CN116070132A - Method for predicting transparency of seawater and sea surface temperature based on multi-source remote sensing data - Google Patents
Method for predicting transparency of seawater and sea surface temperature based on multi-source remote sensing data Download PDFInfo
- Publication number
- CN116070132A CN116070132A CN202211632315.9A CN202211632315A CN116070132A CN 116070132 A CN116070132 A CN 116070132A CN 202211632315 A CN202211632315 A CN 202211632315A CN 116070132 A CN116070132 A CN 116070132A
- Authority
- CN
- China
- Prior art keywords
- data
- transparency
- sea
- surface temperature
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a method for predicting sea water transparency and sea surface temperature based on multi-source remote sensing data, which comprises the steps of firstly collecting enough water color satellite data, carrying out residual correction, inversion of inherent optical quantity of a water body, inversion of sea water transparency and other pretreatment operations on remote sensing reflectivity data, carrying out space-time interpolation on the obtained sea water transparency data and sea surface temperature data, and further manufacturing a training data set; then constructing a forecasting model, and training the forecasting model to determine an optimal forecasting model, wherein the forecasting model is constructed by adopting a mode of combining a convolution long-short-term memory network (ConvLSTM) with a full link; finally, based on the trained forecasting model, forecasting of the transparency of the sea water and the sea surface temperature of the specific sea area is achieved, and accuracy of future sea environment information is guaranteed.
Description
Technical Field
The invention belongs to the field of marine remote sensing technology application, and particularly relates to a method for predicting the transparency and the sea surface temperature of sea water based on multi-source remote sensing data.
Background
Transparency of sea water (Z) sd ) The maximum visible depth (Sechi depth) of a Seck disk immersed in water is usually expressed as a measure of the seawater's water power and is also an important factor affecting the distribution of the underwater light field; sea Surface Temperature (SST), which refers to the temperature of the surface water of the ocean, is an important physical parameter in the knowledge of the body of water of the ocean. The method has very important significance for the prediction and development research of the transparency of the seawater, the marine ecology, the water quality, the marine transportation and the marine culture, and has positive strategic significance for weather research, the marine activities, the aquaculture industry, the fishing industry and the like by mastering the temperature information of the sea surface in the future.
In the past studies of sea water transparency measurement, the past sea water transparency was estimated mainly based on historical data and by constructing an empirical linear model and a semi-analytical algorithm. Kratzer et al (2008) propose a method of estimating the transparency of sea water using the diffusion attenuation coefficient in the vertical direction; prasad et al (2016) found a good correlation between the transparency of seawater and the ratio of the brightness of the two bands of the radiation of the water by analyzing a large amount of water color data, and constructed a linear model for detecting the transparency of the seawater based on the two; doron et al (2011) propose a method of calculating the transparency of seawater using the inherent optical parameters obtained by inversion of the ionizing radiation. In the researches, although the past seawater transparency of a designated sea area can be accurately inverted, only the future development trend of the seawater transparency can be approximately and qualitatively described, and the requirements of the prediction accuracy cannot be met.
For sea level temperature prediction, barnston and Smith (1996) proposed a typical correlation prediction sea level temperature method (canonical correlation analysis), i.e., a set of data is processed according to physical objective facts and a model for predicting sea level temperature is constructed by using the correlation between the two data. The method of combining long-short-term memory neural network model (LSTM) with AdaBoost algorithm is proposed by Xiao et al (1993), so that the moderate-short-term prediction of sea surface temperature is realized, the prediction capability is improved, and the problem of overfitting is avoided. Jia et al (2022) uses the Loess algorithm to perform seasonal weighting treatment on the data, and then uses LSTM to predict sea surface temperature in the eastern sea area of the country. The CFCC-LSTM model proposed by Yang et al (2018) classifies SST predictions as a problem of sequence prediction, and predicts future SST from a spatiotemporal perspective of historical data. In the conventional sea surface temperature prediction research, the method for constructing the prediction model through correlation often has strong dependence on the region, and cannot express the complex relationship between environmental factors influencing the sea surface temperature. In the prediction model constructed by LSTM, the calculation methods of the parameters such as weight in the network structure all adopt a full-link mode of matrix multiplication, the network parameters in the model are too many, the calculation amount is huge, and the space information can not be effectively extracted.
In the prior art, the measurement of the transparency of the sea water and the prediction of the sea surface temperature are mainly realized by constructing a deep learning prediction model by utilizing the existing data, and then predicting the future transparency of the sea water and the sea surface temperature of the predicted point by the model, so that only the point-to-point prediction is realized, and the training data in the point-to-point prediction process often lack the space information around the predicted point, thereby limiting the accuracy of model prediction. The ocean optical remote sensing technology is an emerging detection technology, sea surface water color information can be detected through a detector carried on a remote sensing platform to further obtain the sea information, the rapid development of the water color satellite remote sensing technology enables future sea water transparency and sea surface temperature prediction to be possible by utilizing past water color satellite remote sensing data, and based on the consideration, a prediction method for extracting predicted points and space-time characteristics around the predicted points through a deep learning model based on the water color remote sensing data is needed to be provided, so that the prediction method for the sea water transparency and the sea surface temperature from surface to point is realized.
