Method for predicting seawater transparency and sea surface temperature based on multi-source remote sensing data

Abstract

The application discloses a method for predicting seawater transparency and sea surface temperature based on multi-source remote sensing data, which comprises the following steps: firstly, collecting sufficient water color satellite data, performing residual correction, water body inherent optical quantity inversion, seawater transparency inversion and other preprocessing operations on the remote sensing reflectivity data, and performing space-time interpolation on the obtained seawater transparency data and sea surface temperature data, and then making a training data set; then, a prediction model is constructed, and the prediction model is trained to determine the best prediction model, wherein the prediction model adopts a combination of a convolution long short-term memory network (ConvLSTM) and a full link to construct the prediction model; finally, the seawater transparency and the sea surface temperature of a specific sea area are predicted based on the trained prediction model, so that the accuracy of future marine environment information is ensured.

Description

Technical Field

[0001] This invention belongs to the field of marine remote sensing technology applications, specifically relating to a method for predicting seawater transparency and sea surface temperature based on multi-source remote sensing data. Background Technology

[0002] Seawater transparency (Z) sd Sea surface transparency (SST) is typically expressed as the maximum depth visible when the secant is submerged in water. It is a measure of seawater visibility and a crucial factor influencing the distribution of the underwater light field. SST refers to the temperature of the surface water of the ocean and is an important physical parameter for understanding ocean water bodies. Research on predicting sea surface transparency is of great significance for marine underwater ecology, water quality, navigation, and aquaculture. Furthermore, understanding future SST information has positive strategic implications for meteorological and climatic research, navigation activities, aquaculture, and fishing.

[0003] Previous studies on seawater transparency measurements have primarily relied on historical data and empirical linear models and semi-analytical algorithms to estimate past seawater transparency. Kratzer et al. (2008) proposed a method to estimate seawater transparency using the vertical diffuse attenuation coefficient; Prasad et al. (2016), through analysis of extensive water color data, discovered a strong correlation between seawater transparency and the ratio of water-leaving radiance in two bands, and constructed a linear model for detecting seawater transparency based on these two factors; Doron et al. (2011) proposed a method to calculate seawater transparency using intrinsic optical parameters obtained from water-leaving radiation inversion. While these studies can accurately invert past seawater transparency in a specified sea area, they can only roughly qualitatively describe the future trend of seawater transparency and cannot meet the requirements for accurate prediction and forecasting.

[0004] For sea surface temperature (SST) prediction, Barnston and Smith (1996) proposed canonical correlation analysis, which processes a set of data based on physical facts and uses the correlation between the data and the data to construct a model for predicting SST. Xiao et al. (1993) proposed a method combining a Long Short-Term Memory (LSTM) neural network model with the AdaBoost algorithm to achieve short- to medium-term SST prediction, improving prediction ability and avoiding overfitting. Jia et al. (2022) used the Loess algorithm to first perform seasonal weighting on the data and then used LSTM to predict SST in the East China Sea. Yang et al. (2018) proposed the CFCC-LSTM model, which classifies SST prediction as a sequence prediction problem, predicting future SST from the spatiotemporal perspective of historical data. In previous SST forecasting studies, the method of constructing prediction models through correlation often has a strong regional dependence and cannot express the complex relationships between environmental factors affecting SST. In prediction models built using LSTM, the calculation of parameters such as weights in the network structure is done using a fully linked matrix multiplication method. These models have too many network parameters, resulting in a huge computational load, and they cannot effectively extract spatial information.

[0005] In existing technologies, the measurement of seawater transparency and the prediction of sea surface temperature mainly rely on building deep learning prediction models using existing data. These models then calculate future seawater transparency and sea surface temperature at a predicted point, achieving only point-to-point prediction. However, the training data in this point-to-point prediction process often lacks spatial information about the surrounding area, thus limiting the accuracy of the model's predictions. Ocean optical remote sensing technology is an emerging detection technology that uses detectors mounted on remote sensing platforms to detect sea surface color information and acquire ocean information. The rapid development of ocean color satellite remote sensing technology has made it possible to predict future seawater transparency and sea surface temperature using past ocean color satellite remote sensing data. Based on these considerations, there is an urgent need to propose a method based on ocean color remote sensing data that uses deep learning models to extract the spatiotemporal features of the predicted point and its surrounding area, thereby achieving point-to-point prediction of seawater transparency and sea surface temperature. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting seawater transparency and sea surface temperature based on multi-source remote sensing data. This method utilizes historical marine remote sensing images to predict future seawater transparency and sea surface temperature, providing more continuous, accurate, and advanced marine environmental information for marine activities and marine disaster prevention.

