Method for extracting intrinsic spectral information of near-shore underwater target driven by constrained unmixing model

Abstract

The application discloses a kind of restricted unmixed model driven near-shore underwater target intrinsic spectral information extraction methods, comprising: constructing restricted unmixed model, observed hyperspectral image is represented as the combination of endmember dictionary and abundance matrix, based on prior transformation parameter and target prior spectrum constructs flexible prior constraint;With endmember dictionary, abundance matrix and prior transformation parameter as optimization variable, the optimization process of restricted unmixed model is developed into bidirectional dynamic depth development network, and the optimization variable is updated in multiple stages jointly optimized;Output target endmember in optimization endmember dictionary as target intrinsic spectral information, and output target abundance distribution corresponding to target endmember in optimization abundance matrix as target detection response graph.The application relates to the field of hyperspectral remote sensing image processing technology, and can realize the joint estimation of target intrinsic spectral information and target spatial abundance distribution, so as to provide more reliable target prior and detection response for near-shore underwater target detection.

Description

Technical Field

[0001] This invention relates to the field of hyperspectral remote sensing image processing technology, specifically a method for extracting the intrinsic spectral information of nearshore underwater targets driven by a constrained unmixing model. Background Technology

[0002] Hyperspectral remote sensing images consist of multiple consecutive narrow bands, with each pixel containing a spectral curve that can be used to distinguish different materials or targets. In nearshore underwater target detection scenarios, the inherent spectral information of the target is an important basis for characterizing the target's physical properties. If the inherent spectrum that more closely approximates the actual state of the target can be extracted from real observation images, it can provide more reliable target priors for subsequent target information characterization, target information enhancement, and background suppression detectors.

[0003] However, the nearshore underwater environment differs from the terrestrial hyperspectral detection scenario. Water has different attenuation effects on different wavelengths, changes in water depth alter the spectral transmission path, the underwater background mixes with the target response, path radiation is superimposed on the observed signal, and the target surface condition changes with adhesion and deposition. Therefore, the observed spectrum of the target is not a direct measurement of the target's inherent spectrum, but rather a degraded result after water transmission and background mixing.

[0004] Currently, existing spectroscopic information technologies mainly include three categories: 1) Target prior direct input detection methods: These methods directly input the target prior spectrum measured in the laboratory into spectral angle matching, target information representation, or target information enhancement detectors, and generate detection responses by utilizing the similarity between the prior spectrum and the observed pixels; 2) Traditional hyperspectral unmixing methods: These methods represent the observed hyperspectral image as the product of the endmember dictionary and the abundance matrix, and estimate the mixing ratio of each material endmember and each endmember in the pixel by unmixing; 3) Deep learning target spectrum extraction or detection methods: These methods learn the target spectrum representation or detection response through neural networks, and can improve the adaptability to complex scenes by using a data-driven approach.

[0005] The first type of prior art closest to this invention is a hyperspectral unmixing scheme based on target prior constraints. Its typical process is as follows: input the observed hyperspectral image and the target prior spectrum; construct an endmember dictionary containing target and background endmembers; solve for the abundance matrix by minimizing the reconstruction error; add a similarity constraint with the target prior during the target endmember update process; and finally extract the target endmembers or target abundance as the detection basis. The second type of prior art closest to this invention is an unmixing or detection network scheme based on deep expansion. Its typical process is as follows: first, establish an unmixing or optimization objective function; then, expand several iterative update steps into network layers; learn the step size, threshold, or regularization parameters through end-to-end training; and finally output the endmembers, abundance, or detection response. This type of method attempts to balance the interpretability of the optimization model with the expressive power of the neural network.

[0006] The aforementioned prior art has the following technical defects: Defect 1: The target spectral extraction results are unstable and easily affected by prior mismatch. Because the laboratory target prior differs from the observed state of the real underwater target, and existing methods often directly use this prior or only perform weak corrections, the model cannot determine whether the spectral shift is due to changes in the target's true state or background interference when surface adhesion, sediment cover, or water attenuation causes a spectral shift. This technical logic directly leads to instability in the target spectral extraction results across different aquatic scenarios.

