An estuary salt tide inversion method driven by physical mechanism and data

Abstract

The application discloses an estuary salinity inversion method driven by physical mechanism and data, and belongs to the cross field of remote sensing and water environment monitoring. After high-spectral remote sensing data and spatial coordinates are acquired, a joint spectral encoder is used for adaptive band selection and bidirectional Mamba sequence scanning, and absorption and scattering characteristics are decoupled. Meanwhile, a physical guided spectral decomposition module is used to extract physical priors based on an original salinity response base. Then, a tidal perception spatial encoder is used to extract gradient attenuation characteristics of a relative salt front. Finally, an asymmetric cross-modal attention mechanism is used for space-time-spectral fusion, and an uncertainty prediction head is input. The application introduces a composite loss function with physical prior constraints such as salinity monotonicity and salinity order, solves the problems that a pure data-driven model lacks physical interpretability and is easy to violate hydrodynamic common sense, and realizes water quality inversion with high precision and confidence quantification in large-scale estuary salinity monitoring.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of remote sensing and water environment monitoring, and in particular to a method for inverting estuarine saltwater intrusion driven by both physical mechanisms and data. Background Technology

[0002] Hyperspectral images, containing hundreds of consecutive narrow bands, can capture subtle differences in the spectral absorption and scattering characteristics of water components, giving them an irreplaceable advantage in the retrieval of inland and estuarine water quality parameters (such as conductivity, total phosphorus, and turbidity). Especially in large estuaries (such as the Yangtze River estuary), saltwater intrusion (i.e., the backflow of high-salinity, high-conductivity water into inland rivers) is frequent, posing a significant threat to regional ecosystems and urban water supply security, necessitating high-precision, large-scale active monitoring methods. However, current technologies for hyperspectral retrieval of complex water environments such as saltwater intrusion still face the following significant challenges and technical bottlenecks:

[0003] 1. Long sequence modeling is inefficient and lacks global context:

[0004] Existing deep learning architectures have inherent limitations when processing high-dimensional spectral sequences. Convolutional neural networks (CNNs) have limited receptive fields, making it difficult to capture global correlations across bands; while Transformer-based architectures have global receptive fields, the computational complexity of their self-attention mechanisms increases quadratically with the sequence length, posing a significant computational bottleneck in high-frequency, large-scale remote sensing monitoring.

[0005] 2. Purely data-driven models lack physical interpretability and generalization ability:

[0006] Most current water quality inversion networks are purely "black box" models, relying entirely on data-driven approaches. In estuarine regions, facing strong spatial heterogeneity and high-dimensional noise interference, black box models are highly prone to overfitting. Due to the lack of explicit modeling of water optical mechanisms (such as absorption, scattering, and the slight enhancement effect of electrolyte concentration on refractive index), these models often produce results that violate hydrodynamic principles when faced with small sample limitations (for example, predicting that the conductivity of the upstream channel is actually higher than that of the downstream estuary).

[0007] 3. Fixed spectral indices are difficult to adapt to spatial heterogeneity:

[0008] Traditional remote sensing water quality monitoring relies heavily on manually designed fixed-band spectral indices (such as NDWI and NDSI). However, the composition of estuarine water is extremely complex, and the characteristic combinations of fixed bands cannot be adaptively adjusted according to specific sensor characteristics and environmental dynamics at different tidal periods, making it difficult to capture the nonlinear optical response of complex water bodies. Summary of the Invention

[0009] Based on this, this application provides a method for inverting estuarine salinity under the dual drive of physical mechanism and data. It aims to construct an efficient bidirectional state-space model for global spectral feature representation and innovatively introduces physical guided spectral decomposition (PGSD) and tidal sensing spatial coding mechanism to achieve high-precision and interpretable water quality and conductivity inversion that conforms to the laws of hydrodynamic physics while maintaining extremely low computational overhead.

