Water body transparency remote sensing inversion method, system, medium and device

Abstract

This invention relates to the fields of remote sensing inversion, water color parameter estimation, artificial intelligence, and geospatial information processing. To address the problem that existing methods cannot simultaneously satisfy the accuracy, stability, and interpretability of transparency inversion results for complex water bodies, this invention provides a method, system, medium, and equipment for remote sensing inversion of water body transparency. The method includes constructing training samples; obtaining transparency inversion input features; extracting transparency inversion input features from the training samples, and then training a pre-constructed physical information neural network using a joint loss function; selecting constraint combination features from a physically interpretable feature library using a physical constraint ablation method; extracting transparency inversion input features and constraint combination features from remote sensing reflectance data of the target water body; processing the data using the trained physical information neural network; and generating pixel-level water body transparency inversion results, which can accurately, robustly, and interpretably invert the transparency of complex water bodies.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing inversion, water color parameter estimation, artificial intelligence and geospatial information processing, and in particular to a method, system, medium and equipment for remote sensing inversion of water transparency. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Water transparency is typically characterized by Secchi disk depth (SDD or Zsd), a crucial indicator for describing underwater light fields, water turbidity, and the state of the aquatic ecosystem. Traditional in-situ Secchi disk observations offer high precision, but suffer from limited spatial coverage, restricted observation frequency, and high costs associated with long-term continuous monitoring, making it difficult to meet the needs of large-scale, long-term water environment monitoring and aquatic ecosystem assessment. Satellite remote sensing provides continuous and repeatable water body observations, offering advantages such as large-scale, repeated observations and continuous historical imagery. However, in complex water bodies, suspended particulate matter, phytoplankton, colored soluble organic matter (CDOM), tidal mixing, runoff input, shallow water sediment, and atmospheric correction residuals all influence remote sensing reflectance, resulting in a significantly nonlinear, regionally variable, and sensor-dependent relationship between transparency and multispectral reflectance.

[0004] Existing methods for transparency remote sensing inversion mainly include empirical models, semi-empirical or semi-analytical models, and machine learning models. Empirical models have simple structures but are easily limited by single regions or single-band combinations, and lack cross-regional generalization ability. Semi-analytical models have good physical interpretability but typically require relatively complete intrinsic optical parameters, strong boundary condition assumptions, and sensor calibration information. Purely data-driven models, while capable of fitting complex nonlinear relationships, are prone to producing predictions inconsistent with water color mechanisms in small-sample, cross-water body, and image-based applications. Therefore, a technical solution is needed that can simultaneously improve the accuracy, stability, physical consistency, and image-based deployability of transparency inversion under the limited bandwidth of multispectral sensors. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method, system, medium, and device for remote sensing inversion of water transparency, which can accurately, robustly, and interpretably invert the transparency of complex water bodies.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides a method for remote sensing inversion of water transparency.

[0007] A remote sensing inversion method for water transparency includes: Acquire in-situ transparency observation data, water body remote sensing reflectance data, and auxiliary water quality parameters of water samples to construct training samples; Based on the relative spectral response function of the target sensor, the water body remote sensing reflectance data is convolved into equivalent multispectral bands, and then a physically interpretable feature library is constructed and redundancy control is performed to obtain the transparency inversion input features; Transparency inversion input features are extracted from training samples. When the water transparency is predicted by a pre-constructed physical information neural network, reflectivity, absorption, scattering and attenuation information consistent with the water color mechanism are reconstructed. Then, a joint loss function is constructed and the physical information neural network is trained accordingly. At the same time, constraint combination features are selected from the physical interpretable feature library through the physical constraint ablation method. Transparency inversion input features and constraint combination features are extracted from remote sensing reflectance data of the target water body, and processed using a trained physical information neural network to generate pixel-level water body transparency inversion results. The physical information neural network includes a parallel transparency data regression branch and a water color physical branch, with its input end connected to a shared feature backbone and its output end connected to a feature fusion layer and a fully connected layer; the joint loss function is a weighted sum of supervised transparency regression loss, remote sensing reflectance consistency loss, diffuse attenuation coefficient and transparency consistency loss, absorption spectral morphology loss, and selectable band ratio monotonic loss.

[0008] As one implementation, the physically interpretable feature library includes original band reflectance and its logarithmic transformation, two-band ratio, normalized difference, three-band difference ratio, normalized band reflectance, red-green difference, visible-near-infrared contrast, and turbidity proxy features.

[0009] As one implementation method, the process of redundancy control for the physically interpretable feature library is as follows: Features in the physically interpretable feature library that are more relevant than a set threshold, have opposite signs, or are algebraically equivalent are grouped into the same redundant group. Calculate the comprehensive score of each candidate feature within each redundant group, and retain features with a comprehensive score not lower than a set score threshold and a clear explanation of the water color mechanism as input features for transparency inversion; the comprehensive score of each candidate feature within each redundant group for: ; in, For the j-th candidate feature With transparency label The absolute correlation; Contribute a score to the stability of cross-validation or the validation set; This is a redundancy penalty item; , and The weights are non-negative.

[0010] As one implementation method, the training process for the physical information neural network also includes: Pixel features are extracted from the pixel-level water transparency inversion results. The extrapolation risk is judged based on the distance or quantile range between the pixel features and the training sample distribution. Pixels that exceed the threshold are output with low confidence or quality labels. ; in, pixel features Mahalanobis distance relative to the training sample distribution; μ is the mean vector of the training features; To train the feature covariance matrix; T denotes transpose; Denotes the inverse of the covariance matrix; if If the pixel reflects a value exceeding a preset threshold or the reflectance of a pixel exceeds the quantile range of the training samples, the pixel is marked as having low confidence.

[0011] As one implementation method, during the training process, the pixel-level water transparency inversion result generated by the physical information neural network is as follows: ; ; ; in, For the first Transparency prediction of each sample reduced to physical units; For the first Log-prediction of transparency for a single sample; The preset small parameter value is greater than zero; For the first The adaptive fusion coefficients of each sample; For the first The logarithmic predicted value of transparency output from the regression branch of the data for each sample; For the first Optical transparency proxy for each sample; The maximum contribution limit for the water color physics branch; It is an S-shaped function; and For learnable fusion parameters; for Transpose of; For the first Shared hidden representation of samples.

