Lake water quality multi-parameter deep learning inversion framework based on hyperspectral data

Abstract

The invention relates to the technical field of water environment monitoring, and discloses a lake water quality multi-parameter deep learning inversion framework based on hyperspectral data, and the framework comprises the following steps: S1, data preparation and preprocessing: obtaining the hyperspectral data of lake water quality and corresponding water quality in-situ data, and preprocessing the hyperspectral data and the water quality in-situ data; s2, constructing a feature extraction and parameter inversion model which sequentially comprises a one-dimensional convolutional neural network module, a bidirectional long-short-term memory network module and a three-dimensional attention module, and inputting the preprocessed hyperspectral data into the model. According to the method, loss distribution can be dynamically optimized according to inversion requirements of different water quality parameters, the accuracy and generalization ability of simultaneous inversion of multiple parameters such as chlorophyll a, total suspended solids and transparency are remarkably improved, dependence of a traditional model on specific water body types is broken through, and the method can adapt to lake water bodies with different hydrological and optical characteristics.

Description

Technical Field

[0001] This invention relates to the field of water environment monitoring technology, specifically to a deep learning inversion framework for multi-parameter lake water quality based on hyperspectral data. Background Technology

[0002] Currently, remote sensing monitoring of water quality parameters in inland water bodies such as lakes and reservoirs is an important technical means in the field of water environment protection. Existing water quality parameter inversion techniques can be mainly divided into two categories: traditional models and machine learning-based models. Among them, traditional inversion methods include physical analysis models, semi-analysis models, and empirical algorithms. Physical analysis models are based on radiative transfer theory and achieve parameter inversion by constructing a mathematical model of the radiative transfer process in water bodies. However, this type of model requires extremely high accuracy of the inherent optical properties of water bodies, and these parameters vary significantly in water bodies in different regions and seasons, making it difficult to obtain universally applicable data. This results in limitations in large-scale applications. The applicability of remote sensing models in cross-regional water quality inversion tasks is limited. Semi-analytical models and empirical algorithms complete the inversion by establishing statistical mathematical relationships between remote sensing reflectance and water quality parameters. The former simplifies some optical processes based on the physical model, while the latter directly relies on measured data to fit regression equations. However, both types of models make oversimplified assumptions about the optical properties of water bodies. In Class II water bodies with complex optical properties, their inversion accuracy and applicability are greatly limited, and they cannot accurately characterize the heterogeneous optical behavior of water bodies. In recent years, with the development of artificial intelligence technology, some studies have begun to use machine learning methods to carry out water quality parameter inversion, such as support vector machines (SVR). Methods such as random forests and traditional neural network models have improved the fitting ability of the nonlinear relationship between spectra and water quality parameters through data-driven approaches, thus improving the inversion effect to some extent. However, existing technologies still have several significant drawbacks: First, there is a highly complex nonlinear coupling relationship between water quality parameters and water spectral signals. The remote sensing spectrum of the water surface is the result of the combined effects of multiple water components such as chlorophyll a, total suspended matter, and dissolved organic matter. How to effectively decouple this complex signal and accurately invert the concentration of individual components remains a core challenge for current technologies. Second, most existing machine learning models have relatively simple structures. Traditional neural networks can only capture local spectral features, while support vector machines and random forests have limited processing capabilities for long-sequence data. They are unable to effectively capture the long-range spectral sequence-dependent features contained in hyperspectral data, resulting in insufficient adaptability of the models to different spectral modes and poor generalization performance. Finally, most models are developed for specific regions or specific water body types, and the model parameters are highly bound to the training data, lacking universality. When applied to other water bodies with different hydrological conditions and optical properties, the model performance will drop significantly, exhibiting problems such as low inversion accuracy and poor robustness, and failing to meet the actual needs of cross-regional and multi-type water quality monitoring.

[0003] To address these issues, those skilled in the art have proposed a deep learning inversion framework for multi-parameter lake water quality based on hyperspectral data. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a deep learning inversion framework for multi-parameter lake water quality based on hyperspectral data, which solves the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a deep learning inversion framework for multi-parameter lake water quality based on hyperspectral data, comprising the following steps:

[0006] S1: Data preparation and preprocessing: Obtain hyperspectral data of lake water quality and corresponding in-situ measured data of water quality parameters, and preprocess the hyperspectral data and in-situ measured data of water quality parameters.

