A remote sensing water quality inversion method combining differential learning rate and spectral geometric characteristics

Abstract

The present invention discloses a remote sensing water quality inversion method combining a differential learning rate with spectral geometric characteristics, comprising: collecting satellite images of various sites and obtaining remote sensing reflectance of each site; deriving surface water monitoring site information and water quality index information data from a surface water database; eliminating remote sensing reflectance of sites with obvious abnormalities and constructing a remote sensing reflectance curve set; eliminating abnormal values ​​of water quality indicators; calculating spectral geometric feature data of the remote sensing reflectance curves of each site, merging the feature data into a feature matrix, and dividing the feature matrix into a training set and a test set; merging the water quality indicators after eliminating abnormal water quality index values ​​as a data set to be fitted into an output set, and dividing the output set into a training output set and a test output set; constructing a machine learning model, inputting the training set into the model for training to obtain a trained model; putting the test set into the trained model for testing, and after evaluation, deploying the optimal model online.

Description

Technical Field

[0001] The present invention relates to the technical field of water environment remote sensing monitoring, and in particular to a remote sensing water quality inversion method combining differential learning rate and spectral geometric characteristics. Background Art

[0002] With the rapid development of industry and agriculture and the accelerating process of urbanization, my country's total water consumption has increased significantly, and wastewater discharge has also increased significantly, which has brought a heavy burden to the self-purification process of surface water. In particular, the water quality of urban rivers has deteriorated rapidly, causing serious impacts on social and economic development, urban environmental landscape and human health.

[0003] In recent years, the monitoring and management of inland water bodies has received increasing attention. Monitoring lakes and rivers not only helps us better understand the impacts of environmental change on freshwater ecosystems but also provides rich information for environmental forecasting. my country faces severe freshwater resource shortages, and water pollution has reduced water quality, further exacerbating water scarcity. However, only a small number of water bodies undergo regular and consistent monitoring. Therefore, increasing the scope and frequency of inland water monitoring is an urgent need. Currently, most inland water monitoring is still based on field observations. Conventional water monitoring requires the deployment of numerous monitoring points within a water body, with laboratory analysis of collected water samples to provide information on the spatiotemporal distribution of water quality. While field measurements can provide more detailed (species-level) information, they are time-consuming and labor-intensive, and are limited by weather and hydrological conditions, making long-term continuous observation difficult. Furthermore, unevenly distributed sampling points can lead to significant observation errors. Remote sensing technology, with its real-time, wide-range, and periodic observation capabilities, offers advantages unmatched by traditional monitoring methods. It meets the current needs of inland water monitoring and holds great potential and value.

[0004] Numerous studies have validated the feasibility of artificial neural networks, RBF neural networks, support vector machines, and random forests in water quality inversion. These methods have demonstrated certain advantages and significantly improved inversion accuracy. However, due to limited sample size and limited sampling data, the constructed neural network models are simple in structure and prone to overfitting, resulting in high variance and weak generalization. Leveraging existing data, exploring data features, and improving network learning strategies are crucial for overcoming obstacles in the application of machine learning methods in water quality inversion and are crucial for achieving widespread application of remote sensing in water quality monitoring. Summary of the Invention

[0005] In view of the above problems existing in the prior art, the present invention provides a remote sensing water quality inversion method combining differential learning rate and spectral geometric characteristics, which includes the following steps:

[0006] S1. Collect satellite images of each site, perform radiometric calibration on the collected satellite images, calculate the radiometric brightness, calculate the apparent reflectance from the radiometric brightness, and then calculate the remote sensing reflectance from the apparent reflectance; derive surface water monitoring site information and water quality index information data from the surface water database;

[0007] S2. Eliminate obviously abnormal remote sensing reflectance of sites and construct a set of remote sensing reflectance curves; eliminate abnormal values ​​of water quality indicators;

[0008] S3. Calculate the spectral geometric feature data of each station through the remote sensing reflectance curve, merge the spectral geometric feature data into an m*n feature matrix with each column as a feature and each row as a sample, and divide the feature matrix into a training set and a test set; merge the water quality indicators after removing the outliers as the data set to be fitted into the output set, and divide the output set into a training output set and a test output set;

[0009] S4. Build a machine learning model, input the training set into the model for training, and obtain a trained model;

[0010] S5. Put the test set into the trained model for testing. The results are evaluated using mean relative error, root mean square error, and coefficient of determination. The optimal model is deployed online.

