A small sample water quality parameter inversion method, a storage medium and a computer device

Abstract

The application discloses a small sample water quality parameter inversion method, a storage medium and computer equipment, relates to the technical field of remote sensing water quality monitoring, and mainly comprises the following steps: data acquisition and preprocessing, source domain inherent optical parameter (IOPs) inversion model training, cross-domain generative adversarial enhancement, target domain IOPs prediction and water quality parameter inversion. The small sample water quality parameter inversion method, the storage medium and the computer equipment can improve the accuracy of river water quality inversion under the condition of small samples.

Description

Technical Field

[0001] This invention relates to the field of remote sensing water quality monitoring technology, and more specifically, to a method for inverting water quality parameters in a small sample, a storage medium, and a computer device. Background Technology

[0002] With the rapid development of remote sensing technology, water quality monitoring using multispectral or hyperspectral remote sensing images has become an important means of water environment protection and management. Existing water quality parameter inversion methods can be broadly classified into three categories: empirical / statistical models, semi-empirical / semi-analytical models, and deep learning-based methods.

[0003] Empirical / statistical models often employ empirical relationships between measured water quality parameters and remotely sensed reflectance (Rrs), such as linear regression, exponential regression, random forests, and support vector machines. These methods are simple to implement and easy to generalize. However, they are highly dependent on the sample size and distribution. When the sample size is limited, overfitting can easily occur, resulting in insufficient model generalization ability and limiting their applicability to different water bodies or cross-regional situations.

[0004] Semi-empirical and semi-analytical models are based on the radiative transfer relationship between inherent optical properties (IOPs) and remotely sensed water reflectance. Examples include the algorithm recommended by IOCCG and the QAA algorithm proposed by Lee et al. These methods have clear physical meaning and can link spectral parameters with water quality parameters through IOPs (such as absorption coefficient α and backscattering coefficient bb). However, their model parameters vary significantly across different water body types, requiring a large amount of measured data as constraints in application, thus limiting their applicability under diverse water body conditions.

[0005] In recent years, methods such as convolutional neural networks (CNNs), transfer learning, and ensemble learning have been introduced into water quality parameter inversion. By training with large-scale measured sample data, the complex nonlinear mapping relationship between reflectivity and water quality parameters can be automatically learned, achieving good results in some studies.

[0006] However, all three methods mentioned above implicitly assume that the optical properties of the training domain and the target domain are co-distributed. When the target domain is a river with high turbidity and the training domain is a clean ocean, this assumption fails, leading to uncontrollable model bias.

[0007] In actual river basin monitoring, it is often difficult to obtain a sufficient number of measured water quality samples. On the one hand, field sampling and experimental analysis are costly; on the other hand, river basins have strong spatial heterogeneity and temporal dynamics, resulting in a limited number of samples per basin. At the same time, most publicly available large-scale water quality sample sets are derived from ocean waters, whose optical properties differ significantly from those of river waters. Directly transferring these samples would lead to decreased model accuracy and insufficient applicability. Summary of the Invention

[0008] The purpose of this invention is to provide a method, storage medium, and computer equipment for inverting water quality parameters in a small sample, which can improve the accuracy of river water quality inversion under small sample conditions.

[0009] This invention provides a method for inverting water quality parameters in a small sample, comprising the following steps:

[0010] S1: Acquire the spectral data of the source water body and the corresponding inherent optical property data, the spectral data of the target water body and the corresponding water quality parameter data, perform preprocessing and data enhancement processing on the spectral data of the source water body and the target water body, and obtain the processed spectral data;

[0011] S2: Construct an inversion model of inherent optical properties. Use the spectral data of the processed source water body and the corresponding inherent optical property data to train the inversion model of inherent optical properties to obtain a trained inversion model of inherent optical properties.

[0012] S3: Input the spectral data of the target domain water body into the generative adversarial network, use the spectral data of the source domain water body for adversarial training, and generate target domain enhanced spectral data with the same spectral data distribution as the source domain water body.

[0013] S4: Input the enhanced spectral data of the target domain into the trained intrinsic optical property inversion model to obtain the predicted value of the intrinsic optical property of the target domain;

[0014] S5: Based on the predicted values ​​of the inherent optical properties of the target domain, the predicted results of the water quality parameters of the target domain are obtained using the water quality parameter inversion model.

[0015] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described small sample water quality parameter inversion method.

[0016] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described small sample water quality parameter inversion method.