Disclosure of Invention
The invention aims to provide a method for predicting the transparency and the sea surface temperature of sea water based on multi-source remote sensing data, which utilizes sea remote sensing historical images to forecast the transparency and the sea surface temperature of future sea water and provides more continuous, accurate and advanced sea environment information for sea activity and sea disaster prevention.
The invention is realized by adopting the following technical scheme: a method for predicting the transparency of seawater and the temperature of the sea surface based on multi-source remote sensing data comprises the following steps:
step A, constructing a training data set: collecting water color satellite data, wherein the water color satellite data comprise remote sensing reflectivity data and sea surface temperature data of each wave band for many years, carrying out residual correction, inversion of inherent optical quantity of a water body and inversion operation of sea water transparency on the remote sensing reflectivity data, and then carrying out space-time interpolation on the obtained sea water transparency data and sea surface temperature data to further manufacture a training data set;
and B, constructing a forecasting model and training the forecasting model to determine an optimal forecasting model, wherein the forecasting model has the following structure and principle:
(1) The prediction model is constructed by combining a convolution long-short-term memory network (ConvLSTM) with a full link, and comprises N layers of ConvLSTM networks and a prediction layer based on full-link calculation, wherein N is an integer from 2 to 5;
(2) Inputting a three-dimensional array which is formed by data windows based on N x N pixels and is formed by N x d together every day in the previous d days as a model, acquiring space-time characteristic information of the data windows through forward propagation learning of N layers of ConvLSTM network units, calculating the d+1th day value of the N x N data windows through a forecasting layer, and finally outputting the value of a central pixel of the window, thereby completing one-time prediction of the value of a central pixel; and traversing the whole scene data in a window sliding mode, and executing the operation on each window to predict the sea water transparency and sea surface temperature of the whole scene on day d+1.
And C, forecasting the transparency of the sea water and the sea surface temperature of the specific sea area based on the trained forecasting model.
Further, in the step B: the ConvLSTM model internal convolution calculation formula is as follows:
in the formula, t represents the t day, i, o, f and h represent the convolution calculation mode, i, o, f and h represent the input gate, the output gate, the forgetting gate and the transmission state of the model network respectively, C, H represents the cell state and the hidden state, W, b represents the weight and the bias of corresponding data respectively, X represents the input data set, and sigma represents the activation function.
Further, in the step B, during the training, the number of ConvLSTM layers in the model is adjusted from 2 to 5, when the MAE and loss of each model reach the accuracy requirement, the rate and accuracy of the model calculated under the Convlstm states with 2, 3, 4 and 5 layers are compared, and the ConvLSTM layer number and the relevant super parameter value of the model with highest prediction efficiency are saved as the optimal model;
the calculation formula for determining the average absolute error and the loss function used by the ConvLSTM model super-parameters is as follows:
wherein MAE is the mean absolute error, loss is the loss function, x i Is the ith value, y, of the model prediction i Is the i-th true value and n is the total number of windows in the training sample.
Further, the step a is specifically implemented by the following manner:
step A1, residual error correction: calculating residual errors in the remote sensing reflectivity data based on an inherent optical characteristic data processing system IDAS;
step A2, inversion of inherent optical quantity of water: inputting the remote sensing reflectivity data subjected to residual correction in the step A1 into an inherent optical quantity algorithm NNSAA, and obtaining a water body backscattering coefficient and an absorption coefficient of each wave band, thereby obtaining a total backscattering coefficient and a total absorption coefficient;
step A3, inverting the transparency of the seawater: based on the remote sensing reflectivity data after residual correction in the step A1 and the total backscattering coefficient and the total absorption coefficient obtained in the step A2, calculating the transparency Z of the seawater according to a transparency inversion algorithm sd ；
Step A4, data space-time interpolation: performing spatial interpolation on the transparency data of each scene of sea water and the sea surface temperature data by using a natural clinical interpolation algorithm to improve the spatial coverage rate of the data; and performing time sequence interpolation on the missing seawater transparency data and the seatable temperature data in the required time sequence data by using a linear interpolation algorithm so as to ensure that the data exist every day in the same sea area and ensure the integrity of the sequence data.