[0011] As a further improvement of the present invention, step S1 includes sub-steps including: data collection, data preliminary selection, data cleaning, data matching and data integration.

[0012] The sub-steps of surface water data preprocessing in step S1 include: surface water data export, surface water data preliminary selection, surface water data cleaning and surface water data integration.

[0013] The sub-steps of satellite data preprocessing in step S1 include: radiometric calibration, atmospheric correction, geometric correction, georeferencing, image fusion, image mosaicking, and water area extraction.

[0014] The sub-steps of spatial matching and information extraction in step S1 include: spatial vectorization, coordinate matching, band value extraction, information organization and archiving.

[0015] As a further improvement of the present invention, the outlier removal in step S2 mainly includes: using the Dixon test method to screen and remove outliers of water quality indicators, using the spectral matching method to remove obviously abnormal site remote sensing reflectance, and constructing a remote sensing reflectance curve set based on the remote sensing reflectance of the site after removing the outliers.

[0016] The formula for calculating the spectral distance using the spectral matching method in step S2 is:

[0017]

[0018] Where D i 2 is the spectral distance, Rrs lut (λ i ) is the standard spectrum curve data, Rrs pixel (λ i ) is the remote sensing reflectance obtained at the sample point, λ i The value ranges from 400 to 900 nm, where i is the number of satellite bands.

[0019] As a further improvement of the present invention, the formula for calculating the minimum spectral distance by the spectral matching method in step S2 is:

[0020]

[0021] Where, is the average spectral distance, and n is the number of sites.

[0022] As a further improvement of the present invention, the spectral geometric characteristic data of the remote sensing reflectance curve of each site in step S3 includes: spectral curve area, spectral curve angle, spectral curve direction, spectral curve slope ratio and spectral curve projection length;

[0023] The calculation formula for the spectral curve area is: f area =∫f(λ i )ΔλdΔλ, where Δλ is the wavelength interval, f(λ i ) is the wavelength λ i Remote sensing reflectivity at i is the wavelength corresponding to the i satellite band number;

[0024] The calculation formula of the spectral curve angle is: Where i, j, k = {1, 2, 3, 4} and i≠j≠k; i, j, k are the number of satellite bands; λ i ,λ j and λ k is the wavelength corresponding to different satellite band numbers; Rrs i 、Rrs j and Rrs k is the remote sensing reflectivity corresponding to different satellite band numbers;

[0025] The calculation formula for the spectral curve direction is: Where i, j = {1, 2, 3, 4} and i ≠ j;

[0026] The calculation formula for the slope ratio of the spectral curve is: Where i, j, k = {1, 2, 3, 4} and i ≠ j ≠ k, where Δλ is the wavelength interval, Δλ i,j is the wavelength interval corresponding to the number of satellite bands i and j, Δλk,i is the wavelength interval corresponding to the number of satellite bands i and k;

[0027] The calculation formula of the spectral curve projection length is: f prj =Rrs i -Rrs j , where i, j = {1, 2, 3, 4} and i ≠ j.

[0028] As a further improvement of the present invention, the machine learning model constructed in step S4 is mainly configured with different learning rates according to the number of hidden layers, and the differential learning rate of each layer is calculated according to the following formula:

[0029] η i =η0 / (1-i×α)

[0030] Where i is the number of hidden layers, η i is the learning rate of the i-th layer, η0 is the learning rate of the first layer, and α is the learning rate change rate.