[0017] Implementing the small-sample water quality parameter inversion method, storage medium, and computer equipment provided by this invention has the following beneficial effects:

[0018] This invention addresses the problems of existing water quality parameter inversion methods, such as strong dependence on large-scale sample data, significant differences in optical properties between publicly available marine water quality datasets and high-turbidity river water bodies, and the tendency for models to overfit or have insufficient generalization ability under small sample conditions. Even with extremely limited samples in the target domain, this invention fully utilizes publicly available marine water datasets. It compensates for insufficient samples by modeling the optical properties of individual opacities (IOPs) and using GAN data augmentation. A multi-scale feature fusion deep learning model is constructed using source domain spectral data and measured IOPs. Through multi-scale convolutional feature extraction, feature fusion, and regularization strategies, overfitting is effectively suppressed, improving generalization performance under different water body environments and achieving high-precision IOP inversion. A generative adversarial network (GAN) is used to match the distribution of the target domain spectrum, converting the target domain spectrum into enhanced spectral data consistent with the source domain distribution. This effectively reduces the difference in optical properties between the source and target domains, improving cross-regional applicability. The enhanced spectrum is input into the source domain IOPs model to predict the target domain IOPs. Based on the predicted IOPs values ​​and combined with a small-sample fine-tuning strategy, the water quality parameters are obtained using the water quality parameter inversion model.

[0019] This invention combines physics and data-driven approaches, introducing IOPs as intermediate variables to organically integrate the physical model with the data-driven model, thereby improving the interpretability and robustness of the model. By combining multi-scale feature fusion and fine-tuning strategies, it can better adapt to the high turbidity characteristics of river basins and maintain high prediction accuracy even when there are very few measured samples in the target domain. This enables cross-domain high-precision water quality parameter inversion and has the advantages of strong generalization ability, good interpretability, and wide application range. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0021] Figure 1 This is a flowchart of the small sample water quality parameter inversion method provided by the present invention;

[0022] Figure 2 This is a system architecture flowchart provided by the present invention;

[0023] Figure 3 This is the multi-scale feature fusion structure diagram provided by the present invention;

[0024] Figure 4 This is a flowchart of the CycleGAN domain adaptation process provided by the present invention;

[0025] Figure 5 This is a structural block diagram of the computer device provided by the present invention. Detailed Implementation

[0026] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] Figure 1 A schematic diagram of the small-sample water quality parameter inversion method of this embodiment is shown. In this embodiment, the small-sample water quality parameter inversion method includes the following steps:

[0028] S1: Acquire the spectral data of the source water body and the corresponding inherent optical property data, the spectral data of the target water body and the corresponding water quality parameter data, perform preprocessing and data enhancement processing on the spectral data of the source water body and the target water body, and obtain the processed spectral data;

[0029] In one exemplary embodiment, the spectral data of the source water body and the target water body include imagery, blue, green, red, near-infrared and short-wave infrared reflectance data of the corresponding date and location obtained from multispectral satellite data;

[0030] Inherent optical properties data include absorption coefficient and backscattering coefficient;

[0031] Water quality parameters include chlorophyll concentration, total suspended solids concentration, and absorption coefficient of colored soluble organic matter;

[0032] In one exemplary embodiment, preprocessing includes radiometric calibration and atmospheric correction;

[0033] Data augmentation processing includes random noise perturbation, band substitution, and spectral smoothing.

[0034] S2: Construct an inversion model of inherent optical properties. Use the spectral data of the processed source water body and the corresponding inherent optical property data to train the inversion model of inherent optical properties to obtain a trained inversion model of inherent optical properties.

[0035] In one exemplary embodiment, the intrinsic optical property data inversion model includes a multi-scale convolutional feature extraction module, a feature fusion module, and a fully connected regression module.

[0036] The multi-scale convolutional feature extraction module includes multiple parallel convolutional kernels for parallel convolutional extraction of spectral features in different spectral band ranges; the multiple convolutional kernels are 1×1 convolutional kernels, 3×3 convolutional kernels, and 5×5 convolutional kernels;

[0037] The feature fusion module is a concatenation operation unit or a weighted fusion unit, used to concatenate or weightedly fuse convolutional features of different scales; the weighted fusion unit is a compression-activation module or a convolutional block attention module;

[0038] The fully connected regression module maps fused features to predicted intrinsic optical properties through regression prediction;

[0039] As an exemplary embodiment, in step S2, the loss function uses mean squared error (MSE) or mean absolute error (MAE), and combines L2 regularization, Dropout, and batch normalization to improve the model's generalization ability.