Further, in the step A3, when the sea water transparency inversion is performed, the sea water transparency inversion is performed according to the water turbidity index T d Respectively constructing corresponding seawater transparency calculation formulas, and finally calculating to obtain seawater transparency image data, wherein the method specifically comprises the following steps:
(1) Definition of the turbidity index T of a Water body d Preliminary classification is carried out on the water body:
T d ＝1.8386R rs (667)-R rs (488)
(2) Taking the total backscattering coefficient and the total absorption coefficient calculated in the step A2 as input to construct a method for estimating the low-turbidity seawater Z sd The formula:
(3) Construction based on R rs (748) And R is R rs (869) The formula for estimating the transparency of extremely turbid seawater:
Z st,et ＝0.0036[R rs (748)-R rs (869)] -0.840 ,T d ＞0.014
(4) Combining (2) and (3), and constructing a middle turbid seawater transparency calculation formula:
Z sd,mt ＝W·Z sd,tc +(1-W)Z sd,et ,0.01≤T d ≤0.014 (10)
wherein Z is sd,lt Z is calculated according to the transparency of the low-turbidity sea water sd,et Z is calculated according to the transparency of extremely turbid seawater sd,mt B, calculating a formula for transparency of middle cloudy seawater bw (488) The backscattering coefficient of pure water at 488nm is given by W, which is the weight value of the connection formula.
Further, the step A4 is specifically implemented in the following manner:
(1) The natural clinical interpolation algorithm is adopted to carry out spatial data interpolation filling on the sea water transparency and sea surface temperature image data, and the specific formula is as follows:
wherein f (x) is the value after interpolation at the point x to be interpolated, w i (x) For sample points i, i=1, …, n, weight to x, f that participate in interpolation i A value of sample point i;
the weight calculation formula is as follows:
wherein a is i For the Thiessen multiple area where the sample points involved in interpolation are located, a (x) is the area of the Thiessen polygon where the point to be interpolated x is located.
(2) The linear interpolation algorithm is adopted to interpolate and fill the time series data of the sea water transparency and sea surface temperature image data, and the specific formula is as follows:
wherein f (d) is a value after interpolation on the day d to be interpolated, f (d) 1 ) Is the d day before and nearest to the d day 1 Of (d), f (d) 2 ) Is the d following and nearest to the d day 2 Is a value of (2).
Further, in the step a, when the training data set is manufactured, the following manner is adopted:
the method comprises the steps of taking multi-time-sequence seawater transparency or sea surface temperature as an original data set, constructing a data unit set which is composed of n multiplied by d pixels and takes a three-dimensional array as a storage format as a training sample, namely traversing a satellite remote sensing image in continuous d days in a time scale manner, traversing an n multiplied by n pixel window on data of each day, and finally forming a three-dimensional array set which is composed of a large number of n multiplied by d pixels.
Compared with the prior art, the invention has the advantages and positive effects that:
the scheme predicts the transparency of the sea water and the sea surface temperature by constructing a prediction model, wherein the prediction model is constructed by combining a convolution long-short-term memory network (ConvLSTM) and full links, and comprises N ConvLSTM layers and a prediction layer. In the implementation process of the scheme, the spatial characteristics of data are reserved in a mode of constructing an n multiplied by n data window taking a pixel to be predicted as a center, and meanwhile, a three-dimensional array is constructed in a data window in continuous d days, so that the time sequence characteristics of the data are reserved, and finally, the prediction of the transparency of sea water and the sea surface temperature from surface to point is realized. And taking the three-dimensional array as the input of a model, extracting the time and space characteristic information of a data window through forward propagation learning of N layers of ConvLSTM network units in the model, and calculating the sea water transparency and sea surface temperature predicted value of the pixel to be predicted on day d+1 in a full-connection calculation mode of a prediction layer. And then through carrying out window sliding traversal on the data in the appointed sea area, the operation is carried out on the data respectively, so that the sea water transparency and sea surface temperature of the d+1th day of the whole sea area can be predicted, and the accuracy of future sea environment information is ensured.
Drawings
FIG. 1 is a general flow chart of sea water transparency and sea surface temperature forecast in accordance with an embodiment of the present invention;
FIG. 2 is a flowchart of IDAS calculation of remote sensing reflectivity residual error according to an embodiment of the present invention;
FIG. 3 is a flow chart of NNSAA calculation of inherent optical quantity of water in an embodiment of the invention;
FIG. 4 is a diagram of a ConvLSTM neural network for predicting sea water transparency and sea surface temperature according to an embodiment of the present invention;
FIG. 5 is a schematic diagram showing the variation of MAE and loss in the sea level temperature prediction model training process according to the embodiment of the present invention;
FIG. 6 is a schematic diagram showing the variation of MAE and loss in the seawater transparency prediction model training process according to an embodiment of the present invention;
FIG. 7 is a graph showing the spatial distribution of predicted values and real values of sea surface temperatures according to an embodiment of the present invention;
FIG. 8 is a graph showing the spatial distribution of predicted and actual values of the transparency of sea water according to an embodiment of the present invention;
FIG. 9 is a graph showing the mean square error (MRE) analysis result of the comparison between the sea surface temperature and the sea transparency predicted by the embodiment of the invention and the actual observation real data;
FIG. 10 is a diagram showing the comparison of real time series and predicted time series of sea surface temperatures according to an embodiment of the present invention;
FIG. 11 is a diagram showing the comparison of a real time series and a predicted time series of the transparency of seawater according to an embodiment of the present invention.