[0007] This invention is achieved using the following technical solution: a method for predicting seawater transparency and sea surface temperature based on multi-source remote sensing data, comprising the following steps:

[0008] Step A: Constructing a training dataset: Collect water color satellite data, which includes remote sensing reflectance data and sea surface temperature data for various bands over many years. Perform residual correction, inversion of inherent optical quantities of water bodies, and inversion of seawater transparency on the remote sensing reflectance data. Then, perform spatiotemporal interpolation on the obtained seawater transparency data and sea surface temperature data to create a training dataset.

[0009] Step B: Construct a forecasting model and train it to determine the optimal forecasting model. The structure and principle of the forecasting model are as follows:

[0010] (1) The forecast model is constructed by combining Convolutional Long Short-Term Memory (ConvLSTM) network with fully connected layers, including an N-layer ConvLSTM network and a forecast layer based on fully connected computation, where N is an integer from 2 to 5;

[0011] (2) The model input is a three-dimensional array of type n×n×d, which is composed of n×n pixels for each day of the previous d days. The spatiotemporal feature information of the data window is learned through forward propagation of N layers of ConvLSTM network units. Then, the value of the n×n data window on day d+1 is calculated through the prediction layer. Finally, the value of the center pixel of the window is output, thus completing the prediction of the value of the center pixel. The entire scene data is traversed by sliding the window, and the above operation is performed on each window to predict the seawater transparency and sea surface temperature of the entire scene on day d+1.

[0012] Step C: Based on the trained forecast model, forecast the seawater transparency and sea surface temperature of a specific sea area.

[0013] Furthermore, in step B: the formula for calculating the convolution within the ConvLSTM model is as follows:

[0014]

[0015]

[0016]

[0017]

[0018]

[0019] In the formula, t represents day t, * represents the convolution calculation method, i, o, f, h represent the input gate, output gate, forget gate, and transmission state of the model network, respectively, C and H represent the cell state and hidden state, W and b represent the weights and biases of the corresponding data, X represents the input dataset, and σ represents the activation function.

[0020] Furthermore, in step B, during the training process, by adjusting the number of ConvLSTM layers in the model from 2 to 5, when the MAE and loss of each model reach the accuracy requirements, the calculation speed and accuracy of the model with 2, 3, 4, and 5 ConvLSTM layers are compared, and the ConvLSTM layer number and related hyperparameter values ​​of the model with the highest prediction efficiency are saved as the best model.

[0021] The formulas for calculating the mean absolute error and loss function used to determine the hyperparameters of the ConvLSTM model are as follows:

[0022]

[0023]

[0024] In the formula, MAE is the mean absolute error, loss is the loss function, and x i It is the i-th value predicted by the model, y i It is the i-th true value, and n is the total number of windows in the training samples.

[0025] Furthermore, step A is specifically implemented in the following manner:

[0026] Step A1, Residual Correction: Calculate the residuals in the remote sensing reflectance data based on the Inherent Optical Properties Data Processing System (IDAS);

[0027] Step A2, Inversion of Intrinsic Optical Quantities of Water: Input the remote sensing reflectance data after residual correction in Step A1 into the Intrinsic Optical Quantity Algorithm NNSAA to obtain the backscattering coefficient and absorption coefficient of water for each band, and then obtain the total backscattering coefficient and total absorption coefficient.

[0028] Step A3, Seawater Transparency Inversion: Based on the remote sensing reflectance data after residual correction in Step A1 and the total backscattering coefficient and total absorption coefficient obtained in Step A2, the seawater transparency Z is calculated according to the transparency inversion algorithm. sd ;

[0029] Step A4, Spatiotemporal data interpolation: Spatial interpolation of seawater transparency and sea surface temperature data for each scene is performed using the natural proximity interpolation algorithm to improve the spatial coverage of the data; temporal interpolation of missing seawater transparency and sea surface temperature data in the required time series data is performed using the linear interpolation algorithm to ensure that there is data every day in the same sea area and to ensure the integrity of the series data.

[0030] Furthermore, in step A3, when performing seawater transparency inversion, the turbidity index T is used as the basis. d Each corresponding formula for calculating seawater transparency is constructed, and finally, seawater transparency image data is calculated, including:

[0031] (1) Define the turbidity index T of water bodies d Preliminary classification of water bodies:

[0032] T d =1.8386R rs (667)-R rs (488)

[0033] (2) Using the total backscattering coefficient and total absorption coefficient calculated in step A2 as inputs, construct a system for estimating the Z-value of low-turbidity seawater. sd formula:

[0034]

[0035] (3) Construct a system based on R rs (748) and R rs The formula for estimating the difference in (869) is: [Formula for the transparency of extremely turbid seawater]

[0036] Z st,et =0.0036[R] rs (748)-R rs (869)] -0.840 ,T d >0.014

[0037] (4) Combining (2) and (3), the formula for calculating the transparency of moderately turbid seawater is constructed as follows:

[0038] Z sd,mt =W·Z sd,tc +(1-W)Z sd,et 0.01≤T d ≤0.014 (10)

[0039] In the formula, Z sd,lt The formula for calculating the transparency of low-turbidity seawater, Z sd,et The formula for calculating the transparency of extremely turbid seawater, Z sd,mt The formula for calculating the transparency of moderately turbid seawater, bbw (488) is the backscattering coefficient of pure water at 488 nm, and W is the weight value of the connection formula.

[0040] Furthermore, step A4 is specifically implemented in the following manner:

[0041] (1) Spatial data interpolation and filling of seawater transparency and sea surface temperature image data are performed using the natural proximity interpolation algorithm. The specific formula is as follows:

[0042]

[0043] In the formula, f(x) is the interpolated value at the point x to be interpolated, and w i (x) represents the sample points i, i = 1, ..., n, and their weights on x, f i The value of sample point i;

[0044] The formula for calculating the weight is as follows:

[0045]

[0046] In the formula, a i Let a(x) be the area of ​​the Thiessen polygon where the sample point to be interpolated is located, and let a(x) be the area of ​​the Thiessen polygon where the point x to be interpolated is located.

[0047] (2) A linear interpolation algorithm was used to interpolate and fill in the time series data of seawater transparency and sea surface temperature image data. The specific formula is as follows:

[0048]

[0049] In the formula, f(d) is the value after interpolation on day d, f(d1) is the value of day d1 that is closest to day d before day d, and f(d2) is the value of day d2 that is closest to day d after day d.

[0050] Furthermore, in step A, the training dataset is created using the following method:

[0051] Using multi-time series seawater transparency or sea surface temperature as the original dataset, a data unit set consisting of n×n×d pixels in a three-dimensional array storage format is constructed as the training sample. That is, while traversing the satellite remote sensing images over consecutive days at the time scale, the data of each day is traversed through an n×n pixel window, ultimately forming a three-dimensional array set consisting of a large number of n×n×d pixels.

[0052] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0053] This scheme predicts seawater transparency and sea surface temperature by constructing a forecasting model. The forecasting model is built using a combination of Convolutional Long Short-Term Memory (ConvLSTM) networks and fully connected layers, comprising N ConvLSTM layers and one forecasting layer. In implementation, the spatial characteristics of the data are preserved by constructing an n×n data window centered on the pixel to be predicted. Simultaneously, a three-dimensional array is constructed using data windows spanning consecutive days (d days) to preserve the temporal characteristics of the data, ultimately achieving a point-to-area prediction of seawater transparency and sea surface temperature. This three-dimensional array is used as input to the model. Through forward propagation learning of the N ConvLSTM network units, the temporal and spatial feature information of the data window is extracted. The predicted seawater transparency and sea surface temperature for the pixel on day d+1 are then calculated using the fully connected computation method of the forecasting layer. Furthermore, by sliding the window through the data within a specified sea area and performing the above operations, the prediction of seawater transparency and sea surface temperature for the entire sea area on day d+1 can be achieved, ensuring the accuracy of future marine environmental information. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the overall process of forecasting seawater transparency and sea surface temperature according to an embodiment of the present invention.

[0055] Figure 2 This is a flowchart illustrating the IDAS calculation of remote sensing reflectance residuals according to an embodiment of the present invention.

[0056] Figure 3 This is a flowchart illustrating the NNSAA calculation of inherent optical quantities of water bodies according to an embodiment of the present invention.

[0057] Figure 4 This is a diagram of the ConvLSTM neural network structure used for predicting seawater transparency and sea surface temperature in an embodiment of the present invention.

[0058] Figure 5 This is a schematic diagram illustrating the changes in MAE and loss during the training process of the sea surface temperature prediction model according to an embodiment of the present invention.

[0059] Figure 6 This is a schematic diagram illustrating the changes in MAE and loss during the training process of the seawater transparency prediction model in an embodiment of the present invention.

[0060] Figure 7 This is a spatial distribution comparison diagram of the predicted and actual sea surface temperature values ​​in an embodiment of the present invention.