[0040] S3: Input the spectral data of the target domain water body into the generative adversarial network, use the spectral data of the source domain water body for adversarial training, and generate target domain enhanced spectral data with the same spectral data distribution as the source domain water body.

[0041] In one exemplary embodiment, the generative adversarial network is a cycle-consistent generative adversarial network, a conditional generative adversarial network, or a Pix2Pix network.

[0042] In one exemplary embodiment, the generative adversarial network is a cycle-consistent generative adversarial network, and the loss function for adversarial training is as follows:

[0043] ,

[0044] ,

[0045] ,

[0046] in, , and These are adversarial loss, cycle consistency loss, and spectral reconstruction loss, respectively. , , , These represent the generator, discriminator, source domain data distribution, and target domain data distribution, respectively. , These represent source domain data and target domain data, respectively. Let represent the expected classification confidence of the discriminator for the real target domain sample y; This indicates that logarithmic processing is being performed; This is the classification function of the discriminator, used to distinguish whether the input data is real (from the real distribution) or fake (generated by the generator). Let G(x) represent the expected classification confidence of the discriminator for the generated sample G(x). Let be the mapping function of the generator, representing the transformed sample obtained after inputting the source domain sample x into the generator; These represent the expectations for the source and target domain data, respectively. This represents the target domain to source domain generator; This represents a source-to-target domain generator; Represents the L1 norm; This represents the predicted / generated spectrum, which can be the generator output or the result of the downstream decoder. True spectral label; This represents the L2 norm.

[0047] As an exemplary embodiment, L2 regularization, Dropout, and Batch Normalization are introduced during the model training process in step S3 to improve the model's generalization ability.

[0048] S4: Input the enhanced spectral data of the target domain into the trained intrinsic optical property inversion model to obtain the predicted value of the intrinsic optical property of the target domain;

[0049] S5: Based on the predicted values ​​of the inherent optical properties of the target domain, the predicted results of the water quality parameters of the target domain are obtained using the water quality parameter inversion model;

[0050] In one exemplary embodiment, the water quality parameter inversion model is a multilayer perceptron, convolutional neural network, recurrent neural network, neural network with attention mechanism, random forest, support vector regression, gradient boosting regression tree, partial least squares regression, principal component regression, K nearest neighbor regression, or a semi-analytical model based on the analytical formula of inherent optical properties.

[0051] In one exemplary embodiment, step S5 specifically includes:

[0052] The water quality parameter inversion model is initially trained using water quality parameter samples from the source and target domains to obtain the initially trained water quality parameter inversion model.

[0053] Based on the water quality parameter data of the target water body, the parameters of the initially trained water quality parameter inversion model are fine-tuned using early stopping and learning rate decay strategies to obtain an optimized water quality parameter inversion model.

[0054] The predicted values ​​of the inherent optical properties of the target domain are input into the optimized water quality parameter inversion model to obtain the predicted results of the water quality parameters of the target domain.

[0055] In some embodiments, the above-described method for inverting small sample water quality parameters can also be implemented in the following ways.

[0056] like Figure 2 The diagram shown is a system architecture flowchart. In this embodiment, the small sample water quality parameter inversion method includes the following steps:

[0057] Step 1: Data Acquisition and Preprocessing: Collect multi-band remote sensing reflectance spectral data of source and target water bodies, perform radiometric calibration and atmospheric correction on the spectral data to obtain preprocessed spectral data; among which, the source data includes the corresponding intrinsic optical parameters (IOPs) of the water body, and the target data includes at least the corresponding water quality parameters.

[0058] In one exemplary embodiment, data augmentation is further performed in step 1, including: performing enhancement processing on the spectral data such as random noise perturbation, band replacement, and spectral smoothing.

[0059] Step 2, Source Domain IOPs Inversion Model Training: Based on the source domain spectral data and the corresponding IOPs, construct and train an IOPs inversion model based on multi-scale feature fusion.

[0060] In one exemplary embodiment, the deep learning IOPs inversion model with multi-scale feature fusion includes:

[0061] The multi-scale convolution feature extraction module is used to extract spectral features in parallel convolutions across different spectral bands.

[0062] The feature fusion module is used to concatenate or weightedly fuse convolutional features of different scales.

