A sar-ms adaptive guided curvelet-ihs fusion method for precise monitoring of tidal flats

Abstract

This invention proposes a SAR-MS adaptive guided curve wave-IHS fusion method for precise monitoring of tidal flats, comprising: acquiring and processing SAR images and multispectral images to be fused; performing an orthogonal IHS transform on the processed multispectral image to decompose it into intensity, hue, and saturation components; calculating the local gradient and local variance information of the processed SAR image and intensity components respectively, and adaptively generating a guiding map; under the constraints of the guiding map, decomposing the processed SAR image and intensity components respectively to obtain a base layer, detail layer, and texture layer; adopting differentiated fusion rules according to the differences in physical meaning and noise characteristics of each layer to obtain a fused base layer, fused detail layer, and fused texture layer; reconstructing the fused base layer, fused detail layer, and fused texture layer to obtain new intensity components; and performing an inverse IHS transform on the new intensity, hue, and saturation components to generate a fused tidal flat SAR and multispectral image.

Description

Technical Field

[0001] This invention relates to the field of tidal flat monitoring technology, and in particular to a SAR-MS adaptive guided curved wave-IHS fusion method for precise tidal flat monitoring. Background Technology

[0002] In remote sensing Earth observation, single-modal data, limited by imaging mechanisms and physical conditions, struggles to comprehensively represent the multidimensional information of complex surface scenes. Collaborative analysis of multi-source remote sensing data has become a crucial method for precise surface observation. Image fusion, a key technology in multi-source remote sensing collaborative processing, integrates complementary information from different sensors to generate fused images with richer spatial details and spectral features. Currently, it mainly includes panchromatic and multispectral fusion, multispectral and hyperspectral fusion, and SAR and multispectral image fusion.

[0003] SAR, as an active microwave imaging system, possesses all-weather, all-day observation capabilities and can effectively reflect the geometric structure and scattering characteristics of ground features. However, it is susceptible to speckle noise, which reduces image interpretation accuracy. Multispectral imagery has strong spectral representation and ground feature differentiation capabilities, but it is easily affected by cloud cover, atmospheric conditions, and lighting. The two systems are significantly complementary in their imaging mechanisms and information representation. Fusion can comprehensively utilize the spatial structure information of SAR imagery and the spectral characteristics of multispectral imagery, thereby improving the observation capabilities of complex surface scenes. However, due to the significant differences in the imaging mechanisms of SAR and multispectral imagery, their fusion remains a significant challenge in remote sensing image processing.

[0004] Existing SAR and multispectral image fusion methods mainly include deep learning-based methods, component substitution methods, multi-scale transformation methods, variational optimization methods, and hybrid methods. However, these methods still have certain limitations for the complex scenes of tidal flats: deep learning methods rely on a large number of labeled samples, while tidal flat remote sensing samples are scarce and exhibit strong spatiotemporal heterogeneity, limiting the model's generalization ability; component substitution methods have high computational efficiency but are prone to spectral distortion, affecting the accurate differentiation of vegetation, bare tidal flats, and other land features; multi-scale transformation methods, while possessing good spectral preservation capabilities, are sensitive to SAR speckle noise, and traditional transformations are difficult to effectively characterize linear landform features such as tidal channels and tidal grooves, easily producing edge artifacts and texture distortion; variational optimization methods can comprehensively balance multiple fusion constraints, but have high computational complexity, and parameter selection relies on experience, making it difficult to meet the high-efficiency processing requirements of long-term tidal flat monitoring.

[0005] In summary, existing hybrid fusion methods based on IHS transform still face challenges in complex tidal flat scenarios, including insufficient suppression of speckle noise, inadequate preservation of spectral information, and difficulties in the collaborative optimization of multi-scale spatial-spectral features. Non-uniform speckle noise weakens the details of tidal channels and micro-topography, and spectral distortion easily confuses vegetation, bare beaches, and areas with different water contents, thus affecting the accurate monitoring of the dynamic evolution of tidal flats. Summary of the Invention

[0006] Purpose of the invention: To address the problems faced by existing hybrid fusion methods based on IHS transform in complex tidal flat scenarios, such as insufficient suppression of speckle noise, inadequate preservation of spectral information, and difficulties in the collaborative optimization of multi-scale spatial-spectral features, this invention proposes a SAR and multispectral image fusion method that integrates adaptive guided curvelet multi-scale decomposition and orthogonal IHS transform. The method achieves a dynamic balance between noise suppression and detail enhancement through adaptive multi-scale decomposition and ensures the integrity of spectral information through orthogonal IHS transform, providing technical support for high-precision remote sensing monitoring in complex tidal flat scenarios.

[0007] Technical solution: A SAR-MS adaptive guided curve wave-IHS fusion method for precise monitoring of tidal flats, comprising:

[0008] Step 1: Acquire the SAR image and multispectral image to be fused, and perform spatial registration, size unification, band selection and pixel normalization on the two types of images to obtain the processed SAR image and multispectral image.

