Coastal water iron concentration remote sensing monitoring method based on satellite data and neural network

Abstract

The invention discloses a near-shore water body iron concentration remote sensing monitoring method based on satellite data and a neural network, and relates to the technical field of environment monitoring, and the method comprises the following steps: S1, collecting a Sentinel-2 satellite image, and carrying out atmospheric correction to obtain water surface remote sensing reflectivity data; s2, extracting original input features according to the water surface remote sensing reflectivity data; s3, based on the original input features, training the back propagation neural network model by using the training data set; and S4, inputting water surface remote sensing reflectivity data into the trained back propagation neural network model to obtain an iron concentration value. The technical scheme of the invention is highly innovative, systematic and physical improvement is performed on the input end, the network structure, the neurons and the training target of the neural network for specific technical problems, and a complete and original technical scheme is formed.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology, specifically to a remote sensing monitoring method for iron concentration in nearshore waters based on satellite data and neural networks. Background Technology

[0002] Iron (Fe) is a key element in marine ecosystems, serving as an essential micronutrient for phytoplankton growth and a crucial tracer for measuring terrestrial inputs and environmental pressures from human activities. In semi-enclosed waters such as nearshore bays, iron concentrations exhibit extreme spatiotemporal variability due to factors such as river inputs, industrial wastewater discharge, and urban runoff. Excessive iron input can alter the aquatic environment and threaten the health of marine life. Therefore, long-term, dynamic monitoring of iron concentrations in nearshore waters is of great significance.

[0003] Currently, the main technical solutions for monitoring iron concentration in water bodies and their drawbacks are as follows:

[0004] 1. Traditional field sampling and laboratory analysis: This method involves collecting water samples from a vessel at a pre-designated station, followed by analysis using sophisticated instruments (such as ICP-MS) to obtain precise iron concentration values. This is the "gold standard" for obtaining iron concentration data, offering high accuracy. However, its drawbacks include the fact that "point-like" sampling cannot reflect the distribution of water parameters across the entire surface; furthermore, it is costly, time-consuming, and highly susceptible to weather and sea conditions, resulting in a low observation frequency and difficulty in capturing rapid dynamic changes in iron concentration caused by runoff, tides, and sewage discharge events.

[0005] 2. Remote Sensing Inversion Based on Traditional Empirical Models: To overcome the spatiotemporal limitations of on-site sampling, remote sensing technology has been introduced into water quality monitoring. This method utilizes satellite imagery to establish an empirical or semi-empirical relationship between remotely sensed reflectance and measured iron concentration, such as using specific single bands or band combinations to establish a multiple linear regression (MLR) model with iron concentration. However, the optical properties of nearshore waters are extremely complex, and the relationship between iron concentration and remotely sensed reflectance exhibits a highly nonlinear nature. Traditional models such as MLR assume a linear relationship between variables, making it difficult to accurately characterize this complex relationship, resulting in low inversion accuracy and poor model robustness.

[0006] 3. Remote Sensing Inversion Based on Conventional Neural Networks: In recent years, some studies have attempted to use conventional neural networks (such as BPNNs) to fit the aforementioned nonlinear relationships. However, these methods typically use band reflectance directly as input and employ standard network structures, activation functions, and loss functions for training. This generalized approach does not fully consider prior knowledge in the field of water color remote sensing, such as the synergistic and antagonistic effects between different bands, and the specific numerical ranges (such as high concentrations of pollution) required by the monitoring targets. Therefore, the model's generalization ability and early warning capability for critical pollution events still have room for improvement, and the technical solutions themselves lack targeted structural and algorithmic innovation, resulting in weak creativity. Summary of the Invention

[0007] To address the above problems, this invention proposes a remote sensing monitoring method for iron concentration in nearshore waters based on satellite data and neural networks.

[0008] The technical solution of this invention is: a remote sensing monitoring method for iron concentration in nearshore waters based on satellite data and neural networks, comprising the following steps:

[0009] S1. Acquire Sentinel-2 satellite images and perform atmospheric correction to obtain water surface remote sensing reflectance data;

[0010] S2. Extract the original input features based on the water surface remote sensing reflectance data;

[0011] S3. Based on the original input features, train the backpropagation neural network model using the training dataset;

[0012] S4. Input the water surface remote sensing reflectance data into the trained backpropagation neural network model to obtain the iron concentration value.