[0063] A fully connected regression module is used to map fused features to predicted IOPs, where IOPs include at least one of absorption coefficient (a) and backscattering coefficient (bb);

[0064] like Figure 3 The diagram shown is a multi-scale feature fusion structure diagram.

[0065] Step 3, Cross-Domain Generative Adversarial Enhancement: Input the target domain spectral data into the generative adversarial network, perform adversarial training with the source domain spectral data, and generate target domain enhanced spectral data with the same distribution as the source domain.

[0066] In one exemplary embodiment, the generative adversarial network is at least one of CycleGAN, Conditional Generative Adversarial Networks (cGAN), or Pix2Pix network, and spectral reconstruction loss, domain adversarial loss, and cycle consistency loss are introduced during training to reduce the difference in spectral distribution between the target domain and the source domain.

[0067] like Figure 4 The diagram shown is a flowchart of the CycleGAN domain adaptation process.

[0068] In one exemplary embodiment, data augmentation is further performed in step 3, including introducing L2 regularization, Dropout, and Batch Normalization during model training to improve the model's generalization ability.

[0069] Step 4: Target domain IOPs prediction: Input the enhanced spectral data of the target domain into the source domain IOPs inversion model to obtain the predicted values ​​of the target domain IOPs.

[0070] Step 5: Water quality parameter inversion: Based on the predicted values ​​of IOPs in the target domain, the predicted results of water quality parameters in the target domain are obtained using the water quality parameter inversion model.

[0071] In one exemplary embodiment, the inversion model from IOPs to water quality parameters is at least one of the following: Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN, LSTM, GRU), Neural Network with Attention Mechanism, Random Forest (RF), Support Vector Regression (SVR), Gradient Boosting Regression Tree (GBRT, XGBoost, LightGBM), Partial Least Squares Regression (PLSR), Principal Component Regression (PCR), K Nearest Neighbor Regression (KNN), or a semi-analytical model based on an analytical formula of inherent optical properties; the water quality parameters include, but are not limited to, at least one of chlorophyll concentration (Chl-a), total suspended solids concentration (TSM), and dissolved colored organic matter absorption coefficient (CDOM).

[0072] In one exemplary embodiment, step 5 involves optimizing the IOPs-to-water quality parameter inversion model using a small-sample fine-tuning strategy, including:

[0073] Initial training was performed using water quality parameter samples from the source and target domains.

[0074] The model parameters were fine-tuned using a small number of measured samples in the target domain;

[0075] Early stopping and learning rate decay strategies are used during fine-tuning to prevent overfitting.

[0076] In some embodiments, the above-described method for inverting small sample water quality parameters can also be implemented in the following ways.

[0077] In this embodiment, the method for retrieving water quality parameters from a small sample includes the following steps:

[0078] Step S1: Data acquisition and preprocessing;

[0079] Step S2: Training the source domain IOPs inversion model;

[0080] Step S3: Cross-domain generation of adversarial enhancement;

[0081] Step S4: Predict target domain IOPs;

[0082] Step S5: Water quality parameter inversion;

[0083] The data acquisition in step S1 includes the following:

[0084] Data from n source domain samples, including collection locations Sampling date Imagery for corresponding dates and locations is obtained from multispectral satellite data. It includes at least the reflectance of the blue (B), green (G), red (R), near-infrared (NIR), and short-wave infrared (SWIR) bands, and the measured inherent optical quantities. ,in The absorption coefficient is... The backscattering coefficient is... .

[0085] Data from m target domain samples, including collection locations Sampling date Imagery for corresponding dates and locations is obtained from multispectral satellite data. It includes at least the reflectance of the blue (B), green (G), red (R), near-infrared (NIR), and short-wave infrared (SWIR) bands, and is a measured water quality parameter. Where Chl-a represents chlorophyll concentration, TSM represents total suspended solids concentration, and CDOM is the absorption coefficient of colored soluble organic matter. .

[0086] Step S2, the training of the source domain IOPs inversion model, includes the following:

[0087] 1) Constructing the dataset

[0088] Based on the coordinates of the sample points Extract the reflectance of the corresponding pixel in the red, green, blue, near-infrared, and short-wave infrared bands. , , , , ).

[0089] Then calculate each characteristic index according to the following formula:

[0090]

[0091]

[0092]

[0093]

[0094]

[0095]

[0096]

[0097]

[0098] Among them, NDWI is the Normalized Difference Water Index, NDGI is the Normalized Difference Green Index, EWI is the Environmental Water Index, WRI is the Water Ratio Index, ChI is the Chlorophyll Index, NDMI is the Normalized Difference Moisture Index, LSWI is the Land Surface Water Index, and TI is the Turbidity Index.