[0012] As one implementation method, the expression for the optical transparency proxy is: ; in, This is the visibility scaling factor; It is a collection of visible light wavelengths; This indicates that the band with the strongest penetration and the least diffuse attenuation is selected. wavelength Dispersion attenuation coefficient proxy.

[0013] As one implementation method, the supervised transparency regression loss, remote sensing reflectance consistency loss, diffuse attenuation coefficient and transparency consistency loss, absorption spectral morphology loss, and selectable band ratio monotonicity loss are respectively set as follows: , , , and : ; ; ; ; ; in, This represents the number of training samples; For robust error function; For the sample Measured transparency; The preset small parameter value is greater than zero; For the first Log-prediction of transparency for a single sample; A set of bands that participate in physical constraints; Number of bands; For stable logarithmic or scaling transformations; For the sample In the band The observed remote sensing reflectance at the location; For the sample In the band Remote sensing reflectance proxy reconstructed by the water color physical branch; To express summation; For the first Optical transparency proxy for each sample; For the sample In the band First Class absorption component proxy; For the sample In the band First Class absorption component proxy; They are classified as colored soluble organic matter and non-algal particulate matter, respectively. This is used to penalize only samples that violate the orientation constraint; To satisfy the first Turbidity proxy for each sample Greater than the Turbidity proxy for each sample An ordered set of sample pairs; For the first Log-prediction of transparency for a single sample; For the first Logarithmic prediction of transparency for each sample.

[0014] A second aspect of the present invention provides a remote sensing inversion system for water transparency.

[0015] A remote sensing inversion system for water transparency includes: The training sample construction module is used to acquire on-site transparency observation data, water body remote sensing reflectance data, and auxiliary water quality parameters of water samples to construct training samples; The inversion input feature acquisition module is used to convolve the water body remote sensing reflectance data into equivalent multispectral bands based on the target sensor's relative spectral response function, thereby constructing a physically interpretable feature library and performing redundancy control to obtain transparency inversion input features. The network training and constraint screening module is used to extract transparency inversion input features from training samples. When the water transparency is predicted by the pre-constructed physical information neural network, the reflectivity, absorption, scattering and attenuation information consistent with the water color mechanism are reconstructed. Then, a joint loss function is constructed and the physical information neural network is trained accordingly. At the same time, constraint combination features are screened from the physical interpretable feature library through the physical constraint ablation method. The water transparency inversion module is used to extract transparency inversion input features and constraint combination features from remote sensing reflectance data of the target water body, process them using a trained physical information neural network, and generate pixel-level water transparency inversion results. The physical information neural network includes a parallel transparency data regression branch and a water color physical branch, with its input end connected to a shared feature backbone and its output end connected to a feature fusion layer and a fully connected layer; the joint loss function is a weighted sum of supervised transparency regression loss, remote sensing reflectance consistency loss, diffuse attenuation coefficient and transparency consistency loss, absorption spectral morphology loss, and selectable band ratio monotonic loss.

[0016] A third aspect of the present invention provides a computer-readable storage medium.

[0017] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the water transparency remote sensing inversion method described above.

[0018] A fourth aspect of the present invention provides an electronic device.

[0019] An electronic device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps in the water transparency remote sensing inversion method described above.

[0020] Compared with the prior art, the beneficial effects of the present invention are: The water transparency remote sensing inversion method provided by this invention convolves water remote sensing reflectance data into equivalent multispectral bands, constructs a physically interpretable feature library, and filters out transparency inversion input features. Then, a joint loss function is constructed by weighting the supervised transparency regression loss, remote sensing reflectance consistency loss, diffuse attenuation coefficient and transparency consistency loss, absorption spectral morphology loss, and selectable band ratio monotonicity loss. Based on this, a physical information neural network is trained. Simultaneously, a physical constraint ablation method is used to filter constraint combination features from the physically interpretable feature library. Finally, transparency inversion input features and constraint combination features are extracted from the target water body's remote sensing reflectance data. The trained physical information neural network is then used for processing to generate pixel-level water transparency inversion results, achieving the goal of accurately, robustly, and interpretably inverting the transparency of complex water bodies. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1 This is a flowchart of the water transparency remote sensing inversion method according to an embodiment of the present invention; Figure 2 This is an example image of remote sensing inversion of the transparency of a certain body of water during the high-water season, according to an embodiment of the present invention; wherein, Figure 2 (a) in the figure represents the remote sensing inversion result of the transparency of a certain water body on July 14, 2021, during the high-water season; Figure 2 (b) in the figure shows the remote sensing inversion result of the transparency of a certain water body on August 2, 2022, during the high-water season; Figure 2 (c) in the figure represents the remote sensing inversion result of the transparency of a certain water body on June 2, 2023, during the high-water season; Figure 2 (d) in the figure represents the remote sensing inversion result of the transparency of a certain water body on August 7, 2024, during the high-water season; Figure 3This is an example image of remote sensing inversion of the transparency of a certain water body during the dry season, according to an embodiment of the present invention; wherein, Figure 3 (a) in the figure represents the remote sensing inversion result of the transparency of a certain water body on December 5, 2021, during the dry season; Figure 3 (b) in the figure shows the remote sensing inversion result of the transparency of a certain water body on October 21, 2022, during the dry season; Figure 3 (c) in the figure represents the remote sensing inversion result of the transparency of a certain water body on November 25, 2023, during the dry season; Figure 3 (d) in the figure represents the remote sensing inversion result of the transparency of a certain water body on October 10, 2024, during the dry season; Figure 4 The results of the five-fold cross-validation in this embodiment of the invention are shown; wherein, (a) in 4 is the five-fold cross-validation result of the physical information neural network after removing the monotonic constraint of the optional band ratio; Figure 4 (b) in the figure represents the five-fold cross-validation result of the physical information neural network that retains only the data constraints; Figure 4 (c) in the figure represents the five-fold cross-validation results of the physical information neural network after removing the diffuse decay coefficient and transparency consistency constraints; Figure 4 (d) in the figure represents the five-fold cross-validation result of the physical information neural network after removing the consistency constraint of remote sensing reflectance; Figure 4 (e) represents the five-fold cross-validation result of the physical information neural network after removing the absorption spectral morphology constraints; Figure 4 In the diagram, (f) represents the five-fold cross-validation result of the neural network containing all constraints. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0026] Transparency is influenced by suspended particulate matter, phytoplankton, colored dissolved organic matter (CDOM), as well as sediment and hydrodynamic processes. It can directly reflect changes in water color, sediment transport, and ecological risks in water bodies such as estuaries, nearshore areas, lakes, and reservoirs.