[0007] Defect 2: Numerous false alarms in the background and insufficient separability between target and background responses. Nearshore water background, shallow shoreline, localized strong reflections, and variations in seabed texture can produce spectral responses similar to the target. However, existing unconstrained or weakly constrained unmixing methods do not strictly correlate target endmembers with target abundance. Therefore, some background components enter the target abundance map, resulting in an inflated response in the background region and ultimately leading to a higher false alarm rate.

[0008] Defect 3: Traditional iterative solutions are inefficient, and deep networks lack interpretability. Because constrained unmixed models simultaneously optimize multiple variables with complex constraints, traditional iterative algorithms often require multiple rounds of alternating updates, resulting in computational delays and parameter sensitivity. While ordinary deep networks offer faster inference, they lack explicit physical meaning for endmembers, abundance, and prior transformations. Both suffer from deficiencies in efficiency and interpretability, respectively.

[0009] Defect 4: A single optimization path is ill-suited for complex underwater scenarios. Gradient descent is better suited for stable approximation, while momentum optimization is better suited for rapid advancement; each has different advantages at different stages. Existing fixed update strategies cannot dynamically adjust the update mode based on the network deployment stage and the current variable state, resulting in either slow convergence or oscillations, making it impossible to stably extract the target's intrinsic spectrum.

[0010] In summary, the main shortcomings of existing technologies lie in the lack of a progressive collaborative processing mechanism: first, there is a lack of flexible constraints that allow for reasonable spectral correction of the target prior; second, there is a lack of joint physical modeling of the target endmembers and the target abundance; and finally, there is a lack of a bidirectional dynamic expansion structure that can automatically balance stability and convergence speed according to the optimization stage. Summary of the Invention

[0011] To address the problems in existing technologies such as target prior spectral mismatch, nonlinear degradation of observed spectra, severe mixing of background and water components, and difficulty in stably extracting the inherent spectra of targets in real-world aquatic environments, this invention provides a method for extracting the inherent spectral information of nearshore underwater targets driven by a constrained unmixing model. By incorporating hyperspectral observation images, target prior spectra, endmember dictionaries, abundance matrices, and prior transformation parameters into a unified optimization framework, this method achieves joint estimation of the inherent spectral information and spatial abundance distribution of targets, thereby providing more reliable target priors and detection responses for nearshore underwater target detection. Specifically, the objectives of this invention include the following four aspects, each corresponding to one of the deficiencies described in the background technology: To address the first deficiency, this invention constructs a flexible target prior constraint by introducing prior transformation parameters. A restricted affine transformation is performed on the target prior spectrum to generate a transformed prior, which guides the update direction of the target endmembers. The prior transformation parameters include a linear transformation matrix and a bias term, allowing reasonable spectral shape correction of the target prior within a finite range. Simultaneously, a regularization term in the transformation matrix limits the degree to which the linear transformation matrix deviates from the identity matrix, controlling the correction magnitude of the target prior. This avoids the problem of fixed priors failing to adapt to real underwater spectral shifts and prevents excessive drift of the target endmembers, thus preventing loss of physical directivity.

[0012] To address the second deficiency, this invention unifies target intrinsic spectral extraction and target spatial detection response into a single constrained unmixing framework through endmember-abundance joint modeling and a mask matrix mechanism. The target endmembers in the endmember dictionary correspond to the target's intrinsic spectral information, and the abundance rows of the target endmembers in the abundance matrix correspond to the target detection response map. The introduction of a mask matrix ensures that flexible prior constraints apply only to the target endmembers, while background and water endmembers remain freely updated. The target endmembers maintain the target's physical directionality, and the background endmembers have sufficient degrees of freedom to fit complex backgrounds, thereby reducing confusion between the target and the background and lowering the false alarm rate.

[0013] To address the third defect, this invention employs a deep decomposition approach to expand the alternating optimization process of the constrained unmixing model into a learnable network structure. It expands the iterative processes such as abundance update, endmember update, and prior transformation parameter update into deep network layers, ensuring that the computation of each layer corresponds to a clear optimization step. Simultaneously, through training, it obtains efficient end-to-end inference capabilities, thereby balancing physical interpretability and network inference efficiency.