[0031] As a further improvement of the present invention, the calculation formulas for the model evaluation index in step S5 are as follows:

[0032] Average relative error:

[0033] Root mean square error:

[0034] Coefficient of determination:

[0035]

[0036]

[0037]

[0038] Where, is the predicted value, is the mean of the observed values, y i is the observed value, and n is the number of test sets. Based on the three evaluation indicators, the model with the best inversion effect is selected and applied to water quality remote sensing inversion and monitoring.

[0039] Compared with existing technologies,

[0040] 1. Traditional water environment monitoring methods mainly rely on manual sampling plus laboratory sample analysis and ground-based automatic monitoring stations. Manual sampling has low efficiency, poor timeliness, limited monitoring range and certain personal safety issues. Ground-based automatic monitoring stations have few monitoring indicators, are expensive and require long-term continuous maintenance. The present invention takes advantage of satellite data and considers the advantages of a large amount of multispectral satellite remote sensing data, such as wide monitoring range, high spatial resolution and extensive information. In view of the small area, narrow width and scattered and wide distribution of inland water bodies, a water quality inversion model method based on satellite data is constructed. This can effectively improve the monitoring efficiency of inland water environment, improve timeliness, expand the monitoring range, and reduce monitoring costs.

[0041] 2. This invention fully mines the information hidden behind the data from a geometric perspective, proposes the spectral geometric characteristics of the remote sensing reflectance curve, and through an improved differential learning rate optimization model, obtains results with higher accuracy, stronger robustness, and clearer network features.

[0042] 3. Based on reality, the present invention better utilizes the massive multispectral satellite remote sensing data, deeply mines the information contained in satellite remote sensing images, summarizes a complete set of water quality inversion model construction processes, and provides data guarantee for the construction of water quality parameter inversion models. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The present invention is a flow chart of a remote sensing water quality inversion method combining differential learning rate and spectral geometric characteristics. DETAILED DESCRIPTION

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0045] The present invention is described in further detail below with reference to the accompanying drawings:

[0046] like Figure 1 As shown, the present invention provides a remote sensing water quality inversion method combining differential learning rate and spectral geometric characteristics, comprising the following steps:

[0047] S1. Collect satellite images of each site, perform radiometric calibration on the collected satellite images, calculate the radiometric brightness, calculate the apparent reflectance through the radiometric brightness, and then calculate the remote sensing reflectance through the apparent reflectance; export the surface water monitoring site information and water quality index information data from the surface water database.

[0048] The process of preprocessing satellite data

[0049] Radiometric calibration is performed on the collected satellite images. Radiometric calibration is to establish a quantitative relationship between the sensor output value and the actual radiance corresponding to the sensor based on the pre-determined sensor response value. Specifically, the digital quantized output value DN recorded by the sensor is converted into the radiance at the sensor entrance pupil. The radiance at the sensor entrance pupil can be expressed as:

[0050] L λ =Gain×DN+Bias

[0051] Where, L λ Indicates the radiation brightness value, the unit is W / (cm 2 ·μm·sr); Gain and Bias represent gain and offset, which can be found in the parameter table. The unit is W / (cm 2 ·μm·sr).

[0052] During the imaging process, remote sensing images are affected by Rayleigh scattering, Mie scattering, refraction, and absorption, resulting in distortion of the radiation information received at the satellite payload entrance pupil. Therefore, atmospheric correction is required to eliminate the influence of these errors. This embodiment uses the 6S atmospheric radiation transmission model to eliminate the influence of the atmosphere. This model not only considers the influence of atmospheric scattering but also the influence of the solar altitude angle on the ground information during transmission, resulting in higher accuracy.

[0053] Calculate apparent reflectance in atmospheric correction:

[0054]

[0055] Where, ρ λ is the apparent reflectance; L λ is the radiation brightness value at the entrance pupil of the satellite payload channel, in W / (cm 2 ·μm·sr); d is the distance between the Earth and the Sun, celestial unit; ESUN λ is the solar irradiance; θ s is the solar zenith angle.