[0027] Unless otherwise specified, reflectivity, transparency, band combination, physical proxy parameters and neural network parameters in this invention can be equivalently replaced according to the specific sensor, the water area studied and the sample conditions.

[0028] In this invention, "proxy" refers to a biooptical proxy, also known as an optical substitution index, which is defined as: In water color remote sensing inversion, when the concentration of a certain water quality component (such as CDOM, non-algal particulate matter, etc.) cannot be directly quantitatively observed through satellite remote sensing, characteristic optical parameters of the component at a specific wavelength (such as absorption coefficient, backscattering coefficient) are selected as "proxies" to indirectly obtain the concentration information of the component through the remotely observable spectral signal.

[0029] Figure 1 A schematic diagram illustrating the principle of the remote sensing inversion method for water transparency according to an embodiment of the present invention is provided. Figure 1 The water transparency remote sensing inversion method of this embodiment may include the following steps: S101: Obtain on-site transparency observation data, water body remote sensing reflectance data, and auxiliary water quality parameters of water samples to construct training samples.

[0030] Transparency can be measured using a standard black and white Seisch disk; water body remote sensing reflectance data can come from on-site hyperspectral observations, spectral measurements above the water surface, open-source water color spectral datasets, or satellite-ground synchronous remote sensing samples; auxiliary water quality parameters include, but are not limited to, suspended particulate matter concentration, chlorophyll a concentration, CDOM absorption coefficient, temperature, salinity, water body type, and spatial coordinates.

[0031] The auxiliary water quality parameters serve the following purposes: First, to identify anomalous samples with significant inconsistencies between reflectance, transparency, and water quality components, such as samples with extremely high suspended matter concentrations but abnormally flat reflectance spectra. Second, to characterize the main optical sources of transparency variations, namely, chlorophyll a representing the absorption and scattering contributions of phytoplankton, total suspended matter representing the scattering of particles and the contribution of sediment turbidity, and the CDOM absorption coefficient representing the absorption contribution of colored dissolved organic matter to the ultraviolet-visible band. Third, to support stratified sampling, cross-validation, and graded accuracy evaluation according to estuaries, nearshore areas, lakes, reservoirs, or other water body types. Fourth, to assist in interpreting the absorption proxy, backscattering proxy, and diffuse attenuation proxy of the physical branch output, and can participate in training as an auxiliary task or weakly supervised constraint when corresponding observations are available. Fifth, when applied to image processing, to determine whether the target image exceeds the optical background range of the training samples, thereby generating low-confidence labels.

[0032] It should be noted that auxiliary water quality parameters are mainly used for sample quality control, physical interpretation, stratified evaluation, and constraint screening during the training phase. When the target remote sensing image lacks pixel-by-pixel chlorophyll a, total suspended matter, or CDOM observations, the image inversion still uses the target sensor reflectance and its band combination characteristics as input; if synchronous or prior water quality products exist, they can be used as optional auxiliary inputs or quality control information.

[0033] No. Each training sample consists of observation parameters, sensor standardization parameters, feature parameters, and sample attributes, and can be represented as: (1); in, For the first One training sample; Used as a unique identifier for the sample; The label shows the actual transparency. For the sample At wavelength Hyperspectral remote sensing reflectance at the location; The target sensor obtained by convolution of the spectral response function Reflectivity of one equivalent band; For the reason Constructed band ratios, normalized differences, three-band combinations, and water color proxy characteristics; To supplement water quality and background attribute vectors, chlorophyll a concentration, weather information, and aCDOM440 can be included. n (Absorption coefficient of colored dissolved organic matter at 440nm wavelength), water body type, site, latitude and longitude, and sampling date; For quality control and stratification identification, including valid sample identification, transparency level, water body type stratification, and cross-validation commas.

[0034] S102: Based on the relative spectral response function of the target sensor, the water remote sensing reflectance data is convolved into equivalent multispectral bands, and then a physically interpretable feature library is constructed and redundancy control is performed to obtain the transparency inversion input features.

[0035] The on-site hyperspectral reflectance is screened for validity, and samples with missing measurements, negative values, abnormal peaks, and incomplete spectral coverage are removed. The reflectance is then uniformly resampled to a preset wavelength range. In one embodiment, the hyperspectral data is uniformly set to 400~900nm in 1nm intervals, and valid samples that match the measured SDD (Seidel disk depth) values ​​are retained.

[0036] To ensure that the model training data maintains consistency with real satellite imagery in terms of spectral structure, based on the target sensor's... Relative spectral response function of each band Convolve the hyperspectral reflectance R(λ) into an equivalent multispectral reflectance. : (2); In formula (2), For the target sensor Equivalent remote sensing reflectance of each band; wavelength The on-site hyperspectral remote sensing reflectance at the location; For the first Each band response range; The hyperspectral sampling interval; This represents summation over the band response range. When the spectrum is a continuous function, formula (2) can be equivalently written as an integral weighted average. This step does not introduce learnable parameters and belongs to deterministic spectral uniformity preprocessing. Formula (2) comes from the principle of weighted averaging of sensor band responses. Target sensor The actual signal received in each band is not the reflectivity of a single center wavelength, but rather the reflectivity within the band's response range. With relative spectral response function The integral response. The on-site hyperspectral sample can be converted into the equivalent reflectance of the target multispectral sensor using formula (2).