[0007] S2: Construct a feature extraction and parameter inversion model, which includes a one-dimensional convolutional neural network module, a bidirectional long short-term memory network module, and a three-dimensional attention module. The preprocessed hyperspectral data is input into the model, and the local spectral features are extracted by the one-dimensional convolutional neural network module, the long-range dependencies of the spectral sequence are captured by the bidirectional long short-term memory network module, and the features are adaptively weighted by the three-dimensional attention module.

[0008] S3: Model optimization. An adjustable weighted loss function is used to optimize the feature extraction and parameter inversion model. The adjustable weighted loss function improves the accuracy of multi-parameter synchronous inversion by assigning weights to the loss terms corresponding to different water quality parameters.

[0009] S4: Model training and validation. The optimized model is trained using the partitioned training set, and the performance of the trained model is validated using the test set.

[0010] S5: Water quality parameter inversion output. Input the hyperspectral data of the lake to be inverted into the trained and validated model, and output the inversion results of multiple parameters of lake water quality.

[0011] Preferably, in step S1, the hyperspectral data and in-situ measured water quality parameters are derived from the GLORIA public dataset. The spectral range of the hyperspectral data is 350nm to 900nm, and the spectral resolution is 1nm. The in-situ measured water quality parameters include chlorophyll a data, total suspended matter data, and transparency data.

[0012] Preferably, in step S1, the preprocessing includes: performing minimum-maximum normalization on the hyperspectral data and the in-situ measured water quality parameters respectively to eliminate feature scale differences; dividing the dataset composed of the preprocessed hyperspectral data and the in-situ measured water quality parameters into a training set and a test set in a 7:3 ratio, wherein 70% of the samples are used as the training set and 30% of the samples are used as the test set.

[0013] Preferably, in step S2, the one-dimensional convolutional neural network module includes 64 neurons, uses a convolution kernel of size 3, performs convolution operations on the input hyperspectral data with a stride of 1, and combines the ReLU activation function to extract local spectral features from the hyperspectral data; the input of the one-dimensional convolutional neural network module is hyperspectral reflectance data of 551 bands, and the output is a feature map after local feature extraction.

[0014] Preferably, in step S2, the bidirectional long short-term memory network module includes 64 neurons. The bidirectional long short-term memory network module receives the local spectral feature map output by the one-dimensional convolutional neural network module, and learns the forward and backward features of the spectral sequence through a bidirectional learning mechanism to capture the long-range spectral sequence dependencies in the hyperspectral data and output feature data containing long-range dependency information.

[0015] Preferably, in step S2, the 3D attention module includes a first convolutional layer, a second convolutional layer, a global average pooling layer, and a first fully connected layer connected in sequence. Both the first and second convolutional layers use a convolutional kernel of size 1 and a stride of 1 to convolve the feature data output by the bidirectional long short-term memory network module. The global average pooling layer performs global average pooling on the convolutionally processed feature data to compress the feature dimension. The first fully connected layer calculates feature weights on the pooled feature data and performs element-wise multiplication weighting on the feature weights and the feature data output by the bidirectional long short-term memory network module to enhance attention to key spectral regions and their neighborhood dependencies, while suppressing noise and irrelevant information interference, and outputting optimized feature data.

[0016] Preferably, in step S2, the feature extraction and parameter inversion model further includes a second fully connected layer. The second fully connected layer receives optimized feature data output by the 3D attention module. The number of neurons in the second fully connected layer is adaptively adjusted according to the number of lake water quality parameters to be inverted. When simultaneously inverting the three parameters of chlorophyll a, total suspended matter, and transparency, the second fully connected layer is set to have 3 neurons to output the inversion prediction values ​​corresponding to the three water quality parameters.

[0017] Preferably, in step S3, the weight coefficients of the adjustable weighted loss function are dynamically adjusted according to the inversion accuracy requirements of the lake water quality parameters to be inverted. When it is necessary to improve the inversion accuracy of a certain water quality parameter, the weight coefficient of the corresponding loss term of that water quality parameter is increased to prioritize the optimization of the inversion error of that parameter. When it is necessary to balance the inversion accuracy of multiple water quality parameters, the weight coefficients of the corresponding loss terms of each water quality parameter are configured to be equal or nearly equal values.

