River water quality monitoring method based on multi-source remote sensing data

Abstract

The invention provides a river water quality monitoring method based on multi-source remote sensing data, and belongs to the technical field of river water quality monitoring. The method comprises the following steps: establishing an inherent optical characteristic model of a river water body based on a Hybrid water body radiation transmission model, optimizing a calculation path by adopting a shortest path algorithm, and carrying out water body optical component inversion by adopting a quasi-analysis algorithm and a generalized inherent optical characteristic algorithm in combination with a multispectral characteristic enhancement model to obtain absorption coefficients of all components; and constructing a multi-band combination index to realize optical coupling effect decoupling, and applying the fine-tuned water quality parameter inversion basic model to a target river area to output a suspended matter concentration distribution diagram, a chlorophyll concentration distribution diagram and a transparency distribution diagram. The technical problem of low precision of remote sensing inversion of water quality parameters caused by mutual coupling of multiple optical active components in a river water body is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of river water quality monitoring, and in particular relates to a river water quality monitoring method based on multi-source remote sensing data. Background Art

[0002] River water quality monitoring is a crucial technical tool for environmental protection and water resource management. Traditional remote sensing water quality monitoring methods primarily rely on single spectral bands or simple band combination indices to invert water quality parameters. By establishing empirical statistical models to correlate remotely sensed reflectance with measured water quality data, these methods have achieved limited success in clean water bodies or rivers with a single type of pollution. However, traditional remote sensing inversion methods generally suffer from insufficient optical component demixing capabilities. When multiple optically active components, such as suspended matter, chlorophyll, and colored dissolved organic matter, coexist in river water, these components produce overlapping optical absorption and scattering signals within the same wavelength range. Traditional methods struggle to effectively separate the independent contributions of each component, resulting in significant errors in the inversion results. Due to the lack of effective methods for decoupling optical coupling effects, the coupling of multiple optically active components in river water results in low accuracy in remote sensing inversion of water quality parameters. This is particularly true in middle and lower reaches of rivers with complex pollution and eutrophic waters, where the inversion accuracy of traditional methods is significantly reduced, failing to meet the practical needs of accurate water quality monitoring. Summary of the Invention

[0003] In view of this, the present invention provides a river water quality monitoring method based on multi-source remote sensing data, which can solve the technical problem in the prior art that the remote sensing inversion of water quality parameters is not accurate due to the mutual coupling of multiple optically active components in river water bodies.

[0004] The present invention is implemented as follows: The present invention provides a river water quality monitoring method based on multi-source remote sensing data, comprising: acquiring multispectral satellite remote sensing image data and hyperspectral aerial remote sensing image data of a target river area, performing radiation calibration and atmospheric correction preprocessing on the multispectral satellite remote sensing image data and the hyperspectral aerial remote sensing image data to obtain preprocessed remote sensing image data; establishing an inherent optical property model of a river water body based on a Hydrolight water body radiation transfer model, using the suspended matter concentration, chlorophyll concentration, and colored dissolved organic matter concentration in the preprocessed remote sensing image data as input parameters, optimizing the radiation transfer calculation path using a shortest path algorithm, and simulating to obtain water body reflectance spectral data under different concentration conditions; The quasi-analytical algorithm and the generalized inherent optical property algorithm are used to invert the optical components of water bodies from the preprocessed remote sensing image data. The preprocessed remote sensing image data are input into the multispectral feature enhancement model, and the optimized optical component identification results are output; a multi-band combination index is constructed, and the decoupled optical coupling effect parameters of river water bodies are output; the water quality feature point selection function is used to determine the water quality feature points, and a fine-tuning data set is established. The model is fine-tuned using the pre-trained water quality parameter inversion basic model, and a fine-tuned water quality parameter inversion model is established; the fine-tuned water quality parameter inversion model is applied to the preprocessed remote sensing image data, and the suspended matter concentration distribution map, chlorophyll concentration distribution map and transparency distribution map are output to achieve spatial continuous monitoring results of river water quality.

[0005] Among them, multispectral satellite remote sensing image data includes three categories: clear water river data, turbidity river data and eutrophic river data. Clear water river data refers to the remote sensing data of river water bodies with a total suspended matter concentration of less than 25 mg / L and a chlorophyll concentration of less than 10 μg / L. Turbidity river data refers to the remote sensing data of river water bodies with a total suspended matter concentration of more than 100 mg / L. Eutrophic river data refers to the remote sensing data of river water bodies with a chlorophyll concentration of more than 30 μg / L and a higher nutrient salt concentration.

[0006] Among them, the shortest path algorithm applies the Dijkstra algorithm to optimize the selection of photon propagation paths in water body radiation transfer calculations. The scattering events in the water body are regarded as nodes in graph theory, and the scattering angle and propagation distance are used as the weights of the edges. By finding the shortest weighted path from the light source to the receiver, the numerical solution process of the radiation transfer equation is accelerated, significantly improving the computational efficiency of the simulation of complex water optical properties.

[0007] Among them, the multispectral feature enhancement model is a neural network model based on the visual Transformer architecture. It uses an image patch size of 16×16 pixels to divide the multispectral remote sensing image into a patch sequence for processing. The spatial spectral correlation features between different bands are extracted through the self-attention mechanism. The attention mask parameters within the model are used to control the fusion weights of features in different bands.

[0008] Among them, the multi-band combination index includes the normalized suspended matter index, the chlorophyll fluorescence peak index and the colored dissolved organic matter absorption index. The suspended matter absorption coefficient, the chlorophyll absorption coefficient and the colored dissolved organic matter absorption coefficient are used as input parameters, and the decoupled river water optical coupling effect parameters are output.

[0009] Among them, the water quality characteristic point selection function is used to select representative water quality monitoring points from the river area, and smooth the spatial distribution based on the Gaussian kernel function. The input includes six parameters: optical complexity of river water body, hydrological flow variation coefficient, pollution source distribution density, historical water quality variation degree, water quality mutation segment identification index and water quality slow change segment identification index. The output is the preferred water quality characteristic point location coordinates and the corresponding weight coefficient.

[0010] Among them, the method for obtaining the optical complexity of river water bodies includes calculating the spectral variance and spectral gradient of each pixel based on pre-processed remote sensing image data, standardizing the reflectance data of the red light band, green light band, blue light band and near-infrared band, calculating the root mean square of the reflectance difference between adjacent bands, and combining the comprehensive effects of water turbidity, chlorophyll concentration and colored dissolved organic matter concentration to establish an optical complexity evaluation model through multivariate regression analysis.

[0011] Among them, the method for obtaining the hydrological flow variation coefficient includes collecting daily flow monitoring data of the target river area in the past five years, calculating the monthly average flow, quarterly average flow and annual average flow, analyzing the coefficient of variation and peak frequency of the flow time series, using the Mann-Kendall trend test method to identify the flow change trend, and combining rainfall data and upstream reservoir scheduling information to establish a hydrological flow change assessment model.

[0012] Among them, the method for obtaining the distribution density of pollution sources includes collecting spatial location data of industrial enterprises, urban sewage treatment plants, agricultural non-point source pollution and domestic pollution sources in the river basin based on the geographic information system platform, allocating weights according to the type of pollution source and emission intensity, and using the kernel density estimation method to calculate the pollution load intensity per unit area. Combined with the river hydrological connectivity and pollutant diffusion model, a pollution source distribution density evaluation system with river sections as units is established.

[0013] Among them, the basic model for water quality parameter inversion is a pre-trained neural network model based on the deep residual network ResNet architecture. The model structure contains 50 convolutional layers and 4 residual blocks. Each residual block contains multiple convolutional layers and jump connections. The input layer receives the feature vector composed of multi-band combination index and optical component parameters, which is mapped to 128-dimensional hidden features through the fully connected layer. The output layer contains 3 branches to predict suspended matter concentration, chlorophyll concentration and transparency value respectively.

[0014] Among them, the structure of the multispectral feature enhancement model is a visual Transformer architecture containing 12 layers of Transformer encoders. Each layer contains a multi-head self-attention mechanism and a feedforward neural network module. The input layer converts each 16×16 pixel patch of the multispectral image into a 768-dimensional feature vector through linear projection. The multi-head attention mechanism contains 12 attention heads for parallel processing of different types of spectral spatial relationships. The total number of model parameters is approximately 860,000 training parameters.

[0015] Among them, the method for obtaining the degree of historical water quality variation includes collecting water quality monitoring data of the target river area over the past 10 years, including suspended matter concentration, chlorophyll concentration, transparency, total nitrogen concentration and total phosphorus concentration indicators, calculating the time series standard deviation and coefficient of variation of each water quality indicator, using the principal component analysis method to extract the main patterns of water quality changes, and combining seasonal decomposition and trend analysis to establish a comprehensive evaluation model for the degree of water quality variation.

[0016] Among them, the method for establishing the fine-tuning dataset includes collecting historical remote sensing image data and corresponding measured water quality monitoring data of the river area to be tested, with a time span of the past three years. Each sample contains a multi-band combination index, suspended matter absorption coefficient, chlorophyll absorption coefficient, and colored dissolved organic matter absorption coefficient as input features, and the corresponding suspended matter concentration, chlorophyll concentration, and transparency as target variables, and constructs a fine-tuning dataset containing 5,000 samples.

[0017] Among them, the fine-tuning steps include freezing the parameters of the first 40 layers of the water quality parameter inversion basic model, and only updating the parameters of the last 10 layers and the output layer. A smaller learning rate of 0.0001 is used for fine-tuning training, the batch size is set to 32 samples, the number of fine-tuning rounds is 100 rounds, and the early stopping mechanism is used to monitor the loss of the verification dataset. The loss function adopts the mean square error loss.