[0037] In a Landsat OLI or Landsat OLI-2 land-oriented imager, equivalent reflectances are generated for five water color-related bands: B1 (coastal / aerosol), B2 (blue), B3 (green), B4 (red), and B5 (near-infrared). Thermal infrared, panchromatic, and shortwave infrared bands can be excluded based on water color information contributions and the original spectral coverage. This spectral homogenization step does not introduce training parameters and is a deterministic preprocessing step, which reduces model transfer errors caused by differences in sensor bands.

[0038] For imagery data, low-cloudy remote sensing images of the target sensor during the target period are acquired. These images are then radiometrically calibrated, atmospherically corrected, water-masked, invalid value removed, and corrected for residual solar flares and sky reflectance. The surface reflectance ρ after reflectance correction is converted to remote sensing reflectance using ρ / π. The calculation of imagery features, missing value imputation, and normalization parameters are consistent with those used in the training process.

[0039] The physically interpretable feature library includes raw band reflectance and its logarithmic transformation, two-band ratio, normalized difference, three-band difference ratio, normalized band reflectance, red-green difference, visible-near-infrared contrast, and turbidity proxy features.

[0040] Let the target sensor be the first The reflectance of each band is Candidate feature library Includes: original band reflectivity and its logarithmic transformation, and the ratio of the two bands. Normalized difference Three-band difference ratio Normalized form In addition, water color descriptors such as red-green difference, visible-near-infrared contrast, and turbidity proxy.

[0041] Candidate feature library It can be constructed using the following formula, where ε is a stability constant to avoid a denominator of zero: (3); in, For the candidate feature set; , , These represent the reflectivity of the target sensor at different wavelengths. , , The bands are numbered, and the denominator bands in the ratio or difference do not completely overlap with the numerator bands; The preset small parameter value is greater than zero.

[0042] (4); in, For the sample The One standardized feature; For the sample The One original candidate feature; and The first and second in the training set, respectively The mean and standard deviation of each feature. , The median, which is used to fill in missing values, is estimated solely from the training set and is fixedly invoked at the image inference end. For each sample Candidate feature vectors are obtained. The training phase only computes on a subset of the training data. and and for the first Each feature is standardized.

[0043] Redundancy control is performed on candidate features, removing features with antisymmetric repetition, sign-reversed repetition, or high algebraic correlation. For example (B2) B1) / B4 and (B1) B2) / B4 are simply opposite in sign and not both retained. Subsequently, Pearson correlation, Spearman correlation, cross-validation performance, and the physical meaning of water color are combined to sort and filter the features, forming a compact, stable, and easily interpretable set of input features.

[0044]

[0045] or ;or ;or (5); in, For the first q A redundant feature group; and There are two candidate features; The correlation coefficient is either Pearson or Spearman. The redundancy detection threshold can be set according to the actual situation. It is a constant ratio or a constant offset, and its value can also be preset by human intervention. Formula (5) is used to divide highly correlated, oppositely signd, or algebraically equivalent features into the same redundant group.

[0046] As one implementation method, the process of redundancy control for the physically interpretable feature library is as follows: Features in the physically interpretable feature library that are more relevant than a set threshold, have opposite signs, or are algebraically equivalent are grouped into the same redundant group. For example, (B2-B1) / B4 and (B1-B2) / B4 are only opposite in sign and are not retained at the same time. Calculate the comprehensive score of each candidate feature within each redundant group, and retain features with a comprehensive score not lower than a set score threshold and a clear explanation of the water color mechanism as input features for transparency inversion; the comprehensive score of each candidate feature within each redundant group for: (6); in, For the j-th candidate feature With transparency label The absolute correlation has a value range of [0,1]. To contribute scores to cross-validation stability or the validation set, for example, by selecting j candidate features. The normalized variance is mapped to [0,1]. As a redundancy penalty term, it can reduce the number of candidate features by j. The maximum correlation coefficient with other candidate features in the redundant group is normalized to [0,1]. , and Non-negative weights , , and These can be set to 0.5, 0.3, and 0.2 respectively; in other embodiments, , and Other values ​​can also be set, which will not be detailed here. All parameters in the overall score are dimensionless data. This processing avoids the model acquiring spurious importance due to repeated features and reduces the risk of overfitting under small sample conditions.

[0047] In one embodiment, the candidate features most strongly correlated with transparency in the Landsat OLI equivalent band include (B2) B1) / B4, B3 / B4, B2 / (B1+B4), B2 / B4 and (B5) B3 / B4. Here, B3 / B4 reflects the color contrast between the green and red light bands and can serve as proxy information for the diffuse attenuation coefficient: when the water is relatively clear, red light is suppressed by absorption in pure water while green light is relatively high, resulting in an increase in B3 / B4; when the water is turbid, particle backscattering increases red light reflection, causing B3 / B4 to decrease or flatten. This mechanism is consistent with the relationship between underwater visibility and transparency, which is approximately controlled by the minimum diffuse attenuation coefficient.

[0048] S103: Extract transparency inversion input features from training samples, and when predicting water transparency through a pre-constructed physical information neural network, simultaneously reconstruct reflectivity, absorption, scattering, and attenuation information consistent with the water color mechanism. Then, construct a joint loss function and train the physical information neural network accordingly. At the same time, select constraint combination features from the physically interpretable feature library through the physical constraint ablation method.

[0049] The physical information neural network includes a parallel transparency data regression branch and a water color physical branch. Its input is connected to a shared feature backbone, and its output is connected to a feature fusion layer and a fully connected layer.

[0050] Physical information neural networks receive standardized feature vectors. and target sensor multi-band remote sensing reflectance vector After the shared feature backbone is extracted and uniformly represented, the data is input into the regression branch and the water color physics branch respectively. It is obtained by constructing feature parameters, filtering, missing data imputation, and standardization; It is directly composed of the reflectance of the water color-correlated band of the target sensor. Both originate from the reflectance of the target sensor.

[0051] (7); in, For the sample The model input vector; This indicates vector concatenation; The standardized feature vector; For the sample of One water color-related band remote sensing reflectance vector; For the sample In the band The observed remote sensing reflectance at the location.

[0052] (8); in, For the sample Shared implicit representation; For parameters The shared feature backbone can be implemented by a multi-layer fully connected network, a one-dimensional convolutional network, a Transformer encoder, or a combination thereof.

[0053] In one embodiment, the shared backbone includes three fully connected layers with hidden layer widths of 64, 64 and 32 respectively, and layer normalization, SiLU activation function and regularized Dropout are set after each layer. Generate implicit representations through multi-layer nonlinear mapping : (9); (10); (11); in, W 1. W 2. W 3 and b 1. b 2. b 3 is a learnable parameter. LN Representation layer normalization, SiLU This represents a smooth, non-linear activation function; Dropout is a regularization function. h 1. h 2 represents the first and second hidden representations. A shared backbone serves both the data branch and the physical branch, enabling both to learn transparency-related information within the same spectral feature space. This layered expansion clarifies the formation path of the hidden representations, avoiding the interpretation of the physical information network as a single black-box mapping; the normalization, activation, and random deactivation operations of each layer can be equivalently replaced based on the training sample size and sensor dimension.

[0054] The data regression branch is used to directly predict the logarithmic value (SDD) of transparency to enhance adaptability to different levels of transparency. The water color physics branch is used to output non-negative latent optical parameters, including but not limited to CDOM absorption, non-algal particle absorption, phytoplankton absorption, particle backscattering, and morphological parameters such as CDOM absorption slope, non-algal particle absorption slope, and backscattering spectral index.

[0055] (12); in, The logarithmic predicted value of transparency output by the regression branch for sample n; For the data regression branch with parameter ψ; log prediction is used to reduce the leverage effect of high-transparency samples on training.

[0056] (13); in, This is the non-negative latent optical parameter vector output by the water color physics branch; For physical branch mapping with parameter ω; It is a non-negative mapping function used to ensure that the absorption, scattering, and related scale parameters are non-negative. It may include CDOM absorption proxy at the reference wavelength, non-algal particulate matter absorption proxy, phytoplankton absorption proxy, particle backscattering proxy, absorption spectrum slope, and backscattering spectrum index, etc.

[0057] For the center wavelength or equivalent wavelength of each band of the target sensor, the total absorption, total backscattering, remote sensing reflectance proxy, diffuse attenuation coefficient, and optical transparency proxy reconstructed using the latent optical parameters output by the physical branch are used. Reconstruction can employ water color semi-analysis relations, quasi-single scattering approximations, empirical calibration functions, or their simplified forms. The corresponding coefficients can be taken from empirical values ​​in literature or used as constrained learnable parameters during the training phase. (14); in, wavelength λ Total absorption coefficient agent; It is absorbed by pure water; , and These represent the absorption proxy amplitudes of CDOM, non-algal particles, and phytoplankton at the reference wavelength λ0, respectively. and These represent the slopes of the absorption spectra of CDOM and non-algal particulate matter, respectively. This represents the spectral shape function of phytoplankton absorption.

[0058] (15); in, This is a proxy for the total backscattering coefficient at wavelength λ; It is backscattering of pure water; Reference wavelength The magnitude of the particle backscattering proxy; η is the particle backscattering spectrum index.

[0059] (16); in, The relative contribution of backscattering to the total absorption and backscattering; The preset small parameter value is greater than zero.

[0060] (17); in, Remote sensing reflectance proxy reconstructed for the physics branch; and These are empirical or restricted learnable coefficients used to express a semi-analytical approximation of the relationship between the absorption-backscattering proxy and remote sensing reflectivity.

[0061] (18); in, wavelength λ Diffuse attenuation coefficient proxy; , , and These are non-negative empirical coefficients or restricted learnable coefficients. The square root cross term is used to express the combined effect of absorption and scattering in enhancing light attenuation.

[0062] (19); in, Agent for optical transparency; This is the visibility scaling factor; For visible light bands, B1 to B3 can be selected in Landsat OLI; min indicates the band with the strongest penetration and the least diffuse attenuation; ε is a stability constant to prevent the denominator from being zero.

[0063] During the training process, the pixel-level water transparency inversion results generated by the physical information neural network are as follows: (20); (twenty one); (twenty two); in, For the first The adaptive fusion coefficients of each sample; The maximum contribution limit for the water color physics branch; It is an S-shaped function; and For learnable fusion parameters; For the final logarithmic prediction of transparency; Transparency prediction to be restored to physical units; The preset small parameter value is greater than zero; For the first A shared latent representation for each sample. By restricting... The upper limit is set to prevent incomplete physical proxies from overly dominating predictions under multispectral conditions. Through the above chain, the network is simultaneously forced to reconstruct reflectivity, absorption, scattering, and attenuation information consistent with the water color mechanism when predicting transparency.

[0064] In one embodiment, Setting it to 0.1 ensures that the physical branch primarily functions as a constraint and regularization mechanism, preventing the physical proxy from being misused for inverting complete intrinsic optical parameters when multispectral information is insufficient. It should be noted that in other embodiments, Other values ​​may be set by those skilled in the art based on the actual situation, which will not be detailed here. This design preserves the nonlinear fitting capability of the data branch, while using the water color physics branch to restrict the model from entering a solution space that clearly violates the optical mechanism.

[0065] To ensure that the network learning process simultaneously satisfies data fitting and water color physical consistency, a joint loss function is constructed. The joint loss function is a weighted sum of supervised transparency regression loss, remote sensing reflectance consistency loss, diffuse attenuation coefficient and transparency consistency loss, absorption spectral morphology loss, and selectable band ratio monotonicity loss.