[0009] Step 2: Perform an orthogonal IHS transform on the processed multispectral image to decompose it into intensity component I, hue component H, and saturation component S;

[0010] Step 3: Calculate the local gradient information and local variance information of the processed SAR image and intensity component I respectively, and adaptively generate a guidance map based on the local gradient information and local variance information;

[0011] Step 4: Under the constraints of the guidance map, adaptive guided curvelet multi-scale three-layer decomposition is performed on the processed SAR image and intensity component I to obtain the base layer, detail layer and texture layer.

[0012] Step 5: Based on the differences in physical meaning and noise characteristics of the base layer, detail layer, and texture layer, differentiated fusion rules are applied to obtain the fused base layer, fused detail layer, and fused texture layer.

[0013] Step 6: Reconstruct the blend base layer, blend detail layer, and blend texture layer, and apply dynamic range constraints to obtain a new intensity component I';

[0014] Step 7: Perform an inverse IHS transform on the new intensity component I', hue component H, and saturation component S to generate a tidal flat SAR and multispectral fusion image.

[0015] Furthermore, the orthogonal IHS transformation is performed on the processed multispectral image to decompose it into intensity component I, hue component H, and saturation component S, including:

[0016] After normalizing the RGB visible light bands of the processed multispectral image, the RGB space is mapped to the IHS space using an orthogonal linear IHS transform to obtain the intensity component I, the hue component H, and the saturation component S.

[0017] Furthermore, the base layer includes macroscopic outlines and background brightness, the detail layer includes tidal flat boundaries, tidal channel edges and local structures, and the texture layer includes tidal channels, tidal passages, vegetation patches and fine textures of the beach surface.

[0018] Furthermore, for the base layer, a joint strategy of local energy and adaptive threshold is adopted for fusion to obtain the fused base layer.

[0019] Furthermore, for the base layer, a fusion strategy combining local energy and adaptive threshold is employed to obtain a fused base layer, including:

[0020] The fusion weights are calculated based on the local energy of the SAR base layer and the base layer of intensity component I.

[0021] Weighted fusion is used when the pixel difference between the two is within a threshold range;

[0022] The base layer retains the intensity component I when the pixel difference exceeds the adaptive threshold.

[0023] Furthermore, for the detail layer, an improved PCNN strategy driven by local contrast is used for fusion to obtain the fused detail layer.

[0024] Furthermore, for the detail layer, an improved PCNN strategy driven by local contrast is used for fusion to obtain a fused detail layer, including:

[0025] The local contrast of the detail layer of SAR and the detail layer of intensity component I is used as the basis for adaptive adjustment of PCNN link strength.

[0026] The fusion weights are determined based on the cumulative firing count of neurons in the detail layer of SAR and the detail layer of intensity component I during the neuron iteration process in the PCNN model, thus obtaining the fused detail layer.

[0027] Furthermore, for the texture layer, a joint sparse representation and noise adaptive suppression strategy are used to fuse them, resulting in a fused texture layer.

[0028] Furthermore, for the texture layer, a joint sparse representation and noise adaptive suppression strategy are employed for fusion to obtain a fused texture layer, including:

[0029] By combining the texture layers that represent SAR and intensity component I with an overcomplete dictionary, the amount of texture information is determined based on the L1 norm of the sparse coefficients and fusion weights are assigned. At the same time, an adaptive noise threshold is set to suppress the small amplitude sparse coefficients corresponding to SAR speckle noise, thus obtaining the fused texture layer.

[0030] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0031] (1) In view of the problem that a single SAR or multispectral image cannot simultaneously meet the requirements of tidal flat structure identification and spectral interpretation, this invention improves the ability to express the boundaries of land features, tidal groove textures and water-land interface zones in tidal flat areas by fusing complementary information from the two types of images, providing a clearer image basis for subsequent tidal flat extraction and monitoring.

[0032] (2) The present invention preserves the hue and saturation information of multispectral images well through orthogonal IHS transformation, reduces spectral distortion in the fusion process, and maintains good distinguishability of typical tidal flat features such as water bodies, bare beaches and vegetation, which is conducive to improving the stability of tidal flat classification and boundary interpretation.

[0033] (3) The present invention can retain effective spatial details while suppressing SAR speckle noise through adaptive guided curve multi-scale decomposition, and has a good enhancement effect on tidal channels, beach textures and micro-topographic structures, thus improving the problem of unclear detailed information in complex tidal flat areas.

[0034] (4) By applying differentiated fusion rules to the base layer, detail layer and texture layer respectively, this invention achieves a good balance between spectral preservation, noise suppression and spatial detail enhancement, making the fusion results more suitable for tidal flat range extraction, area calculation and dynamic change analysis.

[0035] (5) Two sets of tidal flat experiments show that the method of the present invention is superior to the comparative method in terms of visual effect and objective indicators, and the average improvement of each indicator is 8.89% compared with the second best method; the ablation experiment also shows that each module has a positive effect on improving the quality of fused images and the applicability of tidal flat monitoring. Attached Figure Description

[0036] Figure 1 A flowchart for an adaptive guided curve wave multi-scale fusion method;

[0037] Figure 2 Flowchart of adaptive guided curve wave multiscale three-level decomposition for SAR and I component;

[0038] Figure 3 A flowchart illustrating the three-layer component differentiation fusion strategy and the reconstruction process of the new I component;

[0039] Figure 4Examples of SAR and MS images;

[0040] Figure 5 The result of image fusion is shown, where, Figure 5 (a1) to (a8) represent the fusion results of Scene 1. Figure 5 (b1) to (b8) in the image represent the fusion results of Scene 2;

[0041] Figure 6 This is a graph showing the results of the ablation experiment. Detailed Implementation

[0042] The technical solution of this embodiment will now be further described in conjunction with the accompanying drawings and examples.