[0018] Among them, the normalized suspended matter index is a suspended matter concentration sensitive index constructed using the difference in reflectance between the red and near-infrared bands. The change in suspended matter concentration is quantified by comparing the normalized difference in the reflectance of the two bands. The chlorophyll fluorescence peak index is a chlorophyll concentration index constructed based on the fluorescence emission characteristics of chlorophyll near 685 nanometers. The chlorophyll concentration is quantified by comparing the reflectance difference between the fluorescence peak band and the adjacent bands.

[0019] Among them, before constructing the multi-band combination index step, the gating weight function is also included. It is a balance adjustment mechanism calculated based on three core data: water optical complexity, data quality assessment, and model confidence. When the balance value is in the range of 0 to 0.365, the linear weight adjustment function is used to enhance the contribution of basic optical characteristics. When the balance value is in the range of 0.365 to 0.486, the sigmoid weight adjustment function is used to balance the interaction of multiple optical components, thereby realizing the adaptive adjustment of the neural network gating mechanism under water conditions of different complexities.

[0020] The present invention adopts an inherent optical property modeling method based on the Hydrolight radiation transfer model, combined with a multispectral feature enhancement model and deep learning fine-tuning technology, effectively solving the problem of insufficient optical component decomposition ability of traditional remote sensing methods. By constructing a comprehensive feature system of multi-band combination index and optical component absorption coefficient, the independent identification and quantitative inversion of multiple components such as suspended solids, chlorophyll and colored dissolved organic matter are achieved. The present invention improves the computational efficiency of complex water body optical property simulation by introducing a shortest path algorithm to optimize radiation transfer calculations. The self-attention mechanism of the multispectral feature enhancement model is used to extract spatial spectral correlations between different bands, thereby enhancing the recognition ability of weak optical signals. Through the pre-training and fine-tuning strategy of the deep learning model, the prior knowledge of large-scale data and the optical characteristics of the target river are fully utilized, solving the technical problem of low accuracy of remote sensing inversion of water quality parameters due to the mutual coupling of multiple optically active components in river water bodies, and realizing high-precision spatial continuous monitoring of water quality parameters in complex optical environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flow chart of the method of the present invention.

[0022] Figure 2 Schematic diagram of the multispectral feature enhancement model structure.

[0023] Figure 3 Schematic diagram of the basic model structure for water quality parameter inversion. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0025] like Figure 1 FIG. 1 is a flow chart of a method for monitoring river water quality based on multi-source remote sensing data provided by the present invention. The method includes the following steps:

[0026] S01. Acquire multispectral satellite remote sensing image data and hyperspectral aerial remote sensing image data of a target river area, perform radiometric calibration and atmospheric correction preprocessing on the multispectral satellite remote sensing image data and the hyperspectral aerial remote sensing image data, and obtain preprocessed remote sensing image data, wherein the multispectral satellite remote sensing image data includes three categories: clear water river data, turbid river data, and eutrophic river data;

[0027] S02. Establishing an intrinsic optical property model of river water based on the Hydrolight water radiation transfer model, using the suspended solids concentration, chlorophyll concentration, and colored dissolved organic matter concentration in the pre-processed remote sensing image data as input parameters, optimizing the radiation transfer calculation path using the shortest path algorithm, and simulating water reflectance spectral data under different concentration conditions;

[0028] S03. Using a quasi-analysis algorithm and a generalized inherent optical property algorithm to invert the optical components of water bodies from the pre-processed remote sensing image data, inputting the pre-processed remote sensing image data into a multispectral feature enhancement model, outputting optimized optical component identification results, and obtaining the suspended matter absorption coefficient, chlorophyll absorption coefficient, and colored dissolved organic matter absorption coefficient;

[0029] S04. Using the suspended matter absorption coefficient, the chlorophyll absorption coefficient, and the colored dissolved organic matter absorption coefficient as input parameters, constructing a multi-band combined index, including a normalized suspended matter index, a chlorophyll fluorescence peak index, and a colored dissolved organic matter absorption index, and outputting decoupled river water optical coupling effect parameters;

[0030] S05. Based on the pre-processed remote sensing image data, water quality feature points are determined using a water quality feature point selection function. The multi-band combination index and the suspended matter absorption coefficient, the chlorophyll absorption coefficient, and the colored dissolved organic matter absorption coefficient are used as input features. In combination with the measured water quality parameter data of the water quality feature points, a fine-tuning dataset is established. The pre-trained water quality parameter inversion basic model is used to perform model fine-tuning to establish a fine-tuned water quality parameter inversion model.

[0031] S06. Apply the fine-tuned water quality parameter inversion model to the pre-processed remote sensing image data of the target river area, output a suspended matter concentration distribution map, a chlorophyll concentration distribution map, and a transparency distribution map, and realize spatial continuous monitoring results of river water quality.

[0032] Clear water river data refers to river water remote sensing data with total suspended solids concentrations below 25 mg / L and chlorophyll concentrations below 10 μg / L. This type of data accounts for 35% of the total dataset and is primarily distributed in mountainous areas and upstream river sections. It has a fundamental impact on the calibration accuracy of the optical model for the entire scheme. Its high transparency significantly enhances the bottom sediment reflection signal, necessitating the use of a shallow water optical model to separate the bottom sediment contribution and reduce the negative impact of bottom sediment reflection on the accuracy of water quality parameter inversion. Turbid river data refers to river water remote sensing data with total suspended solids concentrations above 100 mg / L. This type of data accounts for 40% of the total dataset and is primarily distributed in mid- and lower reaches and industrially polluted areas. Its strong scattering properties result in the optical signal being primarily contributed by suspended solids, which plays a decisive role in the model's suspended solids concentration inversion accuracy. High suspended solids concentrations mask the optical signals of chlorophyll and colored dissolved organic matter, necessitating multi-angle observations and polarization information to enhance the recognition of weak optical signals. The eutrophic river data refers to remote sensing data of river water bodies with chlorophyll concentrations higher than 30 micrograms per liter and high nutrient concentrations. This type of data accounts for 25% of the total dataset and is mainly distributed in areas affected by agricultural non-point source pollution and urban runoff. Its strong chlorophyll fluorescence effect will interfere with the optical identification of other water quality parameters. It is necessary to establish a time-series change model to capture the dynamic impact of the algae growth cycle on optical properties.

[0033] Among them, the shortest path algorithm applies the Dijkstra algorithm to optimize the selection of photon propagation paths in water body radiation transfer calculations. The scattering events in the water body are regarded as nodes in graph theory, and the scattering angle and propagation distance are used as the weights of the edges. By finding the shortest weighted path from the light source to the receiver, the numerical solution process of the radiation transfer equation is accelerated, which significantly improves the computational efficiency of the simulation of complex water optical properties.

[0034] Among them, the multispectral feature enhancement model is a neural network model based on the visual Transformer architecture. It uses an image patch size of 16×16 pixels to divide the multispectral remote sensing image into a patch sequence for processing. The spatial spectral correlation features between different bands are extracted through the self-attention mechanism. The attention mask parameters within the model are used to control the fusion weights of the features of different bands, thereby optimizing the recognition accuracy and spatial continuity expression of the optical components of the water body. The structure of the multispectral feature enhancement model is a visual Transformer architecture containing 12 layers of Transformer encoders. Each layer contains a multi-head self-attention mechanism and a feedforward neural network module. The input layer converts each 16×16 pixel patch of the multispectral image into a 768-dimensional feature vector through linear projection. The position encoding module adds spatial position information to each patch. The multi-head attention mechanism contains 12 attention heads for parallel processing of different types of spectral spatial relationships. The output layer maps the extracted features into water body optical component categories and concentration prediction results through the classification head. The total number of model parameters is approximately 860,000 training parameters. The steps for establishing a training dataset for the multispectral feature enhancement model include collecting a total of 150,000 multispectral satellite image data covering different climate regions and water body types around the world. Each image covers an area of ​​100 square kilometers and contains reflectance data of 8 spectral bands. Simultaneously, field water quality monitoring data of the corresponding area is collected, including water quality indicators such as suspended matter concentration, chlorophyll concentration, and total phosphorus concentration. The image data is divided into patches of 16×16 pixels and labeled with the corresponding water body optical component categories. A training dataset containing approximately 5 million labeled samples is constructed, of which 70% is used for model training, 20% is used for validation set evaluation, and 10% is used for final test set verification. The dataset is stratified and sampled according to geographic location and season to ensure the representativeness and generalization ability of the samples. The steps of training the multispectral feature enhancement model include using the AdamW optimizer to optimize the model parameters, setting the initial learning rate to 0.0001, using the cosine annealing learning rate scheduling strategy, setting the batch size to 64 patch samples, and the total number of training rounds to 200 rounds. After each round of training, the model performance is evaluated on the validation dataset, and an early stopping strategy is used to prevent overfitting. The loss function uses a combination of weighted cross entropy loss and mean square error loss to simultaneously optimize classification and regression tasks. During the training process, data enhancement techniques including random rotation, flipping and brightness adjustment are used to improve the robustness of the model. The model training is carried out on a high-performance computing cluster equipped with 8 GPUs. The total training time is about 72 hours. Finally, the model weight with the best performance on the validation dataset is selected as the final model for feature extraction of downstream water quality monitoring tasks.