[0099] right Perform logarithmic processing .

[0100] Input feature vector:

[0101]

[0102] The training set and the test set are randomly divided in an 8:2 ratio.

[0103] 2) Construction of IOPs inversion model with multi-scale feature fusion

[0104] Spectral data based on the source region and its corresponding We construct and train a deep learning IOPs inversion model based on multi-scale feature fusion, with the following structural design:

[0105] Multi-scale convolutional feature extraction module: Employs multiple parallel convolutional kernels (such as 1×1, 3×3, 5×5) to extract features from spectral data across different spectral bands, capturing multi-scale spectral features. Convolutional kernels of different scales can extract features from local details to global patterns, adapting to the complex variations in water spectra.

[0106] Feature fusion module: This module integrates features extracted from convolutions at different scales into a unified feature representation through concatenation or weighted fusion (such as attention mechanisms). The fusion process can enhance key features and suppress irrelevant information through adaptive weights (such as squeeze-and-excitation (SE) modules or convolutional block attention modules (CBAM)).

[0107] Fully connected regression module: Input the fused features into the fully connected layer and predict the IOPs value through regression, including at least one of the absorption coefficient (a) and backscattering coefficient (bb).

[0108] The loss function uses mean squared error (MSE) or mean absolute error (MAE), and combines L2 regularization, Dropout, and batch normalization to improve the model's generalization ability.

[0109] The cross-domain generative adversarial enhancement in step S3 includes the following:

[0110] The model architecture employs a Cycle-Consistent Generative Adversarial Network (CycleGAN) architecture, including a generator. Generator and two discriminators The following loss function is introduced during training to reduce the difference in spectral distribution between the target domain and the source domain.

[0111] Combat losses:

[0112]

[0113] Cyclic consistency loss:

[0114]

[0115] Spectral reconstruction loss:

[0116]

[0117] The source domain (labeled) and target domain (unlabeled or with few labels) spectral data are input into the GAN for adversarial training. The generator and discriminator are iteratively optimized. Batch normalization and Dropout are used to prevent model overfitting and generate target domain enhanced spectral data with the same distribution as the source domain.

[0118] The target domain IOPs prediction in step S4 includes the following:

[0119] The enhanced spectral data generated in step S3 is standardized (e.g., Z-score normalization) to ensure consistency with the model input format.

[0120] Input the source domain IOPs inversion model trained in step S2 to obtain the target domain IOPs prediction values ​​(such as absorption coefficient a and backscattering coefficient bb).

[0121] The water quality parameter inversion in step S5 includes the following:

[0122] Based on the predicted IOPs values ​​of the target domain obtained in step S4, the water quality parameters of the target domain (such as chlorophyll concentration Chl-a, total suspended solids concentration TSM, and dissolved colored organic matter absorption coefficient) are further predicted using the water quality parameter inversion model.

[0123] For cases with a limited number of experimental samples in the target domain, a fine-tuning strategy is employed to optimize the published model. During fine-tuning, the low-level feature extraction module is frozen, and only the parameters of the high-level regression module are updated to prevent disruption of the original feature distribution. Early stopping and learning rate decay strategies are used to prevent overfitting.

[0124] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the small sample water quality parameter inversion method described above. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium may also include combinations of the above types of memory.

[0125] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the small sample water quality parameter inversion method described above.

[0126] like Figure 5As shown, the computer device 120 may include: at least one processor 121, such as a central processing unit (CPU), at least one communication interface 123, memory 124, and at least one communication bus 122. The communication bus 122 is used to enable communication between these components. The communication interface 123 may include a display screen and a keyboard; optionally, the communication interface 123 may also include a standard wired interface or a wireless interface. The memory 124 may be high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory 124 may also be at least one storage device located remotely from the aforementioned processor 121. The memory 124 stores application programs, and the processor 121 calls the program code stored in the memory 124 to execute any of the aforementioned method steps. The communication bus 122 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The communication bus 122 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5The term 124 is represented by a single line, but this does not imply a single bus or a single type of bus. The memory 124 may include volatile memory, such as random-access memory (RAM); it may also include non-volatile memory, such as flash memory, hard disk drive (HDD), or solid-state drive (SSD); or it may include combinations of the above types of memory. The processor 121 may be a central processing unit (CPU), a network processor (NP), or a combination of a CPU and an NP. The processor 121 may further include hardware chips. These hardware chips may be application-specific integrated circuits (ASICs), programmable logic devices (PLDs), or combinations thereof. The PLD may be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL), or any combination thereof. Optionally, the memory 124 is also used to store program instructions. The processor 121 can call the program instructions to implement the small sample water quality parameter inversion method as described in this embodiment.