Detailed Description
In order that the above objects, features and advantages of the invention will be more readily understood, a further description of the invention will be rendered by reference to specific embodiments thereof which are illustrated in the appended drawings. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention, however, the present invention may be practiced otherwise than as described herein, and therefore the present invention is not limited to the specific embodiments disclosed below.
The embodiment provides a method for forecasting the transparency of sea water and the sea surface temperature based on multi-source remote sensing data and calling a deep learning model, which is used for forecasting the value of the d+1st day of a pixel in the center of a window based on a three-dimensional array formed by window data of continuous d days, so that the prediction of the transparency of sea water and the sea surface temperature from surface to point is realized. In order to improve the quality of data used by the method and ensure the precision of a prediction result, preprocessing operations such as residual correction, inversion of inherent optical quantity of a water body, inversion of seawater transparency, data interpolation and the like are required to be carried out on the collected remote sensing reflectivity data, and the data interpolation operation is carried out on the collected sea surface temperature data, and specifically comprises the following steps:
step A, constructing a training data set: collecting sufficient water color satellite data, wherein the water color satellite data comprises multi-band remote sensing reflectivity data and sea surface temperature data for many years, carrying out preprocessing operations such as residual correction, inversion of inherent optical quantity of a water body, inversion of sea water transparency and the like on the remote sensing reflectivity data, carrying out interpolation of time sequence and space data on the obtained sea water transparency data and sea surface temperature data, and further manufacturing a model training data set;
and B, constructing a forecasting model and training the forecasting model to determine the optimal structure and relevant super parameters of the forecasting model, wherein the structure and principle of the forecasting model are as follows:
the model takes a three-dimensional array consisting of the first 30 days of seawater transparency or sea surface temperature data of a data window with the size of 15 multiplied by 15 as input, firstly, the time sequence characteristic and the space characteristic information of the data window are extracted through forward propagation learning of N layers of ConvLSTM network units, then, the seawater transparency and the sea surface temperature value of the 31 st day of the data window are calculated through the forecasting layer, finally, the value of the central pixel of the data window is output, and finally, the prediction of the seawater transparency and the sea surface temperature of the 31 st day of the pixel is completed. By executing the above process on all the pixels of the whole sea area, the transparency of the sea water and the sea surface temperature on the 31 st day of the whole sea area can be predicted.
In the model training process, the sea water transparency and sea surface temperature data of the window on the 1 st day are abandoned, the sea water transparency and sea surface temperature data of the window on the 31 st day are used as the 30 th data in the newly input three-dimensional array, the sea water transparency and sea surface temperature of the window on the 32 nd day can be predicted by cycling the process once, and further the sea water transparency and sea surface temperature of the 32 nd day predicted by the whole sea area model are obtained.
In the process of determining the optimal structure and related super parameters of the model, the number of ConvLSTM layers in the model is adjusted to be from 2 to 5 when the model is trained, and when the MAE and loss of each model reach the accuracy requirement, the model with highest forecasting efficiency is selected to be optimal by comparing the speed and accuracy of the model respectively predicted under the ConvLSTM states with 2, 3, 4 and 5 layers.
And C, forecasting the transparency of the sea water and the sea surface temperature of the specific sea area based on the trained forecasting model.
In order to more clearly understand the scheme of the invention, the invention is described in detail below by combining a prediction method of sea water transparency and sea surface temperature of a certain sea area, wherein a prediction flow is shown in a figure 1, and specifically comprises the following steps:
1. building a training data set:
(1) Firstly, according to the relation between the remote sensing reflectivity actually received by the known satellite and the remote sensing reflectivity and residual error theoretically received, a formula (1) is obtained.
R rs,t (λ)＝R rs,s (λ)-ΔR rs Iλ)(1)
Wherein R is rs,s (lambda) is lambda-band remote sensing reflectivity actually received by a water color satellite, R rs,t (lambda) is the lambda-band remote sensing reflectivity, delta R, theoretically received by the water-based satellite rs And (lambda) is the residual error contained in the actually received lambda-band remote sensing reflectivity.
(2) Then, R is rs,s (lambda) input neural network-based intrinsic optical quantity calculation model (NQAA) to obtain a (555) and b b (670) The method comprises the steps of carrying out a first treatment on the surface of the At this time, considering that the ocean body lacks suspended particles and is very clean, the total absorption coefficient at 670nm of the ocean body is defined as being equal to the pure water absorption coefficient, i.e., a (670) ≡a w (670)。
(3) Next, ΔR is determined based on equation (1) rs (555) And DeltaR rs (670) Linear empirical relationship, equation (2):
ΔR rs (555)＝m 0 ΔR rs (670)+m 1 (2)
wherein m is 0 And m 1 Is a linear relation empirical constant of remote sensing reflectivity residuals in 555nm and 670nm wave bands.