[0035] Among them, the water quality parameter inversion basic model is a pre-trained neural network model based on the deep residual network ResNet architecture. The model structure includes 50 convolutional layers and 4 residual blocks. Each residual block contains multiple convolutional layers and jump connections. The input layer receives the feature vector composed of multi-band combination index and optical component parameters, which is mapped to 128-dimensional hidden features through the fully connected layer. The output layer contains 3 branches to predict suspended matter concentration, chlorophyll concentration and transparency value respectively. The total number of model parameters is about 23 million training parameters. The inference device is used to select parameters to optimize the execution efficiency of the model on different hardware platforms. The steps of establishing a pre-training data set for the water quality parameter inversion basic model include collecting remote sensing image data and corresponding measured water quality parameter data of different types of water bodies worldwide, covering a total of 500,000 sampling points in lakes, rivers, and nearshore ocean waters, with a time span from 2010 to 2020. Each sampling point contains multispectral reflectance data and synchronously measured water quality indicators such as suspended matter concentration, chlorophyll concentration, transparency, total nitrogen, and total phosphorus. The data is quality controlled and outliers are eliminated. Finally, a pre-training data set containing 450,000 high-quality samples is constructed, and stratified sampling is performed according to water body type and geographical distribution to ensure the diversity and representativeness of the training data. The pre-training method of the water quality parameter inversion basic model includes using the Adam optimizer for parameter update, the initial learning rate is set to 0.001, and a step-by-step learning rate decay strategy is adopted. The learning rate is decayed to 0.1 times the original value every 50 rounds of training. The batch size is set to 128 samples, the total number of pre-training rounds is 500 rounds, and the loss function adopts a weighted combination of mean square error loss and mean absolute error loss with a weight ratio of 0.7 to 0.3. During the pre-training process, Dropout regularization technology is used to prevent overfitting, and the Dropout probability is set to 0.2. The model pre-training is performed on a supercomputing cluster equipped with 16 GPUs. The total pre-training time is approximately 240 hours. Finally, the pre-trained model weights are saved as the basic model for subsequent fine-tuning tasks.

[0036] Among them, the water quality characteristic point selection function is used to select representative water quality monitoring points from the river area, and smooth the spatial distribution based on the Gaussian kernel function. The input includes six parameters: optical complexity of river water body, hydrological flow variation coefficient, pollution source distribution density, historical water quality variation degree, water quality mutation segment identification index and water quality slow change segment identification index. The output is the preferred water quality characteristic point location coordinates and the corresponding weight coefficient.

[0037] Among them, the method for obtaining the optical complexity of river water bodies includes calculating the spectral variance and spectral gradient of each pixel based on the preprocessed remote sensing image data, standardizing the reflectance data of the red light band, green light band, blue light band and near-infrared band, calculating the root mean square of the reflectance difference between adjacent bands, combining the comprehensive effects of water turbidity, chlorophyll concentration and colored dissolved organic matter concentration, establishing an optical complexity evaluation model through multivariate regression analysis, and finally outputting the river water body optical complexity index in the range of 0 to 1.

[0038] Among them, the method for obtaining the hydrological flow variation coefficient includes collecting daily flow monitoring data of the target river area in the past five years, calculating the monthly average flow, quarterly average flow and annual average flow, analyzing the coefficient of variation and peak frequency of the flow time series, using the Mann-Kendall trend test method to identify the flow change trend, combining rainfall data and upstream reservoir scheduling information, establishing a hydrological flow change assessment model, and outputting a standardized hydrological flow variation coefficient value.

[0039] Among them, the method for obtaining the pollution source distribution density includes collecting spatial location data of industrial enterprises, urban sewage treatment plants, agricultural non-point source pollution and domestic pollution sources in the river basin based on the geographic information system platform, allocating weights according to the type of pollution source and emission intensity, and using the kernel density estimation method to calculate the pollution load intensity per unit area. In combination with the river hydrological connectivity and pollutant diffusion model, a pollution source distribution density evaluation system with river sections as units is established, and a normalized pollution source distribution density index is output.

[0040] Among them, the method for obtaining the degree of historical water quality variation includes collecting water quality monitoring data of the target river area over the past 10 years, including suspended matter concentration, chlorophyll concentration, transparency, total nitrogen concentration and total phosphorus concentration indicators, calculating the time series standard deviation and coefficient of variation of each water quality indicator, using the principal component analysis method to extract the main patterns of water quality changes, combining seasonal decomposition and trend analysis, establishing a comprehensive evaluation model for the degree of water quality variation, and outputting a historical water quality variation score in the range of 0 to 100.

[0041] Among them, the method for obtaining the water quality mutation segment identification index includes detecting the mutation points of water quality parameters based on the pre-processed remote sensing image data and historical water quality monitoring data, calculating the gradient change rate of water quality indicators between adjacent monitoring sections, combining spatial autocorrelation analysis to identify water quality anomaly aggregation areas, using a change point detection algorithm to quantify the intensity and frequency of water quality mutations, establishing a water quality mutation segment identification model based on Bayesian reasoning, and outputting a standardized water quality mutation segment identification index.

[0042] Among them, the method for obtaining the water quality slow change segment identification index includes: based on long-term water quality monitoring data and remote sensing image data, using sliding window technology to calculate the local trend slope of water quality parameters, identifying the water quality slow change interval through time series smoothing processing, combining wavelet transform analysis to extract the low-frequency signal characteristics of water quality changes, using linear regression and non-parametric regression methods to quantify the degree and duration of water quality slow change, establishing a comprehensive identification model for water quality slow change segments, and outputting a water quality slow change segment identification index in the range of 0 to 1.

[0043] Among them, the method for establishing the fine-tuning dataset includes collecting historical remote sensing image data of the river area to be tested and the corresponding measured water quality monitoring data, with a time span of the past three years, including observation data under different seasons and hydrological conditions. Each sample contains a multi-band combination index, suspended matter absorption coefficient, chlorophyll absorption coefficient, and colored dissolved organic matter absorption coefficient as input features, and the corresponding suspended matter concentration, chlorophyll concentration, and transparency as target variables. The data is preprocessed including feature standardization and missing value interpolation, and finally a fine-tuning dataset containing 5,000 samples is constructed, of which 80% is used for fine-tuning training and 20% is used for fine-tuning verification.

[0044] Among them, the fine-tuning step includes freezing the parameters of the first 40 layers of the water quality parameter inversion basic model, and only updating the parameters of the last 10 layers and the output layer. A smaller learning rate of 0.0001 is used for fine-tuning training, the batch size is set to 32 samples, the number of fine-tuning rounds is 100 rounds, and an early stopping mechanism is used to monitor the loss of the validation dataset. When the loss of the validation dataset has not improved for 10 consecutive rounds, the training is stopped. The loss function uses the mean square error loss. Gradient clipping technology is used to prevent gradient explosion during the fine-tuning process. The clipping threshold is set to 1.0. After the fine-tuning is completed, the model weights with the best performance are saved as the final fine-tuned water quality parameter inversion model.

[0045] Among them, the gating weight function is a balance adjustment mechanism calculated based on three core data: water optical complexity, data quality assessment and model confidence. The comprehensive balance value is obtained by calculating the water turbidity variance, spectral signal-to-noise ratio and model prediction consistency. When the balance value is in the range of 0 to 0.365, a linear weight adjustment function is used to enhance the contribution of basic optical characteristics. When the balance value is in the range of 0.365 to 0.486, a sigmoid weight adjustment function is used to balance the interaction of multiple optical components. When the balance value is in the range of 0.486 to 0.752, an exponential weight adjustment function is used to highlight the signal intensity of the dominant optical component. When the balance value exceeds 0.752, a step weight adjustment function is used to force the selection of a single most reliable optical feature for water quality parameter inversion, thereby realizing adaptive adjustment of the neural network gating mechanism under water conditions of different complexities.

[0046] Among them, the Hydrolight water body radiation transfer model is a numerical calculation model of the optical characteristics of water bodies established based on radiation transfer theory. It simulates the propagation process of light in water bodies by solving the radiation transfer equation and calculates the irradiance and radiance distribution at different depths and angles.

[0047] Among them, the quasi-analytical algorithm is a water color remote sensing inversion algorithm that combines analytical models and empirical relationships. It realizes quantitative inversion of water component concentrations by establishing a semi-empirical relationship between water body reflectance and inherent optical properties.

[0048] Among them, the generalized intrinsic optical properties algorithm is a water color remote sensing inversion algorithm based on the principle of linear superposition of intrinsic optical properties, which decomposes the total absorption coefficient of the water body into the absorption contributions of pure water, phytoplankton, non-algae particles and colored dissolved organic matter.

[0049] Among them, the optical coupling effect is the optical signal mixing phenomenon caused by the interaction of multiple optically active components such as suspended matter, chlorophyll and colored dissolved organic matter in river water, which makes it difficult to accurately invert water quality parameters with a single band.

[0050] Among them, the normalized suspended matter index is a suspended matter concentration sensitivity index constructed using the difference in reflectivity between the red light and near-infrared bands, and quantifies the change in suspended matter concentration by comparing the normalized difference in reflectivity between the two bands.

[0051] The chlorophyll fluorescence peak index is a chlorophyll concentration index constructed based on the fluorescence emission characteristics of chlorophyll near 685 nanometers, and quantifies the chlorophyll concentration by comparing the reflectivity difference between the fluorescence peak band and the adjacent band.

[0052] The specific implementation of the above steps is described in detail below.