[0010] This application provides a method for inverting estuarine salinity based on both physical mechanisms and data, the method comprising:

[0011] Step S1: Obtain the remote sensing hyperspectral sequence data matrix of the target water area. and the corresponding water surface spatial standardized coordinate data matrix ;

[0012] Step S2: The remote sensing hyperspectral sequence data matrix The input is fed into a joint spectral encoder, where weighted spectral features are extracted using an adaptive band gating mechanism. Simultaneously, bidirectional state-space model sequence scanning and multi-scale spectral convolutional pyramid feature extraction are performed, decoupled to output absorption features specific to the water body. With scattering characteristics ;

[0013] Step S3: The remote sensing hyperspectral sequence data matrix The input is fed into the physics-guided spectral decomposition module, which predicts decomposition coefficients and extracts physical prior features based on a pre-defined physical dictionary that includes a salinity response basis. ;

[0014] Step S4: Standardize the spatial coordinate data matrix of the water surface. The data is input into the tidal sensing spatial encoder, and multi-scale radial basis function gradient decay features are extracted by calculating the relative distance along the flow from the coordinate point to the dynamic salt front reference point. ;

[0015] Step S5: Apply the absorption feature The scattering characteristics With the aforementioned physical prior features The spliced ​​spectral modal features and the gradient attenuation features The input is fed into an asymmetric cross-modal attention module, where spatial features guide the attention of the spectral channels to obtain spatiotemporal-spectral fusion features. ;

[0016] Step S6: Combine the spatiotemporal-spectral fusion features Input to the uncertainty prediction head, output the inversion prediction mean of the target water quality parameters. and the corresponding log-variance of cognitive uncertainty And employing Gaussian likelihood loss Saltwater intrusion monotonicity constraint loss and salinity order constraint loss The composite loss function includes Perform model optimization.

[0017] Optionally, the feature extraction calculation process in step S2, which extracts weighted spectral features through an adaptive band gating mechanism, specifically includes:

[0018] ;

[0019] ;

[0020] in, This is the mapping weight matrix for the adaptive band gating mechanism. and The weight matrix is ​​a learnable linear mapping. and For activation function, This is a Hadamard product operation, used to suppress noise bands outside the atmospheric window and output weighted spectral features. ;

[0021] The bidirectional state-space model sequence scan includes a forward scan path and a reverse scan path. The specific calculation formula includes:

[0022] Forward scan feature output ;

[0023] Reverse scan feature output ;

[0024] Bidirectional global feature fusion ;

[0025] in, This represents a sequence modeling operation performed using discretized continuous state-space equations. Indicates the dimension along the band sequence The flip operation, This indicates a splicing operation at the channel level. Presentation layer normalization operation, This represents the learnable channel feature merging weight matrix.

[0026] Optionally, the multi-scale spectral convolution pyramid in step S2 utilizes three one-dimensional convolutions with different receptive fields to capture spectral morphological features. The calculation process specifically includes:

[0027]

[0028] in, Indicates the kernel size as One-dimensional convolution operations are used to capture abrupt spectral transitions, medium-broad absorption peaks, and broad envelope features, respectively. Used to compress multi-scale outputs to a uniform dimension.

[0029] Optionally, step S2 further includes calculating a learnable water quality spectral index by defining two sets of soft band selector weights. and Through the formula:

[0030]

[0031] Normalized difference characteristics ;

[0032] Dynamic ratio characteristics ;

[0033] in, , These are the water quality characteristic spectral indices obtained through weighted aggregation using an adaptive soft band selector. It is a very small constant. This is the preset clamping threshold.

[0034] Optionally, the physical-guided spectral decomposition module in step S3 specifically includes the following physical reconstruction and feature extraction calculation process:

[0035]

[0036]

[0037] in, For physically reconstructed hyperspectral sequence data matrix, The output mapping weight matrix of the physics-guided spectral decomposition module. , These are the preset absorption substrate and scattering substrate, respectively; For modeling the electrolytic optical effect on a salinity-responsive substrate; , , This is the corresponding non-negative decomposition coefficient matrix.