[0018] Preferably, in step S4, the model training adopts a strategy combining three-fold cross-validation and five-fold cross-validation. The model is trained through multiple rounds of iterative training using the training set. After each round of training, the inversion performance of the model is verified using the test set. The verification indicators include the coefficient of determination, mean absolute error, and root mean square error. When the model's verification indicators meet the preset threshold, the model is deemed to have passed the training.

[0019] Preferably, the method further includes a visualization step of the inversion results: the multi-parameter inversion results of lake water quality output in step S5 are correlated with the geospatial information of the lake to generate a spatial distribution map of water quality parameters. The spatial distribution map can intuitively show the numerical range and distribution differences of chlorophyll a concentration, total suspended matter concentration and transparency in different areas of the lake.

[0020] This invention provides a deep learning inversion framework for multi-parameter lake water quality based on hyperspectral data. It offers the following advantages:

[0021] 1. This invention organically combines a one-dimensional convolutional neural network module, a bidirectional long short-term memory network module, and a three-dimensional attention module. It can accurately extract local spectral features from hyperspectral data using the one-dimensional convolutional neural network module, fully capture long-range dependencies in spectral sequences using the bidirectional long short-term memory network module, and adaptively weight key spectral features using the three-dimensional attention module to suppress noise interference. Simultaneously, with an adjustable weighted loss function, it can dynamically optimize the loss allocation according to the inversion requirements of different water quality parameters, significantly improving the accuracy and generalization ability of multi-parameter inversion such as chlorophyll a, total suspended matter, and transparency. This breaks the dependence of traditional models on specific water body types and can adapt to lakes with different hydrological and optical characteristics, overcoming the limitations of poor generalization and insufficient robustness in existing technologies.

[0022] 2. This invention ensures the stability of model training and the reliability of inversion results through standardized data preparation and preprocessing procedures and reliable model training and verification strategies. Furthermore, it adds a visualization step for inversion results, which can correlate water quality parameter inversion results with geospatial information to generate spatial distribution maps, intuitively presenting the distribution differences of water quality parameters in different areas of the lake, facilitating water environment monitoring personnel to quickly grasp the water quality status. In addition, this framework can be adapted to inversion of hyperspectral remote sensing images after atmospheric correction, and can be extended to marine water parameter inversion by adapting to multispectral data, making it widely applicable and providing efficient technical support for large-scale, high-precision water environment monitoring. Attached Figure Description

[0023] Figure 1 This is a graph showing the relationship between the concentration of various substances in the water and the radiance of sunlight at the water-air interface according to the present invention.

[0024] Figure 2 This is an overall framework diagram of the present invention;

[0025] Figure 3 This is a scatter plot showing the correlation between the predicted chlorophyll a (Chl-a) values ​​and the in-situ measured values ​​in this invention.

[0026] Figure 4 This is a graph showing the results of the DLMWQR independent inversion of water quality parameters according to the present invention;

[0027] Figure 5 This is a graph showing the results of the DLMWQR simultaneous inversion of three water quality parameters according to the present invention;

[0028] Figure 6 This is a visualization result of the independent inversion of water quality parameters based on the multifunctional DLMWQR model and GF-5 imagery in this invention;

[0029] Figure 7 This is a diagram showing the results of the present invention inverting transparency (SDD) and total suspended matter (TSS) based on the DLMWQR model under GF-5 imagery;

[0030] Figure 8 This is a visualization of the chlorophyll a concentration in the Gulf of Mexico from February to April 2012, retrieved from MODIS data, in an application example of this invention.

[0031] Figure 9 This is a diagram illustrating the impact of different weight configurations of the adjustable weighted loss function on the results of simultaneously inverting three water quality parameters in this invention; wherein, Figure 9 (a) is a graph showing the root mean square error of chlorophyll a inversion as a function of weights. Figure 9 (b) is a graph showing the change of the root mean square error of total suspended matter inversion with weights. Figure 9 (c) is a graph showing the change of root mean square error in transparency inversion with weights;

[0032] Figure 10 This is a diagram illustrating the influence of different training parameters (training period, batch size, L2 regularization coefficient) on the water quality parameter inversion results in this invention; wherein, Figure 10 (a) shows the results of independent inversion of each parameter. Figure 10 (b) shows the result of simultaneously inverting the three parameters. Detailed Implementation

[0033] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Please see the appendix Figure 1 - Appendix Figure 10 This invention provides a deep learning inversion framework for multi-parameter lake water quality based on hyperspectral data, comprising the following steps:

[0035] S1: Data preparation and preprocessing. Obtain hyperspectral data of lake water quality and corresponding in-situ measured water quality parameters. Preprocess the hyperspectral data and in-situ measured water quality parameters. In S1, the hyperspectral data and in-situ measured water quality parameters are from the GLORIA public dataset. The spectral range of the hyperspectral data is 350nm~900nm, and the spectral resolution is 1nm. The in-situ measured water quality parameters include chlorophyll a data, total suspended matter data, and transparency data.