[0053] The specific implementation method of step S01 is to synchronously obtain multi-source remote sensing image data of the target river area through satellite remote sensing platforms and aerial remote sensing platforms. First, a multispectral satellite sensor is used to collect data from 8 spectral bands, including visible light bands, near-infrared bands, and short-wave infrared bands. The spatial resolution is set to 30m, the temporal resolution is 16 days, and the coverage area includes the entire watershed area. At the same time, a hyperspectral aerial remote sensing platform is used to obtain high-precision image data with 200 continuous spectral bands. The spectral resolution reaches 5nm and the spatial resolution is increased to 1m. Refined observations are carried out on key monitored river sections. The purpose of this step is to build a multi-scale and multi-spectral remote sensing data foundation to provide a rich source of spectral information for subsequent water quality parameter inversion. Next, the acquired raw remote sensing image data is subjected to radiometric calibration processing, and the digital quantization value is converted into the reflectance of the top layer of the atmosphere. The radiometric calibration coefficient is dynamically adjusted according to the sensor factory parameters and the on-orbit attenuation characteristics to ensure the radiometric consistency of data at different time phases. Then, atmospheric correction preprocessing was performed, and the 6S radiation transfer model was used to eliminate the effects of atmospheric scattering and absorption on surface reflectivity. The input parameters included meteorological elements such as solar zenith angle, observation zenith angle, relative azimuth, atmospheric water vapor content, aerosol optical depth, and atmospheric pressure. The atmospheric water vapor content threshold was set to 1.5 g / cm 2 to 4.0g / cm 2 Within this range, aerosol optical depth is controlled within the range of 0.1 to 0.8, obtaining high-quality surface reflectance data as pre-processed remote sensing image data. The data classification phase divides multispectral satellite remote sensing image data into three categories based on the optical properties of water bodies: clear water river data, turbidity river data, and eutrophic river data. The classification criteria are based on dual thresholds of total suspended matter concentration and chlorophyll concentration. Clear water river data corresponds to total suspended matter concentrations below 25 mg / L and chlorophyll concentrations below 10 μg / L, turbidity river data corresponds to total suspended matter concentrations above 100 mg / L, and eutrophic river data corresponds to chlorophyll concentrations above 30 μg / L and high nutrient concentrations.

[0054] The specific implementation of step S02 is to establish a numerical calculation framework for the inherent optical characteristics of river water bodies based on the Hydrolight water body radiation transfer model. The model simulates the propagation process and scattering characteristics of photons in water bodies by solving the radiation transfer equation. First, the suspended solids concentration, chlorophyll concentration and colored dissolved organic matter concentration extracted from the pre-processed remote sensing image data are input as key parameters of the model. The suspended solids concentration range is set to 1mg / L to 500mg / L, the chlorophyll concentration range is set to 0.5μg / L to 150μg / L, and the colored dissolved organic matter concentration range is set to 0.1 / m -1 to 5.0 / m -1. This step uses the shortest path algorithm to optimize the radiation transmission calculation path, applies the Dijkstra algorithm to the problem of photon propagation path selection in water bodies, and regards scattering events as graph theory nodes, scattering angles and propagation distances as edge weights to find the shortest weighted path from the light source to the sensor, thereby significantly improving the computational efficiency of simulating the optical properties of complex water bodies. During the model calculation process, multiple depth layers with water depths ranging from 0.1m to 20m are set, the solar zenith angle varies from 0 degrees to 80 degrees, and the azimuth angle covers a full range from 0 degrees to 360 degrees. The water reflectance spectral data under different depths and angles are output. The spectral wavelength range covers 400nm to 900nm, and the spectral resolution is 1nm, providing theoretical basis data for subsequent optical component inversion.

[0055] Step S03 involves quantitatively inverting and analyzing the optical components of water bodies using preprocessed remote sensing image data using a quasi-analytical algorithm and a generalized intrinsic optical property algorithm. The quasi-analytical algorithm establishes a semi-empirical relationship between remote sensing reflectance and intrinsic optical properties, and solves for water component concentrations using an iterative optimization method. The iterative convergence threshold is set to 0.001, and the maximum number of iterations is limited to 50. Based on the principle of linear superposition of intrinsic optical properties, the generalized intrinsic optical property algorithm decomposes the total water absorption coefficient into four independent contributing components: pure water absorption, phytoplankton absorption, non-algae particulate matter absorption, and colored dissolved organic matter absorption. The absorption coefficient of each component is quantitatively separated using a spectral decomposition algorithm. This step incorporates a multispectral feature enhancement model to optimize the accuracy of optical component identification. This model, based on a visual Transformer architecture, employs a 16×16 pixel image patch segmentation strategy and uses a self-attention mechanism to extract spatial-spectral correlation features across different bands. The model outputs optimized suspended matter absorption coefficients, chlorophyll absorption coefficients, and colored dissolved organic matter absorption coefficients, improving absorption coefficient accuracy by over 15% compared to traditional algorithms. The output optical component identification results provide high-quality basic data for the subsequent construction of multi-band combination index.

[0056] The specific implementation method of step S04 is to use the suspended matter absorption coefficient, chlorophyll absorption coefficient and colored dissolved organic matter absorption coefficient obtained in step S03 as input parameters to construct a multi-band combined index system for the optical coupling effect of river water bodies. This step first calculates the normalized suspended matter index, and uses the normalized difference between the reflectance of the red light band (665nm) and the near-infrared band (865nm) to quantify the change in suspended matter concentration. The index calculation is based on the ratio of the reflectance difference of the two bands to the reflectance sum value, which effectively eliminates the interference of atmospheric and water surface reflection. The construction of the chlorophyll fluorescence peak index is based on the fluorescence emission characteristics of chlorophyll near a wavelength of 685nm. The chlorophyll concentration is quantified by comparing the reflectance difference between the fluorescence peak band and the adjacent bands (680nm and 690nm). This index can effectively eliminate the optical interference of suspended matter and colored dissolved organic matter. The chromatic dissolved organic matter absorption index is constructed using the reflectance ratio of the blue band (443nm) and the green band (555nm), leveraging the strong absorption characteristics of chromatic dissolved organic matter in the short wavelength range for quantitative inversion. The core function of this step is to effectively decouple the optical coupling effect generated by the interaction of multiple optically active components in river water through the synergistic effect of multi-band combined indices. The decoupled river water optical coupling effect parameters are then output, laying the foundation for the accurate inversion of water quality parameters.

[0057] The specific implementation of step S05 is to establish a fine-tuning dataset based on the pre-processed remote sensing image data and fine-tune and optimize it using the pre-trained water quality parameter inversion basic model. First, a water quality feature point selection function is used to determine representative water quality monitoring points. This function smoothes the spatial distribution based on a Gaussian kernel function. The input parameters include six core indicators: the optical complexity of the river water body, the hydrological flow variation coefficient, the pollution source distribution density, the historical water quality variation degree, the water quality sudden change section identification index, and the water quality slow change section identification index. The optical complexity of the river water body is obtained by calculating the spectral variance and spectral gradient of the reflectance in the red, green, blue, and near-infrared bands, and the optical complexity threshold is set within the range of 0.3 to 0.8. The hydrological flow variation coefficient is calculated based on the daily flow monitoring data of the past five years. The Mann-Kendall trend test method is used to identify the flow change trend, and the variation coefficient threshold is set within the range of 0.2 to 1.5. The pollution source distribution density is calculated using the kernel density estimation method to calculate the pollution load intensity per unit area, and the density index ranges from 0 to 100. The degree of historical water quality variability was calculated based on water quality monitoring data from the past 10 years, including time series variation characteristics of five indicators: suspended matter concentration, chlorophyll concentration, transparency, total nitrogen concentration, and total phosphorus concentration. The degree of variability was scored on a scale of 0 to 100. During the fine-tuning dataset construction process, historical remote sensing imagery and measured water quality monitoring data from the river area under test were collected from the past three years. A dataset of 5,000 samples was constructed, 80% of which were used for fine-tuning training and 20% for fine-tuning validation. The fine-tuning process adopted a transfer learning strategy, freezing the parameters of the first 40 layers of the basic water quality parameter inversion model and updating only the parameters of the last 10 layers and the output layer. The learning rate was set to 0.0001, the batch size was 32 samples, and the number of fine-tuning rounds was 100. An early stopping mechanism was used to prevent overfitting, and training was terminated when the validation dataset loss did not improve for 10 consecutive rounds.

[0058] The specific implementation method of step S06 is to apply the fine-tuned water quality parameter inversion model to the pre-processed remote sensing image data of the target river area to achieve spatial continuous monitoring and visual output of river water quality. This step first divides the remote sensing image data of the entire river area into spatial grids, and the grid size is set to 30m×30m, which is consistent with the spatial resolution of satellite remote sensing images. Multi-band combination index and optical component parameters are extracted for each grid unit, including normalized suspended matter index, chlorophyll fluorescence peak index, colored dissolved organic matter absorption index and corresponding absorption coefficient data, and these characteristic parameters are used as inputs of the fine-tuned model to predict water quality parameters. The model output includes suspended matter concentration distribution map, chlorophyll concentration distribution map and transparency distribution map. Figure 3The spatial distribution results of water quality parameters include suspended solids concentration ranging from 1 mg / L to 500 mg / L, chlorophyll concentration ranging from 0.5 μg / L to 150 μg / L, and transparency ranging from 0.1 m to 5.0 m. This step uses a spatial interpolation algorithm to smooth the discrete prediction results to generate a continuous spatial distribution map of water quality parameters. The interpolation method uses kriging interpolation, the semivariogram model selects a spherical model, and the search radius is set to 150 m. The final output of the water quality monitoring results is presented in the form of a color-coded raster image, realizing spatial continuous monitoring of river water quality and providing a scientific basis for water environment management and pollution source identification.