[0043] To address the challenge of accurate monitoring in complex tidal flat scenarios, this invention employs adaptive guided curvelet multi-scale decomposition as its core, combined with orthogonal IHS transform to construct a fusion method. The core mechanism is as follows: building upon the advantages of traditional curvelet transform in capturing anisotropic features, an adaptive guidance mechanism based on local image gradients and variance is introduced to dynamically optimize the scale, orientation, and support domain parameters of the curvelet basis. This resolves the inherent contradiction in traditional fixed-parameter multi-scale decomposition, where noise suppression results in lost details, while retaining details leaves residual noise, thus achieving accurate separation of SAR speckle and effective structural features.

[0044] This invention constructs a complete framework of spectral decoupling, multi-scale decomposition, hierarchical feature fusion, and image reconstruction, which consists of four core steps: First, the MS image is decoupled from the RGB space into intensity component I, hue component H, and saturation component S through orthogonal IHS transformation. The H and S components remain unchanged throughout the fusion process to avoid spectral distortion at its source. Next, adaptive guided curvelet multi-scale decomposition is performed simultaneously on the registered SAR image and the I component, dividing it into a base layer corresponding to macroscopic contours, a detail layer corresponding to edge structures, and a texture layer corresponding to fine ground features, simultaneously completing speckle suppression and brightness detail enhancement. Adaptive fusion rules are designed to address the differences in the physical properties of the three components, resulting in a new fused I component. Finally, image reconstruction is completed through inverse IHS transformation, generating a fused image with both high spectral fidelity and high spatial resolution. The process is as follows: Figure 1 As shown, it includes the following steps:

[0045] Step 1: Acquire the SAR image and multispectral image to be fused, and perform spatial registration, size unification, band selection and pixel normalization on the two types of images.

[0046] Step 2: After normalizing the RGB visible light bands of the multispectral image, an orthogonal linear IHS transform is used to map the RGB space to the IHS space. The orthogonal IHS transform is then performed, decomposing the image into an intensity component (I), a hue component (H), and a saturation component (S). The intensity component (I) serves as the primary component for fusion with the SAR image. The hue component (H) and saturation component (S) carry the main spectral information and remain unchanged during the fusion process, thereby reducing the risk of color shift and spectral distortion during fusion.

[0047] As one implementation method, for the input multispectral image, the pixel values ​​of its R, G, and B visible light bands are first normalized to the [0,1] interval, and then the conversion from RGB color space to IHS space is completed through orthogonal linear IHS forward transform. The transformation formula is as follows:

[0048] (1)

[0049] In the formula, I is the intensity component, the core component associated with SAR image fusion; H is the hue component and S is the saturation component, both carrying the core spectral information of the multispectral image, remaining constant throughout the fusion process to avoid spectral distortion at its source. The transformation matrix used in this study is an orthogonal matrix, which can achieve lossless decoupling between components, and the transformation is completely reversible without loss of original spectral information.

[0050] Step 3: Calculate the local gradient magnitude and local variance for the SAR image and intensity component I, respectively. The local gradient magnitude is used to characterize edge intensity, and the local variance is used to characterize texture complexity. The two are combined through adaptive weights to form a guiding map. In regions with more complex textures and more prominent edges, the guiding map tends to strengthen edge constraints; in flat regions, it emphasizes noise suppression and background stabilization.

[0051] The guide diagram is used to dynamically adjust the curvebow basis parameters, balancing noise suppression and structure preservation. The formula is as follows:

[0052] (2)

[0053] (3)

[0054] (4)

[0055] In the formula, For pixel coordinates, For A 5x5 rectangular neighborhood window centered on the center. The total number of pixels within the window; gradient magnitude. Calculated using the Sobel operator, the corresponding formula (3) is... , These are the first-order partial derivatives of the pixel in the x and y directions, respectively, used to reflect the edge strength of the pixel; local variance. The corresponding formula (4) is as follows: The average pixel value within the window, used to characterize the gray-level dispersion of a local region; balance coefficient. Adaptive value selection strategy: That is, the richer the local texture, The larger the value, the more the guiding graph emphasizes the constraint effect of edge information.

[0056] Step 4: Under the constraints of the guiding map, curvelet multi-scale decomposition is performed on the SAR image and intensity component I, and the decomposition results are divided into a base layer, a detail layer, and a texture layer. The base layer mainly represents the macroscopic contour and background brightness, the detail layer mainly represents the tidal flat boundary, tidal channel edge, and local structure, and the texture layer mainly represents the tidal channel, tidal passage, vegetation patches, and fine texture of the beach surface. This process balances speckle noise suppression and effective structure preservation by adaptively adjusting the curvelet basis scale, orientation, and support domain.