So far, the remote sensing reflectivity of 670nm is applied to the formula (1), and the remote sensing reflectivity actually received by the satellite and the remote sensing reflectivity residual error can be used for representing the remote sensing reflectivity theoretically received by the satellite.
(4) Combining a (555) and b obtained in the previous steps b (670)、ΔR rs (555) Absorption coefficient of pure water (a) w (670) By substituting 0.439 into formula (3), the product containing b alone can be obtained b (555) And DeltaR rs (670) Equation of two unknownsGroup, simultaneous solution to obtain ΔR rs (670)：
In equation (3), the left hand side of the equation is based on the reflectance (r) at the sea interface rs ) And inherent optical quantity (a and b b ) The radiation transmission relation between the two, namely formula (4); to the right of the equation is based on the reflectance (r rs ) The relationship between the reflection rate of the sea-air interface received by the satellite and the reflection rate of the sea-air interface received by the satellite is obtained, namely, a formula (5); g in the formula 0 And g 1 0.089sr respectively -1 And 0.1245sr -1 Is a fixed empirical constant of (a).
To this end, a remote sensing reflectivity residual ΔR at 670nm has been obtained rs (670)。
(5) Finally, according to the spectrum relation of the data residual errors, namely a formula (6), the remote sensing reflectivity residual error of each wave band is calculated:
where S represents the spectral slope coefficient of the residual, C is a constant, this embodiment by applying a constant to ΔR rs The data set was counted and set to C0.04.
Step 4, inverting the transparency of the seawater: r based on residual error correction rs (lambda) data and the total back-scattering coefficient data and total absorption coefficient data calculated in step 3, and calculating the transparency (Z) of seawater according to a transparency inversion algorithm sd )；
Considering the situation that the algorithm used for calculating the inherent optical quantity in the step 3 has better effect in medium-low turbid water bodies and is not ideal in extremely turbid water bodies with complex optical properties. This example first defines the water turbidity index (T d ) Preliminary classification of the water body is carried out as shown in formula (7), and then according to T d And respectively constructing corresponding seawater transparency calculation formulas, and finally calculating to obtain seawater transparency image data.
In this step, a (λ) and b calculated in step 3 b (lambda) as input, a method for estimating the low-turbidity sea water Z is constructed sd The algorithm of (2) is shown in formula (8). R-based construction is performed by utilizing the characteristic that the sensitivity of extremely turbid water body to water body optical activity component change in near infrared band is higher than that of blue-green light band rs (748) And R is R rs (869) The algorithm for the specific estimation of band differences for extremely cloudy seawater transparency is shown in equation (9). And the transparency calculation formula of the turbid seawater in the construction of the two algorithms is shown as a formula (10).
T d ＝1.8386R rs (667)-R rs (488) (7)
Z st,et ＝0.0036[R rs (748)-R rs (869)] -0.840 ,T d ＞0.014 (9)
Z sd,mt ＝W·Z sd,tc +(1-W)Z sd,et ,0.01≤T d ≤0.014 (10)
Wherein Z is sd,lt Z is calculated according to the transparency of the low-turbidity sea water sd,et Z is calculated according to the transparency of extremely turbid seawater sd,mt B, calculating a formula for transparency of middle cloudy seawater bw (488) Is pureThe backscattering coefficient of water at 488nm, W is the weight of the connectivity formula.
(1) In order to avoid errors caused by the lack of spatial data to the subsequent model training, the embodiment adopts a natural clinical interpolation algorithm to carry out Z sd And performing spatial data interpolation filling on the SST image data, wherein the specific formula is as follows:
wherein f (x) is the value after interpolation at the point x to be interpolated, w i (x) For weighting x for sample points i (i=a, …, n) participating in interpolation, f i Is the value of sample point i.
The weight calculation formula is as follows:
wherein a is i For the Thiessen multiple area where the sample points involved in interpolation are located, a (x) is the area of the Thiessen polygon where the point to be interpolated x is located.
(2) In order to ensure the integrity of the time series data, the embodiment adopts a linear interpolation algorithm to interpolate and fill the time scale of the seawater transparency and the sea surface temperature time series data, and the specific formula is as follows:
wherein f (x) is the value after interpolation on the x-th day to be interpolated, f (x) 1 ) Is the x th day before and nearest to the x th day 1 Values of the days, f (x 2 ) Is the x th after and nearest to the x th day 2 Day value.
2. Constructing a forecasting model and calling the forecasting model:
step 7, model construction and training: convLSTM structure was designed using the python language and model training was performed. By comparing the calculation efficiency and the precision under the models containing different ConvLSTM layers, ensuring that each model is trained until the average absolute error (MAE) and loss function (loss) between the predicted value and the true value reach the precision requirement, and saving the prediction model structure with the optimal calculation efficiency and precision and the related super parameters as the optimal values to be used as the specified sea Z for the subsequent prediction sd Or model invocation of SST.