[0036] In S1, the preprocessing includes: performing minimum-maximum normalization on the hyperspectral data and the in-situ measured water quality parameters to eliminate feature scale differences; dividing the dataset composed of the preprocessed hyperspectral data and the in-situ measured water quality parameters into a training set and a test set in a 7:3 ratio, with 70% of the samples used as the training set and 30% of the samples used as the test set.

[0037] Specifically, in the data preparation and preprocessing stage, the GLORIA public dataset was selected as the source of hyperspectral data and in-situ measured water quality parameters because this dataset covers lake samples from multiple regions and time periods, ensuring the representativeness and diversity of the data. The hyperspectral data was selected in the spectral range of 350nm to 900nm with a spectral resolution of 1nm to cover the characteristic spectral range of components such as chlorophyll a and total suspended matter in the water, accurately capturing subtle spectral differences to support subsequent feature extraction. Minimum-maximum normalization was performed on the hyperspectral data and in-situ measured water quality parameters to eliminate scale differences between different features and ensure balanced feature weights during model training. The dataset was divided into training and test sets in a 7:3 ratio. The training set allows the model to learn the intrinsic relationship between spectra and water quality parameters, while the test set verifies the model's generalization ability on data not used in the training, thereby avoiding model overfitting and ensuring the reliability of the inversion framework.

[0038] S2: Construct a feature extraction and parameter inversion model. The model includes a one-dimensional convolutional neural network module, a bidirectional long short-term memory network module, and a three-dimensional attention module. The preprocessed hyperspectral data is input into the model, and the one-dimensional convolutional neural network module extracts local spectral features, the bidirectional long short-term memory network module captures the long-range dependencies of the spectral sequence, and the three-dimensional attention module adaptively weights the features. The one-dimensional convolutional neural network module includes 64 neurons, uses a convolution kernel of size 3, performs convolution operations on the input hyperspectral data with a stride of 1, and combines the ReLU activation function to extract local spectral features from the hyperspectral data. The input of the one-dimensional convolutional neural network module is hyperspectral reflectance data of 551 bands, and the output is the feature map after local feature extraction.

[0039] Specifically, in constructing the feature extraction and parameter inversion model, a sequential structure of "one-dimensional convolutional neural network (1D-CNN) module → bidirectional long short-term memory network (BiLSTM) module → three-dimensional attention module" is adopted. This design is based on the sequential characteristics of hyperspectral data and the precise requirements for feature extraction in multi-parameter water quality inversion. Hyperspectral data is distributed in a one-dimensional sequence with bands as the dimension. The 1D-CNN module is adapted to this data form. It is set with 64 neurons to ensure the dimensionality and richness of feature extraction. It uses a convolution kernel of size 3 and a stride of 1 for convolution operations, which can accurately capture local spectral units composed of three adjacent bands. Combined with the ReLU activation function, nonlinear transformation can be introduced to avoid the problem of insufficient fitting of linear models to complex spectral features. The input of 551 bands of hyperspectral reflectance data is a reasonable selection of bands in the effective spectral range (350nm~900nm) after preprocessing. The final output feature map can realize the extraction of key local information of the original spectrum and the initial filtering of redundant information. Then, BiLSTM is introduced. The BiLSTM module is designed because 1D-CNN can only extract local features, while different bands in a hyperspectral sequence may have long-range correlations related to water quality parameters. BiLSTM, through a bidirectional learning mechanism, can simultaneously capture such long-range dependencies from both the forward and reverse directions of the spectral sequence, compensating for the deficiency of 1D-CNN in capturing global sequence correlations and providing more comprehensive spectral feature support for subsequent multi-parameter inversion. Finally, a 3D attention module is added because the features extracted by the preceding modules may contain noise or secondary features unrelated to water quality parameters. Through adaptive weighting, the model can dynamically focus on the spectral feature regions that are more critical for the inversion of chlorophyll a, total suspended matter, and transparency, while suppressing the interference of noise and useless features, further improving the effectiveness and relevance of the features. The overall module design forms a complete logical chain of "local feature extraction → long-range correlation capture → key feature enhancement", effectively solving the problem that traditional models have difficulty simultaneously processing local details and global correlations in hyperspectral data and have insufficient feature effectiveness, laying a feature foundation for high-precision multi-parameter inversion of water quality.