[0038] Optionally, the tidal sensing spatial encoder in step S4 has a relative distance along the flow. Radial basis attenuation characteristics The calculation process specifically includes:

[0039]

[0040]

[0041] in, For learnable flow direction angle parameters, As a learnable dynamic salt front reference longitude parameter, and These are the standardized spatial coordinates; The scale parameter is learned in logarithmic space and is used to model the physical properties of salinity exhibiting multi-scale exponential decay as it moves upstream along the river channel.

[0042] Optionally, the asymmetric cross-modal attention module in step S5 selectively activates spectral modal features based on spatial modal features, and its attention weights... With fusion features The calculation process specifically includes:

[0043]

[0044]

[0045] in, For spectral modal characteristics, All are mapping weight matrices. , where is the dimension of the attention hidden layer.

[0046] Optionally, the uncertainty prediction head in step S6 specifically includes the following output calculation process:

[0047]

[0048]

[0049] in, To share the weights of the feature extraction layer, and These are the weight matrices for the mean head and variance head, respectively.

[0050] Optionally, the composite loss function in step S6 Includes task likelihood loss based on Gaussian negative log-likelihood. The calculation process specifically includes:

[0051]

[0052] in, The number of samples used in the calculation. The actual observed values ​​of the target water quality parameters;

[0053] The composite loss function It also includes the salinity monotonicity constraint loss. This is used to penalize anomalies where the predicted conductivity of the upstream is higher than that of the downstream. The calculation process specifically includes:

[0054]

[0055] in, For random sampling of spatial point pairs within a batch Total quantity This is an indicator function that takes the value when the condition is true. Otherwise take .

[0056] Optionally, the composite loss function It also includes salinity order constraint loss. The salinity response coefficient used to constrain the extraction Compared with the actual conductivity label The order of magnitude and rank are positively correlated, and the calculation process specifically includes:

[0057]

[0058] in, The number of point pairs sampled. It is a symbolic function.

[0059] The beneficial effects of the technical solutions provided in this application include at least the following:

[0060] (1) Breaking through the bottleneck of long sequence modeling: This application constructs a bidirectional Mamba scanning mechanism, which solves the problem of unidirectional context loss while maintaining linear computational complexity, and realizes efficient global reasoning of large-scale spectral images.

[0061] (2) Physically guided spectral decomposition: The unique "salinity response substrate" is introduced to accurately model the optical effect of saltwater intrusion, breaking the black box limitation of pure data-driven model and greatly improving the physical interpretability of features and the generalization ability of the model.

[0062] (3) Perfectly fits the laws of hydrodynamics: Combines tidal sensing spatial coding with physical composite loss function for correction, forces the model to follow the common sense of saline space decay, and ensures industrial-grade robustness under small sample conditions.

[0063] (4) Provide quantitative uncertainty assessment: By simultaneously outputting the spatial distribution map of the predicted mean and uncertainty variance through Gaussian likelihood modeling, it provides a direct quantitative decision basis for business departments to assess the reliability of remote sensing data. Attached Figure Description

[0064] To more clearly illustrate the embodiments of this application or the technical solutions in the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0065] Figure 1 This is a schematic diagram of the structure of a combined spectral encoder (including bidirectional Mamba) provided in an embodiment of this application;

[0066] Figure 2 A schematic diagram of the Physically Guided Spectral Decomposition (PGSD) module provided in an embodiment of this application. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0068] In the description of this application, the terms "comprising," "having," and any variations thereof are intended to cover non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those steps or units that are expressly listed, but may also include other steps or units that are not expressly listed but are inherent to these processes, methods, products, or apparatuses, or steps or units added based on further optimizations conceived in this application.

[0069] Please refer to Figure 1 The document illustrates a flowchart of a method for retrieving estuarine saltwater intrusion driven by both physical mechanisms and data, as provided in an embodiment of this application. This method may include the following steps:

[0070] Step S1: Obtain the remote sensing hyperspectral sequence data matrix of the target water area. and the corresponding water surface spatial standardized coordinate data matrix .