[0014] To address defect four, this invention constructs a bidirectional dynamic deep unfolding network, setting up parallel robustness-first branches and convergence-first branches. Both branches take the initial optimization variables as input, and each branch contains multiple cascaded optimization modules. Each optimization module contains several optimization layers, and each optimization layer updates variables using either gradient descent or momentum optimization. In the robustness-first branch, the number of optimization layers in gradient descent decreases as the optimization module progresses, while the number of optimization layers in momentum optimization increases, allowing the network to first search the reasonable solution space in a stable manner and then gradually accelerate the convergence speed. In the convergence-first branch, the number of optimization layers in momentum optimization decreases as the optimization module progresses, while the number of optimization layers in gradient descent increases, allowing the network to first quickly traverse flat regions or escape unfavorable local regions, and then finely approximate the stable region. Finally, the outputs of the two branches are weighted and fused, simultaneously utilizing the complementary advantages of stable search and rapid progress.

[0015] To achieve the above objectives, this invention provides a method for extracting intrinsic spectral information of nearshore underwater targets driven by a constrained unmixing model, characterized by comprising the following steps: Step 1: Acquire the nearshore underwater hyperspectral image to be processed and the target prior spectrum, and perform preprocessing on them respectively to obtain the preprocessed observed hyperspectral image and the target prior spectrum after band alignment; Step 2, constructing a constrained unmixing model, including representing the observed hyperspectral image as a combination of an endmember dictionary and an abundance matrix, and constructing flexible prior constraints based on prior transformation parameters and the target prior spectrum, so that they act on the target endmembers of the endmember dictionary; Step 3: Using the endmember dictionary, the abundance matrix, and the prior transformation parameters as optimization variables, construct the joint optimization objective function of the restricted unmixing model, and initialize the endmember dictionary, the abundance matrix, and the prior transformation parameters to obtain the initial optimization variables; Step 4: Expand the optimization process of the restricted unmixing model into a bidirectional dynamic deep expansion network, and use the initial optimization variables as the starting point to perform multi-stage joint optimization and update of the optimization variables based on the bidirectional dynamic deep expansion network; Step 5: Output the target endmembers in the optimized backend meta-dictionary as the target's inherent spectral information, and output the target abundance distribution of the corresponding target endmembers in the optimized abundance matrix as the target detection response map.

[0016] Compared with the prior art, the present invention has the following beneficial technical effects: 1. This invention, through a flexible target prior constraint mechanism, allows the target prior to be adaptively corrected within a limited range while maintaining the target's physical orientation. It can achieve a balance between "maintaining the physical relevance of laboratory priors" and "adapting to changes in the real underwater environment," effectively mitigating the prior mismatch caused by water transport, attached organisms, bottom sedimentation, and changes in target attitude, while also preventing the target endmember from being absorbed by the complex background. 2. This invention unifies target intrinsic spectrum extraction and target abundance detection into an endmember-abundance joint optimization framework, so that the model no longer simply outputs a black box detection result, but simultaneously provides target intrinsic spectrum and spatial abundance response with physical meaning. The target endmember is used to characterize the target's own spectral properties, and the target abundance is used to indicate the spatial distribution of the target in the image. The two constrain each other, reducing the logical disconnect between the target spectrum and the detection response, and reducing background false alarms. 3. This invention adopts a bidirectional dynamic depth unfolding backbone, combining the stability of gradient descent and the speed of momentum optimization. By adaptively fusing and dynamically adjusting the solution strategy at different stages, it balances solution accuracy and convergence efficiency. 4. This invention is driven by a constrained unmixed physical model, with each layer of the network corresponding to a clear optimization step, which has good interpretability. Moreover, it adopts an unsupervised training method, does not rely on pixel-level spatial annotation, and is suitable for underwater target detection scenarios where annotations are scarce. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the method for extracting intrinsic spectral information of nearshore underwater targets driven by a constrained unmixing model in an embodiment of the present invention.