[0061] Figure 8 This is a spatial distribution comparison diagram of the predicted and actual values ​​of seawater transparency in an embodiment of the present invention;

[0062] Figure 9This is a schematic diagram of the mean square error (MRE) analysis results comparing the predicted sea surface temperature and seawater transparency with the actual observed data in an embodiment of the present invention.

[0063] Figure 10 This is a schematic diagram comparing the actual time series and the predicted time series of sea surface temperature in an embodiment of the present invention;

[0064] Figure 11 This is a schematic diagram comparing the actual time series and the predicted time series of seawater transparency in an embodiment of the present invention. Detailed Implementation

[0065] To better understand the above-described objects, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the present invention; however, the present invention may be practiced in other ways than those described herein, and therefore, the present invention is not limited to the specific embodiments disclosed below.

[0066] This embodiment proposes a method for predicting seawater transparency and sea surface temperature based on multi-source remote sensing data and a deep learning model. It predicts the value of the central pixel on day d+1 using a three-dimensional array composed of continuous d-day window data, achieving a point-to-area prediction of seawater transparency and sea surface temperature. To improve the quality of the data used in this method and ensure the accuracy of the prediction results, preprocessing operations such as residual correction, inversion of inherent optical quantities of water bodies, inversion of seawater transparency, and data interpolation are required for the collected remote sensing reflectance data. Data interpolation is performed on the collected sea surface temperature data, specifically including the following steps:

[0067] Step A: Constructing a training dataset: Collect sufficient water color satellite data, including multi-band remote sensing reflectance data and sea surface temperature data from multiple years. Perform preprocessing operations such as residual correction, inversion of inherent optical quantities of water bodies, and inversion of seawater transparency on the remote sensing reflectance data. Then, perform time series and spatial data interpolation on the obtained seawater transparency data and sea surface temperature data to create a model training dataset.

[0068] Step B: Construct a forecast model and train it to determine the optimal structure and related hyperparameters of the forecast model. The structure and principle of the forecast model are as follows:

[0069] The forecasting model employs a combination of Convolutional Long Short-Term Memory (ConvLSTM) networks and a fully connected layer, comprising N ConvLSTM layers and one forecasting layer. The model takes a 3D array of seawater transparency or sea surface temperature data for the previous 30 days (a 15×15 data window) as input. First, the N ConvLSTM network units undergo forward propagation to extract the temporal and spatial features of the data window. Then, the forecasting layer calculates the seawater transparency and sea surface temperature values ​​for the 31st day of the data window. Finally, the value of the center pixel of the data window is output, thus completing the prediction of the seawater transparency and sea surface temperature for that pixel on the 31st day. By performing the above process on all pixels across the entire sea area, the seawater transparency and sea surface temperature for the entire sea area on the 31st day can be predicted.

[0070] During model training, the seawater transparency and sea surface temperature data for day 1 of the window are discarded, while the seawater transparency and sea surface temperature data for day 31 of the window are used as the 30th data in the new input three-dimensional array. By repeating the above process once, the seawater transparency and sea surface temperature for day 32 of the window can be predicted, thereby obtaining the seawater transparency and sea surface temperature for day 32 predicted by the model for the entire sea area.

[0071] In determining the optimal model structure and related hyperparameters, the number of ConvLSTM layers in the model was adjusted from 2 to 5 during model training. When the MAE and loss of each model reached the accuracy requirements, the prediction rate and accuracy of the models with 2, 3, 4, and 5 ConvLSTM layers were compared, and the model with the highest prediction efficiency was selected as the best.

[0072] Step C: Forecast the seawater transparency and sea surface temperature of a specific sea area based on the trained forecast model.

[0073] To better understand the present invention, the following detailed description is provided using a method for predicting seawater transparency and sea surface temperature in a specific sea area. The forecasting process is as follows: Figure 1 As shown, the specific steps include:

[0074] I. Constructing the training dataset:

[0075] Step 1: Data Acquisition: Collect sufficient ocean color satellite data, including remote sensing reflectance and sea surface temperature data for various bands over many years. In this embodiment, secondary and tertiary data from the Moderate Resolution Imaging Spectroradiometer (MODIS) for a certain sea area from 2011 to 2021 are collected, namely, remote sensing reflectance (R... rs (λ)) and sea surface temperature data (SST);

[0076] Step 2, Residual Correction: Calculate R based on the Inherent Optical Properties Data Processing System (IDAS). rs (λ) Residuals in the data provide high-quality data for subsequent seawater transparency inversion. The specific process is as follows: Figure 2 As shown, the process is as follows:

[0077] (1) First, based on the known relationship between the actual remote sensing reflectance received by the satellite, the theoretically received remote sensing reflectance, and the residual, formula (1) is obtained.