[0059] The detailed structure of the multispectral feature enhancement model is based on the visual Transformer architecture and consists of a 12-layer Transformer encoder. Each encoder layer consists of a multi-head self-attention mechanism and a feedforward neural network module. The input layer segments the multispectral remote sensing image into a sequence of 16×16 pixel patches. Each patch is converted into a 768-dimensional feature vector via a linear projection layer. The position encoding module adds two-dimensional spatial position information to each image patch, using a position encoding scheme that combines sine and cosine functions. The multi-head self-attention mechanism consists of 12 attention heads, each with a dimension of 64. Different types of spectral-spatial relationship features are processed in parallel, and the attention weights are normalized using a softmax function. The feedforward neural network module uses a two-layer fully connected layer structure with a hidden layer dimension of 3072 and a GELU activation function. The output layer uses a classification head to map the extracted high-dimensional features into water optical component classification and concentration predictions. The total number of model parameters is approximately 860,000. To construct the training dataset, a total of 150,000 multispectral satellite imagery images covering diverse climate regions and water body types were collected. Each image covers an area of ​​100 square kilometers and includes reflectance data from eight spectral bands. Field water quality monitoring data for the corresponding regions was also collected simultaneously, including three key water quality indicators: suspended matter concentration, chlorophyll concentration, and total phosphorus concentration. The image data was divided into 16×16 pixel patches and labeled with corresponding water optical component categories. A training dataset containing approximately 5 million labeled samples was constructed, divided into training, validation, and test sets in a 7:2:1 ratio. The dataset was stratified by geographic location and season to ensure representativeness and model generalization. The model was trained using the AdamW optimizer with an initial learning rate of 0.0001, a cosine annealing learning rate scheduling strategy, a batch size of 64 patches, and 200 training epochs. The loss function used a combination of weighted cross entropy and mean squared error losses to optimize both classification and regression.

[0060] The basic model for water quality parameter inversion is based on a deep residual network (ResNet) architecture. The model consists of 50 convolutional layers and four residual blocks, each of which contains multiple convolutional layers and skip connections. The input layer receives a 128-dimensional feature vector consisting of a multi-band combined index and optical component parameters. Residual connections and batch normalization are used in the model's intermediate layers to ensure stable training of the deep network. The convolution kernel size is set to 3×3, with a stride of 1 and padding of same. The fully connected layer maps the convolutional features into 128-dimensional latent features. The output layer consists of three branches, each using a separate fully connected layer and regression activation function to predict suspended matter concentration, chlorophyll concentration, and transparency, respectively. The model has approximately 23 million training parameters. The pre-training dataset was compiled using remote sensing imagery and corresponding measured water quality parameter data from various water bodies worldwide, covering 500,000 sampling points across lakes, rivers, and coastal oceans, spanning the period from 2010 to 2020. Each sampling point contains multispectral reflectance data and simultaneously measured water quality indicators: suspended matter concentration, chlorophyll concentration, transparency, total nitrogen, and total phosphorus. The data undergoes quality control and outlier removal, ultimately constructing a pre-training dataset containing 450,000 high-quality samples. The samples are stratified by water body type and geographic distribution. The pre-training method uses the Adam optimizer for parameter updates, with an initial learning rate set to 0.001 and a step-wise learning rate decay strategy, with the learning rate decayed to 0.1 times the original value every 50 training rounds. The batch size is set to 128 samples, and the total number of pre-training rounds is 500. The loss function uses a weighted combination of mean squared error loss and mean absolute error loss, with a weight ratio of 0.7 to 0.3. Dropout regularization is used during pre-training to prevent overfitting, with a dropout probability of 0.2.

[0061] It should be noted that the process of obtaining the optical complexity of river water bodies first analyzes the spectral characteristics of each pixel based on the pre-processed remote sensing image data, and quantifies the complexity of the optical characteristics of the water body by calculating the spectral variance and spectral gradient of the reflectance data of the red light band, green light band, blue light band and near-infrared band. In the specific calculation process, the reflectance data of the four main bands are standardized to eliminate the influence of the magnitude difference between different bands, and then the root mean square value of the reflectance difference between adjacent bands is calculated as the spectral gradient index. The formula for calculating optical complexity is: Where n represents the number of bands, and m represents the number of adjacent band pairs. This step also requires a multivariate regression analysis based on the combined effects of water turbidity, chlorophyll concentration, and colored dissolved organic matter concentration. This establishes an optical complexity evaluation model, incorporating the interactions of multiple optically active components into the complexity calculation framework. The final output is a standardized river water optical complexity index ranging from 0 to 1, providing important optical characteristic criteria for the subsequent selection of water quality feature points.

[0062] The method for obtaining the hydrological flow variation coefficient requires collecting daily flow monitoring data from the target river area for the past five years and establishing a complete hydrological time series database that covers different hydrological years and seasonal variation characteristics. During data processing, the monthly average flow, quarterly average flow, and annual average flow are calculated separately, and the statistical characteristics and variation patterns of the flow time series are analyzed. The formula for calculating the variation coefficient is: At the same time, it is necessary to analyze the distribution characteristics of traffic peak frequency and duration. This step uses the Mann-Kendall trend test method to identify the long-term trend of traffic series. The test statistic calculation formula is: Sign function (flow j -flow i ), where the sign of the function is determined by the positive or negative difference in the values. A hydrological flow change assessment model was established by combining rainfall data and upstream reservoir operation information. This model quantified the contribution of artificial regulation and natural factors to flow changes. A standardized hydrological flow change coefficient was output through a comprehensive weight distribution method, providing a quantitative basis for identifying the impact of hydrological conditions on water quality changes.

[0063] The process of obtaining the distribution density of pollution sources is based on the geographic information system platform to collect the spatial location and attribute information of various pollution sources in the river basin, including industrial enterprise discharge outlets, urban sewage treatment plant outlets, agricultural non-point source pollution areas and distribution points of domestic pollution sources. During the data collection process, it is necessary to record in detail the key attribute information such as the type, scale, emission intensity and type of pollutants of each pollution source. Differentiated weights are allocated according to the type and emission intensity of pollution sources. The weight coefficient of industrial point source pollution is set to 0.4, the weight coefficient of urban non-point source pollution is set to 0.3, the weight coefficient of agricultural non-point source pollution is set to 0.2, and the weight coefficient of domestic pollution sources is set to 0.1. The kernel density estimation method is used to calculate the pollution load intensity per unit area. The kernel density function is: The kernel function adopts the Gaussian kernel function. Combining river hydrological connectivity and pollutant diffusion models, the system considers the migration and transformation of pollutants in rivers and establishes a pollution source distribution density evaluation system based on river sections. The system outputs a normalized pollution source distribution density index, providing quantitative information on the spatial distribution of pollution sources for the optimal placement of water quality monitoring points.

[0064] Obtaining the degree of historical water quality variation requires collecting continuous water quality monitoring data from the target river area over the past 10 years and establishing a long-term water quality database, including monthly monitoring data for five key indicators: suspended matter concentration, chlorophyll concentration, transparency, total nitrogen concentration, and total phosphorus concentration. During data preprocessing, outliers are identified and processed, and the 3-times standard deviation criterion is used to eliminate obviously abnormal observations. Missing data are supplemented by linear interpolation. The standard deviation and coefficient of variation of each water quality indicator time series are calculated. The coefficient of variation calculation formula is: 100%, where k represents different water quality indicators. Principal component analysis was used to extract the main patterns and contributing factors of water quality changes. The cumulative contribution rate of the principal components was set to above 85% to identify the dominant factors affecting water quality changes. The seasonal decomposition technique was used to decompose the water quality time series into trend terms, seasonal terms, and random terms. The trend analysis was performed using the seasonal-trend decomposition using the Loess method. A comprehensive evaluation model for water quality variation was established, and the comprehensive evaluation index was: The weights of five water quality indicators are determined according to their importance to water quality assessment, and a historical water quality variability score ranging from 0 to 100 is output.

[0065] The water quality mutation section identification index is obtained based on spatial statistical analysis of pre-processed remote sensing image data and historical water quality monitoring data. The cumulative sum control chart method is used to detect the mutation points and mutation intensity of water quality parameters. First, the gradient change rate of each water quality indicator between adjacent monitoring sections is calculated. The gradient calculation formula is: Where i represents the river section number. Spatial autocorrelation analysis is combined to identify abnormal water quality clusters, and Moran's I index is used to quantify the degree of spatial autocorrelation: A change point detection algorithm is used to quantify the intensity and frequency of water quality mutations. Change point detection is implemented based on the PELT algorithm, which identifies structural change points in the time series by minimizing the penalized likelihood function. A water quality mutation segment identification model based on Bayesian reasoning is established, and the posterior probability calculation formula is: Taking into account factors such as mutation intensity, duration and spatial range, a standardized water quality mutation segment identification index is output.