[0019] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0022] This embodiment discloses a method for extracting intrinsic spectral information of nearshore underwater targets driven by a constrained unmixing model. It is used to extract intrinsic spectra that more closely approximate the actual state of the target from real nearshore underwater hyperspectral observation images, or for tasks such as shallow water sediment identification, underwater facility detection, coastal zone target monitoring, and spectral extraction of complex background materials. (Reference) Figure 1 The method specifically includes the following steps: Step 1: Acquire the nearshore underwater hyperspectral image to be processed and the target prior spectrum, and perform preprocessing on them respectively to obtain the preprocessed observed hyperspectral image and the target prior spectrum after band alignment; Step 2: Construct a constrained unmixing model, which includes representing the observed hyperspectral image as a combination of an endmember dictionary and an abundance matrix, and constructing flexible prior constraints based on prior transformation parameters and the target prior spectrum, so that they act on the target endmembers of the endmember dictionary. Step 3: Using the endmember dictionary, abundance matrix, and prior transformation parameters as optimization variables, construct the joint optimization objective function of the constrained unmixing model, and initialize the endmember dictionary, abundance matrix, and prior transformation parameters to obtain the initial optimization variables; Step 4: Expand the optimization process of the constrained unmixing model into a bidirectional dynamic deep expansion network, and start from the initial optimization variables, perform multi-stage joint optimization and update of the optimization variables based on the bidirectional dynamic deep expansion network; Step 5: Output the target endmembers in the optimized backend meta-dictionary as the target's inherent spectral information, and output the target abundance distribution of the corresponding target endmembers in the optimized abundance matrix as the target detection response map.

[0023] In this embodiment, the acquired nearshore underwater hyperspectral image is specifically an observation matrix containing L spectral bands and N spatial pixels; the target prior spectrum is a one-dimensional spectral vector, which can come from laboratory measurements, historical sample databases, on-site calibration or confirmed physical simulation spectral databases, target pixel average spectra, target material spectral databases, etc.

[0024] The preprocessing of nearshore underwater hyperspectral images includes invalid band removal, outlier handling, radiative amplitude normalization, band scale unification, and data matrixing to obtain preprocessed observational hyperspectral images. This eliminates numerical interference caused by differences in the dimensions and amplitudes of different bands on subsequent joint optimization of endmembers and abundance. Preprocessing of the target prior spectrum includes band alignment and scale unification. Specifically, if the target prior band range is not completely consistent with the observed image (i.e., the nearshore underwater hyperspectral image or the preprocessed observational hyperspectral image), interpolation, cropping, or resampling based on the observed image bands is required to obtain the band-aligned target prior spectrum. This ensures a one-to-one correspondence between the target prior spectrum and the observed hyperspectral image in the band dimension, guaranteeing that subsequent target prior constraints can directly apply to the target endmembers in the endmember dictionary.

[0025] In this embodiment, the construction process of the restricted unmixing model specifically includes: First, establish an endmember-abundance unmixing representation: observe hyperspectral images Represented as an end-member dictionary With abundance matrix The combination of , where the end-member dictionary Including target endmembers, background endmembers, and water-related endmembers, abundance matrix This represents the mixing ratio of each endmember in different spatial pixels; under the endmember-abundance unmixing representation, the extraction of the target's intrinsic spectral information corresponds to solving the target endmembers in the endmember dictionary, and the target detection corresponds to estimating the spatial abundance distribution of the target endmembers; Then, specify the target endmember location: to avoid confusion between target and background endmembers during unconstrained demixing, the endmember dictionary is... The first column is set as the target end cell and the target prior spectrum As the physical guiding signal for the target endmember, other endmembers are used to fit the underwater background, water path radiation, and other non-target components. Then, flexible target prior constraints are constructed: considering that the spectrum of the real underwater target will shift compared to the laboratory prior, a prior transformation parameter is introduced to affect the target prior spectrum. A constrained transformation is performed to generate a priori transformations, the priori transformation parameters of which include the linear transformation matrix. and bias Through linear transformation matrix and bias Prior spectrum of the target Perform affine transformation It allows for reasonable spectral shape corrections to the target prior within a limited range, and does not require the target endmember to be correct. With the target prior spectrum Completely identical; Finally, a mask matrix is ​​introduced. Controlling the scope of prior constraints: mask matrix Select only end-member dictionary The position of the target endmember corresponds to the position of the target endmember, so that the flexible prior constraint mainly acts on the target endmember, while the background endmember and water endmember are kept to be updated freely. This allows the target endmember to maintain the physical orientation of the target, while the background endmember still has enough degrees of freedom to fit the complex background.