[0040] The bidirectional long short-term memory network module includes 64 neurons. The bidirectional long short-term memory network module receives the local spectral feature map output by the one-dimensional convolutional neural network module, and learns the forward and backward features of the spectral sequence through a bidirectional learning mechanism to capture the long-range spectral sequence dependencies in the hyperspectral data and output feature data containing long-range dependency information.

[0041] The 3D attention module comprises a first convolutional layer, a second convolutional layer, a global average pooling layer, and a first fully connected layer connected in sequence. The first and second convolutional layers both use a convolutional kernel of size 1 and a stride of 1 to convolve the feature data output by the bidirectional long short-term memory network module. The global average pooling layer performs global average pooling on the convolutionally processed feature data to compress the feature dimension. The first fully connected layer calculates the feature weights on the pooled feature data and performs element-wise multiplication weighting on the feature weights and the feature data output by the bidirectional long short-term memory network module to enhance the focus on key spectral regions and their neighborhood dependencies, while suppressing the interference of noise and irrelevant information, and outputting optimized feature data.

[0042] The feature extraction and parameter inversion model also includes a second fully connected layer. The second fully connected layer receives the optimized feature data output by the 3D attention module. The number of neurons in the second fully connected layer is adaptively adjusted according to the number of lake water quality parameters to be inverted. When simultaneously inverting the three parameters of chlorophyll a, total suspended matter and transparency, the second fully connected layer is set to 3 neurons to output the inversion prediction values ​​of the corresponding three water quality parameters.

[0043] S3: Model optimization. An adjustable weighted loss function is used to optimize the feature extraction and parameter inversion model. The adjustable weighted loss function improves the accuracy of multi-parameter synchronous inversion by assigning weights to the loss terms corresponding to different water quality parameters.

[0044] The weighting coefficients of the adjustable weighted loss function are dynamically adjusted according to the inversion accuracy requirements of the lake water quality parameters to be inverted. When it is necessary to improve the inversion accuracy of a certain water quality parameter, the weighting coefficient of the corresponding loss term of that water quality parameter is increased to prioritize the optimization of the inversion error of that parameter. When it is necessary to balance the inversion accuracy of multiple water quality parameters, the weighting coefficients of the corresponding loss terms of each water quality parameter are configured to be equal or nearly equal values.

[0045] S4: Model training and validation. The optimized model is trained using the partitioned training set, and the performance of the trained model is validated using the test set.

[0046] The model training adopts a strategy that combines three-fold cross-validation and five-fold cross-validation. The model is trained in multiple rounds using the training set. After each round of training, the inversion performance of the model is validated using the test set. The validation metrics include the coefficient of determination, mean absolute error, and root mean square error. When the model's validation metrics meet the preset thresholds, the model is considered to have passed the training.

[0047] S5: Water quality parameter inversion output. Input the hyperspectral data of the lake to be inverted into the trained and validated model, and output the inversion results of multiple parameters of lake water quality.

[0048] It also includes a visualization step for the inversion results: the multi-parameter inversion results of lake water quality output in step S5 are correlated with the geospatial information of the lake to generate a spatial distribution map of water quality parameters. The spatial distribution map can intuitively show the numerical range and distribution differences of chlorophyll a concentration, total suspended matter concentration and transparency in different areas of the lake.

[0049] Specifically, in the water quality parameter inversion output stage, the hyperspectral data of the lake to be inverted is input into a trained and validated model. The model outputs inversion results based on the nonlinear correlation between the learned hyperspectral features and water quality parameters (chlorophyll a, total suspended matter, transparency, etc.). The visualization step of the inversion results generates a spatial distribution map of water quality parameters by associating the inversion results with the geospatial information of the lake. This visually presents the numerical range and distribution differences of water quality parameters in different areas of the lake, providing intuitive and quantitative spatial information support for water environment monitoring and management decisions.