In this embodiment, when MAE and loss tend to stabilize in the training process, the super-parameters in this training are determined as the optimal model parameters, as shown in fig. 4 and 5. In this example, Z for the first 30 days of a data window consisting of 15×15 pixels sd And SST, and outputting Z of the central pixel of the 31 st day of the window after calculation by N ConvLSTM layers and 1 forecast layer sd And SST.
The invention adopts a mode of combining a convolution long-short-term memory network (ConvLSTM) and full links to construct a prediction model, and the network of the model consists of N ConvLSTM layers and 1 prediction layer. The model takes the sea water transparency and sea surface temperature data of the first 30 days of a 15 multiplied by 15 data window as input, extracts space-time characteristic information of the data window through forward propagation learning of N layers of ConvLSTM network units, calculates sea water transparency and sea surface temperature values of the 31 st day of the window through a forecasting layer, and finally outputs the sea water transparency and sea surface temperature values of a central pixel of the window as a prediction result, so that one-time prediction of sea water transparency and sea surface temperature in a single pixel is completed. By traversing the above process for all pixels, the transparency of sea water and the sea surface temperature of the 31 st day of the whole sea area can be predicted. After the whole sea area is predicted once, the first day data of the input three-dimensional array is abandoned, the 31 st day value is used as the 30 th day data of the newly input three-dimensional data to form the newly input three-dimensional array, the multi-time sequence seawater transparency and the sea surface temperature can be predicted by cycling the above processes for a plurality of times, and the specific structure of the invention is shown in figure 3.
Wherein, convLSTM internal convolution calculation formulas are shown as formula (13) to formula (16):
where t represents the t-th day, i, o, f, h represent convolution calculations, i, o, f, h represent input, output, forget, and transmission states of the model network, C, H represent cell states and hidden states, W, b represent weights and biases of corresponding data, X represents the input dataset, and σ represents an activation function.
In order to find out the structure and relevant super parameters of the optimal prediction model, the embodiment adjusts the number of ConvLSTM layers in the model from 2 to 5 when training the model, compares the speed and the precision of the model when the MAE and loss of each model reach the precision requirement under the ConvLSTM states with 2, 3, 4 and 5 layers, and designates the ConvLSTM layers and relevant super parameters of the optimal model under the condition that the model prediction efficiency is highest. In the embodiment, the optimal ConvLSTM model for forecasting the transparency of the sea water and the sea surface temperature is determined by training the model and comprises 3 ConvLSTM layers. The calculation formulas of MAE and loss used for determining the optimal model structure and related super parameters are as follows:
wherein x is i Is the model predictive ith value, y i Is the i-th true value and n is the total number of windows in the training sample.
the embodiment calls the trained forecast model parameters and is applied to the south China sea SST and Z sd Is used for the prediction of (1). By comparing SST and Z before and after prediction sd Spatial distribution of (FIGS. 7, 8), SST and Z sd Mean square error (MRE) of true and predicted values (FIG. 9), and multi-temporal SST and Z sd Trends in predicted and actual values (fig. 10, 11), it can be seen that the accuracy of the model meets the requirements.
Wherein the calculation formula of the mean square error MRE is as follows:
wherein x is i Is the value of the model predictive day i, y i Is the true value on day i, n is the amount of data for comparison.
The present invention is not limited to the above-mentioned embodiments, and any equivalent embodiments which can be changed or modified by the technical content disclosed above can be applied to other fields, but any simple modification, equivalent changes and modification made to the above-mentioned embodiments according to the technical substance of the present invention without departing from the technical content of the present invention still belong to the protection scope of the technical solution of the present invention.
Claims (7)
1. The method for predicting the transparency of the seawater and the sea surface temperature based on the multi-source remote sensing data is characterized by comprising the following steps of:
step A, constructing a training data set: collecting water color satellite data, wherein the water color satellite data comprise remote sensing reflectivity data and sea surface temperature data of each wave band for many years, carrying out residual correction, inversion of inherent optical quantity of a water body and inversion operation of sea water transparency on the remote sensing reflectivity data, and then carrying out space-time interpolation on the obtained sea water transparency data and sea surface temperature data to manufacture a training data set;
and B, constructing a forecasting model and training the forecasting model to determine an optimal forecasting model, wherein the forecasting model has the following structure and principle:
(1) The prediction model is constructed by combining a convolution long-short-term memory network ConvLSTM with a full link and comprises N ConvLSTM layers and a prediction layer based on full link calculation, wherein N is more than or equal to 2 and less than or equal to 5;
(2) Inputting a three-dimensional array of N multiplied by d, which is formed by data windows formed by N multiplied by N pixels every day in the previous d days, as a model, acquiring space-time characteristic information of the data windows through forward propagation learning of N layers of ConvLSTM network units, calculating the d+1th day value of the N multiplied by N data windows through a forecasting layer, and finally outputting a forecasting value of a central pixel of the window, thereby completing primary forecasting of a central pixel value; traversing the whole scene data in a window sliding mode, and executing the operation on each window to realize the prediction of the sea water transparency and sea surface temperature of the whole scene on day d+1.