[0013] Furthermore, in S1, the water surface remote sensing reflectance data The expression is:

[0014] ;

[0015] in, For band , For band spectral response function, The starting wavelength of the band, The wavelength at the end of the band. For high spectral reflectance, λ is the wavelength.

[0016] Furthermore, in S2, reflectance values ​​of several spectral bands with a Pearson correlation coefficient greater than a set threshold with respect to water iron concentration are selected from the water surface remote sensing reflectance data as the original input features.

[0017] Furthermore, S3 includes the following sub-steps:

[0018] S31. Collect water iron concentration data and satellite remote sensing reflectance data to form a training dataset;

[0019] S32. Process the training dataset;

[0020] S33. Construct a backpropagation neural network model based on the original input features;

[0021] S34. Train the backpropagation neural network model using the processed training dataset.

[0022] Furthermore, in S32, the expression for processing the training dataset is:

[0023] ;

[0024] in, The normalized value. Original reflectivity The minimum value of the feature in the training set. The maximum value of the feature in the training set. The preset scaling factor. It is a logarithmic function.

[0025] Furthermore, the backpropagation neural network model includes an input layer, a spectral co-enhancement layer, a hidden layer, a Dropout layer, and an output layer connected in sequence;

[0026] The spectral synergistic enhancement layer includes a first feature extraction branch and a second feature extraction branch set in parallel. The first feature extraction branch is used to extract the logarithmic response features between bands, and the second feature extraction branch is used to extract the difference ratio features between bands. The logarithmic response features and the difference ratio features are linearly superimposed and fused.

[0027] Furthermore, the expression for the linear superposition and fusion of the spectral synergistic enhancement layers is as follows:

[0028] ;

[0029] in, The enhancement features obtained by the spectral synergistic enhancement layer As the first trainable weight, As the second trainable weight, For bias, It is a logarithmic function. It is a non-linear activation function. As the first input feature, This is the second input feature.

[0030] Furthermore, the activation functions of neurons in the hidden layer The expression is:

[0031] ;

[0032] in, This represents the net input to the neuron. For learnable gain parameters, For exponents.

[0033] Furthermore, the loss function of the backpropagation neural network model The expression is:

[0034] ;

[0035] in, The number of training samples, Positive weighting coefficients This represents the actual iron concentration value. These are the model's predicted values.

[0036] The beneficial effects of this invention are:

[0037] (1) By introducing a spectral synergistic enhancement layer, the present invention can actively learn and construct spectral features that are better than the original band. Combined with a customized activation function and loss function, it can effectively capture the highly nonlinear relationship between iron concentration and spectral signal in optically complex water bodies.

[0038] (2) This invention is more sensitive to high pollution risk. The high concentration sensitive weighted loss function is used so that the model focuses on optimizing the prediction accuracy of high concentration samples during training. This has important practical significance for environmental supervision and pollution early warning, and improves the practical value of the method.

[0039] (3) The technical solution of the present invention is highly innovative. It systematically and physically improves the input end (normalization), network structure (enhancement layer), neurons (activation function) and training target (loss function) of the neural network in response to specific technical problems, forming a complete and original technical solution.

[0040] (4) This invention has high spatiotemporal resolution and outstanding cost-effectiveness. It inherits the advantages of remote sensing monitoring and uses free Sentinel-2 data to achieve high-frequency observation with a spatial resolution of 10 meters and a revisit cycle of 5 days. At the same time, the standardized process is easy to automate and has the potential for operational application. Attached Figure Description

[0041] Figure 1 A flowchart of a remote sensing monitoring method for iron concentration in nearshore waters based on satellite data and neural networks;

[0042] Figure 2 This is a schematic diagram of the improved BP neural network model structure used in this invention;

[0043] Figure 3 This is a schematic diagram illustrating how the loss function value of the improved BPNN model changes with the number of iterations during the training process.

[0044] Figure 4(a) is a scatter plot comparing the performance of multiple linear regression (MLR) models;

[0045] Figure 4(b) Scatter plot comparing the performance of conventional BPNN;

[0046] Figure 4(c) is a scatter plot comparing the performance of the improved BPNN model of this invention;

[0047] Figure 5 This is an example diagram showing the spatial distribution of Fe concentration as of January 1, 2025, retrieved using the method of this invention.

