced to map the original features to a high-dimensional space. Specifically, for nonlinear separability problems (in classification tasks, if the data cannot be completely separated into different categories by a linear hyperplane (such as a line or plane) in the original feature space, it is called a nonlinear separable problem), a Gaussian kernel function is introduced. Formula (4) is used to map the original features (referring to the initial features extracted directly from the data and not subjected to nonlinear transformation) to a high-dimensional space to improve classification capabilities. In collaboration with S21, high-precision classification of complex objects is achieved, laying the foundation for subsequent spatiotemporal fusion.

[0057]

[0058] Among them, x i with x j is the sample eigenvector, which has a different physical meaning from the spatial coordinates (X, Y) in the subsequent steps; K represents the inner product kernel function for solving nonlinear problems; δ represents the bandwidth of the Gaussian radial basis kernel function; ||x i -x j || represents the Euclidean distance between two samples in the feature space, characterizing the similarity of spectral features; g represents the Gaussian radial basis function parameter.

[0059] S23: Based on the classification results of the support vector machine, the surface cover type ratio is output using formula (5), that is, the proportion of high-resolution pixels of each feature category within the low-resolution pixel. This provides basic data for subsequent temporal change prediction.

[0060] f c (X i ,Y i )=N c (X i ,Y i ) / q (5);

[0061] Where X and Y are the horizontal and vertical coordinates of a pixel. c (X i ,Y i ) represents low-resolution pixels (X i ,Y i ) is the percentage of high-resolution pixels of feature category c; N c (X i ,Y i ) means (X i ,Y i ) corresponds to the low spatial resolution pixel belonging to the feature category c; q represents the total number of high spatial resolution pixels corresponding to all feature categories.

[0062] In a specific embodiment, S3 specifically includes:

[0063] S31: Screening reference time t a To the predicted time t b For pixels where the type of the feature has not changed, the linear model is used, and the specific formula is formula (6) to calculate the temporal change of the low-resolution pixel;

[0064]

[0065] Where l is the total number of classifications obtained by the support vector machine; ΔF(c) represents the reference time t a To the prediction time t b The change information of the c-th type of land feature, ΔC(X i ,Y i ) is the temporal variation of low-resolution pixels.

[0066] S32: Superimpose the temporal variation of the low-resolution pixel with the reference moment image and generate the temporal variation prediction value using formula (7) Preliminarily reflect the dynamic change trend of ground objects and predict the temporal changes of high-resolution images.

[0067]

[0068] Among them, X ij ,Y ij F represents the horizontal and vertical coordinates of the jth high-resolution pixel in the i-th low-resolution pixel. a (X ij ,Y ij ) is the reference moment image at the corresponding coordinate.

[0069] In a specific embodiment, S4 specifically includes:

[0070] S41: The residual is defined as the difference between the true value and the predicted value of the time series change. The residual is used to reflect the local error in spatiotemporal fusion, as shown in formula (8);

[0071]

[0072] Among them, E(X i ,Y i ) represents low-resolution pixels (X i ,Y i ), m represents the number of high-resolution pixels contained in the low-resolution pixel.

[0073] S42: A spatial prediction model is constructed using thin plate spline function, as shown in formula (9), and the residual distribution under spatial dependence is optimized by minimizing the difference between the residual and the predicted value;

[0074]

[0075] in, is the prediction time t b The high-resolution pixel spatial prediction value, a0, a1, a2 are linear term coefficients, b i is the nonlinear term coefficient, r i is the distance between pixels, and N is the total number of control points (low-resolution pixel centers). r i 2 =(XX i ) 2 +(YY i ) 2 ;when When the minimum value is taken, the above formula (9) is the optimal solution.

[0076] S43: Introduce the homogeneity index to quantify the uniformity of the landform types within the pixel neighborhood; construct a dynamic weight function to adaptively adjust the residual allocation weight according to the homogeneity index value. Areas with high HI values ​​(uniform surface) prioritize relying on spatial prediction results, while areas with low HI values ​​(heterogeneous surface) are combined with temporal prediction residuals to improve the fusion accuracy of areas with sudden landform changes. The homogeneity index and weight function are defined as shown in formulas (10) and (11).