[0056] Calculate remote sensing reflectivity:

[0057]

[0058] Where: ρ s is the surface reflectivity to be determined, also known as remote sensing reflectivity; ρ λ is the apparent reflectivity; ρ0 is the equivalent emissivity of the atmospheric path radiation term; T(μ s ) is the total transmittance of downlink radiation, T(μ v) is the total transmittance of uplink radiation, μ s =cos(θ s ), μ v =cos(θ v ),θ s ,θ v , φ are the solar zenith angle, observation zenith angle and relative azimuth respectively; S is the hemispherical reflectivity of the lower boundary of the atmosphere.

[0059] The process of preprocessing surface water data

[0060] Pure water body satellite image data was extracted using the Normalized Difference Water Index (NDWI). Specifically, the Normalized Difference Water Index (NDWI) formula was entered into the Band Math tool, and the green band and near-infrared band as well as the corresponding thresholds were specified. The Band Math tool then processed the pure water body satellite image according to the NDWI formula, the specified bands, and the thresholds to obtain a water area boundary vector file. The water area vector boundary was then corrected through visual interpretation and human-computer interaction, and the pure water body satellite image data was then obtained through cropping.

[0061] The normalized water index is achieved through the following formula:

[0062]

[0063] Where Rrs green Indicates the reflectivity of the green band, Rrs nir Represents the reflectivity in the near-infrared band.

[0064] Surface water monitoring station information and water quality indicator information data such as dissolved oxygen, chemical oxygen demand, biochemical oxygen demand, permanganate index, ammonia nitrogen, total phosphorus, total nitrogen, turbidity, pH, transparency, suspended matter concentration, and chlorophyll a were exported from the surface water database. Then, spatial vectorization was performed based on the station coordinates in the station information and spatially matched with the remote sensing image data. Finally, the remote sensing reflectance of the corresponding station location was extracted, organized and archived, and uploaded to the database.

[0065] S2. Eliminate the remote sensing reflectance of sites with obvious abnormalities and construct a set of remote sensing reflectance curves; eliminate abnormal values ​​of water quality indicators.

[0066] The process of removing outliers from surface water data using the Dixon test

[0067] The specific steps are as follows:

[0068] 1. Arrange n surface water data from small to large as x1, x2, ..., x n-1 , x n , n∈[31,100].

[0069] 2. Calculate the high-end outlier value D according to the corresponding formula n and test low-end outlier D n '

[0070]

[0071] 3. Determine the significance level α by looking up the critical value table of the one-sided Dixon test in Table C.1 of GB / T 4883-2008 and finding the critical value D. 1-α (n).

[0072] 4. Check for high-end outliers. When D n >D 1-α (n), determine x (n) is an outlier, otherwise it is judged that no outlier is found; test the low-end outlier, when D n ′>D 1-α (n), determine x (1) is an outlier, otherwise it is judged that no outlier is found.

[0073] 5. For the detected outlier x (1) or x (n) , determine the elimination level α * , find the critical value in the lookup table Test for high-end outliers, when When x (n) is a statistical outlier, otherwise it is judged that no x is found (n) is a statistical outlier (i.e. x (n) is the divergence value); test the low-end outlier value, when When x (1) is a statistical outlier, otherwise it is judged that no x is found (1) is a statistical outlier (i.e. x (1) is the divergence value).

[0074] The process of removing obviously abnormal site remote sensing reflectance using spectral matching method

[0075] The specific steps are as follows:

[0076] 1. Calculate the spectral distance D i 2

[0077]

[0078] Among them Rrs lut (λ i ) is the standard spectrum curve data, Rrs pixel (λ i ) is the remote sensing reflectivity obtained at the site, λ iThe value ranges from 400 to 900 nm, where i is the number of satellite bands.

[0079] 2. Calculate the average spectral distance

[0080] n is the number of sites.

[0081] 3. Compare the spectral distance with the spectral average distance. The stations with spectral distance greater than 1.5 times the spectral average distance are outliers x 异常值 , Remove outliers.