[0127] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for inverting water quality parameters in a small sample, characterized in that, Includes the following steps: S1: Acquire the spectral data of the source water body and the corresponding inherent optical property data, the spectral data of the target water body and the corresponding water quality parameter data, and perform preprocessing and data enhancement processing on the spectral data of the source water body and the target water body to obtain the processed spectral data; S2: Construct an intrinsic optical property inversion model. The inversion model is trained using the processed spectral data of the source water body and the corresponding intrinsic optical property data. The trained intrinsic optical property inversion model includes a multi-scale convolutional feature extraction module, a feature fusion module, and a fully connected regression module. The multi-scale convolutional feature extraction module includes multiple parallel convolutional kernels for parallel convolutional extraction of spectral features across different spectral bands. These kernels are 1×1, 3×3, and 5×5 kernels. The feature fusion module is a splicing operation unit or a weighted fusion unit, used to splice or weightedly fuse convolutional features at different scales. The weighted fusion unit is a compression-excitation module or a convolutional block attention module. The fully connected regression module maps the fused features to predicted intrinsic optical property values ​​through regression prediction. S3: Input the spectral data of the target water body into the generative adversarial network, use the spectral data of the source water body for adversarial training, and generate target domain enhanced spectral data with the same spectral data distribution as the source water body; S4: Input the enhanced spectral data of the target domain into the trained intrinsic optical property inversion model to obtain the predicted value of the intrinsic optical property of the target domain; S5: Based on the predicted values ​​of the inherent optical properties of the target domain, the predicted results of the water quality parameters of the target domain are obtained using the water quality parameter inversion model.

2. The method for inverting small sample water quality parameters according to claim 1, characterized in that, The spectral data of the source and target water bodies include imagery, blue, green, red, near-infrared, and short-wave infrared reflectance data for the corresponding dates and locations obtained from multispectral satellite data; the intrinsic optical property data include absorption coefficient and backscattering coefficient; and the water quality parameter data include chlorophyll concentration, total suspended solids concentration, and absorption coefficient of colored soluble organic matter.

3. The method for inverting small sample water quality parameters according to claim 1, characterized in that, The preprocessing includes radiometric calibration and atmospheric correction; the data augmentation includes random noise perturbation, band substitution, and spectral smoothing.

4. The method for inverting small sample water quality parameters according to claim 1, characterized in that, The generative adversarial network is a cycle-consistent generative adversarial network, and the loss function for adversarial training is as follows: ， ， ， in, , and These are adversarial loss, cycle consistency loss, and spectral reconstruction loss, respectively. , , , These represent the generator, discriminator, source domain data distribution, and target domain data distribution, respectively. , These represent source domain data and target domain data, respectively. This represents the expected classification confidence of the discriminator for the true target domain sample y; This indicates that logarithmic processing is being performed; This is the classification function of the discriminator, used to distinguish whether the input data is real or fake; This represents the expected classification confidence of the discriminator for the generated sample G(x); Let be the mapping function of the generator, representing the transformed sample obtained after inputting the source domain sample x into the generator; These represent the expectations for the source and target domain data, respectively. This represents the target domain to source domain generator; This represents a source-to-target domain generator; Represents the L1 norm; This represents the predicted / generated spectrum, which can be the generator output or the result of the downstream decoder. True spectral label; This represents the L2 norm.

5. The method for inverting small sample water quality parameters according to claim 1, characterized in that, Step S5 specifically includes: The water quality parameter inversion model is initially trained using water quality parameter samples from the source and target domains to obtain the initially trained water quality parameter inversion model. Based on the water quality parameter data of the target water body, the parameters of the initially trained water quality parameter inversion model are fine-tuned using an early stopping and learning rate decay strategy to obtain an optimized water quality parameter inversion model. The predicted values ​​of the inherent optical properties of the target domain are input into the optimized water quality parameter inversion model to obtain the predicted results of the water quality parameters of the target domain.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the small sample water quality parameter inversion method as described in any one of claims 1-5.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the small sample water quality parameter inversion method as described in any one of claims 1-5.

Images