[0057] For SAR images and multispectral I component An adaptive guided curvelet multi-scale decomposition method is employed to achieve accurate separation of the base layer, detail layer, and texture layer. This decomposition combines the anisotropy advantages of curvelet transform with an adaptive guidance mechanism to dynamically match the feature differences between the two types of images. The specific flowchart is shown below. Figure 2 As shown.

[0058] As one implementation approach, the scale of the curvebow basis is iteratively optimized using the guiding graph G as a constraint. ,direction and support domain This achieves three-layer information separation. The discrete form of the curvelet transform is:

[0059] (5)

[0060] in, For the curve coefficient, The curvebow basis function satisfies the characteristics of high directional resolution at fine scales and high spatial resolution at coarse scales. Based on formula (5), the specific formulas for the three-level decomposition of the two types of images are as follows:

[0061] SAR image decomposition

[0062] (6)

[0063] The core parameters and corresponding noise suppression mechanisms for curve wave decomposition are set as follows: Scale parameters These correspond to low-frequency, mid-high-frequency, and mid-frequency components, respectively. Larger scales cover a smaller spatial range and focus more intently on ground details and texture features; directional parameters The number of directions increases with scale; at fine scales, a multi-directional configuration is used to ensure accurate capture of the anisotropy of texture and detail; support domain The resolution is synchronously reduced as the scale increases, thus balancing the spatial resolution and computational efficiency of the decomposition. To address the inherent speckle noise in SAR images, a layered noise suppression function is designed, where the detail layer suppression function is... In the formula The standard deviation of speckle noise in SAR images can be used to simultaneously achieve effective edge coefficient enhancement and noise figure suppression; the texture layer suppression function is... This can further suppress the weak speckle noise remaining in the texture layer and improve the signal-to-noise ratio of the decomposition results.

[0064] I component decomposition

[0065] (7)

[0066] For the curvelet multiscale decomposition of the multispectral I component, its scale, orientation, and core parameters of the support domain are completely consistent with those of the SAR image decomposition, ensuring strict matching of the decomposition levels of the two types of images; at the same time, a detail enhancement function is designed to complement the I component. The enhancement coefficient of this function increases synchronously with the increase of pixel edge intensity, and the maximum enhancement factor is set to 1.5 times, which can completely preserve the original brightness details of the multispectral image. Since the noise level of the I component is extremely low after multispectral preprocessing, no additional noise suppression operation is set, and its original texture information is directly retained to participate in subsequent fusion.

[0067] Step 5: Based on the differences in physical meaning and noise characteristics among the base layer, detail layer, and texture layer, differentiated fusion rules are applied. Each fusion rule includes three steps: feature measurement, weight allocation, and fusion decision. A detailed flowchart is shown below. Figure 3 As shown, a base layer, a detail layer, and a texture layer are obtained through fusion; the specific operations include:

[0068] The base layer fusion employs a joint strategy of local energy and adaptive threshold. The fusion weights are calculated based on the local energies of the SAR base layer and the I-component base layer. When the pixel difference between the two is within a reasonable range, weighted fusion is used; when the pixel difference exceeds the adaptive threshold, the I-component base layer information is retained first to ensure background brightness and spectral stability.

[0069] As an implementation method, the base layer carries the macroscopic contours of ground features and background brightness information of two types of images. The core goal of fusion is to fully preserve the original brightness background of the multispectral I component, integrate the stable ground feature structure features of the SAR image, and suppress the interference of outliers on the fusion result.

[0070] First, structural stability is measured based on local energy characteristics. Local energy can effectively characterize the stability of structures within a region; higher energy corresponds to more reliable ground features. The calculation formula is as follows:

[0071] (8)

[0072] (9)

[0073] In the formula, the calculation window is a 3×3 neighborhood. It is a Gaussian weighted matrix used to highlight the contribution of the center pixel and reduce the interference of the edge pixels; , These represent the local energy of the SAR image and the I-component base layer, respectively.

[0074] Based on this, adaptive weight allocation is performed using the local energy ratio. The weights are dynamically adjusted according to the stability of the regional structure, as shown in the following formula:

[0075] (10)

[0076] In the formula, To avoid the denominator being zero, the weights are dynamically adjusted according to local energy, with larger weights given to regions with more stable structures.

[0077] Finally, an adaptive threshold constraint is used to complete the fusion decision, where the adaptive threshold T is twice the global standard deviation of the I component's base layer. Based on the statistical characteristics of the image, reasonable structural differences and outliers are distinguished. The fusion rules are as follows:

[0078] (11) (12)

[0079] When the pixel difference between the two base layers is within the threshold range, a weighted fusion method is used to take into account the structural information of both; when the difference exceeds the threshold, the pixel value of the I component is retained first to ensure the stability of the background brightness of the fused image.

[0080] The detail layer fusion employs an improved PCNN strategy driven by local contrast. Local contrast is used to characterize detail saliency and serves as the basis for adaptive adjustment of PCNN link strength. The fusion weights are determined based on the cumulative firing count of neurons in the SAR detail layer and the I-component detail layer, thereby enhancing salient edge structures while suppressing false activations due to low-amplitude noise.