[0082] According to the remote sensing reflectivity of the site after removing outliers, a set of remote sensing reflectivity curves of the site is constructed, in which the remote sensing reflectivity is the Y coordinate and the wavelength is the X coordinate.

[0083] S3. Calculate the spectral geometric feature data of each station through the remote sensing reflectance curve, merge the spectral geometric feature data into an m*n feature matrix with each column as a feature and each row as a sample, and divide the feature matrix into a training set and a test set; after removing the abnormal values ​​of water quality indicators, the water quality indicators such as dissolved oxygen, chemical oxygen demand, biochemical oxygen demand, permanganate index, ammonia nitrogen, total phosphorus, total nitrogen, turbidity, pH, transparency, suspended matter concentration, chlorophyll a, etc. are used as the data set to be fitted and merged into the output set, and the output set is divided into a training output set and a test output set.

[0084] The process of calculating the geometric characteristics of remote sensing reflectivity curves

[0085] The spectral geometric characteristic data of the extracted remote sensing reflectance curve of each station is calculated, including geometric characteristics such as spectral curve area, curve angle, curve direction, curve slope ratio, and projection length. The specific calculation formulas for each characteristic are as follows:

[0086] Spectral curve area:

[0087] f area =∫f(λ i )ΔλdΔλ, where Δλ is the wavelength interval, f(λ i ) is the wavelength λ i Remote sensing reflectivity at i is the wavelength corresponding to the i satellite band number.

[0088] Spectral curve angle:

[0089] Where i, j, k = {1, 2, 3, 4} and i≠j≠k; i, j, k are the number of satellite bands; λ i ,λ j and λ k is the wavelength corresponding to different satellite band numbers; Rrsi 、Rrs j and Rrs k is the remote sensing reflectivity corresponding to different satellite band numbers.

[0090] Spectral curve direction:

[0091] Where i, j = {1, 2, 3, 4} and i ≠ j.

[0092] Spectral curve slope ratio:

[0093] Where i, j, k = {1, 2, 3, 4} and i ≠ j ≠ k, where Δλ is the wavelength interval, Δλ i,j is the wavelength interval corresponding to the number of satellite bands i and j, Δλ k,i is the wavelength interval corresponding to the satellite band number i and k.

[0094] Spectral curve projection length:

[0095] f prj =Rrs i -Rrs j Where i, j = {1, 2, 3, 4} and i ≠ j.

[0096] The process of building a feature set

[0097] Merge the spectral geometric feature data to form a feature set: , the spectral geometric feature data is merged into an m*n feature matrix with each column as a feature and each row as a sample to improve the computational efficiency and speed up the model training, thus saving resources and time. Similarly, the water quality indicators such as dissolved oxygen, chemical oxygen demand, biochemical oxygen demand, permanganate index, ammonia nitrogen, total phosphorus, total nitrogen, turbidity, pH, transparency, suspended solids concentration, chlorophyll a, etc. are combined as the data sets to be fitted into the output set: y = [w1,w2,w3,...,w n ] T , where w1 is the dissolved oxygen data set, w2 is the chemical oxygen demand data set, w3 is the biochemical oxygen demand data set, w4 is the permanganate index data set, w5 is the ammonia nitrogen data set, w6 is the total phosphorus data set, w7 is the total nitrogen data set, w8 is the turbidity data set, w9 is the pH data set, and w 10 is the transparency data set, w 11 is the suspended matter concentration data set, w 12 is the chlorophyll a data set and w n Set up other data collections as needed.

[0098] Use singular value decomposition (SVD) to decompose each eigenvector into N subvectors, and select the largest r singular values ​​and the corresponding left and right singular vectors as the transformed eigenvectors, where the sum of the largest r singular values ​​should account for more than 90% of the sum of all singular values.

[0099] The feature matrix is ​​divided into training set and test set in the ratio of 8:2, and the output set is divided into training output set and test output set according to the same rule.

[0100] S4. Build a machine learning model, input the training set into the model for training, and obtain a trained model.