[0071] Step S2: Convert the remote sensing hyperspectral sequence data matrix The input is fed into a joint spectral encoder, where weighted spectral features are extracted using an adaptive band gating mechanism. Simultaneously, bidirectional state-space model sequence scanning and multi-scale spectral convolutional pyramid feature extraction are performed, decoupled to output absorption features specific to the water body. With scattering characteristics .

[0072] In this step, the bidirectional state-space model sequence scan in step S2 includes a forward scan path and a reverse scan path constructed using the discretized continuous state-space equations. The reverse scan path scans the sequence features along the band dimension. Perform two flipping operations This allows each band to be given complete bidirectional global context information after being spliced ​​together in the channel dimension.

[0073] Specifically, the feature extraction calculation process for extracting weighted spectral features through an adaptive band gating mechanism includes:

[0074] ;

[0075] ;

[0076] in, This is the mapping weight matrix for the adaptive band gating mechanism. and The weight matrix is ​​a learnable linear mapping. and For activation function, This is a Hadamard product operation, used to suppress noise bands outside the atmospheric window and output weighted spectral features. ;

[0077] The bidirectional state-space model sequence scan includes a forward scan path and a reverse scan path. The specific calculation formula includes:

[0078] Forward scan feature output ;

[0079] Reverse scan feature output ;

[0080] Bidirectional global feature fusion ;

[0081] in, This represents a sequence modeling operation performed using discretized continuous state-space equations. Indicates the dimension along the band sequence The flip operation, This indicates a splicing operation at the channel level. Presentation layer normalization operation, This represents the learnable channel feature merging weight matrix.

[0082] In this step, the multi-scale spectral convolution pyramid captures spectral morphological features using three one-dimensional convolutions with different receptive fields. The calculation process specifically includes:

[0083]

[0084] in, Indicates the kernel size as One-dimensional convolution operations are used to capture abrupt spectral transitions, medium-broad absorption peaks, and broad envelope features, respectively. Used to compress multi-scale outputs to a uniform dimension.

[0085] It also includes the calculation of learnable water quality spectral indices, by defining two sets of soft band selector weights. and Through the formula:

[0086]

[0087] Normalized difference characteristics ;

[0088] Dynamic ratio characteristics ;

[0089] in, , These are the water quality characteristic spectral indices obtained through weighted aggregation using an adaptive soft band selector. It is a very small constant. This is the preset clamping threshold.

[0090] Step S3: Convert the remote sensing hyperspectral sequence data matrix The input is fed into the physics-guided spectral decomposition module, which predicts decomposition coefficients and extracts physical prior features based on a pre-defined physical dictionary that includes a salinity response basis. .

[0091] In this step, the physical reconstruction calculation of the physics-guided spectral decomposition module is as follows: ;

[0092] in, For a salinity-responsive substrate specifically designed for modeling the optical effects of electrolytes, This is the predicted salinity response coefficient.

[0093] The physics-guided spectral decomposition module, specifically includes the following physical reconstruction and feature extraction calculation processes:

[0094]

[0095]

[0096] in, For physically reconstructed hyperspectral sequence data matrix, The output mapping weight matrix of the physics-guided spectral decomposition module. , These are the preset absorption substrate and scattering substrate, respectively; For modeling the electrolytic optical effect on a salinity-responsive substrate; , , This is the corresponding non-negative decomposition coefficient matrix.

[0097] Step S4: Standardize the spatial coordinate data matrix of the water surface. The data is input into the tidal sensing spatial encoder, and multi-scale radial basis function gradient decay features are extracted by calculating the relative distance along the flow from the coordinate point to the dynamic salt front reference point. .

[0098] In this step, the relative distance along the current calculated by the tidal sensing spatial encoder in step S4 is: ,

[0099] in For the flow direction angle parameter, For dynamic salt front parameters; using the formula Extract multi-scale spatial decay features.