[0078] R rs,t (λ)=R rs,s (λ)-ΔR rs Iλ)(1)

[0079] In the formula, R rs,s (λ) represents the actual λ-band remote sensing reflectance received by the water color satellite, R rs,t (λ) represents the theoretically received λ-band remote sensing reflectance from the water color satellite, ΔR rs (λ) represents the residual contained in the actual received λ-band remote sensing reflectance.

[0080] (2) Then, R rs,s (λ) Input the neural network-based intrinsic optical quantity calculation model (NQAA) to obtain a(555) and b b (670); At this point, considering that the ocean water is very clean due to the lack of suspended particles, the total absorption coefficient of the ocean water at 670 nm is defined as equal to the absorption coefficient of pure water, that is, a(670)≈a w (670)

[0081] (3) Next, based on formula (1), determine ΔR. rs (555) and ΔR rs (670) Linear empirical relation, i.e., formula (2):

[0082] ΔR rs (555)=m0ΔR rs (670)+m1(2)

[0083] Where m0 and m1 are empirical constants for the linear relationship between the remote sensing reflectance residuals in the 555nm and 670nm bands.

[0084] Therefore, by applying the 670nm remote sensing reflectance to formula (1), the remote sensing reflectance theoretically received by the satellite can be represented by the remote sensing reflectance actually received by the satellite and the remote sensing reflectance residual.

[0085] (4) Combining the a(555) and b obtained in the previous steps b (670), ΔR rs(555) and pure water absorption coefficient (a w Substituting (670)=0.439) into formula (3), we can obtain the result containing only b. b (555) and ΔR rs (670) The system of equations with two unknowns can be solved simultaneously to obtain ΔR. rs (670):

[0086]

[0087]

[0088]

[0089] In formula (3), the left side of the equation is based on the reflectance (r) at the air-sea interface. rs ) and inherent optical quantities (a and b) b The equation is derived from the radiative transfer relationship between the air and sea, i.e., formula (4); the right side of the equation is based on the reflectivity (r) at the air-sea interface. rs The relationship between g0 and g1 on the air-sea interface received by the satellite is obtained, i.e., formula (5); where g0 and g1 are 0.089sr and g1 respectively. -1 and 0.1245sr -1 The fixed empirical constant.

[0090] At this point, the remote sensing reflectance residual ΔR at 670 nm has been obtained. rs (670)

[0091] (5) Finally, based on the spectral relationship of the data residuals, i.e., formula (6), the remote sensing reflectance residuals for each band are calculated:

[0092]

[0093] In the formula, S represents the spectral slope coefficient of the residual, and C is a constant. In this embodiment, ΔR is used to... rs After statistical analysis of the dataset, C is set to 0.04.

[0094] Step 3, Inversion of inherent optical quantities of water: The residual-corrected R... rs (λ) The total backscattering coefficient b for each band is obtained using the input intrinsic optics algorithm (NNSAA, a known and mature algorithm). b (λ) and the total absorption coefficient a(λ), the specific process is as follows Figure 3 As shown, we will not elaborate further here.

[0095] Step 4, Seawater Transparency Inversion: Based on the R after residual correction rsUsing the (λ) data and the total backscattering coefficient and total absorption coefficient data calculated in step 3, the seawater transparency (Z) is calculated according to the transparency inversion algorithm. sd );

[0096] Considering that the algorithm used in step 3 to calculate the inherent optical quantities works well in low to medium turbidity water bodies, but is not ideal in extremely turbid water bodies with more complex optical properties, this embodiment first defines the water turbidity index (T). d The water bodies are initially classified as shown in formula (7), and then classified according to T. d The corresponding formulas for calculating seawater transparency were constructed, and the seawater transparency image data were finally calculated.

[0097] In this step, a(λ) and b are calculated in step 3. b (λ) is used as input to construct a function for estimating Z in low-turbidity seawater. sd The algorithm is shown in formula (8). Taking advantage of the fact that extremely turbid water bodies are more sensitive to changes in optically active components in the near-infrared band than in the blue-green band, an algorithm based on R... rs (748) and R rs (869) The algorithm for specifically estimating the transparency of extremely turbid seawater using band difference is shown in formula (9). Combining the above two algorithms, the formula for calculating the transparency of moderately turbid seawater is shown in formula (10).