[0081] As an implementation method, the detail layer carries the core mid-to-high frequency spatial structure information such as the edges and corners of ground objects in the two types of images. It is a key link to improve the spatial resolution of the fused image. The core goal of this fusion layer is to maximize the preservation of effective detail features of SAR images and multispectral I components, enhance the saliency of ground object edges, and at the same time avoid the amplification of SAR speckle noise, so as to achieve synergistic optimization of detail enhancement and noise suppression.

[0082] First, detail saliency is measured based on local contrast. Local contrast can effectively characterize the degree of detail prominence in the region to which a pixel belongs. The higher the contrast, the stronger the saliency of effective details such as the edges and textures of ground features. The calculation formula is as follows:

[0083] (13)

[0084] In the formula, the calculation window is a 5×5 neighborhood, the numerator is the maximum and minimum values ​​of the pixels within the window, and the calculation result is normalized to the [0,1] interval to avoid numerical overflow.

[0085] Building upon this, an improved PCNN model driven by local contrast is constructed to achieve effective adaptive differentiation between details and noise. The model uses normalized local contrast as the link strength coefficient, and the neuron iterative update formula is as follows:

[0086] (14)

[0087] In the formula, n is the number of iterations, and n is taken as 5. For neuron feedback input, Enter for the link. This is the synaptic connection weight matrix within a 3×3 neighborhood. These are the internal activity terms of neurons. This represents the link strength coefficient and the local contrast at the corresponding location. Direct binding. All model-related parameters are adaptively calibrated based on local contrast: initial threshold. Take 0.8, attenuation coefficient Set the value to 0.9 and reset the threshold. By setting the parameter to 0.2 and verifying it through theoretical boundary constraints and parameter sensitivity pre-experiments, this parameter combination can achieve hierarchical activation of effective details with different saliency within 5 iteration cycles, while suppressing low-amplitude interference corresponding to SAR speckle noise throughout the process, thus solving the problems of noise misfire and loss of weak details that are prone to occur in traditional PCNN with fixed parameters.

[0088] The final fusion decision is based on the cumulative number of firings during the neuron iteration process. The more firings, the stronger the detail saliency of the corresponding pixel, and the higher the fusion weight. The fusion formula is as follows:

[0089] (15)

[0090] In the formula, , These represent the cumulative firing counts of neurons for corresponding pixels in the SAR image and the I-component detail layer, respectively. This refers to the detail layer components after fusion.

[0091] Texture layer fusion employs a joint sparse representation and adaptive noise suppression strategy. By jointly representing the texture layers of the two types of images using an overcomplete dictionary, the amount of texture information is determined based on the L1 norm of the sparse coefficients, and fusion weights are assigned. At the same time, an adaptive noise threshold is set to suppress small-amplitude sparse coefficients corresponding to SAR speckle noise, thereby obtaining a more stable tidal flat texture representation.

[0092] As an implementation method, the texture layer carries the mid-frequency fine texture information of two types of images and is a key link in depicting the fine landform features such as tidal channels and vegetation patches in the tidal flat area. The core objective of this link is to accurately extract the effective texture structure of the SAR image, fully preserve the original texture consistency of the multispectral I component, and further suppress the residual SAR speckle noise, so as to achieve the synergistic optimization of texture feature fidelity and noise suppression.

[0093] First, a joint overcomplete dictionary adapted to the texture features of the two types of images is constructed. The dictionary is then learned using the K-SVD algorithm. The dictionary has an atomic dimension of K=6 and a number of atoms L=256. The training samples are taken from noise-free uniform regions of the two texture layers. The dictionary update formula is as follows:

[0094] (16)

[0095] In the formula, For the sample matrix, It is a sparse coefficient matrix. This is a sparse regularization parameter used to balance dictionary reconstruction error and sparsity.

[0096] Based on this, sparse coding is performed on the SAR image and the texture layer of the I component, and the optimal sparse coefficients are solved using the alternating direction multiplier method. The solution formula is as follows:

[0097] (17)

[0098] Subsequently, adaptive weight allocation is performed based on the L1 norm of the sparse coefficients, and a noise adaptive suppression mechanism is designed simultaneously. The L1 norm of the sparse coefficients can effectively characterize the texture information of the corresponding region; the larger the norm, the more significant the texture features. The fusion weights are then assigned based on this, and the weight calculation formula is as follows:

[0099] (18)

[0100] Simultaneously, an adaptive noise threshold is set to remove small-amplitude sparse coefficients corresponding to speckle noise, further suppressing noise residue in the texture layer. The fusion coefficient constraint rules are as follows:

[0101] (19)

[0102] In the formula, is the fusion coefficient, and is the noise adaptive threshold, which is set based on the statistical characteristics of speckle noise in SAR images.

[0103] Finally, the texture layer is reconstructed based on the fused sparse coefficients and the joint overcomplete dictionary, resulting in the fused texture layer components. The reconstruction formula is as follows:

[0104] (20)

[0105] Step 6: Reconstruct the blend base layer, blend detail layer, and blend texture layer, and apply dynamic range constraints to obtain a new intensity component I', and normalize it to the dynamic range of the original intensity component I.