[0100] Step S5: Absorption characteristics scattering characteristics With physical prior characteristics Spectral modal characteristics and gradient attenuation characteristics formed by splicing The input is fed into an asymmetric cross-modal attention module, where spatial features guide the attention of the spectral channels to obtain spatiotemporal-spectral fusion features. .

[0101] In this step, the asymmetric cross-modal attention module in step S5 selectively activates spectral modal features based on spatial modal features, and its attention weights... With fusion features The calculation process specifically includes:

[0102]

[0103]

[0104] in, For spectral modal characteristics, All are mapping weight matrices. , where is the dimension of the attention hidden layer.

[0105] Step S6: Fusion of spatiotemporal-spectral features Input to the uncertainty prediction head, output the inversion prediction mean of the target water quality parameters. and the corresponding log-variance of cognitive uncertainty And employing Gaussian likelihood loss Saltwater intrusion monotonicity constraint loss and salinity order constraint loss The composite loss function includes Perform model optimization.

[0106] In this step, the uncertainty prediction head's output calculation process specifically includes:

[0107]

[0108]

[0109] in, To share the weights of the feature extraction layer, and These are the weight matrices for the mean head and variance head, respectively.

[0110] The composite loss function in step S6 Includes task likelihood loss based on Gaussian negative log-likelihood. The calculation process specifically includes:

[0111]

[0112] in, The number of samples used in the calculation. The actual observed values ​​of the target water quality parameters;

[0113] The composite loss function It also includes the salinity monotonicity constraint loss. This is used to penalize anomalies where the predicted conductivity of the upstream is higher than that of the downstream. The calculation process specifically includes:

[0114]

[0115] in, For random sampling of spatial point pairs within a batch Total quantity This is an indicator function that takes the value when the condition is true. Otherwise take .

[0116] Composite loss function It also includes salinity order constraint loss. The salinity response coefficient used to constrain the extraction Compared with the actual conductivity label The order of magnitude and rank are positively correlated, and the calculation process specifically includes:

[0117]

[0118] in, The number of point pairs sampled. It is a symbolic function.

[0119] The following is an optional embodiment of this application, specifically:

[0120] Model training and large-scale part based on the measured aggregate dataset of the Yangtze River estuary

[0121] To verify the effectiveness and high accuracy of this application in the actual complex estuarine hydrodynamic environment, this embodiment details the entire process of model training and remote sensing deployment based on the field measurement data collected from January 7 to 8, 2026, near the Yangtze River estuary.

[0122] Step C1: Continuous and synchronous acquisition of multi-source data in the estuary field**

[0123] During the typical tidal period of saltwater intrusion from January 7 to 8, 2026, a two-day simultaneous underwater experiment was conducted at the confluence of the Yangtze River estuary and upstream.

[0124] 1. Spectral mode acquisition: Acquiring hyperspectral remote sensing data of the water surface using a UAV hyperspectral imager. It covers the visible light to near-infrared band;

[0125] 2. Spatial Modal Acquisition: Equipped with a high-precision GPS positioning system, it records the precise latitude and longitude coordinates of each sampling point in real time over two days. ;

[0126] 3. True Water Quality Measurement: A multi-parameter water quality analyzer is used to simultaneously and in-situ test the physicochemical parameters of the water surface, focusing on collecting conductivity (ms / cm), turbidity (NTU), and pH value as true labels for supervised training. .

[0127] Step C2: Multi-source data alignment and spatiotemporal complete set construction

[0128] The hyperspectral images collected on January 7th and 8th were spatiotemporally matched with the extracted points, GPS coordinates, and true water quality values. After removing invalid values, they were merged to form a complete spatiotemporal measured dataset.

[0129] The summarized spectral features and spatial coordinates were then Z-score standardized. Subsequently, the complete dataset was randomly shuffled and divided into sets according to a preset ratio (80% as the training set mask and 20% as the test set mask) to ensure that both the training and test sets contained water samples from different tidal states and spatial locations over two days.