[0098] T d =1.8386R rs (667)-R rs (488) (7)

[0099]

[0100] Z st,et =0.0036[R] rs (748)-R rs (869)] -0.840 ,T d >0.014 (9)

[0101] Z sd,mt =W·Z sd,tc +(1-W)Z sd,et 0.01≤T d ≤0.014 (10)

[0102] In the formula, Z sd,lt Z is the formula for calculating the transparency of low-turbidity seawater. sd,et The formula for calculating the transparency of extremely turbid seawater, Z sd,mt The formula for calculating the transparency of moderately turbid seawater, b bw(488) is the backscattering coefficient of pure water at 488 nm, and W is the weight value of the connection formula.

[0103] Step 5, Spatiotemporal data interpolation: For each day's Z-axis, use the natural nearest point interpolation algorithm. sd Spatial interpolation is performed on the data and SST data to improve the spatial coverage of the data; linear interpolation algorithms are used to interpolate and fill in the time series data of seawater transparency and sea surface temperature to ensure the continuity and integrity of the time series.

[0104] (1) To avoid spatial data gaps causing errors in subsequent model training, this embodiment uses a natural proximity interpolation algorithm for Z. sd Spatial data interpolation and filling is performed using SST image data, with the following specific formula:

[0105]

[0106] In the formula, f(x) is the interpolated value at the point x to be interpolated, and w i (x) represents the weights of the sample points i (i = a, ..., n) participating in the interpolation with respect to x, f i Let i be the value of sample point i.

[0107] The formula for calculating the weight is as follows:

[0108]

[0109] In the formula, a i Let a(x) be the area of ​​the Thiessen polygon where the sample point to be interpolated is located, and let a(x) be the area of ​​the Thiessen polygon where the point x to be interpolated is located.

[0110] (2) To ensure the integrity of the time series data, this embodiment uses a linear interpolation algorithm to interpolate and fill in the time scale of the seawater transparency and sea surface temperature time series data. The specific formula is as follows:

[0111]

[0112] In the formula, f(x) is the value after interpolation on day x, f(x1) is the value of day x1, which is the closest day to day x before day x, and f(x2) is the value of day x2, which is the closest day to day x after day x.

[0113] Step 6: Create the training dataset: using multi-time series Z-series data. sdAlternatively, using SST as the original dataset, data units consisting of n×n×d pixels are constructed as training samples for the deep learning model. That is, while traversing the images over multiple days temporally, each day's data is traversed through a window of n×n pixels, ultimately forming a three-dimensional array set of a large number of n×n×d pixels. In this embodiment, the data window size is designed to be 15×15, and the time range d of the three-dimensional array set is 30 days. The three-dimensional array set of a large number of 15×15×30 pixels obtained through the above steps is used as the input to the neural network model, and the value of the center pixel of the corresponding 15×15 window on the 31st day is used as the model's output value. Finally, the training samples are determined to be 300,000 sets of data, and the validation samples are 50,000 sets of data.

[0114] II. Constructing and calling the forecast model:

[0115] Step 7, Model Building and Training: Design a ConvLSTM structure using Python and train the model. Compare the computational efficiency and accuracy of models with different numbers of ConvLSTM layers, ensuring that the mean absolute error (MAE) and loss function of each model meet accuracy requirements. Save the forecast model structure and related hyperparameters with the best computational efficiency and accuracy as the optimal values ​​for subsequent predictions of the specified sea area Z. sd Or SST model calls.

[0116] In this embodiment, during training, when the MAE and loss tend to stabilize, the hyperparameters for that training iteration are determined as the optimal model parameters, such as... Figure 4 and Figure 5 As shown. In this example, the Z-values ​​of the first 30 days of a data window consisting of 15×15 pixels are displayed. sd The three-dimensional array consisting of the SST and the ΣSST is used as input. After being processed by N ConvLSTM layers and 1 forecast layer, the Z-axis of the center pixel on the 31st day of the window is output. sd And SST.

[0117] This invention employs a combination of Convolutional Long Short-Term Memory (ConvLSTM) networks and fully connected layers to construct a forecasting model. The model consists of N ConvLSTM layers and one forecasting layer. The model takes the seawater transparency and sea surface temperature data for the first 30 days of a 15×15 data window as input. After forward propagation through N ConvLSTM network units, it extracts the spatiotemporal features of the data window. Then, the forecasting layer calculates the seawater transparency and sea surface temperature values ​​for the 31st day of the window. Finally, it outputs the seawater transparency and sea surface temperature values ​​of the center pixel of the window as the prediction result, thus completing one prediction of seawater transparency and sea surface temperature within a single pixel. By iterating through all pixels, the seawater transparency and sea surface temperature for the entire sea area on the 31st day can be predicted. After one prediction for the entire sea area, the first day's data in the input 3D array is discarded, and the value of the 31st day is used as the 30th day's data in the new input 3D array, forming a new input 3D array. Repeating this process multiple times allows for the prediction of multi-time series seawater transparency and sea surface temperature. The specific structure of this invention is as follows: Figure 3 As shown.