[0101] Feedforward Neural Network Model Design Using Differential Learning Rates

[0102] Use differential learning rate to build a machine learning model, input the training set into the model for training, and then verify the trained model with the test set.

[0103] The steps to build a machine learning model are as follows:

[0104] 1. Build a neural network with multiple layers of neurons;

[0105] 2. Set the hidden layer activation function to ReLU function and the output layer activation function to linear function;

[0106] 3. Set different learning rates according to the number of hidden layers. The learning rate of each layer is calculated according to the following formula: η i =η0 / (1-i×α), where i is the number of hidden layers, η i is the learning rate of the i-th layer, η0 is the learning rate of the first layer, and α is the learning rate change rate.

[0107] The training steps of the feedforward neural network model are as follows:

[0108] a. Initialize model parameters;

[0109] b. Input training sample data;

[0110] c. Calculate the predicted values ​​of water quality indicators;

[0111] d. Calculate the loss function;

[0112] e. Determine whether the error has reached the minimum. If not, update the model parameters and repeat steps b, c, and d until the error is minimized. Stop iteration, record the minimum error, and save the model.

[0113] The model loss function calculation formula is as follows:

[0114]

[0115] Where yi is the water quality index of the i-th sample point, is the predicted value of the water quality index of the i-th sample point, and L is the loss function of the model.

[0116] The calculation formula for model parameter update is as follows:

[0117]

[0118] in, is the gradient of neurons in layer i, η i is the learning rate of the i-th layer, ε i is the model parameter before updating, ε i+1 are the updated model parameters.

[0119] S5. Put the test set into the trained model for testing. The results are evaluated using mean relative error, root mean square error, and coefficient of determination. The optimal model is deployed online.

[0120] The process of testing the model

[0121] The test set is placed into the trained model for testing. The results are evaluated using mean relative error, root mean square error, and coefficient of determination, and the optimal model is deployed online.

[0122] The results of model inversion are a set of continuous real numbers, and the unit dimensions and absolute values ​​of different indicators vary greatly. Therefore, we should focus on the average relative error rather than the absolute error.

[0123] The root mean square error is used to measure the inversion accuracy of the model, while the coefficient of determination is used to measure the goodness of fit of the model.

[0124] Average relative error (ARE): The smaller the ARE value, the better the model quality and the more accurate the prediction.

[0125] Root Mean Square Error (RMSE): The smaller the RMSE, the better the model inversion accuracy.

[0126] Coefficient of determination (R 2 ):

[0127]

[0128]

[0129]

[0130] Where, is the predicted value, is the mean of the observed values, yi is the observed value, and n is the number of test sets. 2 The closer the value of R is to 1, the better the regression line fits the observed value; on the contrary, 2 The smaller the value of , the worse the regression line fits the observed values.

[0131] The best model is selected based on the three evaluation indicators.

[0132] in conclusion:

[0133] A specific embodiment provides a remote sensing water quality inversion method that combines differential learning rate and spectral geometric characteristics, which solves the problem of how to construct a high-precision water quality parameter inversion model method under small samples, better utilizes the value of high-resolution multispectral satellite imagery, and provides a solution to the problem of water environment pollution monitoring.