[0106] As an implementation method, the core of this step is to complete the multi-scale reconstruction and dynamic range constraint of the three-layer components after fusion, so as to ensure the spectral fidelity of the subsequent IHS inverse transform. First, the fused base layer L F Detail layer D F Texture layer T F Linear superposition and reconstruction yield the initial new intensity components, as shown in formula (21); where the detail layer weights Texture layer weights Preliminary experiments have shown that a balance can be struck between spatial detail enhancement and spectral fidelity. Subsequently, the reconstructed image is normalized from minimum to maximum. Component mapping to original Dynamic range of components To avoid spectral distortion in subsequent inverse transformations, the normalization formula is shown in (22).

[0107] (twenty one)

[0108] (twenty two)

[0109] Step 7: Perform an inverse IHS transformation on the new intensity component I' and the original hue component H and saturation component S to generate a tidal flat SAR and multispectral fusion image with good spectral fidelity and spatial detail representation.

[0110] As an implementation method, this step relies on the invertibility of the orthogonal IHS transform to complete the final image generation, closing the entire technical path of this paper. Since the IHS forward transform used in this paper is a standard orthogonal linear transform, its inverse transform matrix is ​​the transpose of the forward transform matrix, which can minimize the loss of spectral information during the transformation process. The inverse transform formula is shown in (23). The inverse transform input is the fused new intensity component. With original hue component saturation component , integration of the entire fixed process , The components remain unchanged to ensure spectral fidelity; the output is the RGB three-band R of the fused image. F G F B F Finally, the pixel values ​​are normalized to the [0, 255] range for subsequent analysis and application.

[0111] (twenty three)

[0112] To further illustrate the beneficial effects of the method in the embodiments of the present invention, comparative experiments and ablation verification were conducted. In the ablation experiments, the fusion method proposed in the embodiments of the present invention was compared with five advanced fusion methods. Specifically, the methods compared included PCA, VSMWLS, IFCNN, SwinFusion, and DRF. Subsequently, a comprehensive and detailed evaluation of the fusion results was performed, with the evaluation system consisting of both quantitative and qualitative parts. In terms of qualitative evaluation, the visual effect of the image, noise suppression capability, and the degree of integration of effective information were mainly examined. In terms of quantitative evaluation, six representative evaluation indicators were selected, including error of relative global dimensionless synthesis (ERGAS), root mean square error (RMSE), correlation coefficient (CC), mutual information (MI), peak signal-to-noise ratio (PSNR), and universal image quality index (UIQI). ERGAS and RMSE measure spectral and spatial reconstruction errors, CC and UIQI evaluate structural consistency and image quality, MI reflects information integration capability, and PSNR characterizes the reconstruction signal-to-noise ratio. This index system covers five core dimensions: spectral fidelity, spatial detail, structural consistency, information fusion, and visual quality, and can comprehensively and objectively evaluate the performance of various fusion methods.

[0113] Finally, to further verify the effectiveness of the proposed method, ablation experiments were conducted, including removing the IHS module, replacing the traditional adaptive-guided three-layer curvelet decomposition with a non-adaptive approach, and removing the targeted multilayer fusion strategy while adopting a single fusion strategy. These experiments systematically evaluated the innovation and practicality of the proposed method from multiple dimensions.

[0114] The experimental data were derived from publicly available Sentinel series remote sensing data released by the European Space Agency's Copernicus program. SAR imagery used C-band synthetic aperture radar data acquired by the Sentinel-1 satellite, while multispectral imagery used optical multispectral data acquired by the MSI sensor on the Sentinel-2 satellite. All data underwent radiometric calibration, atmospheric correction, and spatial registration preprocessing, and were uniformly cropped to the same size to meet the requirements of subsequent data fusion experiments.

[0115] Figure 5 A comparison of SAR and multispectral image fusion results for tidal flat scene 1, in which... Figure 5 (a1) in the image represents the original SAR image. Figure 5 (a2) in the image represents the original multispectral image. Figure 5 In the table, (a3)-(a7) represent the fusion results of five mainstream comparison methods, respectively. Figure 5 (a8) in the figure represents the fusion result of the method proposed in the embodiment of the present invention. Figure 5In (a1) to (a8), the PCA, SwinFusion, and VSMWLS methods, lacking a collaborative optimization mechanism for multispectral and SAR spatial information, either exhibit severe spectral distortion (e.g., PCA's pale colors and blurred ground features) or significant noise residue and detail loss (e.g., SwinFusion's color cast and VSMWLS's texture blur). Visual observation reveals color shifts and speckled noise in building edges, vegetation textures, and bare ground patches, significantly reducing the ground feature recognition of the original multispectral image. The natural green hues of vegetation and the clear outlines of buildings cannot be effectively interpreted due to distortion and noise. While the IFCNN method is slightly better than the former in preserving spectral information, it still suffers from local detail blurring and insufficient contrast due to the lack of adaptive fusion rules for different feature layers. The DRF method performs well in spectral preservation and detail extraction, but subtle texture noise is visible in vegetation areas, indicating room for improvement in the layered fusion of cross-modal features. In contrast, the method proposed in this embodiment of the invention, based on adaptive guided curvelet multi-scale decomposition and IHS transform framework, achieves multi-scale three-layer targeted fusion, synergizing the complementary advantages of the two types of data. Visually, this method not only fully preserves the natural colors of the multispectral image but also accurately integrates the spatial details of the SAR image, while effectively suppressing noise and spectral distortion. Ground feature information can be efficiently interpreted, and the final fused image possesses both high-fidelity spectral features and high-resolution spatial details.