[0066] (twenty three); in, For the joint loss function; To monitor the transparency regression loss, the error between predicted transparency and measured transparency can be measured using Smooth-L1, mean squared error, or mean absolute error. The remote sensing reflectance consistency loss is used to constrain the reconstructed remote sensing reflectance of the physical branch to be consistent with the observed remote sensing reflectance. The diffuse decay coefficient and transparency consistency loss are used to constrain the predicted transparency to be consistent with the transparency proxy derived from the diffuse decay coefficient. To compensate for the loss of absorption spectral morphology, non-aquatic absorption spectral morphology that abnormally increases with wavelength is used to make CDOM and non-algal particulate matter absorption exhibit attenuation characteristics consistent with the water color mechanism; The optional band ratio monotonic loss is used to constrain the decrease in transparency as turbidity proxies increase in a specific water body. , , and The weights are the physical loss values; the sum of these physical weights is 1, and they can be set to 0.25, 0.25, 0.25, and 0.25 respectively; those skilled in the art can also set them to other values ​​according to the actual situation; each loss in the joint loss function is a dimensionless value.

[0067] The monitoring loss for transparency regression, remote sensing reflectance consistency loss, diffuse attenuation coefficient and transparency consistency loss, absorption spectral morphology loss, and selectable band ratio monotonicity loss are as follows: (twenty four); (25); (26); (27); (28); in, This represents the number of training samples; For robust error function; For the sample Measured transparency; The preset small parameter value is greater than zero; For the first Log-prediction of transparency for a single sample; A set of bands that participate in physical constraints; Number of bands; For stable logarithmic or scaling transformations; For the sample In the band The observed remote sensing reflectance at the location; For the sample In the band Remote sensing reflectance proxy reconstructed by the water color physical branch; To express summation; For the first Optical transparency proxy for each sample; For the sample In the band First Class absorption component proxy; For the sample In the band First Class absorption component proxy; They are classified as colored soluble organic matter and non-algal particulate matter, respectively. This is used to penalize only samples that violate the orientation constraint; To satisfy the first Turbidity proxy for each sample Greater than the Turbidity proxy for each sample An ordered set of sample pairs; For the first Log-prediction of transparency for a single sample; For the first Logarithmic prediction of transparency for each sample.

[0068] The turbidity proxy can be (B4+B5) / (B3+ε) or other proven effective water color proxies. Both formulas (27) and (28) are truncated before squaring, i.e., ReLU(·). 2 This preserves directional constraints and only penalizes samples that violate the agreed-upon direction (abnormal increase in absorption spectrum with wavelength or increase in transparency instead of decrease when turbidity increases).

[0069] The effectiveness of the constraint ablation was diagnosed through ablation experiments and cross-validation. Based on any combination of RMSE (root mean square error), MAE (mean absolute error), MAPE (mean absolute percentage error), R² (coefficient of determination), and bias and folding stability on the validation and test sets, it was determined whether each constraint improved generalization performance, and the constraint combinations effective for the current water optical conditions were retained. The specific steps of constraint ablation are as follows: First, train a model that includes all constraints; Then let each , , , =0; Then set all physical loss weights to 0 but retain the bi-branch structure; Finally, the mean error and inter-fold variance of each model on the same stratified five-fold scale were compared.

[0070] Five-fold cross-validation based on 229 estuarine and nearshore samples showed that Figure 4 (a) in the figure represents the five-fold cross-validation results of the physical information neural network after removing the monotonic constraint of the selectable band ratio, with R²=0.79 and RMSE=-56.01cm; Figure 4 (b) in the figure represents the five-fold cross-validation results of the physical information neural network that retains only the data constraints, with R²=0.78 and RMSE=-58.16cm; Figure 4 (c) in the figure represents the five-fold cross-validation results of the physical information neural network after removing the diffuse decay coefficient and transparency consistency constraints, with R²=0.78 and RMSE=-58.35cm; Figure 4 In the figure, (d) represents the five-fold cross-validation result of the physical information neural network after removing the remote sensing reflectance consistency constraint, with R²=0.78 and RMSE=-58.64cm; Figure 4In the figure (e), the results of the five-fold cross-validation of the physical information neural network after removing the absorption spectrum morphology constraints are: R²=0.76, RMSE=-60.25cm. Figure 4 In the diagram, (f) represents the five-fold cross-validation result of the physical information neural network containing all constraints, with R²=0.74 and RMSE=-62.29cm. If removing a constraint reduces the error or improves the stability, it indicates that the constraint is too strong or unsuitable under the current optical conditions, and its weight should be turned off or reduced in the final model.

[0071] The samples are stratified according to transparency level or water body type to construct training, validation, and test sets. Feature selection, standardized parameter fitting, and hyperparameter search are performed only within the training set or training-validation set to avoid information leakage. The model can be trained using Adam (daptive moment estimation), stochastic gradient descent, or other optimization algorithms, and strategies such as early stopping, gradient clipping, regularized dropout, and weight decay are employed to reduce the risk of overfitting. (29); in, This is the set of model parameters after training. Learnable parameters include shared trunk, data regression branch, physical branch, and fusion branch.

[0072] The training process for a physical information neural network also includes: Pixel features are extracted from the pixel-level water transparency inversion results. The extrapolation risk is judged based on the distance or quantile range between the pixel features and the training sample distribution. Pixels that exceed the threshold are output with low confidence or quality labels. ; in, pixel features Mahalanobis distance relative to the training sample distribution; μ is the mean vector of the training features; To train the feature covariance matrix; T denotes transpose; Denotes the inverse of the covariance matrix; if If the pixel reflects a value exceeding a preset threshold or the reflectance of a pixel exceeds the quantile range of the training samples, the pixel is marked as having low confidence.

[0073] During the batch image inversion stage, the satellite images to be processed undergo the same water masking, reflectivity scale conversion, invalid value processing, feature construction, missing value imputation, and standardization processes as in the training stage. Subsequently, the pixel features are input into the trained physical information network to obtain pixel-level transparency prediction results. The system can further generate transparency thematic maps, statistical tables, time-series variation maps, and early warning results for abnormally low transparency by date, season, water area zone, or hydrological period.