[0077]

[0078] The range of HI is 0 to 1, and the larger the value, the more uniform the surface type. k =1, it means that the central pixel and the similar pixels have the same ground object type. k =0, it means that the types of land features are completely different.

[0079]

[0080] Among them, CW(X ij ,Y ij ) represents the dynamic weight function, HI(X ij ,Y ij ) represents the homogeneity index, R(X i ,Y i ) is a correction term representing the time prediction residual.

[0081] In a specific embodiment, S5 specifically includes:

[0082] S51: Normalize the dynamic weight function to obtain a new weight function W(X ij ,Y ij ), the residuals are normalized and distributed, and the specific formula is shown in formula (12); the residuals after normalization distribution are combined with the change information ΔF(c) to reconstruct the change value ΔF(x ij ,y ij ), the specific formula is shown in formula (13);

[0083] r(X ij ,Y ij )=m×R(X i ,Y i )×W(X ij ,Y ij ) (12);

[0084] ΔF(X ij ,Y ij )=r(X ij ,Y ij )+ΔF(c) (13);

[0085] Among them, r(X ij ,Y ij ) represents the residual assigned to the high-resolution pixel.

[0086] S52: Use the STARFM method to screen pixels with similar spectra and space, and calculate the spatial distance weight D k, and calculate the normalized weight of the reference pixel based on the spatial distance weight; use the normalized weight to perform weighted fusion on the change value of the high-resolution pixel to obtain the weighted residual; the initial moment image F a (X ij ,Y ij ) is superimposed with the weighted residual to output a high temporal and spatial resolution fused image F b (X ij ,Y ij );

[0087]

[0088] Among them, w c represents the normalized weight of the cth reference pixel, which is used for weighted fusion, and n represents the total number of similar pixels; ΔF(X c ,Y c ) represents the residual correction of the c-th reference pixel.

[0089] S53: Based on high temporal and spatial resolution fused images, identify potential geological disaster areas by analyzing changes in landform types and terrain characteristics.

[0090] In a specific embodiment, the method of the present invention is used to Figure 2 The remote sensing images of the study area are fused uncontrolled, as shown in Figures 3(a), 3(b), 3(c), 3(d), and 3(e). Figure 3(a) is the real image of GF-1, Figure 3(b) is the real image of GF-6, Figure 3(c) is the real image of Sentinel-2, and Figure 3(d) is the fusion result of the high-resolution images. This result combines the data features of GF-1 and GF-6, showing more detailed ground feature categories. The real image of Sentinel-2 and the fusion result image are used as input, and after fusion by the method of the present invention, the predicted image of Sentinel-2 is obtained (Figure 3(e).

[0091] The results of the SVM-FSDAF spatiotemporal fusion of the method of the present invention, after local magnification, are shown in Figures 4(a), 4(b), 4(c), 4(d), and 4(e). Among them, Figures 4(a) and 4(b) are the magnified results of some areas of the real images of GF-1 and GF-6, respectively, Figure 4(c) is the magnified result of the real image of Sentinel-2 data, Figure 4(d) is the fusion result of the SVM-FSDAF algorithm with GF-1 and GF-6 data as input and the corresponding area magnification result, and Figure 4(e) is the fusion result of the GF-1 and GF-6 fusion results and Sentinel-2 data as input and the corresponding area magnification result.

[0092] It can be seen that at the same magnification, the actual images from GF-1 and GF-6 are darker. The accuracy of the spatiotemporal fusion results in a structural similarity (SSIM) of 0.928, a relative global error (ERGAS) of 26.03%, and an R² and RMSE of 0.692 and 8.53, respectively. This indicates that the resolution of the fused Sentinel forecast data is consistent with that of the fused GF imagery. The fusion of GF-1 and GF-6 data also enhances image clarity and feature richness, ultimately resulting in a high-temporal-series, high-spatial-resolution imagery dataset. This spatiotemporal fusion process, while integrating the Sentinel data, mitigates the ambiguity of low-spatial-resolution pixels and improves the temporal resolution of the GF series satellites, maintaining a spatial resolution of 2 meters and a temporal resolution of 1-3 days.