[0116] Figure 5 Images (b1) to (b8) in the table show a comparison of the fusion results for scene 2 of the tidal flats. Figure 5 (b1) in the image is the original SAR image of the scene. Figure 5 (b2) in the image is the original multispectral image. Figure 5 (b3)-(b7) in the table represent the fusion results of five mainstream comparison methods, respectively. Figure 5 (b8) represents the fusion result of the method proposed in this embodiment of the invention. Figure 5 (b1) in Figure 5 In the comparison of fusion results shown in (b8), the PCA, SwinFusion, and VSMWLS methods, lacking a collaborative optimization mechanism for cross-modal features, exhibit hue and noise levels highly similar to the SAR image. Severe spectral distortion and speckle noise significantly interfere with the interpretation of ground cover information. Visual observation reveals that the PCA method... Figure 5 The colors of the fused image (b6) deviate completely from the original multispectral image. Figure 5 The natural tones in (b2) are completely blurred by noise in the bare and vegetated areas; the SwinFusion method Figure 5In (b7), there is obvious color cast and loss of detail; the outline of the river edge becomes blurred due to distortion. VSMWLS method Figure 5 The fused image (b4) is generally grayish, the texture of the terrain is masked by noise, and the spectral information of the original multispectral image is almost impossible to effectively identify. IFCNN method Figure 5 While method (b5) is slightly better than the aforementioned methods in preserving spectral information, it still suffers from local detail blurring and insufficient contrast due to the lack of adaptive fusion rules for different feature layers. In contrast, the Ours method proposed in this paper... Figure 5 Among them, (b8) shows the best performance in fusion effect. This method combines IHS transform with an adaptive guided curvelet multilayer decomposition framework, and designs adaptive fusion rules for the base layer, detail layer, and texture layer respectively. It not only fully preserves the natural tones of the multispectral image, but also accurately integrates the spatial details of the SAR image. Visually, the fused image of Ours method not only completely suppresses obvious speckle noise, but also eliminates subtle texture noise, and the boundaries of ground features are clearly distinguishable. It achieves a better balance between spectral information preservation and spatial detail enhancement.

[0117] Table 1 shows the effects of the six indicators. Figure 5 Quantitative evaluation results of the fused images. In the table, the optimal and near-optimal evaluation metrics are shown in bold and underlined, respectively. ERGAS, RMSE, CC, MI, PSNR, and UIQI metrics measure the quality of the fused image from the perspectives of spatial accuracy, pixel-level accuracy, structural similarity, information content, noise level, and overall quality method, respectively. In Table 1, the method proposed in this embodiment of the invention achieves optimal values ​​for all evaluation metrics. Specifically, compared to the near-optimal method, the improvements in ERGAS, RMSE, CC, MI, PSNR, and UIQI metrics reach 17.21%, 36.59%, 4.42%, 8.77%, 9.52%, and 4.43%, respectively. Particularly noteworthy are the significant improvements in RMSE and ERGAS, further confirming the superiority of the method in this embodiment of the invention in terms of fused image quality.

[0118] Table 1 Figure 5 Quantitative evaluation results from (a1) to (a8)

[0119]

[0120] Table 2 shows that the proposed method achieved optimal results in the metrics, with ERGAS, RMSE, CC, MI, PSNR, and UIQI improving by 5.30%, 5.75%, 2.78%, 0.49%, 8.31%, and 3.06% respectively compared to the suboptimal values. The improvements in ERGAS, RMSE, and PSNR were particularly significant. The proposed method and the DRF method obtained optimal and suboptimal quantitative evaluation results, respectively, indicating that both can achieve high fusion quality. This is consistent with the qualitative analysis results presented earlier, jointly validating the effectiveness of the proposed method.

[0121] Table 2 shows the quantitative evaluation results from (b1) to (b8) in Figure 5.

[0122]

[0123] To comprehensively evaluate the effectiveness of each module of the proposed fusion strategy, four sets of control experiments were set up. The first set, Group-A1 (Group-B1), involved removing the IHS module. If severe spectral distortion occurred after removal, it proves that this module is crucial for ensuring spectral fidelity. The second set, Group-A2 (Group-B2), replaced the adaptively guided three-level curvelet decomposition with a traditional curvelet decomposition without adaptive guidance. If the spatial detail extraction and noise suppression capabilities of the method significantly decreased after replacement, it proves that the adaptive guidance module is the core key to the superior performance of this method compared to traditional multi-scale decomposition methods. The third set, Group-A3 (Group-B3), involved removing the targeted multi-level fusion strategy and adopting a single fusion strategy. If severe spectral distortion or insufficient spatial information occurred after removal, it proves that the differentiated strategy is key to balancing spectral and spatial information. The final set, Group-A4 (Group-B4), represents the fusion method proposed in this paper. Specific experimental results are as follows: Figure 6 As shown.