[0118] The formulas for calculating the internal convolution of ConvLSTM are shown in formulas (13) to (16):

[0119]

[0120]

[0121]

[0122]

[0123]

[0124] In the formula, t represents day t, * represents convolution calculation, i, o, f, h represent the input gate, output gate, forget gate, and transport state of the model network, respectively, C and H represent the cell state and hidden state, W and b represent the weights and biases of the corresponding data, X represents the input dataset, and σ represents the activation function.

[0125] To determine the optimal forecasting model structure and related hyperparameters, this embodiment adjusts the number of ConvLSTM layers in the model from 2 to 5 during training. Once the MAE and loss of each model reach the required accuracy, the forecasting rate and accuracy of the models with 2, 3, 4, and 5 ConvLSTM layers are compared. The model with the highest forecasting efficiency is identified as the optimal number of ConvLSTM layers and related hyperparameters. This embodiment determines that the optimal ConvLSTM model for forecasting seawater transparency and sea surface temperature contains 3 ConvLSTM layers after model training. The formulas for calculating MAE and loss used to determine the optimal model structure and related hyperparameters are as follows:

[0126]

[0127]

[0128] In the formula, x i The model predicts the i-th value, y. i It is the i-th true value, and n is the total number of windows in the training samples.

[0129] Step 8: Apply the trained best forecast model to the designated sea area to achieve the desired seawater transparency Z in that area. sd With sea surface temperature (SST) forecast;

[0130] This embodiment calls the trained forecast model parameters and applies them to the SST and Z-axis forecasts in the South China Sea. sd In the forecast, by comparing the SST and Z before and after the forecast. sd Spatial distribution ( Figure 7 , Figure 8 SST and Z sd Mean squared error (MRE) between the true and predicted values Figure 9 ), and multi-time SST and Z sd The trend between predicted and actual values ​​( Figure 10 , Figure 11 As can be seen, the accuracy of the model meets the requirements.

[0131] The formula for calculating the mean squared error (MRE) is as follows:

[0132]

[0133] In the formula, x i The model predicts the value for day i, y. i is the actual value on day i, and n is the amount of data used for comparison.

[0134] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for predicting seawater transparency and sea surface temperature based on multi-source remote sensing data, characterized in that, Includes the following steps: Step A: Constructing a training dataset: Collect water color satellite data, which includes remote sensing reflectance data and sea surface temperature data for various bands over many years. Perform residual correction, inversion of inherent optical quantities of water, and inversion of seawater transparency on the remote sensing reflectance data. Then, perform spatiotemporal interpolation on the obtained seawater transparency data and sea surface temperature data to create a training dataset. Step A is specifically implemented in the following way: Step A1, Residual Correction: Calculate the residuals in the remote sensing reflectance data based on the Inherent Optical Properties Data Processing System (IDAS); Step A2, Inversion of Intrinsic Optical Quantities of Water: Input the remote sensing reflectance data after residual correction in Step A1 into the Intrinsic Optical Quantity Algorithm NNSAA to obtain the backscattering coefficient and absorption coefficient of water for each band, and then obtain the total backscattering coefficient and total absorption coefficient. Step A3, Seawater Transparency Inversion: Based on the remote sensing reflectance data after residual correction in Step A1 and the total backscattering coefficient and total absorption coefficient obtained in Step A2, the seawater transparency Z is calculated according to the transparency inversion algorithm. sd ; Step A4, Spatiotemporal data interpolation: Spatial interpolation of seawater transparency and sea surface temperature data for each scene is performed using the natural proximity interpolation algorithm; temporal interpolation of missing seawater transparency and sea surface temperature data in the required time series data is performed using the linear interpolation algorithm. In step A3, when performing seawater transparency inversion, the turbidity index T is used as a reference. d Each corresponding formula for calculating seawater transparency is constructed, and finally, seawater transparency image data is calculated, including: (1) Define the turbidity index T of water bodies d Preliminary classification of water bodies: (2) Using the total backscattering coefficient and total absorption coefficient calculated in step A2 as inputs, construct a system for estimating the Z-value of low-turbidity seawater. sd formula: (3) Construct a system based on R rs (748) and R rs The formula for estimating the difference in (869) is: [Formula for the transparency of extremely turbid seawater] (4) Combining (2) and (3), the formula for calculating the transparency of moderately turbid seawater is constructed as follows: In the formula, Z sd,lt The formula for calculating the transparency of low-turbidity seawater, Z sd,et The formula for calculating the transparency of extremely turbid seawater, Z sd,mt The formula for calculating the transparency of moderately turbid seawater, b bw (488) is the backscattering coefficient of pure water at 488 nm, and W is the weight value of the connection formula; Step B: Construct a forecasting model and train it to determine the optimal forecasting model. The structure and principle of the forecasting model are as follows: (1) The forecast model is constructed by combining Convolutional Long Short-Term Memory (ConvLSTM) network with a fully connected layer, including N ConvLSTM layers and a forecast layer based on fully connected computation, where 2 ≤ N ≤ 5; (2) The three-dimensional array of n×n×d, which is composed of n×n pixels for each day of the previous d days, is used as the model input. The spatiotemporal feature information of the data window is learned through forward propagation of N layers of ConvLSTM network units. Then, the value of the n×n data window on day d+1 is calculated through the prediction layer. Finally, the predicted value of the center pixel of the window is output, thus completing one prediction of the value of the center pixel. The entire scene data is traversed by sliding the window, and the above operation is performed on each window to realize the prediction of the seawater transparency and sea surface temperature of the entire scene on day d+1. Step C: Based on the trained forecast model, forecast the seawater transparency and sea surface temperature of a specific sea area.