[0050] This invention proposes a deep learning-based multi-parameter water quality inversion framework, DLMWQR, which integrates a one-dimensional convolutional neural network, a bidirectional long short-term memory network, and a three-dimensional attention mechanism to construct an end-to-end spectral-parameter mapping model.

[0051] Figure 2 The overall structure of DLMWQR is demonstrated, comprising three core components: a convolutional module, a BiLSTM module, and a 3D attention module. The input is hyperspectral remote sensing reflectance data across 551 bands. First, local spectral features are extracted using a 1D-CNN layer (64 neurons, kernel size 3). Then, a BiLSTM layer (64 neurons) captures long-range spectral sequence dependencies. Finally, a 3D attention mechanism adaptively weights the features, highlighting the characteristic responses of key spectral regions.

[0052] The parameter configurations and functional descriptions of each network layer are shown in Table 1 below:

[0053] Table 1 Summary of DLMWQR Model Parameters

[0054]

[0055]

[0056] The framework employs an adjustable weight loss function to optimize multi-parameter inversion performance, and its mathematical expression is as follows:

[0057] L total =αL Chl-a +βL TSS +γL SDD

[0058] The weighting coefficients satisfy α+β+γ=1

[0059] The individual loss function uses a non-linearly weighted mean square error:

[0060]

[0061] In the formula, N is the sample size, y i These are measured values. is the predicted value, and c is user-defined. This design effectively alleviates the vanishing gradient problem and improves the sensitivity to low concentration parameters.

[0062] For applications of hyperspectral remote sensing imagery, this invention obtains surface reflectance through atmospheric correction and establishes an inversion model by combining it with synchronously measured data. This framework can also be extended to marine environments, enabling parameter inversion across water body types by adapting to multispectral data such as MODIS.

[0063] Table 2 shows the accuracy comparison results between the DLMWQR model and the machine learning algorithm:

[0064] Table 2

[0065]

[0066]

[0067] Specifically, as can be seen from the table above, the DLMWQR model of this invention achieves a high coefficient of determination (R²) in the independent parameter inversion of chlorophyll a, total suspended matter, and transparency. 2 The DLMWIOR model generally outperforms SVR, XGBoost, CNN, and RNN models, with significantly lower mean absolute error (MAE) and root mean square error (RMSE). In simultaneous inversion of multiple parameters (chlorophyll a, total suspended matter, and transparency), its fit and error metrics are also superior to SVR and XGBoost models. Although the DLMWIOR model has a relatively long computation time, its advantages in accuracy and overall performance in multi-parameter lake water quality inversion are significant, fully demonstrating the technical value of this deep learning inversion framework in capturing the complex relationship between hyperspectral and water quality parameters and improving the accuracy of simultaneous multi-parameter inversion.

[0068] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A hyperspectral data-based lake water quality multi-parameter deep learning inversion framework, characterized in that, The method comprises the following steps: S1: data preparation and preprocessing, obtaining hyperspectral data and corresponding in-situ water quality parameter measurement data of the lake water quality, and preprocessing the hyperspectral data and in-situ water quality parameter measurement data; S2: constructing a feature extraction and parameter inversion model, the model comprises a one-dimensional convolutional neural network module, a bidirectional long short-term memory network module and a three-dimensional attention module in sequence, inputting the preprocessed hyperspectral data into the model, extracting local spectral features through the one-dimensional convolutional neural network module, capturing long-range dependence of spectral sequences through the bidirectional long short-term memory network module, and adaptively weighting the features through the three-dimensional attention module; S3: model optimization, using an adjustable weighted loss function to optimize the feature extraction and parameter inversion model, the adjustable weighted loss function assigns weights to loss terms corresponding to different water quality parameters to improve the simultaneous inversion accuracy of multiple parameters; S4: model training and verification, training the optimized model using the divided training set, and verifying the performance of the trained model using the test set; S5: water quality parameter inversion output, inputting the lake hyperspectral data to be inverted into the trained and verified model, and outputting the inversion results of the lake water quality multi-parameters.

2. The lake water quality multi-parameter deep learning inversion framework based on hyperspectral data according to claim 1, characterized in that, In S1, the hyperspectral data and in-situ water quality parameter measurement data are from the GLORIA public data set, the spectral range of the hyperspectral data is 350nm-900nm, and the spectral resolution is 1nm; the in-situ data includes chlorophyll-a data, total suspended solids data and transparency data.