[0124] Table 3 Figure 6 Quantitative evaluation results from (A1) to (A4)

[0125]

[0126] Table 4 Figure 6 Quantitative evaluation results from (B1) to (B4)

[0127]

[0128] Experimental results show that, compared with the removal of the IHS module, the traditional three-layer curvelet decomposition without adaptive guidance, and the removal of the targeted multilayer fusion strategy, the proposed method achieves significant performance improvements on six representative quantitative evaluation indicators, fully demonstrating the positive role of different fusion strategies in improving fusion quality.

Claims

1. A SAR-MS adaptive guided curve wave-IHS fusion method for precise monitoring of tidal flats, characterized in that: include: Step 1: Acquire the SAR image and multispectral image to be fused, and perform spatial registration, size unification, band selection and pixel normalization on the two types of images to obtain the processed SAR image and multispectral image. Step 2: Perform an orthogonal IHS transform on the processed multispectral image to decompose it into intensity component I, hue component H, and saturation component S; Step 3: Calculate the local gradient information and local variance information of the processed SAR image and intensity component I respectively, and adaptively generate a guidance map based on the local gradient information and local variance information; Step 4: Under the constraints of the guidance map, adaptive guided curvelet multi-scale three-layer decomposition is performed on the processed SAR image and intensity component I to obtain the base layer, detail layer and texture layer. Step 5: Based on the differences in physical meaning and noise characteristics of the base layer, detail layer, and texture layer, differentiated fusion rules are applied to obtain the fused base layer, fused detail layer, and fused texture layer. Step 6: Reconstruct the blend base layer, blend detail layer, and blend texture layer, and apply dynamic range constraints to obtain a new intensity component I'; Step 7: Perform an inverse IHS transform on the new intensity component I', hue component H, and saturation component S to generate a tidal flat SAR and multispectral fusion image.

2. The SAR-MS adaptive guided curve wave-IHS fusion method for precise monitoring of tidal flats according to claim 1, characterized in that: The process of performing an orthogonal IHS transform on the processed multispectral image to decompose it into an intensity component I, a hue component H, and a saturation component S includes: After normalizing the RGB visible light bands of the processed multispectral image, the RGB space is mapped to the IHS space using an orthogonal linear IHS transform to obtain the intensity component I, the hue component H, and the saturation component S.

3. The SAR-MS adaptive guided curve wave-IHS fusion method for precise monitoring of tidal flats according to claim 1, characterized in that: The base layer includes the macro outline and background brightness, the detail layer includes the tidal flat boundary, the edge of the tidal channel and local structure, and the texture layer includes the tidal channel, the tidal channel, vegetation patches and fine texture of the beach surface.

4. The SAR-MS adaptive guided curve wave-IHS fusion method for precise monitoring of tidal flats according to claim 1, characterized in that: For the base layer, a fusion strategy combining local energy and adaptive threshold is adopted to obtain the fused base layer.

5. The SAR-MS adaptive guided curved wave-IHS fusion method for precise monitoring of tidal flats according to claim 4, characterized in that: The base layer is fused using a joint strategy of local energy and adaptive threshold, resulting in a fused base layer, including: The fusion weights are calculated based on the local energy of the SAR base layer and the base layer of intensity component I. Weighted fusion is used when the pixel difference between the two is within a threshold range; The base layer retains the intensity component I when the pixel difference exceeds the adaptive threshold.

6. The SAR-MS adaptive guided curve wave-IHS fusion method for precise monitoring of tidal flats according to claim 1, characterized in that: For the detail layer, an improved PCNN strategy driven by local contrast is used for fusion to obtain the fused detail layer.

7. The SAR-MS adaptive guided curve wave-IHS fusion method for precise monitoring of tidal flats according to claim 6, characterized in that: The aforementioned detailed layer employs a locally contrast-driven improved PCNN strategy for fusion, resulting in a fused detailed layer, including: The local contrast of the detail layer of SAR and the detail layer of intensity component I is used as the basis for adaptive adjustment of PCNN link strength. The fusion weights are determined based on the cumulative firing count of neurons in the detail layer of SAR and the detail layer of intensity component I during the neuron iteration process in the PCNN model, thus obtaining the fused detail layer.

8. The SAR-MS adaptive guided curved wave-IHS fusion method for precise monitoring of tidal flats according to claim 1, characterized in that: For the texture layer, a joint sparse representation and noise adaptive suppression strategy are used to fuse them, resulting in a fused texture layer.

9. The SAR-MS adaptive guided curve wave-IHS fusion method for precise monitoring of tidal flats according to claim 8, characterized in that: For the texture layer, a joint sparse representation and noise adaptive suppression strategy are used for fusion to obtain a fused texture layer, including: By combining the texture layers that represent SAR and intensity component I with an overcomplete dictionary, the amount of texture information is determined based on the L1 norm of the sparse coefficients and fusion weights are assigned. At the same time, an adaptive noise threshold is set to suppress the small amplitude sparse coefficients corresponding to SAR speckle noise, thus obtaining the fused texture layer.

Images