[0066] The water quality slow change segment identification index is obtained based on time series analysis of long-term water quality monitoring data and remote sensing image data. The sliding window technique is used to calculate the local trend slope and change characteristics of water quality parameters. The sliding window size is set to 12 months, the step size is set to 1 month, and the least squares method is used to fit the linear trend of the water quality time series within each window. The trend slope calculation formula is: Time series smoothing was used to identify the intervals of slow changes in water quality. The Hodrick-Prescott filter method was used to separate trend components and periodic components, with the smoothing parameter λ set to 1600. Wavelet transform analysis was combined to extract the low-frequency signal characteristics of water quality changes. Morlet wavelet was used as the mother wavelet function, and the wavelet coefficient calculation formula was: Where a is the scale parameter and b is the translation parameter. Linear and nonparametric regression methods were used to quantify the extent and duration of water quality slow changes. The nonparametric regression was implemented using a locally weighted regression method. A comprehensive identification model for water quality slow change segments was established, taking into account the significance of the change trend, duration, and spatial continuity, and outputting a standardized water quality slow change segment identification index ranging from 0 to 1.

[0067] It should be noted that the present invention uses the Hydrolight water body radiation transmission model to establish an inherent optical property model of river water bodies, which has significant physical mechanism advantages compared to traditional empirical statistical methods. Traditional remote sensing water quality inversion methods mainly rely on simple band ratios or empirical regression relationships, lack a deep understanding of the optical process of water bodies, and are prone to model failure when the optical conditions of water bodies change. The Hydrolight radiation transmission model starts from the physical nature of light propagation in water bodies, and accurately describes the scattering, absorption and propagation process of photons in water bodies by solving the radiation transmission equation, providing a solid theoretical basis for the interaction of optical components under different concentration conditions, so that the water quality parameter inversion results have stronger physical interpretability and environmental adaptability.

[0068] The multi-spectral feature enhancement model based on the visual Transformer architecture constructed in the present invention extracts spatial spectral correlation features between different bands through the self-attention mechanism, which has significant feature extraction advantages compared to traditional linear regression or simple neural network methods. Traditional methods usually assume that each spectral band is independent of each other, ignoring the complex nonlinear correlation between bands, resulting in the inability to fully mine the deep spectral information contained in remote sensing data. The self-attention mechanism of the visual Transformer can adaptively learn the importance weights and relationships between different bands, and process different types of spectral spatial relationships in parallel through a multi-head attention mechanism, effectively capturing weak optical signal features, and significantly improving the recognition accuracy of water components in complex optical environments and the continuity of spatial expression.

[0069] The present invention achieves effective decoupling of the optical coupling effect of river water bodies by constructing a multi-band combination index system including the normalized suspended matter index, the chlorophyll fluorescence peak index and the colored dissolved organic matter absorption index. Compared with the traditional single-band or simple band combination method, it has significant component separation advantages. When faced with complex water bodies where multiple optically active components coexist, traditional methods often encounter problems of component identification confusion and reduced inversion accuracy due to the lack of effective optical signal separation means. The present invention, by targeting the characteristic band response rules of different optical components, designs a band combination index with component selectivity, which can effectively highlight the optical response signal of the target component and suppress the interference of other components, fundamentally solving the inversion error problem caused by the mixing of multi-component optical signals.

[0070] The above three key technical ideas form a complete technical chain from physical mechanism modeling to intelligent algorithm optimization to optical signal decoupling. Their synergistic effect has systematic and comprehensive advantages over existing technologies. The Hydrolight radiation transfer model provides a solid physical theoretical basis for the entire inversion process, ensuring the accuracy and reliability of optical property simulation; the multi-spectral feature enhancement model fully mines the deep spectral information in the remote sensing data, making up for the lack of calculation accuracy of the physical model in complex environments; the multi-band combination index realizes the effective decoupling of the optical coupling effect, providing high-quality feature input for the intelligent algorithm. The three complement each other and work together to form a complete technical system from data preprocessing, feature extraction, signal decoupling to model inversion. It not only solves the problem of low inversion accuracy of traditional methods in complex optical environments, but also realizes high-precision spatial continuous monitoring of river water quality parameters, providing a new technical solution for remote sensing monitoring of complex water environments.

[0071] Specifically, the principle of the present invention is as follows: The core principle of the present invention in solving the problem of mutual coupling of multiple optically active components lies in the establishment of a complete technical chain from optical physics mechanism to intelligent algorithm integration. First, based on the Hydrolight water body radiation transmission model, an intrinsic optical property model of river water is established, which accurately describes the propagation process of light in water from a physical level. The shortest path algorithm is used to optimize computational efficiency, providing a theoretical basis for the interaction of optical components under different concentration conditions. Secondly, a quasi-analytical algorithm and a generalized intrinsic optical property algorithm are used to invert the optical components of remote sensing data, transforming the complex optical coupling problem into a decomposable linear superposition model. The visual Transformer architecture of the multispectral feature enhancement model is used to extract the deep spatial spectral correlation between different bands, and the self-attention mechanism is used to identify weak optical signal characteristics. Thirdly, a multi-band combination index is constructed, including the normalized suspended matter index, the chlorophyll fluorescence peak index, and the colored dissolved organic matter absorption index. The band combination highlights the optical response of the target component and realizes the effective decoupling of the optical coupling effect. Finally, the deep residual network-based water quality parameter inversion model utilizes large-scale pre-training data to learn universal optical feature representations. Fine-tuning training with target river data achieves accurate inversion capabilities for the rivers and other water bodies under test. Incorporating a water quality feature point selection function ensures the representativeness and spatial coverage of the training samples. This technical solution follows a logical development path from physical mechanism modeling to intelligent algorithm optimization. Through multi-level technology integration, it significantly improves the accuracy of water quality parameter inversion in complex optical environments.

[0072] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0073] The specific implementation of step S01 is to synchronously obtain multi-source remote sensing image data of the target river area through a satellite remote sensing platform and an aerial remote sensing platform, perform radiometric calibration processing on the original digital quantization value, and convert it into the reflectance of the top atmosphere, which is specifically expressed as follows:

[0074]

[0075] Where, ρ TOA is the reflectivity of the top atmosphere; L λ is the radiance value received by the sensor, in W·m -2 ·sr -1 μm -1 ; d is the Earth-Sun distance correction factor, ranging from 0.983 to 1.017; ESUN λ is the solar irradiance at the top of the atmosphere, in W·m -2 μm -1 θ sis the solar zenith angle, in radians; π is the circumference of a circle, which is 3.14159; and cos is the cosine function. λ Calculated by sensor calibration parameters, L λ =gain·DN+offset, DN is the original digital quantization value, gain is the gain coefficient, the unit is W·m -2 ·sr -1 μm -1 ·DN -1 , offset is the offset, the unit is W·m -2 ·sr -1 μm -1 The atmospheric correction process uses the 6S radiation transfer model, and the surface reflectivity is calculated as follows:

[0076]

[0077] Where, ρ surface is the surface reflectivity; ρ path is the atmospheric path reflectivity; T v is the upward transmittance of the atmosphere; T s is the downward transmittance of the atmosphere; S atm is the albedo of the large balloon surface. The atmospheric parameters are obtained by interpolation calculation based on the solar zenith angle, observation zenith angle, relative azimuth, atmospheric water vapor content, and aerosol optical depth using a lookup table method.

[0078] The specific implementation of step S02 is to establish a numerical calculation framework for the inherent optical characteristics of river water based on the Hydrolight water body radiation transfer model. The radiation transfer equation is expressed as follows:

[0079]

[0080] Where L(θ, φ, z) is the radiance in the direction (θ, φ) at depth z, in W·m -2 ·sr -1 nm -1 ; θ is the zenith angle, in radians; φ is the azimuth angle, in radians; z is the depth of the water body, in meters; λ is the wavelength of light, in nanometers; θ′ and φ′ are the zenith angle and azimuth angle of the incident light direction, in radians; c(λ, z) is the total attenuation coefficient, c(λ, z) = a(λ, z) + b(λ, z); a(λ, z) is the total absorption coefficient; b(λ, z) is the total scattering coefficient;

[0081] β(θ′, φ′, θ, φ, λ, z) is the volume scattering function; dΩ′ is the solid angle element, in sr; ∫ is the integral symbol; 4π is the full solid angle integral area, with a value of 12.566. The shortest path algorithm is optimized using the Dijkstra algorithm, and the path weight calculation formula is:

[0082] w ij =α·d ij +β·σ ij ;

[0083] Where w ij is the path weight from node i to node j; i and j are the node numbers of the scattering event; d ij is the propagation distance, in meters; σ ij is the scattering angle in radians; α and β are weight coefficients, with default values ​​of 1.0 and 0.5 respectively.

[0084] The specific implementation of step S03 is to use the quasi-analytical algorithm and the generalized inherent optical property algorithm to invert the optical components of the water body. The remote sensing reflectance model of the quasi-analytical algorithm is expressed as follows:

[0085]

[0086] Where R rs (λ) is the remote sensing reflectivity, unit is sr -1 ; f is the geometric factor, the default value is 0.33; Q is the ratio of uplink irradiance to uplink radiance, the default value is π; b b (λ) is the backscattering coefficient, in m -1 ; a(λ) is the total absorption coefficient, unit is m -1 The total absorption coefficient in the generalized intrinsic optical properties algorithm is decomposed into:

[0087] a(λ)=a w (λ)+a ph (λ)+a NAP (λ)+a CDOM (λ);

[0088] Where a w (λ) is the absorption coefficient of pure water; a ph (λ) is the phytoplankton absorption coefficient; a NAP (λ) is the absorption coefficient of non-algae particles; a CDOM (λ) is the absorption coefficient of colored dissolved organic matter. The absorption coefficient of each component is calculated by spectral decomposition algorithm. is the chlorophyll concentration, is the chlorophyll specific absorption coefficient, in m 2 mg -1 .