[0026] In practical applications, constructing flexible target prior constraints is not limited to using affine transformation with linear transformation and bias. It can also use diagonal scaling, low-rank transformation, piecewise linear transformation, band weight transformation, spectral smoothing transformation, or small learnable mapping networks, as long as the constraint that limits excessive target prior offset is retained.

[0027] The joint optimization objective function in this embodiment includes an observation reconstruction error term, an objective prior constraint term, an abundance sparse regularization term, and a transformation matrix regularization term, specifically expressed as follows:

[0028] The observation reconstruction error term is used to explain the true observation image by considering endmembers and abundance. These are prior constraints for the target, used to ensure that the target endmembers do not deviate from the target's physical properties; This is an abundance sparsity regularization term used to make the target response more spatially concentrated and reduce the residue of background endmembers in the target abundance map; The transformation matrix regularization term is used to limit the magnitude of prior corrections and prevent the model from erroneously absorbing background changes into the target spectral correction process; the joint optimization objective function also includes... , And constraints such as the sum of the abundance of each pixel being 1. Among them, These are the prior constraint weights for the target, used to control the strength of the prior constraint on the target. is the abundance sparsity regularization coefficient, used to control the spatial sparsity of abundance; These are the regularization coefficients of the transformation matrix, used to restrict the linear transformation matrix. Deviation from identity matrix The degree of [something]. During application, it is preferable to set it to [something]. , , Alternatively, other parameters can be used in the design.

[0029] In this embodiment, the process of initializing the endmember dictionary is as follows: the target prior spectrum is placed in the first column of the endmember dictionary as the initial value of the target endmember. At the same time, the background dominant endmembers are extracted from the observed hyperspectral image using an endmember extraction algorithm to form a complete initial endmember dictionary, thus avoiding the network from getting stuck in random solutions lacking physical meaning in the early stages of training. The endmember extraction algorithm can employ algorithms such as VCA, N-FINDR, and K-means, or it can use methods such as Robust-OSP, PCA-assisted initialization, random initialization, or sampling from non-target regions.

[0030] In this embodiment, the process of initializing the abundance matrix is ​​as follows: after determining the initial set of endmembers, the non-negative least squares method is used to invert the entire observed hyperspectral image to obtain the initial abundance distribution of each pixel corresponding to each endmember, providing an initial estimate that satisfies non-negativity and the physical meaning of the mixing ratio for subsequent depth unfolding optimization.

[0031] In this embodiment, the process of initializing the prior transformation parameters is as follows: considering the linear transformation matrix and bias It is mainly used to describe the shift of the actual observed target spectrum relative to the target prior. The linear transformation matrix is ​​initialized as an identity matrix and the bias term is initialized as a zero vector or zero matrix, indicating that no significant deformation of the target prior is assumed in the initial state, and the reasonable correction amount is then learned gradually by data-driven approach.

[0032] In this embodiment, the bidirectional dynamic deep unfolding network includes parallel robustness-first branches and convergence-first branches. Both robustness-first and convergence-first branches take the initial optimization variables as input, and each includes multiple cascaded optimization modules. Each optimization module includes several cascaded optimization layers, and the optimization layers update the endmember dictionary, abundance matrix, and prior transformation parameters using gradient descent or momentum optimization modes.

[0033] The update process of gradient descent mode is as follows: First, fix the current endmember dictionary and prior transformation parameters, and update the abundance matrix by gradient according to the observation reconstruction error term and abundance sparsity regularization term in the joint optimization objective function. Combine sparse near-end mapping to suppress irrelevant background response. Then apply non-negative projection and sum-to-one projection to the abundance matrix so that the abundance of each pixel still has the meaning of physical mixing ratio. Then, the updated abundance matrix and the current prior transformation parameters are fixed, and the endmember dictionary is updated according to the gradient of the observation reconstruction error term and the target prior constraint term in the joint optimization objective function. Since the mask matrix only acts on the target endmember, the target endmember will be guided by the target prior, while the background endmember is mainly adjusted freely according to the observation reconstruction error. After the update, a non-negative projection is applied to the endmember dictionary to ensure the physical rationality of the spectral response. Finally, the updated endmember dictionary and abundance matrix are fixed, and the prior transformation parameters are updated according to the target prior constraint term and transformation matrix regularization term in the joint optimization objective function, so that the target prior can learn the necessary spectral shift in the real underwater scene but will not arbitrarily drift into the background spectrum.