[0130] Step C3: Physical-data joint closed-loop training based on the entire set of actual measurements

[0131] The pre-defined training set samples are input into the PhysioMamba system for multiple rounds of optimization. During training, considering the limited sample size of the actual measured data, the composite physical prior loss function proposed in this application is used for depth constraint.

[0132] Saltwater intrusion monotonicity constraint: Within the training batch, the system randomly compares upstream and downstream points in the aggregated dataset. If the predicted conductivity of the upstream section is higher than that of the downstream section, the model immediately applies a monotonicity penalty gradient, forcing the model to learn the hydrodynamic spatial decay law of conductivity decreasing upstream along the river channel;

[0133] Triggering salinity order constraints: The system forces the "salinity response coefficient" extracted by the Physical Spectral Decomposition (PGSD) module. "It maintains a strict positive correlation with the measured conductivity readings of the multi-parameter water quality meter."

[0134] After each iteration, the model is evaluated using 20% ​​of the test set, and the coefficient of determination is saved. ) Optimal network weights.

[0135] Step C4: Large-scale saltwater intrusion evolution map and confidence quantification output

[0136] Using the optimal model weights obtained from training on the two-day aggregated dataset, the entire large-scale hyperspectral remote sensing mosaic image is divided into multiple spatial data blocks and input into the system for pixel-level forward inference:

[0137] 1. The model outputs a high-precision conductivity inversion map of the entire survey area, accurately capturing the upstream intrusion boundary of the saltwater tongue and the spatial gradient changes caused by tidal action;

[0138] 2. The uncertainty prediction head synchronously outputs an uncertainty variance spatial distribution map. In pixels with significant ship wake interference, cloud shadow occlusion, or distributions outside the model training data, the system outputs extremely high cognitive uncertainty values. .

[0139] The comparison between the large-scale saltwater inversion results and the spatial distribution of cognitive uncertainty in the experimental area of ​​the Yangtze River estuary shows that the saltwater tongue exhibits a clear gradient decay trend from the estuary upstream, which conforms to the spatial laws of hydrodynamics. Specifically, in the spatial distribution of the log-variance of cognitive uncertainty synchronously output by the uncertainty prediction head, high uncertainty values ​​are concentrated in pixel positions affected by ship wake interference, cloud shadow occlusion, or exceeding the distribution of model training data, while low uncertainty values ​​correspond to the main water area with higher confidence in the inversion results. It can be seen that the present invention can provide quantitative reliability assessment basis for business departments while achieving high-precision capture of saltwater intrusion boundaries.

[0140] This application also provides an estuarine saltwater inversion system driven by both physical mechanisms and data, which may include:

[0141] The acquisition module is used to acquire the remote sensing hyperspectral sequence data matrix of the target water area. and the corresponding water surface spatial standardized coordinate data matrix ;

[0142] The decoupled output module is used to convert remote sensing hyperspectral sequence data matrices. The input is fed into a joint spectral encoder, where weighted spectral features are extracted using an adaptive band gating mechanism. Simultaneously, bidirectional state-space model sequence scanning and multi-scale spectral convolutional pyramid feature extraction are performed, decoupled to output absorption features specific to the water body. With scattering characteristics ;

[0143] The prediction module is used to generate a matrix of remote sensing hyperspectral sequence data. The input is fed into the physics-guided spectral decomposition module, which predicts decomposition coefficients and extracts physical prior features based on a pre-defined physical dictionary that includes a salinity response basis. ;

[0144] Step S4: Standardize the spatial coordinate data matrix of the water surface. The data is input into the tidal sensing spatial encoder, and multi-scale radial basis function gradient decay features are extracted by calculating the relative distance along the flow from the coordinate point to the dynamic salt front reference point. ;

[0145] Step S5: Absorption characteristics scattering characteristics With physical prior characteristics Spectral modal characteristics and gradient attenuation characteristics formed by splicing The input is fed into an asymmetric cross-modal attention module, where spatial features guide the attention of the spectral channels to obtain spatiotemporal-spectral fusion features. ;

[0146] Step S6: Fusion of spatiotemporal-spectral features Input to the uncertainty prediction head, output the inversion prediction mean of the target water quality parameters. and the corresponding log-variance of cognitive uncertainty And employing Gaussian likelihood loss Saltwater intrusion monotonicity constraint loss and salinity order constraint loss The composite loss function includes Perform model optimization.