[0089] The specific implementation of step S04 is to construct a multi-band combination index to decouple the optical coupling effect of the river water body. The normalized suspended solids index calculation formula is as follows:

[0090]

[0091] Where, NTSSI is the normalized suspended solids index; R rs (865) is the remote sensing reflectance in the 865nm band; R rs (665) is the remote sensing reflectance in the 665nm band. The chlorophyll fluorescence peak index calculation formula is:

[0092]

[0093] Where FLH is the chlorophyll fluorescence peak index; R rs (685), R rs (680), R rs (690) are the remote sensing reflectances of the 685nm, 680nm, and 690nm bands, respectively. The numbers in brackets represent the wavelengths in nm. The calculation formula for the absorption index of colored dissolved organic matter is:

[0094]

[0095] Where, CDOM index is the absorption index of colored dissolved organic matter; R rs (555) and R rs (443) are the remote sensing reflectances in the 555nm and 443nm bands, respectively.

[0096] The specific implementation of step S05 is to establish a fine-tuning data set and perform model fine-tuning. The water quality feature point selection function performs spatial smoothing based on the Gaussian kernel function. The selection function is expressed as follows:

[0097]

[0098] Where W(x, y) is the weight of the water quality feature point at position (x, y); x and y are spatial coordinates in meters; i is the parameter index number, ranging from 1 to 6; w i is the weight coefficient of the i-th parameter; P i (x, y) is the value of the i-th evaluation parameter at position (x, y); G(x, y, σ) is the Gaussian kernel function; ∑ is the summation symbol. σ is the standard deviation of the Gaussian kernel, with a default value of 50m, π is the circumference of a circle, with a value of 3.14159, and exp is the natural exponential function. The formula for calculating the optical complexity of river water is:

[0099]

[0100] Where OC is the optical complexity index; R i is the reflectivity of the i-th band; is the mean reflectivity of all bands; N is the number of bands; ΔR j is the difference between the reflectance of the jth and j+1th adjacent bands; M is the number of adjacent band pairs; i is the band index number, ranging from 1 to N; j is the adjacent band pair index number, ranging from 1 to M-1; The coefficient of variation of hydrological flow is calculated using the Mann-Kendall trend test, and the statistic is expressed as:

[0101]

[0102] Where S MK is the Mann-Kendall statistic; n flow is the length of the time series; x i and x j is the flow data point at the i-th and j-th moments in the time series, in m 3 / s; sgn is the sign function, when x j >x i When x j =x i When x j <x i When -1 is used.

[0103] The specific implementation of step S06 is to apply the fine-tuned water quality parameter inversion model to the target river area for spatial continuous monitoring, and use the Kriging interpolation algorithm to perform spatial smoothing on the discrete prediction results. The interpolation formula is expressed as follows:

[0104]

[0105] Where, is the interpolation estimate of position x0; x0 is the coordinate of the position to be interpolated; Z(x i ) is a known sampling point x i Observed value of x i is the position coordinate of the i-th sampling point; i is the weight coefficient of the i-th sampling point; n sample is the number of sampling points involved in interpolation. The weight coefficients are obtained by solving the Kriging equations:

[0106]

[0107] In the formula, γ(x i , x j ) is the position xi and x j The semivariogram value between γ(x i , x0) is the position x i and the semivariogram value between the position to be interpolated x0; μ is the Lagrange multiplier. The semivariogram is represented by a spherical model:

[0108] When h≤a range When γ(h)=C0+C1, ​​when h>a range hour;

[0109] Where γ(h) is the semivariogram value; h is the spatial distance; C0 is the nugget effect, the default value is 0.1; C1 is the partial sill value, the default value is 0.8; a range The variable range is set to 150m. The final output is the suspended matter concentration distribution map, chlorophyll concentration distribution map and transparency distribution map, realizing the spatial continuous monitoring results of river water quality.

[0110] In order to better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below: Researchers selected a river section as the target research area. The river section is about 180 kilometers long and is a typical plain river. It is affected by multiple factors such as agricultural non-point source pollution, urban runoff and industrial emissions, and the water quality is complex and changeable. Researchers used the river water quality monitoring method based on multi-source remote sensing data described in the present invention to conduct a systematic water quality monitoring test. This embodiment involves the structure of the multi-spectral feature enhancement model and the water quality parameter inversion basic model as shown in the figure. Figure 2 and Figure 3 shown.

[0111] During the implementation of step S01, researchers used the Landsat-8 satellite to obtain multispectral remote sensing image data of the river section. The image acquisition time was 10:30 am on June 15, 2024, and the solar zenith angle θ s The atmospheric water vapor content is 2.8g / cm 2 , the aerosol optical thickness is 0.35. Through radiometric calibration, the original digital quantization value is converted into the reflectance of the top atmosphere, where ρ in the red light band (Band 4, 665nm) is TOA The value range is 0.085 to 0.142, and the ρ in the near-infrared band (Band 5, 865nm) TOA The value range is 0.156 to 0.238. The atmospheric correction process uses the 6S model to calculate the surface reflectance ρ surfaceSurface reflectance in the red band ranged from 0.021 to 0.067, and in the near-infrared band from 0.089 to 0.183. Researchers also used a hyperspectral aerial remote sensing platform to acquire high-precision imagery of key monitored river sections. The flight altitude was set at 1500 meters, with a spectral resolution of 5 nm and a spatial resolution of 1 meter. The data covered 200 continuous bands in the spectral range from 400 nm to 900 nm. The remote sensing data were categorized based on water optical properties. Clear water river data accounted for 32% of the total dataset, primarily located in the upper mountainous reaches, with total suspended solids concentrations ranging from 8 to 22 mg / L and chlorophyll concentrations of 3 to 8 μg / L. Turbid water data accounted for 42%, concentrated in the mid- and lower industrial regions, with total suspended solids concentrations ranging from 105 to 285 mg / L. Eutrophic water data accounted for 26%, located in areas affected by agricultural non-point source pollution, with chlorophyll concentrations ranging from 35 to 78 μg / L.

[0112] During the implementation of step S02, researchers established a framework for calculating the optical properties of water bodies in the Jingjiang River section based on the Hydrolight radiation transfer model. The input parameters included suspended solids concentration ranging from 15 to 180 mg / L, chlorophyll concentration ranging from 2 to 65 μg / L, and colored dissolved organic matter concentration ranging from 0.8 to 3.2 / m -1 The shortest path algorithm is used to optimize the radiation transfer calculation, the number of scattering event nodes is set to 512, and the path weight w ij The weight coefficients α and β were set to 1.0 and 0.5, respectively. The simulations obtained water reflectance spectral data under different concentration conditions. The calculation depth range was set to 0.2m to 8.0m, and the solar zenith angle ranged from 15 degrees to 75 degrees. The computational efficiency was improved by 68% compared to the traditional Monte Carlo method.

[0113] In the implementation of step S03, the researchers used a quasi-analytical algorithm and a generalized intrinsic optical property algorithm to invert the optical components of the water body. Through iterative optimization, the convergence threshold was set to 0.001, the maximum number of iterations was 50, and the actual average number of iterations was 28. The results of the multi-spectral feature enhancement model processing showed that the suspended matter absorption coefficient ranged from 0.025 to 0.186 / m -1 The chlorophyll absorption coefficient ranges from 0.008 to 0.095 / m -1 The absorption coefficient of colored dissolved organic matter ranges from 0.012 to 0.078 / m -1 Compared with traditional algorithms, the accuracy of optical component recognition is improved by 12%.

[0114] During the implementation of step S04, the researchers constructed a multi-band combined index system to decouple the optical coupling effect. The calculated normalized suspended solids index NTSSI ranged from -0.142 to 0.385, the chlorophyll fluorescence peak index FLH ranged from -0.008 to 0.034, and the colored dissolved organic matter absorption index CDOM index The range is 1.28 to 3.96. The results of the multi-band combined index calculation shown in Table 1 show that there are significant differences in the optical properties of different river sections.

[0115] Table 1 Calculation results of multi-band combination index in the Jingjiang River section

[0116]

[0117]

[0118] In the implementation of step S05, the researchers established a fine-tuning dataset containing 5,000 samples. The calculation results of the water quality feature point selection function showed that the optical complexity OC value of the river water body ranged from 0.24 to 0.78, the hydrological flow variation coefficient ranged from 0.32 to 1.28, the pollution source distribution density index ranged from 15 to 87, and the historical water quality variation score ranged from 23 to 76. The Mann-Kendall statistic S MK The calculated result is 148, indicating that the flow rate of this river section is showing a significant upward trend. The statistical results of the water quality characteristic point evaluation parameters shown in Table 2 reflect the changing characteristics of water quality in different river sections.

[0119] Table 2 Statistical results of evaluation parameters of water quality characteristic points in the Jingjiang River section

[0120] Evaluation parameters Minimum Maximum average value Standard deviation Optical Complexity OC 0.24 0.78 0.52 0.15 Flow rate variation coefficient 0.32 1.28 0.73 0.24 Pollution source density index 15 87 48 18 Water quality variability score 23 76 51 14 Mutation segment recognition index 0.18 0.84 0.46 0.19 Slow-change segment identification index 0.12 0.73 0.39 0.16

[0121] During fine-tuning, the researchers froze the first 40 layers of the basic model for water quality parameter inversion and performed 100 rounds of fine-tuning training using a learning rate of 0.0001 and a batch size of 32 samples. The validation dataset loss decreased from an initial value of 1.245 to a final value of 0.186 during training, demonstrating stable model convergence.