[0074] S104: Extract transparency inversion input features and constraint combination features from the remote sensing reflectance data of the target water body, process them using a trained physical information neural network, and generate pixel-level water body transparency inversion results. Figure 2 This is an example image of remote sensing inversion of the transparency of a certain body of water during the high-water season, according to an embodiment of the present invention; wherein, Figure 2 (a) in the figure represents the remote sensing inversion result of the transparency of a certain water body on July 14, 2021, during the high-water season; Figure 2 (b) in the figure shows the remote sensing inversion result of the transparency of a certain water body on August 2, 2022, during the high-water season; Figure 2 (c) in the figure represents the remote sensing inversion result of the transparency of a certain water body on June 2, 2023, during the high-water season; Figure 2 (d) in the figure represents the remote sensing inversion result of the transparency of a certain water body on August 7, 2024, during the high-water season; Figure 3 This is an example image of remote sensing inversion of the transparency of a certain water body during the dry season, according to an embodiment of the present invention; wherein, Figure 3 (a) in the figure represents the remote sensing inversion result of the transparency of a certain water body on December 5, 2021, during the dry season; Figure 3 (b) in the figure shows the remote sensing inversion result of the transparency of a certain water body on October 21, 2022, during the dry season; Figure 3 (c) in the figure represents the remote sensing inversion result of the transparency of a certain water body on November 25, 2023, during the dry season; Figure 3 (d) in the figure represents the remote sensing inversion result of the transparency of a certain water body on October 10, 2024, during the dry season.

[0075] pass Figure 2 (a) to Figure 3 As can be seen from (d) in the present invention, the embodiments of the present invention can perform batch inversion of multi-temporal images to generate spatial distribution of transparency, seasonal comparison and regional statistical results, which are applicable to turbid water bodies near the estuary, lakes and reservoirs, ecological protection areas and water environment supervision scenarios.

[0076] This invention also provides a water transparency remote sensing inversion system, comprising: The training sample construction module is used to acquire on-site transparency observation data, water body remote sensing reflectance data, and auxiliary water quality parameters of water samples to construct training samples; The inversion input feature acquisition module is used to convolve the water body remote sensing reflectance data into equivalent multispectral bands based on the target sensor's relative spectral response function, thereby constructing a physically interpretable feature library and performing redundancy control to obtain transparency inversion input features. The network training and constraint screening module is used to extract transparency inversion input features from training samples. When the water transparency is predicted by the pre-constructed physical information neural network, the reflectivity, absorption, scattering and attenuation information consistent with the water color mechanism are reconstructed. Then, a joint loss function is constructed and the physical information neural network is trained accordingly. At the same time, constraint combination features are screened from the physical interpretable feature library through the physical constraint ablation method. The water transparency inversion module is used to extract transparency inversion input features and constraint combination features from remote sensing reflectance data of the target water body, process them using a trained physical information neural network, and generate pixel-level water transparency inversion results. The physical information neural network includes a parallel transparency data regression branch and a water color physical branch, with its input end connected to a shared feature backbone and its output end connected to a feature fusion layer and a fully connected layer; the joint loss function is a weighted sum of supervised transparency regression loss, remote sensing reflectance consistency loss, diffuse attenuation coefficient and transparency consistency loss, absorption spectral morphology loss, and selectable band ratio monotonic loss.

[0077] It should be noted that each module in the water transparency remote sensing inversion system of this invention corresponds one-to-one with each step in the water transparency remote sensing inversion method of the above embodiments, and their specific implementation processes are the same, so they will not be repeated here.

[0078] The structure of the electronic device according to an embodiment of the present invention will be described in detail below. The electronic device provided in this embodiment includes: at least one processor, a memory, a user interface, and at least one network interface. The various components in the water transparency remote sensing inversion system are coupled together through a bus system. It can be understood that the bus system is used to realize the connection and communication between these components. In addition to a data bus, the bus system also includes a power bus, a control bus, and a status signal bus.

[0079] It is understood that the memory can be volatile memory or non-volatile memory, or both. The memory in this embodiment of the invention is capable of storing data to support the operation of the terminal. Examples of this data include any computer programs used to operate on the terminal, such as operating systems and applications. The operating system includes various system programs, such as the framework layer, core library layer, driver layer, etc., used to implement various basic services and handle hardware-based tasks. Applications can include various applications.

[0080] In some embodiments, the water transparency remote sensing inversion system provided by the present invention can be implemented using a combination of hardware and software. For example, the water transparency remote sensing inversion system provided by the present invention can be a processor in the form of a hardware decoding processor, which is programmed to execute the water transparency remote sensing inversion method provided by the present invention. For instance, the processor in the form of a hardware decoding processor can employ one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0081] As an example, a processor can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where a general-purpose processor can be a microprocessor or any conventional processor, etc.

[0082] As an example of the hardware implementation of the water transparency remote sensing inversion system provided in this embodiment of the invention, the device provided in this embodiment of the invention can be directly executed by a processor in the form of a hardware decoding processor. For example, it can be executed by one or more application specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components to implement the water transparency remote sensing inversion method provided in this embodiment of the invention.

[0083] The memory in this embodiment of the invention is used to store various types of data to support the operation of the water transparency remote sensing inversion system, or to store data for execution. Figure 1The program code for the method shown. Examples of this data include: any executable instructions for operating on a water transparency remote sensing inversion system, such as executable instructions that can be included in the executable instructions to implement the water transparency remote sensing inversion method of the present invention.

[0084] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including functions for executing... Figure 1 The program code for the method shown. In such an embodiment, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by the central processing unit, it performs the various functions defined in the apparatus of this application.

[0085] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0086] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for remote sensing inversion of water transparency, characterized in that, include: Acquire in-situ transparency observation data, water body remote sensing reflectance data, and auxiliary water quality parameters of water samples to construct training samples; Based on the relative spectral response function of the target sensor, the water body remote sensing reflectance data is convolved into equivalent multispectral bands, and then a physically interpretable feature library is constructed and redundancy control is performed to obtain the transparency inversion input features; Transparency inversion input features are extracted from training samples. When the water transparency is predicted by a pre-constructed physical information neural network, reflectivity, absorption, scattering and attenuation information consistent with the water color mechanism are reconstructed. Then, a joint loss function is constructed and the physical information neural network is trained accordingly. At the same time, constraint combination features are selected from the physical interpretable feature library through the physical constraint ablation method. Transparency inversion input features and constraint combination features are extracted from remote sensing reflectance data of the target water body, and processed using a trained physical information neural network to generate pixel-level water body transparency inversion results. The physical information neural network includes a parallel transparency data regression branch and a water color physical branch, with its input end connected to a shared feature backbone and its output end connected to a feature fusion layer and a fully connected layer; the joint loss function is a weighted sum of supervised transparency regression loss, remote sensing reflectance consistency loss, diffuse attenuation coefficient and transparency consistency loss, absorption spectral morphology loss, and selectable band ratio monotonic loss.