[0034] The update process of the momentum optimization mode is as follows: Construct the extrapolated state based on the current endmember dictionary, abundance matrix, prior transformation parameters, and corresponding variables from the previous optimization layer; The complete update process of gradient descent mode is executed in the extrapolated state. The update step size is controlled by a learnable step size or sufficient descent condition. Historical update direction is used to accelerate convergence while suppressing update oscillation.

[0035] It should be noted that in specific applications, gradient descent can be replaced by proximal gradient, projected gradient, quasi-Newton update or Adam-type update, and momentum optimization can be replaced by Nesterov momentum, heavy ball momentum or adaptive momentum mechanism.

[0036] In the robustness-first approach, the number of optimization layers in gradient descent mode decreases as the optimization module progresses, while the number of optimization layers in momentum optimization mode increases as the optimization module progresses. Specifically, all optimization layers in the first optimization module use gradient descent mode, allowing the network to search the reasonable solution space in a stable manner first; all optimization layers in the last optimization module use momentum mode to improve the overall convergence speed; in the intermediate optimization modules, the number of optimization layers in gradient descent mode decreases module by module, while the number of optimization layers in momentum optimization mode increases module by module, achieving a smooth transition from stable search to rapid progress.

[0037] In the convergence-first branch, the number of optimization layers in the momentum optimization mode decreases as the optimization module progresses, while the number of optimization layers in the gradient descent mode increases as the optimization module progresses. Specifically, in the first optimization module, all optimization layers use the momentum mode to prioritize rapid traversal of flat regions or escape from undesirable local regions; in the last optimization module, all optimization layers use the gradient descent mode to finely approximate the solution to the stable region; in the intermediate optimization modules, the number of optimization layers in the momentum optimization mode decreases module by module, while the number of optimization layers in the gradient descent mode increases module by module, achieving a smooth transition from rapid progress to stable convergence.

[0038] After all optimization modules of the robustness-first branch and the convergence-first branch have been executed, the endmember dictionary, abundance matrix, and prior transformation parameters output by the robustness-first branch, as well as the endmember dictionary, abundance matrix, and prior transformation parameters output by the convergence-first branch, are obtained. The endmember dictionaries output by the two branches are weighted and fused to obtain the optimized endmember dictionary; the abundance matrices output by the two branches are weighted and fused to obtain the optimized abundance matrix. The weights for weighting can be learnable parameters or generated from the current variable state, and can be implemented using scalar weights, channel weights, band weights, gating networks, or attention mechanisms.

[0039] In the specific implementation, the bidirectional dynamic depth unfolding network is trained in an unsupervised manner. Since hyperspectral underwater target detection tasks often only provide the target's prior spectrum and lack reliable pixel-level spatial annotations, the training process in this embodiment does not rely on manually annotated target masks. Instead, it trains the model by constraining the difference between the input hyperspectral image and the network reconstruction result. The training loss includes composite reconstruction loss, target prior constraint loss, abundance sparsity constraint loss, and transformation regularization loss. Specifically, the composite reconstruction loss measures the difference between the observed hyperspectral image and the reconstructed image from both Euclidean space and spectral perspectives, ensuring the model maintains both numerical reconstruction accuracy and spectral shape consistency. The target prior constraint loss is the target prior constraint term in the joint optimization objective function, used to drive the target endmembers to approach the target's physical spectrum. The abundance sparsity constraint loss is the abundance sparse regularization term in the joint optimization objective function, used to suppress target abundance residues in the background region. The transformation matrix regularization term in the joint optimization objective function is used to limit the magnitude of prior corrections.

[0040] In practical applications, the reconstruction error metric can also use Frobenius norm, L1 norm, Huber loss, spectral angle distance, weighted band reconstruction loss or a combination of multiple losses. In terms of abundance constraints, Lp sparsity, L1 sparsity, group sparsity, total variational spatial smoothing, target region connectivity constraints or spatial neighborhood consistency constraints can be used.