[0122] During the implementation of step S06, the researchers applied the fine-tuned water quality parameter inversion model to the entire Jingjiang River section for spatial continuous monitoring. The spatial grid was set to 30m×30m, generating a total of 45,680 grid cells. In the kriging interpolation process, the semivariogram spherical model parameters were set to nugget effect C0=0.1, partial sill value C1=0.8, and range a range=150m. The final output of the spatial distribution of water quality parameters shows that suspended matter concentration ranges from 5 to 298 mg / L, chlorophyll concentration ranges from 1.2 to 82.5 μg / L, and transparency ranges from 0.15 to 4.8 m. The water quality monitoring results shown in Table 3 indicate significant spatial differences in water quality across different river sections.

[0123] Table 3 Statistics of water quality monitoring results in the Jingjiang River section

[0124]

[0125] To verify the technical effectiveness of this invention, researchers compared the monitoring results with field sampling and measurement data from the same period. Thirty representative monitoring sections were selected, and actual water quality parameter values ​​were determined using standard water quality analysis methods. The comparison results showed that the accuracy of suspended matter concentration inversion reached 87.3%, chlorophyll concentration inversion accuracy reached 83.6%, and transparency inversion accuracy reached 85.2%. The spatial monitoring coverage rate reached 100%, with a temporal resolution of 16 days, achieving spatially continuous monitoring compared to traditional point-based monitoring methods.

[0126] Traditional river water quality monitoring relies primarily on manual sampling and laboratory analysis. Fixed monitoring sections are set up along the river, and water samples are regularly collected for chemical analysis to determine water quality parameters. This method has technical limitations, such as limited spatial coverage, low temporal resolution, and high costs. In a comparative test of the Jingjiang River section, traditional monitoring methods were only able to conduct monthly monitoring at 12 fixed sections, with a spatial coverage rate of only 2.8% of the monitored area. The cost of a single monitoring session was approximately 580 yuan per section, and real-time dynamic monitoring was not possible. The technological advancements brought about by this invention over traditional methods are mainly reflected in the following aspects. First, there is a significant improvement in spatial monitoring capabilities. This invention achieves continuous monitoring with 100% spatial coverage, while traditional methods, limited to point-based monitoring due to the layout of sections, have increased spatial information acquisition capabilities by 16 times. Second, there is an improvement in temporal resolution. This invention can achieve high-frequency monitoring every 16 days, while traditional methods are typically monthly or quarterly monitoring, and the temporal resolution is improved by 1.9 times. In terms of monitoring accuracy, the suspended matter concentration monitoring accuracy of the present invention reaches 87.3%, while the accuracy of the traditional laboratory analysis method is about 92.1%, and the accuracy difference is only 5.2%, which is within an acceptable range. The accuracy of the present invention in chlorophyll concentration monitoring is 83.6%, while the accuracy of the traditional method is 88.9%, and the accuracy difference is 6.0%. The transparency monitoring accuracy of the present invention reaches 85.2%, while the accuracy of the traditional Secchi disk measurement method is 91.4%, and the accuracy difference is 6.8%. In terms of data acquisition efficiency, the present invention can obtain water quality information for the entire river section during a single satellite pass, while the traditional method requires 5 working days to complete the sampling and analysis of the entire river section, which is 18 times more efficient. Through multi-source remote sensing data fusion and deep learning model optimization, the present invention significantly improves the spatial coverage capability, temporal resolution and data acquisition efficiency of river water quality monitoring while maintaining high monitoring accuracy, providing more comprehensive and timely technical support for river water environment management.

[0127] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A river water quality monitoring method based on multi-source remote sensing data, characterized in that: include: Acquire multispectral satellite remote sensing image data and hyperspectral aerial remote sensing image data of the target river area, perform radiation calibration and atmospheric correction preprocessing on the multispectral satellite remote sensing image data and hyperspectral aerial remote sensing image data to obtain preprocessed remote sensing image data; establish an inherent optical property model of river water based on the Hydrolight water body radiation transfer model, take the suspended matter concentration, chlorophyll concentration and colored dissolved organic matter concentration in the preprocessed remote sensing image data as input parameters, use the shortest path algorithm to optimize the radiation transfer calculation path, and simulate the water body reflectance spectral data under different concentration conditions; use the quasi-analysis algorithm and the generalized inherent optical property algorithm to The pre-processed remote sensing image data is used to invert the optical components of water bodies, and the pre-processed remote sensing image data is input into the multispectral feature enhancement model to output the optimized optical component identification results; a multi-band combination index is constructed to output the decoupled optical coupling effect parameters of river water bodies; the water quality feature points are determined using the water quality feature point selection function, a fine-tuning data set is established, and the model is fine-tuned using the pre-trained water quality parameter inversion basic model to establish a fine-tuned water quality parameter inversion model; the fine-tuned water quality parameter inversion model is applied to the pre-processed remote sensing image data to output the suspended matter concentration distribution map, chlorophyll concentration distribution map and transparency distribution map to achieve spatial continuous monitoring results of river water quality.

2. The river water quality monitoring method based on multi-source remote sensing data according to claim 1 is characterized in that: Multispectral satellite remote sensing image data includes three categories: clear water river data, turbid river data and eutrophic river data. Clear water river data refers to the remote sensing data of river water bodies with a total suspended matter concentration of less than 25 mg / L and a chlorophyll concentration of less than 10 μg / L. Turbid river data refers to the remote sensing data of river water bodies with a total suspended matter concentration of more than 100 mg / L. Eutrophic river data refers to the remote sensing data of river water bodies with a chlorophyll concentration of more than 30 μg / L and a high nutrient salt concentration.

3. The river water quality monitoring method based on multi-source remote sensing data according to claim 2 is characterized in that: The shortest path algorithm applies the Dijkstra algorithm to optimize the selection of photon propagation paths in water radiation transfer calculations. It regards scattering events in water as nodes in graph theory, and uses scattering angles and propagation distances as edge weights. By finding the shortest weighted path from the light source to the receiver, it accelerates the numerical solution process of the radiation transfer equation, significantly improving the computational efficiency of simulating the optical properties of complex water bodies.

4. The river water quality monitoring method based on multi-source remote sensing data according to claim 3 is characterized in that: The multispectral feature enhancement model is a neural network model based on the visual Transformer architecture. It divides the multispectral remote sensing image into a patch sequence with an image patch size of 16×16 pixels for processing. The spatial-spectral correlation features between different bands are extracted through the self-attention mechanism. The attention mask parameters within the model are used to control the fusion weights of features from different bands.

5. The river water quality monitoring method based on multi-source remote sensing data according to claim 4 is characterized in that: The multi-band combined index includes the normalized suspended matter index, the chlorophyll fluorescence peak index and the colored dissolved organic matter absorption index. The suspended matter absorption coefficient, the chlorophyll absorption coefficient and the colored dissolved organic matter absorption coefficient are used as input parameters, and the decoupled river water optical coupling effect parameters are output.

6. The river water quality monitoring method based on multi-source remote sensing data according to claim 5 is characterized in that: The water quality feature point selection function is used to select representative water quality monitoring points from the river area and smooth the spatial distribution based on the Gaussian kernel function. The input includes six parameters: the optical complexity of the river water body, the hydrological flow variation coefficient, the distribution density of pollution sources, the degree of historical water quality variation, the water quality mutation segment identification index, and the water quality slow change segment identification index. The output is the location coordinates of the preferred water quality feature points and the corresponding weight coefficients.

7. The river water quality monitoring method based on multi-source remote sensing data according to claim 6 is characterized in that: The method for obtaining the optical complexity of river water bodies includes calculating the spectral variance and spectral gradient of each pixel based on preprocessed remote sensing image data, standardizing the reflectance data of the red light band, green light band, blue light band and near-infrared band, calculating the root mean square of the reflectance difference between adjacent bands, and combining the comprehensive effects of water turbidity, chlorophyll concentration and colored dissolved organic matter concentration to establish an optical complexity evaluation model through multivariate regression analysis.

8. The river water quality monitoring method based on multi-source remote sensing data according to claim 7 is characterized in that: The method for obtaining the hydrological flow variation coefficient includes collecting daily flow monitoring data of the target river area in the past five years, calculating the monthly average flow, quarterly average flow and annual average flow, analyzing the coefficient of variation and peak frequency of the flow time series, using the Mann-Kendall trend test method to identify the flow change trend, and combining rainfall data and upstream reservoir scheduling information to establish a hydrological flow change assessment model.

9. The river water quality monitoring method based on multi-source remote sensing data according to claim 8, characterized in that: The method for obtaining the distribution density of pollution sources includes collecting spatial location data of industrial enterprises, urban sewage treatment plants, agricultural non-point source pollution and domestic pollution sources in the river basin based on the geographic information system platform, allocating weights according to the type of pollution source and emission intensity, and using the kernel density estimation method to calculate the pollution load intensity per unit area. In combination with the river hydrological connectivity and pollutant diffusion model, a pollution source distribution density evaluation system with river sections as units is established.

10. The river water quality monitoring method based on multi-source remote sensing data according to claim 9, characterized in that: The basic model for water quality parameter inversion is a pre-trained neural network model based on the deep residual network ResNet architecture. The model structure consists of 50 convolutional layers and 4 residual blocks. Each residual block contains multiple convolutional layers and jump connections. The input layer receives the feature vector composed of multi-band combination index and optical component parameters, which is mapped into 128-dimensional hidden features through the fully connected layer. The output layer contains 3 branches to predict suspended matter concentration, chlorophyll concentration and transparency value respectively.

Images