3. The lake water quality multi-parameter deep learning inversion framework based on hyperspectral data according to claim 1, characterized in that, In S1, the preprocessing includes: performing minimum-maximum value normalization processing on the hyperspectral data and in-situ water quality parameter measurement data respectively to eliminate feature scale differences; dividing the data set composed of the preprocessed hyperspectral data and in-situ data into a training set and a test set in a ratio of 7:3, wherein 70% of the samples are used as the training set and 30% of the samples are used as the test set.

4. The lake water quality multi-parameter deep learning inversion framework based on hyperspectral data according to claim 1, characterized in that, In S2, the one-dimensional convolutional neural network module includes 64 neurons, uses a convolution kernel with a size of 3, performs convolution operation on the input hyperspectral data with a step size of 1, and extracts local spectral features in the hyperspectral data in combination with a ReLU activation function; the input of the one-dimensional convolutional neural network module is 551-band hyperspectral reflectance data, and the output is a feature spectrum after local feature extraction.

5. The hyperspectral data-based lake water quality multi-parameter deep learning inversion framework according to claim 1, characterized in that, In S2, the bidirectional long short-term memory network module includes 64 neurons, the bidirectional long short-term memory network module receives the local spectral feature spectrum output by the one-dimensional convolutional neural network module, learns the forward features and backward features of the spectral sequence respectively through a bidirectional learning mechanism, to capture the long-range spectral sequence dependence in the hyperspectral data, and outputs feature data containing long-range dependence information.

6. The hyperspectral data-based lake water quality multi-parameter deep learning inversion framework according to claim 1, characterized in that, In the S2, the 3D attention module comprises a first convolution layer, a second convolution layer, a global average pooling layer and a first full connection layer connected in sequence; the first convolution layer and the second convolution layer both adopt a convolution kernel with a size of 1 and a step of 1 to perform convolution processing on the feature data output by the bidirectional long short-term memory network module; the global average pooling layer performs global average pooling on the feature data after the convolution processing to compress the feature dimension; the first full connection layer calculates the feature weight of the feature data after the pooling, and performs element-level multiplication weighting on the feature data output by the bidirectional long short-term memory network module, so as to enhance the attention to the key spectral region and its neighborhood dependency relationship, suppress the interference of noise and irrelevant information, and output the optimized feature data.

7. The hyperspectral data-based lake water quality multi-parameter deep learning inversion framework according to claim 1, characterized in that, In the S2, the feature extraction and parameter inversion model further comprises a second full connection layer, which receives the optimized feature data output by the 3D attention module, and the number of neurons of the second full connection layer is adaptively adjusted according to the number of lake water quality parameters to be inverted; when simultaneously inverting chlorophyll-a, total suspended solids and transparency, the second full connection layer is provided with 3 neurons to output the inversion prediction values of the corresponding three water quality parameters.

8. The hyperspectral data-based lake water quality multi-parameter deep learning inversion framework according to claim 1, characterized in that, In the S3, the weight coefficient of the adjustable weighted loss function is dynamically adjusted according to the inversion accuracy requirement of the lake water quality parameters to be inverted; when it is required to mainly improve the inversion accuracy of a water quality parameter, the weight coefficient of the loss term corresponding to the water quality parameter is increased to prioritize the optimization of the inversion error of the parameter; when it is required to balance the inversion accuracy of multiple water quality parameters, the weight coefficients of the loss terms corresponding to the water quality parameters are configured to be equal or close to equal values.

9. The hyperspectral data-based lake water quality multi-parameter deep learning inversion framework according to claim 1, characterized in that, In the S4, the model training adopts a strategy combining three-fold cross-validation and five-fold cross-validation, and the model is trained by multiple rounds of iteration using the training set; after each round of training, the inversion performance of the model is verified using the test set, and the verification indexes include the coefficient of determination, the mean absolute error and the root mean square error; when the verification indexes of the model meet the preset threshold, the model training is determined to be qualified.

10. The hyperspectral data-based lake water quality multi-parameter deep learning inversion framework according to claim 1, characterized in that, The method further comprises an inversion result visualization step: associating the lake water quality multi-parameter inversion result output in the step S5 with the geographic spatial information of the lake to generate a spatial distribution map of the water quality parameters, and the spatial distribution map can intuitively display the numerical range and distribution difference of the chlorophyll-a concentration, the total suspended solids concentration and the transparency in different regions of the lake.

Images