[0041] After training, the target endmember output by the network serves as a new target spectral prior. This target endmember is not a simple copy of the original laboratory prior, but rather an adaptive target intrinsic spectrum obtained under the combined effects of real observation images, water transport effects, background mixing information, and flexible prior constraints.

[0042] After bidirectional dynamic expansion and weighted fusion, the network outputs an optimized endmember dictionary and abundance matrix. The first column of the endmember dictionary corresponds to the extracted intrinsic spectral information of the target, and the first row of the abundance matrix corresponds to the abundance distribution of target endmembers at various spatial locations. The target endmembers are output as the intrinsic spectrum of the target, which can be used as prior input for different types of downstream hyperspectral underwater target detection methods such as TUTDF, SSDF, and NCSS-Net to improve target characterization, target information enhancement, or background suppression. The target abundance distribution is output as a target detection response map; a higher abundance value indicates a greater contribution of the target endmembers in that spatial pixel. Subsequent steps include threshold segmentation, sorting and filtering, connected component analysis, or fusion with downstream detectors to obtain the final nearshore underwater target detection results. Output quality can be evaluated using metrics such as detection rate, false alarm rate, ROC curve, and AUC.

[0043] The above description is only a preferred embodiment of the present invention and does not limit the scope of protection of the present invention. All equivalent structural transformations made under the inventive concept of the present invention using the contents of the present invention specification and drawings, or direct / indirect applications in other related technical fields, are included within the scope of protection of the present invention.

Claims

1. A method for extracting intrinsic spectral information of nearshore underwater targets driven by a constrained unmixing model, characterized in that, Includes the following steps: Step 1: Acquire the nearshore underwater hyperspectral image to be processed and the target prior spectrum, and perform preprocessing on them respectively to obtain the preprocessed observed hyperspectral image and the target prior spectrum after band alignment; Step 2, constructing a constrained unmixing model, including representing the observed hyperspectral image as a combination of an endmember dictionary and an abundance matrix, and constructing flexible prior constraints based on prior transformation parameters and the target prior spectrum, so that they act on the target endmembers of the endmember dictionary; Step 3: Using the endmember dictionary, the abundance matrix, and the prior transformation parameters as optimization variables, construct the joint optimization objective function of the restricted unmixing model, and initialize the endmember dictionary, the abundance matrix, and the prior transformation parameters to obtain the initial optimization variables; Step 4: Expand the optimization process of the restricted unmixing model into a bidirectional dynamic deep expansion network, and use the initial optimization variables as the starting point to perform multi-stage joint optimization and update of the optimization variables based on the bidirectional dynamic deep expansion network; Step 5: Output the target endmembers in the optimized backend meta-dictionary as the target's inherent spectral information, and output the target abundance distribution of the corresponding target endmembers in the optimized abundance matrix as the target detection response map.

2. The method for extracting intrinsic spectral information of nearshore underwater targets driven by a constrained unmixing model according to claim 1, characterized in that, Step 1, specifically the preprocessing of the nearshore underwater hyperspectral image and the target prior spectrum, includes: The nearshore underwater hyperspectral image is processed by removing invalid bands, handling outliers, normalizing radiation amplitude, unifying band scales, and matrixing data to obtain the preprocessed observation hyperspectral image. Band alignment and scale uniformity are performed on the target prior spectrum. If the band range of the target prior spectrum is not completely consistent with the observed image, interpolation, cropping or resampling are performed based on the bands of the observed image to obtain the band-aligned target prior spectrum.

3. The method for extracting intrinsic spectral information of nearshore underwater targets driven by a constrained unmixing model according to claim 1, characterized in that, Step 2, the construction process of the restricted unmixing model specifically includes: The observed hyperspectral image is represented as a combination of an endmember dictionary and an abundance matrix, wherein the endmember dictionary includes target endmembers, background endmembers and water-related endmembers, and the abundance matrix represents the mixing ratio of each endmember in different spatial pixels; The first column of the endmember dictionary is set as the target endmember, and the target prior spectrum is constructed as a flexible target prior constraint based on the prior transformation parameters. A mask matrix is ​​introduced, which selects only the position of the target endmember in the endmember dictionary, so that the flexible target prior constraint only applies to the target endmember, while maintaining the free updating of the background endmember and the water-related endmember.