[0048] Figure 6 This is a schematic diagram of the spatial distribution of Fe concentration in Zhanjiang Bay from 2017 to 2024, generated using the method of this invention.

[0049] Figure 7 This is a graph showing the trend of annual average iron concentration in Zhanjiang Bay from 2017 to 2024, based on the method of this invention. Detailed Implementation

[0050] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0051] like Figure 1 As shown, this invention provides a remote sensing monitoring method for iron concentration in nearshore waters based on satellite data and neural networks, comprising the following steps:

[0052] S1. Acquire Sentinel-2 satellite images and perform atmospheric correction to obtain water surface remote sensing reflectance data;

[0053] S2. Extract the original input features based on the water surface remote sensing reflectance data;

[0054] S3. Based on the original input features, train the backpropagation neural network model using the training dataset;

[0055] S4. Input the water surface remote sensing reflectance data into the trained backpropagation neural network model to obtain the iron concentration value.

[0056] The spectral bands selected in step S2 are the four bands of the Sentinel-2 multispectral imager: B2 (center wavelength 493nm), B3 (center wavelength 560nm), B4 (center wavelength 665nm), and B5 (center wavelength 704nm).

[0057] In this embodiment of the invention, in S1, the water surface remote sensing reflectance data The expression is:

[0058] ;

[0059] in, For band , For band spectral response function, The starting wavelength of the band, The wavelength at the end of the band. For high spectral reflectance, λ is the wavelength.

[0060] In this embodiment of the invention, in step S2, reflectance values ​​of several spectral bands with a Pearson correlation coefficient greater than a set threshold with respect to water iron concentration are selected from the water surface remote sensing reflectance data and used as the original input features.

[0061] In this embodiment of the invention, S3 includes the following sub-steps:

[0062] S31. Collect water iron concentration data and satellite remote sensing reflectance data to form a training dataset;

[0063] S32. Process the training dataset;

[0064] S33. Construct a backpropagation neural network model based on the original input features;

[0065] S34. Train the backpropagation neural network model using the processed training dataset.

[0066] A backpropagation neural network model was trained to learn and fit the nonlinear mapping relationship between enhancement features and iron concentration.

[0067] In this embodiment of the invention, in S32, the expression for processing the training dataset is:

[0068] ;

[0069] in, The normalized value. Original reflectivity The minimum value of the feature in the training set. The maximum value of the feature in the training set. The preset scaling factor. It is a logarithmic function.

[0070] The preset scaling factor is used to adjust the degree of stretching for low-value areas of the data.

[0071] In this embodiment of the invention, the backpropagation neural network model includes an input layer, a spectral co-enhancement layer, a hidden layer, a Dropout layer, and an output layer connected in sequence.

[0072] The spectral synergistic enhancement layer includes a first feature extraction branch and a second feature extraction branch set in parallel. The first feature extraction branch is used to extract the logarithmic response features between bands, and the second feature extraction branch is used to extract the difference ratio features between bands. The logarithmic response features and the difference ratio features are linearly superimposed and fused.

[0073] To fully explore the physical meaning of spectral data, the model structure of this invention does not adopt the traditional single-chain stacking, but instead designs a feature enhancement topology with physical constraints. Specifically, the spectral co-enhancement layer is designed as a dual-channel parallel structure: Channel 1 (logarithmic response channel) uses a logarithmic function to simulate the nonlinear decay relationship between water component concentration and reflectivity according to the Beer-Lambert law; Channel 2 (ratio enhancement channel) uses the normalized difference index (NDI) concept to construct the relative variation features between bands. This parallel topology allows the network to independently learn the absolute magnitude features and relative spectral shape features of optics, and finally perform deep fusion in the hidden layer.

[0074] In this embodiment of the invention, the expression for the linear superposition and fusion of the spectral synergistic enhancement layer is as follows:

[0075] ;

[0076] in, The enhancement features obtained by the spectral synergistic enhancement layer As the first trainable weight, As the second trainable weight, For bias, It is a logarithmic function. It is a non-linear activation function. As the first input feature, This is the second input feature.

[0077] The first term in the formula is summed. The output corresponding to the first feature extraction branch represents the logarithmic response characteristics between bands; the second term in the formula is a summation. The output of the second feature extraction branch represents the difference ratio feature between bands.