2. The water transparency remote sensing inversion method as described in claim 1, characterized in that, The physically interpretable feature library includes original band reflectance and its logarithmic transformation, two-band ratio, normalized difference, three-band difference ratio, normalized band reflectance, red-green difference, visible-near-infrared contrast, and turbidity proxy features.

3. The remote sensing inversion method for water transparency as described in claim 1, characterized in that, The process of redundancy control for the physically interpretable feature library is as follows: Features in the physically interpretable feature library that are more relevant than a set threshold, have opposite signs, or are algebraically equivalent are grouped into the same redundant group. Calculate the comprehensive score of each candidate feature within each redundant group, and retain features with a comprehensive score not lower than a set score threshold and a clear explanation of the water color mechanism as input features for transparency inversion; the comprehensive score of each candidate feature within each redundant group for: ； in, For the j-th candidate feature With transparency label The absolute correlation; Contribute a score to the stability of cross-validation or the validation set; This is a redundancy penalty item; , and The weights are non-negative.

4. The water transparency remote sensing inversion method as described in claim 1, characterized in that, The training process for a physical information neural network also includes: Pixel features are extracted from the pixel-level water transparency inversion results. The extrapolation risk is judged based on the distance or quantile range between the pixel features and the training sample distribution. Pixels that exceed the threshold are output with low confidence or quality labels. ； in, pixel features Mahalanobis distance relative to the training sample distribution; μ is the mean vector of the training features; To train the feature covariance matrix; T denotes transpose; Denotes the inverse of the covariance matrix; if If the pixel reflects a value exceeding a preset threshold or the reflectance of a pixel exceeds the quantile range of the training samples, the pixel is marked as having low confidence.

5. The water transparency remote sensing inversion method as described in claim 1, characterized in that, During the training process, the pixel-level water transparency inversion results generated by the physical information neural network are as follows: ； ； ； in, For the first Transparency prediction of each sample reduced to physical units; For the first Log-prediction of transparency for a single sample; The preset small parameter value is greater than zero; For the first The adaptive fusion coefficients of each sample; For the first The logarithmic predicted value of transparency output from the regression branch of the data for each sample; For the first Optical transparency proxy for each sample; The maximum contribution limit for the water color physics branch; It is an S-shaped function; and For learnable fusion parameters; for Transpose of; For the first Shared hidden representation of samples.

6. The water transparency remote sensing inversion method as described in claim 5, characterized in that, The expression for the optical transparency proxy is: ; in, This is the visibility scaling factor; It is a collection of visible light wavelengths; This indicates the selection of the band with the strongest penetration and the least diffuse attenuation. wavelength Dispersion attenuation coefficient proxy.

7. The water transparency remote sensing inversion method as described in claim 1, characterized in that, The monitoring loss for transparency regression, remote sensing reflectance consistency loss, diffuse attenuation coefficient and transparency consistency loss, absorption spectral morphology loss, and selectable band ratio monotonicity loss are respectively set as follows: , , , and : ； ； ； ； ； in, This represents the number of training samples; For robust error function; For the sample Measured transparency; The preset small parameter value is greater than zero; For the first Log-prediction of transparency for a single sample; A set of bands that participate in physical constraints; Number of bands; For stable logarithmic or scaling transformations; For the sample In the band The observed remote sensing reflectance at the location; For the sample In the band Remote sensing reflectance proxy reconstructed by the water color physics branch; To express summation; For the first Optical transparency proxy for each sample; For the sample In the band First Class absorption component proxy; For the sample In the band First Class absorption component proxy; They are classified as colored soluble organic matter and non-algal particulate matter, respectively. This is used to penalize only samples that violate the orientation constraint; To satisfy the first Turbidity proxy for each sample Greater than the Turbidity proxy for each sample An ordered set of sample pairs; For the first Log-prediction of transparency for a single sample; For the first Logarithmic prediction of transparency for each sample.

8. A remote sensing inversion system for water transparency, characterized in that, The water transparency remote sensing inversion method based on any one of claims 1-7 includes: The training sample construction module is used to acquire on-site transparency observation data, water body remote sensing reflectance data, and auxiliary water quality parameters of water samples to construct training samples; The inversion input feature acquisition module is used to convolve the water body remote sensing reflectance data into equivalent multispectral bands based on the target sensor's relative spectral response function, thereby constructing a physically interpretable feature library and performing redundancy control to obtain transparency inversion input features. The network training and constraint screening module is used to extract transparency inversion input features from training samples. When the water transparency is predicted by the pre-constructed physical information neural network, the reflectivity, absorption, scattering and attenuation information consistent with the water color mechanism are reconstructed. Then, a joint loss function is constructed and the physical information neural network is trained accordingly. At the same time, constraint combination features are screened from the physical interpretable feature library through the physical constraint ablation method. The water transparency inversion module is used to extract transparency inversion input features and constraint combination features from remote sensing reflectance data of the target water body, process them using a trained physical information neural network, and generate pixel-level water transparency inversion results. The physical information neural network includes a parallel transparency data regression branch and a water color physical branch, with its input end connected to a shared feature backbone and its output end connected to a feature fusion layer and a fully connected layer; the joint loss function is a weighted sum of supervised transparency regression loss, remote sensing reflectance consistency loss, diffuse attenuation coefficient and transparency consistency loss, absorption spectral morphology loss, and selectable band ratio monotonic loss.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the water transparency remote sensing inversion method as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps in the water transparency remote sensing inversion method as described in any one of claims 1-7.

Images