2. The method for predicting seawater transparency and sea surface temperature based on multi-source remote sensing data according to claim 1, characterized in that: In step B, the formula for calculating the convolution within the ConvLSTM model is as follows: In the formula, t represents day t, * represents the convolution calculation method, i, o, f, h represent the input gate, output gate, forget gate, and transmission state of the model network, respectively, C and H represent the cell state and hidden state, W and b represent the weights and biases of the corresponding data, X represents the input dataset, and σ represents the activation function.

3. The method for predicting seawater transparency and sea surface temperature based on multi-source remote sensing data according to claim 1, characterized in that: In step B, during the training process, the number N of ConvLSTM layers in the model is adjusted from 2 to 5. When the MAE and loss of each model reach the accuracy requirements, the computation speed and accuracy of the models with 2, 3, 4 and 5 ConvLSTM layers are compared. The ConvLSTM layer number and related hyperparameter values ​​of the model with the highest prediction efficiency are saved as the best model. The formulas for calculating the mean absolute error and loss function used to determine the hyperparameters of the ConvLSTM model are as follows: In the formula, MAE is the mean absolute error, loss is the loss function, and x is the mean absolute error. i It is the i-th value predicted by the model, y i It is the i-th true value, and n is the total number of windows in the training samples.

4. The method for predicting seawater transparency and sea surface temperature based on multi-source remote sensing data according to claim 1, characterized in that: Step A4 is implemented in the following manner: (1) The natural proximity interpolation algorithm is used to perform spatial data interpolation and filling of seawater transparency and sea surface temperature image data. The specific formula is as follows: In the formula, f(x) is the interpolated value at the point x to be interpolated, and w i (x) represents the sample points i, i=1,…,n, participating in the interpolation, and their weights on x, f i The value of sample point i; The formula for calculating the weight is as follows: In the formula, a i Let x be the area of ​​the Thiessen polygon where the sample point to be interpolated is located, and a(x) be the area of ​​the Thiessen polygon where the point x to be interpolated is located. (2) A linear interpolation algorithm is used to interpolate and fill in the time series data of seawater transparency and sea surface temperature image data. The specific formula is as follows: In the formula, f(d) is the value after interpolation on day d, f(d1) is the value of day d1, which is the closest day to day d before day d, and f(d2) is the value of day d2, which is the closest day to day d after day d.

5. The method for predicting seawater transparency and sea surface temperature based on multi-source remote sensing data according to claim 1, characterized in that: In step A, the following method is used when creating the training dataset: Using multi-time series seawater transparency or sea surface temperature as the original dataset, a data unit set consisting of n×n×d pixels in a three-dimensional array storage format is constructed as the training sample. That is, while traversing the satellite remote sensing images over consecutive days at the time scale, the data of each day is traversed through an n×n pixel window, ultimately forming a three-dimensional array set consisting of a large number of n×n×d pixels.

Images