[0147] Specific limitations regarding the estuarine salinity inversion system driven by both physical mechanisms and data can be found in the limitations of the estuarine salinity inversion method driven by both physical mechanisms and data mentioned above, and will not be repeated here. Each module in the aforementioned estuarine salinity inversion system driven by both physical mechanisms and data can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of the processor, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0148] In one embodiment, an electronic device, which may be a computer, is provided. The electronic device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device is used for estuarine salinity inversion data. The network interface of the computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an estuarine salinity inversion method.

[0149] In one embodiment of this application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the above-described method for inverting estuary saline intrusion under the dual drive of physical mechanism and data.

[0150] In one embodiment of this application, a computer program product is provided, including a computer program / instructions, which, when executed by a processor, implements the steps of the above-described method for inverting estuarine salinity under the dual drive of physical mechanism and data.

[0151] The computer-readable storage medium and computer program product provided in this embodiment are similar in implementation principle and technical effect to the above method embodiments, and will not be repeated here.

[0152] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0153] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0154] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for retrieving estuarine saltwater intrusion driven by both physical mechanisms and data, characterized in that, The method includes: Step S1: Obtain the remote sensing hyperspectral sequence data matrix of the target water area. and the corresponding water surface spatial standardized coordinate data matrix ; Step S2: The remote sensing hyperspectral sequence data matrix The input is fed into a joint spectral encoder, where weighted spectral features are extracted using an adaptive band gating mechanism. Simultaneously, bidirectional state-space model sequence scanning and multi-scale spectral convolutional pyramid feature extraction are performed, decoupled to output absorption features specific to the water body. With scattering characteristics ; Step S3: The remote sensing hyperspectral sequence data matrix The input is fed into the physics-guided spectral decomposition module, which predicts decomposition coefficients and extracts physical prior features based on a pre-defined physical dictionary that includes a salinity response basis. ; Step S4: Standardize the spatial coordinate data matrix of the water surface. The data is input into the tidal sensing spatial encoder, and multi-scale radial basis function gradient decay features are extracted by calculating the relative distance along the flow from the coordinate point to the dynamic salt front reference point. ; Step S5: Apply the absorption feature The scattering characteristics With the aforementioned physical prior features The spliced ​​spectral modal features and the gradient attenuation features The input is fed into an asymmetric cross-modal attention module, where spatial features guide the attention of the spectral channels to obtain spatiotemporal-spectral fusion features. ; Step S6: Combine the spatiotemporal-spectral fusion features Input to the uncertainty prediction head, output the inversion prediction mean of the target water quality parameters. and the corresponding log-variance of cognitive uncertainty And employing Gaussian likelihood loss Saltwater intrusion monotonicity constraint loss and salinity order constraint loss The composite loss function includes Perform model optimization.

2. The estuarine saltwater inversion method driven by both physical mechanisms and data as described in claim 1, characterized in that, The feature extraction calculation process in step S2, which extracts weighted spectral features through an adaptive band gating mechanism, specifically includes: ； ； in, This is the mapping weight matrix for the adaptive band gating mechanism. and The weight matrix is ​​a learnable linear mapping. and For activation function, This is a Hadamard product operation, used to suppress noise bands outside the atmospheric window and output weighted spectral features. ; The bidirectional state-space model sequence scan includes a forward scan path and a reverse scan path. The specific calculation formula includes: Forward scan feature output ; Reverse scan feature output ; Bidirectional global feature fusion ; in, This represents a sequence modeling operation performed using discretized continuous state-space equations. Indicates the dimension along the band sequence The flip operation, This indicates a splicing operation at the channel level. Presentation layer normalization operation, This represents the learnable channel feature merging weight matrix.