4. The method for extracting intrinsic spectral information of nearshore underwater targets driven by a constrained unmixing model according to claim 3, characterized in that, In step 3, the joint optimization objective function is specifically as follows: in, To observe hyperspectral images, For end-word dictionary, For abundance matrix, For the mask matrix, The linear transformation matrix in the prior transformation parameters, The bias in the prior transformation parameters. For the target prior spectrum, It is the identity matrix. The prior constraint weights for the target This is the abundance sparsity regularization coefficient. These are the regularization coefficients of the transformation matrix.

5. The method for extracting intrinsic spectral information of nearshore underwater targets driven by a constrained unmixing model according to claim 4, characterized in that, Step 3, specifically the initialization process of the endmember dictionary, the abundance matrix, and the prior transformation parameters, is as follows: The target prior spectrum is placed in the first column of the endmember dictionary as the initial value of the target endmember. At the same time, the background dominant endmember is extracted from the observed hyperspectral image using the endmember extraction algorithm to form the initial endmember dictionary. The initial abundance matrix was obtained by inverting the observed hyperspectral images using the non-negative least squares method. The linear transformation matrix in the prior transformation parameters is initialized to an identity matrix, and the bias is initialized to a zero vector or a zero matrix.

6. The method for extracting intrinsic spectral information of nearshore underwater targets driven by a constrained unmixing model according to claim 4 or 5, characterized in that, The bidirectional dynamic deep unfolding network includes parallel robustness-first branches and convergence-first branches. Both the robustness-first branch and the convergence-first branch take the initial optimization variable as input, and both the robustness-first branch and the convergence-first branch include multiple cascaded optimization modules. Each optimization includes several cascaded optimization layers, and the optimization layers update the optimization variable using gradient descent mode or momentum optimization mode. By weighted and fused together the endmember dictionary output by the robustness-first branch and the convergence-first branch, and the abundance matrix optimization variables, we obtain the optimized endmember dictionary and the optimized abundance matrix.

7. The method for extracting intrinsic spectral information of nearshore underwater targets driven by a constrained unmixing model according to claim 6, characterized in that, In the robustness-first branch, the number of optimization layers in the gradient descent mode decreases as the optimization module progresses, while the number of optimization layers in the momentum optimization mode increases as the optimization module progresses. In the convergence priority branch, the number of optimization layers in the momentum optimization mode decreases as the optimization module progresses, while the number of optimization layers in the gradient descent mode increases as the optimization module progresses.

8. The method for extracting intrinsic spectral information of nearshore underwater targets driven by a constrained unmixing model according to claim 6, characterized in that, The gradient descent mode update process specifically includes: First, fix the current endmember dictionary and prior transformation parameters, and perform gradient updates on the abundance matrix based on the observation reconstruction error term and abundance sparse regularization term in the joint optimization objective function. Then, apply non-negative projection and one-projection to the abundance matrix. Then, with the updated abundance matrix and the current prior transformation parameters fixed, the endmember dictionary is updated according to the gradients of the observation reconstruction error term and the objective prior constraint term in the joint optimization objective function, and a non-negative projection is applied to the endmember dictionary. Finally, the updated endmember dictionary and abundance matrix are fixed, and the prior transformation parameters are updated according to the objective prior constraint term and transformation matrix regularization term in the joint optimization objective function.

9. The method for extracting intrinsic spectral information of nearshore underwater targets driven by a constrained unmixing model according to claim 8, characterized in that, The update process of the momentum optimization mode specifically includes: First, construct the extrapolated state based on the current endmember dictionary, abundance matrix, prior transformation parameters, and corresponding variables from the previous optimization layer; Then, the gradient descent mode update process is performed on the extrapolated state.

10. The method for extracting intrinsic spectral information of nearshore underwater targets driven by a constrained unmixing model according to claim 8, characterized in that, The bidirectional dynamic deep unfolding network is trained in an unsupervised manner, and the training loss includes composite reconstruction loss, target prior constraint loss, abundance sparsity constraint loss and transformation regularization loss. The composite reconstruction loss measures the difference between the observed hyperspectral image and the reconstructed image from both Euclidean space and spectral perspectives.

Images