[0093] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0094] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A remote sensing image geological disaster monitoring method based on supervised classification and spatiotemporal fusion, characterized in that: The specific steps are as follows: S1: Generate high-resolution multispectral image data by registering and fusing multi-source remote sensing images; S2: Using a support vector machine to perform supervised classification on the high-resolution multispectral image data to obtain the spatial distribution ratio of each object category; S3: Based on the theory of spectral linear decomposition and the spatial distribution ratio of each feature category, the temporal variation of unchanged pixels is extracted to predict the temporal variation of high-resolution images. S4: Determine the residual calculation formula based on the time series changes, use the thin plate spline function and optimize the residual distribution under the spatial dependence relationship by minimizing the difference between the residual and the predicted value; introduce the homogeneity index and construct a dynamic weight function; S5: Based on the STARFM method, similar pixels are screened and fused by distance and spectral weight to generate a high-temporal-resolution fused image; based on the high-temporal-resolution fused image, potential geological disaster areas are identified by analyzing changes in land feature types and terrain characteristics.

2. The method for monitoring geological disasters using remote sensing images based on supervised classification and spatiotemporal fusion according to claim 1, characterized in that: The S2 specifically includes: S21: Construct the optimal classification hyperplane of the support vector machine and use the minimized weight norm ||ω|| and slack variable ζ i The weighted sum of maximizing the classification interval of samples; S22: For the nonlinearly separable original features in the support vector machine, introducing a Gaussian kernel function to map the original features to a high-dimensional space; S23: Based on the classification result of the support vector machine, output the proportion of high-resolution pixels of each object category within the low-resolution pixels.

3. The method for monitoring geological disasters using remote sensing images based on supervised classification and spatiotemporal fusion according to claim 1, characterized in that: The S3 specifically includes: S31: Filter pixels whose ground feature types have not changed between the reference time and the prediction time, and calculate the temporal variation of low-resolution pixels using a linear model; S32: Superimposing the time variation of the low-resolution pixel with the reference moment image to generate a time variation prediction value, and predicting the temporal variation of the high-resolution image.

4. The method for monitoring geological disasters using remote sensing images based on supervised classification and spatiotemporal fusion according to claim 1, wherein: The S4 specifically includes: S41: defining a residual as the difference between a true value and a predicted value of the time series change, and reflecting a local error in spatiotemporal fusion through the residual; S42: A spatial prediction model is constructed using thin plate spline function, and the residual distribution under spatial dependence is optimized by minimizing the difference between the residual and the predicted value; S43: Introducing a homogeneity index to quantify the uniformity of the landform types within the pixel neighborhood; constructing a dynamic weight function to adaptively adjust the residual allocation weight according to the homogeneity index value.

5. The method for monitoring geological disasters using remote sensing images based on supervised classification and spatiotemporal fusion according to claim 1, characterized in that: The S5 specifically includes: S51: normalizing the dynamic weight function to obtain a new weight function, performing normalized distribution on the residuals; and reconstructing the change value of the high-resolution pixel using the normalized distribution residuals combined with the category change amount; S52: Using the STARFM method to screen pixels with similar spectra and space, calculating spatial distance weights, and calculating normalized weights of reference pixels based on the spatial distance weights; using the normalized weights to perform weighted fusion on the change values ​​of the high-resolution pixels to obtain weighted residuals; superimposing the initial moment image with the weighted residuals to output a high spatiotemporal resolution fused image; S53: Based on the high temporal and spatial resolution fused image, potential geological disaster areas are identified by analyzing changes in land feature types and terrain characteristics.

Images