3. The estuarine saltwater inversion method driven by both physical mechanisms and data as described in claim 1, characterized in that, The multi-scale spectral convolution pyramid in step S2 uses three one-dimensional convolutions with different receptive fields to capture spectral morphological features. The calculation process specifically includes: in, Indicates the kernel size as One-dimensional convolution operations are used to capture abrupt spectral transitions, medium-broad absorption peaks, and broad envelope features, respectively. Used to compress multi-scale outputs to a uniform dimension.

4. The estuarine saltwater inversion method driven by both physical mechanisms and data as described in claim 1, characterized in that, Step S2 also includes the calculation of a learnable water quality spectral index, by defining two sets of soft band selector weights. and Through the formula: Normalized difference characteristics ; Dynamic ratio characteristics ; in, , These are the water quality characteristic spectral indices obtained through weighted aggregation using an adaptive soft band selector. It is a very small constant. This is the preset clamping threshold.

5. The estuarine saltwater inversion method driven by both physical mechanisms and data as described in claim 1, characterized in that, The physical-guided spectral decomposition module in step S3 specifically includes the following physical reconstruction and feature extraction calculation process: in, For physically reconstructed hyperspectral sequence data matrix, The output mapping weight matrix of the physics-guided spectral decomposition module. , These are the preset absorption substrate and scattering substrate, respectively; For modeling the electrolytic optical effect on a salinity-responsive substrate; , , This is the corresponding non-negative decomposition coefficient matrix.

6. The estuarine saltwater inversion method driven by both physical mechanisms and data according to claim 1, characterized in that, The tidal sensing spatial encoder in step S4 has a relative distance along the flow. Radial basis attenuation characteristics The calculation process specifically includes: in, For learnable flow direction angle parameters, As a learnable dynamic salt front reference longitude parameter, and These are the standardized spatial coordinates; The scale parameter is learned in logarithmic space and is used to model the physical properties of salinity exhibiting multi-scale exponential decay as it moves upstream along the river channel.

7. The estuarine saltwater inversion method driven by both physical mechanisms and data according to claim 1, characterized in that, The asymmetric cross-modal attention module in step S5 selectively activates spectral modal features based on spatial modal features, and its attention weights... With fusion features The calculation process specifically includes: in, For spectral modal characteristics, All are mapping weight matrices. , where is the dimension of the attention hidden layer.

8. The method for estuarine saltwater inversion driven by both physical mechanisms and data according to claim 1, characterized in that, The uncertainty prediction head in step S6 specifically includes the following output calculation process: in, To share the weights of the feature extraction layer, and These are the weight matrices for the mean head and variance head, respectively.

9. The method for estuarine saltwater inversion driven by both physical mechanisms and data according to claim 1, characterized in that, The composite loss function in step S6 Includes task likelihood loss based on Gaussian negative log-likelihood. The calculation process specifically includes: in, The number of samples used in the calculation. The actual observed values ​​of the target water quality parameters; The composite loss function It also includes the salinity monotonicity constraint loss. This is used to penalize anomalies where the predicted conductivity of the upstream is higher than that of the downstream. The calculation process specifically includes: in, For random sampling of spatial point pairs within a batch Total quantity This is an indicator function that takes the value when the condition is true. Otherwise take .

10. The estuarine saltwater inversion method driven by both physical mechanisms and data according to claim 1, characterized in that, Composite loss function It also includes salinity order constraint loss. The salinity response coefficient used to constrain the extraction Compared with the actual conductivity label The order of magnitude and rank are positively correlated, and the calculation process specifically includes: in, The number of point pairs sampled. It is a symbolic function.

Images