[0134] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A remote sensing water quality inversion method combining differential learning rate and spectral geometric characteristics, characterized in that: The method comprises the following steps: S1. Collect satellite images of each site, perform radiometric calibration on the collected satellite images, calculate the radiometric brightness, calculate the apparent reflectance from the radiometric brightness, and then calculate the remote sensing reflectance from the apparent reflectance; derive surface water monitoring site information and water quality index information data from the surface water database; S2. Eliminate obviously abnormal remote sensing reflectance of sites and construct a set of remote sensing reflectance curves; eliminate abnormal values ​​of water quality indicators; S3. Calculate the spectral geometric feature data of each station through the remote sensing reflectance curve, merge the spectral geometric feature data into an m*n feature matrix with each column as a feature and each row as a sample, and divide the feature matrix into a training set and a test set; merge the water quality indicators after removing the outliers as the data set to be fitted into the output set, and divide the output set into a training output set and a test output set; S4. Build a machine learning model, input the training set into the model for training, and obtain a trained model. Building a machine learning model includes: 1) Construct a neural network with multiple layers of neurons; 2) Set the hidden layer activation function to the ReLU function and the output layer activation function to the linear function; 3) Set different learning rates according to the number of hidden layers. The learning rate of each layer is calculated according to the following formula: η i =η0 / (1-i×α), where i is the number of hidden layers, η i is the learning rate of the i-th layer, η0 is the learning rate of the first layer, and α is the learning rate change rate; The training steps of the feedforward neural network model are as follows: a. Initialize model parameters; b. Input training sample data; c. Calculate the predicted values ​​of water quality indicators; d. Calculate the loss function; e. Determine whether the error is minimized. If not, update the model parameters and repeat steps b, c, and d until the error is minimized. Stop iteration, record the minimum error, and save the model. The model loss function calculation formula is as follows: Where y i is the water quality index of the i-th sample point, is the predicted value of the water quality index of the i-th sample point, and L is the loss function of the model; The calculation formula for model parameter update is as follows: in, is the gradient of neurons in layer i, η i is the learning rate of the i-th layer, ε i is the model parameter before updating, ε i+1 are the updated model parameters; S5. Put the test set into the trained model for testing. The results are evaluated using mean relative error, root mean square error, and coefficient of determination. The optimal model is deployed online.

2. The remote sensing water quality inversion method combining differential learning rate and spectral geometric characteristics according to claim 1 is characterized in that: In step S2, the Dixon test method is used to remove abnormal values ​​of water quality indicators, and the spectral matching method is used to remove obviously abnormal site remote sensing reflectance; The formula for calculating the spectral distance using the spectral matching method in step S2 is: Where D i 2 is the spectral distance, Rrs lut (λ i ) is the standard spectrum curve data, Rrs pixel (λ i ) is the remote sensing reflectance obtained at the sample point, λ i The value ranges from 400 to 900 nm, where i is the number of satellite bands; The formula for calculating the minimum spectral distance using the spectrum matching method in step S2 is: Where, is the average spectral distance, and n is the number of sites.

3. The remote sensing water quality inversion method combining differential learning rate and spectral geometric characteristics according to claim 1 is characterized in that: The spectral geometric characteristic data of the remote sensing reflectance curve in step S3 includes: spectral curve area, spectral curve angle, spectral curve direction, spectral curve slope ratio and spectral curve projection length; The calculation formula for the spectral curve area is: f area =∫f(λ i )ΔλdΔλ, where Δy is the wavelength interval, f(λ i ) is the wavelength λ i Remote sensing reflectivity at The calculation formula of the spectral curve angle is: Where i, j, k = {1, 2, 3, 4} and i ≠ j ≠ k; i, j, k are the number of satellite bands; λ i ,λ j and λ k is the wavelength corresponding to different satellite band numbers; Rrs i 、Rrs j and Rrs k is the remote sensing reflectivity corresponding to different satellite band numbers; The calculation formula for the spectral curve direction is: Where i, j = {1, 2, 3, 4} and i ≠ j; The calculation formula for the slope ratio of the spectral curve is: Where i, j, k = {1, 2, 3, 4} and i ≠ j ≠ k; where Δλ i,j is the wavelength interval corresponding to the number of satellite bands i and j, Δλ k,i is the wavelength interval corresponding to the number of satellite bands i and k; The calculation formula of the spectral curve projection length is: f prj =Rrs i -Rrs j , where i, j = {1, 2, 3, 4} and i ≠ j.

4. The remote sensing water quality inversion method combining differential learning rate and spectral geometric characteristics according to claim 1 is characterized in that: The calculation formulas for the average relative error, root mean square error and determination coefficient in step S5 are as follows: Average relative error: Root mean square error: Coefficient of determination: Where, is the predicted value, is the mean of the observed values, y i is the observed value, and n is the number of test sets.

Images