And C, forecasting the transparency of the sea water and the sea surface temperature of the specific sea area based on the trained forecasting model.
2. The method for predicting sea water transparency and sea surface temperature based on multi-source remote sensing data of claim 1, wherein: in the step B: the ConvLSTM model internal convolution calculation formula is as follows:
in the formula, t represents the t day, i, o, f and h represent the convolution calculation mode, i, o, f and h represent the input gate, the output gate, the forgetting gate and the transmission state of the model network respectively, C, H represents the cell state and the hidden state, W, b represents the weight and the bias of corresponding data respectively, X represents the input data set, and sigma represents the activation function.
3. The method for predicting sea water transparency and sea surface temperature based on multi-source remote sensing data of claim 1, wherein: in the step B, in the training process, the number N of ConvLSTM layers in the model is regulated from 2 to 5, when the MAE and loss of each model reach the precision requirement, the calculation speed and the precision of the model in the ConvLSTM states with 2, 3, 4 and 5 layers are compared, and the ConvLSTM layers and relevant super-parameter values of the model with highest forecasting efficiency are stored to be used as the optimal model;
the calculation formula for determining the average absolute error and the loss function used by the ConvLSTM model super-parameters is as follows:
wherein MAE average absolute error, loss is a loss function, x i Is the ith value, y, of the model prediction i Is the i-th true value and n is the total number of windows in the training sample.
4. The method for predicting sea water transparency and sea surface temperature based on multi-source remote sensing data of claim 1, wherein: the step A is specifically realized by the following steps:
step A1, residual error correction: calculating residual errors in the remote sensing reflectivity data based on an inherent optical characteristic data processing system IDAS;
step A2, inversion of inherent optical quantity of water: inputting the remote sensing reflectivity data subjected to residual correction in the step A1 into an inherent optical quantity algorithm NNSAA, and obtaining a water body backscattering coefficient and an absorption coefficient of each wave band, thereby obtaining a total backscattering coefficient and a total absorption coefficient;
step A3, inverting the transparency of the seawater: based on the remote sensing reflectivity data after residual correction in the step A1 and the total backscattering coefficient and the total absorption coefficient obtained in the step A2, calculating the transparency Z of the seawater according to a transparency inversion algorithm sd ；
Step A4, data space-time interpolation: performing spatial interpolation on the transparency data of each scene of sea water and the sea surface temperature data by a natural clinical interpolation algorithm; and performing time sequence interpolation on the missing seawater transparency data and the missing sea surface temperature data in the required time sequence data by using a linear interpolation algorithm.
5. The method for predicting sea water transparency and sea surface temperature based on multi-source remote sensing data of claim 4, wherein: in the step A3, when the sea water transparency inversion is carried out, the sea water transparency inversion is carried out according to the water turbidity index T d Respectively constructing corresponding seawater transparency calculation formulas, and finally calculating to obtain seawater transparency image data, wherein the method specifically comprises the following steps:
(1) Definition of the turbidity index T of a Water body d Preliminary classification is carried out on the water body:
T d ＝1.8386R rs 667-R rs 488
(2) Taking the total backscattering coefficient and the total absorption coefficient calculated in the step A2 as input to construct a method for estimating the low-turbidity seawater Z sd The formula:
(3) Construction based on R rs (748) And R is R rs (869) The formula for estimating the transparency of extremely turbid seawater:
Z st,et ＝0.0036R rs 748-R rs 869 -0.840 ,T d ＞0.014
(4) Combining (2) and (3), and constructing a middle turbid seawater transparency calculation formula:
Z sd,mt ＝W·Z sd,tc +1-WZ sd,et ,0.01≤T d ≤0.014(10)
wherein Z is sd,lt Z is calculated according to the transparency of the low-turbidity sea water sd,et Z is calculated according to the transparency of extremely turbid seawater sd,mt B, calculating a formula for transparency of middle cloudy seawater bw (488) The backscattering coefficient of pure water at 488nm is given by W, which is the weight value of the connection formula.
6. The method for predicting sea water transparency and sea surface temperature based on multi-source remote sensing data of claim 4, wherein: the step A4 is specifically realized by adopting the following modes:
(1) The natural clinical interpolation algorithm is adopted to carry out spatial data interpolation filling on the sea water transparency and sea surface temperature image data, and the specific formula is as follows:
wherein f (x) is the value after interpolation at the point x to be interpolated, w i (x) For sample points i, i=1, …, n, weight to x, f that participate in interpolation i A value of sample point i;
the weight calculation formula is as follows:
wherein a is i For the Thiessen multiple area where the sample points involved in interpolation are located, a (x) is the area of the Thiessen polygon where the point to be interpolated x is located.