[0078] In this embodiment of the invention, the activation function of neurons in the hidden layer The expression is:

[0079] ;

[0080] in, This represents the net input to the neuron. For learnable gain parameters, For exponents.

[0081] Learnable gain parameters are used to adaptively adjust the slope of the activation function during training.

[0082] In this embodiment of the invention, the loss function of the backpropagation neural network model... The expression is:

[0083] ;

[0084] in, The number of training samples, Positive weighting coefficients This represents the actual iron concentration value. These are the model's predicted values.

[0085] This invention aims to provide a remote sensing monitoring method for iron concentration in nearshore waters based on Sentinel-2 satellite data and an improved neural network incorporating customized modules. Its objectives are: 1) to achieve continuous, high-frequency, and high-resolution monitoring of the spatiotemporal distribution of iron concentration over a large area of ​​sea; 2) to significantly improve monitoring accuracy, particularly the ability to capture high-concentration anomalies, by designing a neural network model that includes a spectral synergistic enhancement layer and a target-oriented loss function, thereby deeply exploring the complex nonlinear relationship between remote sensing reflectance and iron concentration; and 3) to propose a standardized technical process from data preprocessing, feature enhancement, model building to the generation of long-term time-series data products, providing a solution with significant technological innovation and application value for remote sensing monitoring of metal element concentration in nearshore waters.

[0086] The process comprises four core steps: data acquisition and preprocessing, feature band selection, neural network model construction and training, and concentration inversion and product generation. First, long-term Sentinel-2 satellite imagery is acquired and atmospherically corrected, while a log-linear composite normalization method is used for data preprocessing. Then, through Pearson correlation analysis, the B2, B3, B4, and B5 bands, which are most sensitive to iron concentration, are selected as the original inputs for the model. Next, an improved BP neural network model with a spectral synergistic enhancement layer is constructed. This enhancement layer can nonlinearly fuse the input band information to generate more discriminative features. During the training phase, a loss function weighted for high-concentration samples and a variable-gain activation function are used to ensure the model learning process focuses more on the prediction accuracy of high-concentration pollution events. Finally, the trained optimal model is applied to all preprocessed satellite imagery to generate a spatiotemporal dynamic dataset of iron concentration from 2017 to 2024.

[0087] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The present invention proposes a remote sensing monitoring method for iron concentration in nearshore waters based on Sentinel-2 satellite data and an improved BP neural network. The specific implementation steps are as follows:

[0088] Step 1: Data Acquisition and Preprocessing

[0089] 1. Satellite Data Acquisition and Atmospheric Correction: Collect and download long-term Sentinel-2A / B satellite L1C-level imagery covering the study area (e.g., Zhanjiang Bay) from 2017 to 2024. Atmospheric correction was performed on all imagery using Acolite software to obtain high-quality water surface remote sensing reflectance (Rrs) products.

[0090] 2. On-site data acquisition and spectral matching: Within a window period synchronized or quasi-synchronized with the satellite transit time (e.g., within ±3 hours), surface water samples are collected and total iron (Fe) concentration is determined. Simultaneously, a hyperspectral radiometer is used to measure continuous... The spectrum was analyzed, and the spectral response function (SRF) of each band of Sentinel-2 was used to convert it into the satellite's equivalent reflectance. .

[0091] 3. Dataset Construction: Match the Fe concentration collected on-site with the reflectance of preprocessed, spatiotemporally nearest satellite image pixels to establish a satellite-ground matching dataset for model training and validation.

[0092] 4. Data Normalization: To eliminate the influence of different dimensions between features and to adapt to the nonlinear characteristics of water reflectance data, this invention employs a logarithmic-linear composite normalization method to map the data to the [0,1] interval. This method effectively stretches the data distribution in low reflectance regions while compressing high reflectance regions, enhancing the model's sensitivity to subtle changes in water color.

[0093] Step 2: Construction and Optimization of the Inversion Model

[0094] 1. Selection of characteristic bands: Pearson correlation coefficient (r) is used to quantitatively assess the linear correlation between reflectance and Fe concentration in each band.

[0095] The analysis results show that the four bands B2 (493nm), B3 (560nm), B4 (665nm), and B5 (704nm) are highly correlated with Fe concentration. Therefore, their Rrs values ​​are selected as the original input feature vector X=[Rrs(B2),Rrs(B3),Rrs(B4),Rrs(B5).

[0096] 2. Improved BP Neural Network Model Construction and Training: This invention employs a BP neural network with optimized structure and algorithm, the model structure of which is as follows: Figure 2 As shown.

[0097] The core innovation of the model lies in the introduction of a spectral co-enhancement layer. This layer is located between the 4-node input layer and the 7-node hidden layer, employing a dual-path parallel structure. It receives the original 4-band input, extracts features through parallel logarithmic response and difference ratio branches, and finally fuses these enhanced features and passes them to the subsequent hidden layers.

[0098] 3. Forward Propagation: After passing through the spectral co-enhancement layer, its output enhanced feature vector It is fed into a traditional multilayer perceptron structure. The hidden layer... The output of each neuron The expression is:

[0099] ;

[0100] in, For the hidden layer The linearly weighted input sum of neurons, For the previous layer The node is connected to the current hidden layer. The connection weights of each neuron For the previous layer The output value of each node, For the first Bias terms for each neuron, This invention uses a variable gain activation function. This is a learnable gain parameter associated with the neuron, allowing the network to automatically adjust the steepness of each neuron's activation function during training to adapt to different feature responses. The output layer is linear, yielding the predicted value.

[0101] 4. Backpropagation and weight update: To make the model more focused on the prediction of high-concentration pollution events, this invention adopts a high-concentration-sensitive weighted mean square error loss function.

[0102] The model was trained using a gradient descent-based backpropagation algorithm, with the Adam optimizer being the preferred choice. After approximately 10,000 iterations of training, the loss function changed as follows: Figure 3 As shown, the loss value drops rapidly in the early stages of training, then decreases slowly and eventually stabilizes, indicating that the model has learned sufficiently and converged.

[0103] Step 3: Model Performance Validation

[0104] To verify the superiority of the proposed improved BP neural network (NewBPNN) in nearshore water iron concentration inversion, this embodiment compares it with the traditional multiple linear regression (MLR) model and the unmodified conventional BP neural network model (StandardBPNN). All three models were trained using the same training set and validated on completely independent test sets. The validation results are shown in Figure 4.

[0105] Figure 4 shows a scatter plot comparing the three models on the validation set, with the black dashed line representing the 1:1 reference line.

[0106] 1. As shown in Figure 4(a), the coefficient of determination (R2) of the MLR model is only 0.70, the root mean square error (RMSE) is 71.80µg / L, and the mean absolute percentage error (MAPE) is 22.70%, indicating that the simple linear model is difficult to capture the complex nonlinear relationship between iron concentration and remote sensing reflectance in nearshore waters, and the prediction accuracy is low.

[0107] 2. As shown in Figure 4(b), although the unmodified conventional BPNN model achieved an R² of 0.80, its prediction error was extremely high, with an RMSE as high as 80.85 µg / L and a MAPE as high as 33.97%. Observing the scatter distribution, it can be seen that the predicted values ​​deviate significantly from the measured values ​​and exhibit obvious instability. This indicates that conventional network structures, in the absence of specific prior constraints, are prone to getting trapped in local optima or overfitting, resulting in weak generalization ability.

[0108] 3. As shown in Figure 4(c), the improved BPNN model proposed in this invention exhibits the best overall performance. While maintaining good correlation (R... 2 While achieving a mean square error (MSE) of 0.79, this invention also achieved high accuracy. Its RMSE was significantly reduced to 49.94 µg / L, and its MAPE was only 14.41%. Compared with conventional BPNN models, the root mean square error of this invention was reduced by approximately 38%, and the mean absolute percentage error was reduced by approximately 57%. The data points were more closely distributed around the 1:1 line, demonstrating the effective extraction of key features by the spectral synergistic enhancement layer and the positive effect of the weighted loss function on model optimization, greatly improving the robustness and accuracy of the model.

[0109] Step 4: Long-term concentration inversion and product generation

[0110] The optimized BPNN model trained in the preceding steps was applied to all preprocessed long-term Sentinel-2 images (e.g., 2017-2024) to perform pixel-by-pixel Fe concentration inversion calculations, ultimately generating a rasterized spatiotemporal distribution dataset of Fe concentration covering the entire study period. This dataset can intuitively display the spatial distribution pattern and temporal evolution trend of Fe concentration, such as... Figure 5 The image shows the spatial distribution of Fe concentration at a certain moment, and as shown below... Figure 6 The diagram shows a long-term spatial distribution.

[0111] Furthermore, to visually demonstrate the long-term evolution of water quality in the study area, this invention, based on the generated long-term time-series dataset, statistically analyzed the annual average iron concentration from 2017 to 2024. The results are as follows: Figure 7 As shown in the figure, the iron concentration in this sea area was relatively low (approximately 261-270 µg / L) in 2017-2018, then increased significantly in 2019-2020, reaching a peak of 338.10 µg / L in 2020. Subsequently (2021-2024), it showed a fluctuating downward trend, decreasing to 289.42 µg / L in 2024. This interannual trend can provide data support for assessing the effectiveness of local environmental governance.

[0112] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A remote sensing monitoring method for iron concentration in nearshore waters based on satellite data and neural networks, characterized in that, Includes the following steps: S1. Acquire Sentinel-2 satellite images and perform atmospheric correction to obtain water surface remote sensing reflectance data; S2. Extract the original input features based on the water surface remote sensing reflectance data; S3. Based on the original input features, train the backpropagation neural network model using the training dataset; S4. Input the water surface remote sensing reflectance data into the trained backpropagation neural network model to obtain the iron concentration value.

2. The remote sensing monitoring method for nearshore water iron concentration based on satellite data and neural networks according to claim 1, characterized in that, In S1, the water surface remote sensing reflectance data The expression is: ； in, For band , For band spectral response function, The starting wavelength of the band, The wavelength at the end of the band. For high spectral reflectance, λ is the wavelength.

3. The remote sensing monitoring method for nearshore water iron concentration based on satellite data and neural networks according to claim 1, characterized in that, In step S2, reflectance values ​​of several spectral bands with a Pearson correlation coefficient greater than a set threshold with respect to water iron concentration are selected from the water surface remote sensing reflectance data and used as the original input features.

4. The remote sensing monitoring method for iron concentration in nearshore waters based on satellite data and neural networks according to claim 1, characterized in that, S3 includes the following sub-steps: S31. Collect water iron concentration data and satellite remote sensing reflectance data to form a training dataset; S32. Process the training dataset; S33. Construct a backpropagation neural network model based on the original input features; S34. Train the backpropagation neural network model using the processed training dataset.

5. The remote sensing monitoring method for nearshore water iron concentration based on satellite data and neural networks according to claim 4, characterized in that, In step S32, the expression for processing the training dataset is: ； in, The normalized value. Original reflectivity The minimum value of the feature in the training set. The maximum value of the feature in the training set. The preset scaling factor. It is a logarithmic function.

6. The remote sensing monitoring method for nearshore water iron concentration based on satellite data and neural networks according to claim 4, characterized in that, The backpropagation neural network model includes an input layer, a spectral co-enhancement layer, a hidden layer, a Dropout layer, and an output layer connected in sequence. The spectral synergistic enhancement layer includes a first feature extraction branch and a second feature extraction branch set in parallel; the first feature extraction branch is used to extract the logarithmic response features between bands, and the second feature extraction branch is used to extract the difference ratio features between bands, and the logarithmic response features and the difference ratio features are linearly superimposed and fused.

7. The remote sensing monitoring method for nearshore water iron concentration based on satellite data and neural networks according to claim 6, characterized in that, The expression for the linear superposition and fusion of the spectral synergistic enhancement layer is as follows: ； in, The enhancement features obtained by the spectral synergistic enhancement layer As the first trainable weight, As the second trainable weight, For bias, It is a logarithmic function. It is a non-linear activation function. As the first input feature, This is the second input feature.

8. The remote sensing monitoring method for iron concentration in nearshore waters based on satellite data and neural networks according to claim 6, characterized in that, The activation function of neurons in the hidden layer The expression is: ； in, This represents the net input to the neuron. For learnable gain parameters, It is an index.

9. The remote sensing monitoring method for iron concentration in nearshore waters based on satellite data and neural networks according to claim 6, characterized in that, The loss function of the backpropagation neural network model The expression is: ； in, The number of training samples, Positive weighting coefficients This represents the actual iron concentration value. These are the model's predicted values.

Images