(2) The linear interpolation algorithm is adopted to interpolate and fill the time series data of the sea water transparency and sea surface temperature image data, and the specific formula is as follows:
wherein f (d) is a value after interpolation on the day d to be interpolated, f (d) 1 ) To be before day dAnd the d th most adjacent to the d th day 1 Of (d), f (d) 2 ) Is the d following and nearest to the d day 2 Is a value of (2).
7. The method for predicting sea water transparency and sea surface temperature based on multi-source remote sensing data of claim 1, wherein: in the step a, when the training data set is manufactured, the following manner is adopted:
the method comprises the steps of taking multi-time-sequence seawater transparency or sea surface temperature as an original data set, constructing a data unit set which is composed of n multiplied by d pixels and takes a three-dimensional array as a storage format as a training sample, namely traversing a satellite remote sensing image in continuous d days in a time scale manner, traversing an n multiplied by n pixel window on data of each day, and finally forming a three-dimensional array set which is composed of a large number of n multiplied by d pixels.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211632315.9A CN116070132A (en) | 2022-12-19 | 2022-12-19 | Method for predicting transparency of seawater and sea surface temperature based on multi-source remote sensing data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211632315.9A CN116070132A (en) | 2022-12-19 | 2022-12-19 | Method for predicting transparency of seawater and sea surface temperature based on multi-source remote sensing data |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116070132A true CN116070132A (en) | 2023-05-05 |
Family
ID=86176041
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211632315.9A Pending CN116070132A (en) | 2022-12-19 | 2022-12-19 | Method for predicting transparency of seawater and sea surface temperature based on multi-source remote sensing data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116070132A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117115621A (en) * | 2023-10-24 | 2023-11-24 | 中国海洋大学 | Satellite cloud image prediction method based on improved U-Net network |
-
2022
- 2022-12-19 CN CN202211632315.9A patent/CN116070132A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117115621A (en) * | 2023-10-24 | 2023-11-24 | 中国海洋大学 | Satellite cloud image prediction method based on improved U-Net network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109740485B (en) | Reservoir or small reservoir identification method based on spectral analysis and deep convolutional neural network | |
| CN110749568A (en) | MODIS remote sensing inversion method for phytoplankton absorption characteristics of high-turbidity eutrophic lakes | |
| CN111598156A (en) | PM based on multi-source heterogeneous data fusion2.5Prediction model | |
| CN112666328B (en) | Chlorophyll a remote sensing inversion method and device based on genetic neural network model | |
| CN113408644B (en) | Satellite data reconstruction method and method for detecting response of upper ocean to typhoon | |
| CN114707688A (en) | Photovoltaic power ultra-short-term prediction method based on satellite cloud chart and space-time neural network | |
| CN116070132A (en) | Method for predicting transparency of seawater and sea surface temperature based on multi-source remote sensing data | |
| CN116187203A (en) | Watershed water quality prediction method, system, electronic equipment and storage medium | |
| CN115933010A (en) | Radar echo extrapolation near weather prediction method | |
| Matsui et al. | Improving the resolution of UAV-based remote sensing data of water quality of Lake Hachiroko, Japan by neural networks | |
| CN116403103A (en) | Remote sensing image analysis and cyanobacteria bloom prediction method based on four-dimensional generation countermeasure network | |
| CN114813651A (en) | Remote sensing water quality inversion method combining difference learning rate and spectrum geometric characteristics | |
| CN116879192B (en) | Water bloom prediction method, device, equipment and medium based on satellite remote sensing data | |
| Pozdnyakov et al. | An advanced algorithm for operational retrieval of water quality from satellite data in the visible | |
| CN114186483B (en) | Inversion method fusing buoy data and ocean satellite remote sensing image | |
| Liu et al. | Evaluation of the effectiveness of multiple machine learning methods in remote sensing quantitative retrieval of suspended matter concentrations: A case study of Nansi Lake in North China | |
| Wang et al. | Strawberry ripeness classification method in facility environment based on red color ratio of fruit rind | |
| CN112528869B (en) | Phase-free data imaging method based on complex neural network | |
| CN111723479B (en) | Process-oriented eutrophic lake algae total quantity remote sensing estimation method | |
| CN112085779A (en) | Wave parameter estimation method and device | |
| Zhang et al. | An optical mechanism-based deep learning approach for deriving water trophic state of China's lakes from Landsat images | |
| CN117239744B (en) | Ultra-short-term photovoltaic power prediction method integrating wind cloud No. 4 meteorological satellite data | |
| Kempeneers et al. | Retrieval of oceanic constituents from ocean color using simulated annealing | |
| CN118115824A (en) | Water quality variable concentration prediction method and system | |
| CN115659841A (en) | Atmospheric degradable water yield inversion method and device, electronic equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |