Wetland water body information acquisition method based on deep fusion of hydrodynamic model and remote sensing image

Abstract

The invention discloses a wetland water body information acquisition method based on deep fusion of a hydrodynamic model and a remote sensing image, and the method comprises the steps: collecting remote sensing data, topographic data, meteorological data and hydrological data of an investigation region, constructing a hydrodynamic model based on the above data, and obtaining a daily water body space distribution pattern; calculating normalized water body index raster data from the collected remote sensing data, extracting an accurate water surface spatial pattern, and then, taking a daily water body spatial distribution pattern obtained by the hydrodynamic model as a feature and the accurate water surface spatial pattern extracted by the remote sensing data as an output; and performing space-time fitting regression on the water surface space patterns of the two kinds of data by using a convolution-long short-term memory neural network, and further predicting an accurate daily water body space distribution pattern through daily water body space distribution obtained by a hydrodynamic model based on a regression relationship. According to the method, data of the hydrodynamic model and the remote sensing image are fused through a deep learning method, and accurate and high-frequency wetland water body data can be obtained.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing technology - hydrological science, and particularly to a method for obtaining wetland water body information by deeply integrating a hydrodynamic model and remote sensing images. Background Art

[0002] As one of the important ecosystems on the earth, the water body information of wetlands is of crucial significance for the research, protection of wetland ecosystems, and rational utilization of water resources. Information such as the area, water level, and water volume of wetland water bodies not only directly affects the habitat, reproduction, and migration of organisms in the wetland, but also is closely connected with ecological processes such as the material cycle and energy flow in the wetland. At the same time, it also plays an important role in functions such as climate regulation and flood storage in the surrounding areas. The hydrodynamic model has unique advantages in obtaining wetland water body information. It can simulate the movement and change of water bodies based on physical laws and mathematical equations, thereby obtaining dynamic information of wetland water bodies, and to a certain extent, it can achieve high-frequency simulation and prediction, which helps to deeply understand the short-term change trend of wetland water bodies. However, in practical applications, due to the strong interference of human activities, such as artificially regulating the water level to meet the needs of agricultural irrigation, hydropower development, or flood control, the change law of wetland water bodies deviates from the natural state. At the same time, the complex and diverse topography and landforms of wetlands, such as crisscrossing river branches and numerous pits and depressions, further increase the difficulty of accurately depicting the change process of wetland water bodies by the hydrodynamic model, resulting in a large deviation between its prediction results and the actual situation.

[0003] Remote sensing technology provides another important way to obtain wetland water body information. It can observe wetland water bodies from a macroscopic perspective. Through specific spectral characteristics and image processing algorithms, it can more accurately depict information such as the spatial distribution pattern of wetland water bodies and area changes to a certain extent. However, due to the limitations of its own characteristics, the shooting frequency of satellite or aerial platforms is relatively fixed, which is difficult to meet the needs of high-frequency monitoring of wetland water bodies. Moreover, cloud cover is a very difficult problem in the process of obtaining remote sensing data. The presence of clouds will cause the lack of remote sensing images in some periods, seriously hindering the acquisition of continuous and high-frequency wetland water body data, thus affecting the complete and accurate analysis of the dynamic change process of wetland water bodies. To sum up, the existing hydrodynamic models and remote sensing technologies both have their own limitations in obtaining wetland water body information. There is an urgent need for a new method to integrate the advantages of the two and overcome their deficiencies to achieve accurate and high-frequency acquisition of wetland water body information. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a method for obtaining wetland water body information by deeply integrating a hydrodynamic model and remote sensing images, and the method includes:

[0005] Step S1: Obtain remote sensing image data, hydrological data, topographic data, and meteorological data through the investigation of the research area;

[0006] Step S2: Construct a hydrodynamic model based on the hydrological data, topographic data, and meteorological data, extract the daily water depth data of all grids in the hydrodynamic model, and then obtain the daily water body spatial distribution pattern;

[0007] Step S3: Calculate the water body normalized index according to the remote sensing image data, and extract the daily water body spatial distribution pattern within a specific time period through the threshold division method;

[0008] Step S4: Construct a convolutional neural network, standardize the remote sensing image data and hydrodynamic model data corresponding to the time points within a specific time period, and obtain the encoded time features within a specific time period. Combine the standardized hydrodynamic model data and the encoded features to obtain the synthetic data used as the input of the convolutional neural network. Combine the standardized hydrodynamic model data and the encoded features to obtain the label used as the output of the convolutional neural network. Input the synthetic data into the convolutional neural network. The convolutional neural network outputs a probability map, and the value of each pixel in the probability map represents the probability that the pixel is a water body. Construct an MSE loss function and optimize the convolutional neural network by minimizing the MSE loss function; Obtain the input and corresponding output of the optimal convolutional neural network;

[0009] Step S5: Construct a long short-term memory neural network, stack the dimensions of the input and corresponding output of the optimal convolutional neural network, use the stacked input of the optimal convolutional neural network as the input of the long short-term memory neural network, and use the stacked output of the optimal convolutional neural network as the output of the long short-term memory neural network. Construct an MSE loss function and optimize the long short-term memory neural network by minimizing the MSE loss function; Obtain the optimal long short-term memory neural network as the spatio-temporal relationship model;

[0010] Step S6: Use the daily water body spatial distribution pattern and the corresponding timestamp information obtained in real time by the hydrodynamic model as independent variables, and input them into the spatio-temporal relationship model constructed in Step S5 to obtain the accurate daily-scale water body spatial distribution pattern.

[0011] Further, the specific process of "through the investigation of the research area" in Step S1 is as follows:

[0012] Select a specific lake area as the research area, obtain hydrological, topographic, and meteorological data at the daily scale through regional investigation. The hydrological data includes the observed values of the flow rates of the inflowing and outflowing rivers and the water levels of the stations within the lake area. The topographic data includes the digital elevation model of the lake area. The meteorological data includes the precipitation and wind speed of the lake area; Obtain the moderate resolution imaging spectrometer surface reflectance data of the lake area within a specific time period, that is, the remote sensing image data, and the remote sensing image data covers the near-infrared band and the short-infrared band.

[0013] Further, in step S2, a hydrodynamic model is constructed based on hydrological data, topographic data, and meteorological data, and the daily water depth data of all grids in the hydrodynamic model is extracted. The specific process of obtaining the high-frequency water body spatial distribution pattern on a daily basis is as follows:

[0014] Step S21: Perform grid discretization on the study area containing hydrological data, topographic data, and meteorological data, complete the interpolation of the grids based on the topographic data, set the upper and lower boundaries of the study area respectively, clarify the roughness parameter of the hydrodynamic model, and construct a hydrodynamic model based on the hydrological data and meteorological data to obtain the water depth of all grids in the hydrodynamic model within a specific time period;

[0015] Step S22: Extract the water depth of the grids in the hydrodynamic model that have the same position as the stations in the study area as the simulated values, and use the measured water depth of the corresponding stations in the study area as the measured values. Compare the simulated values and the measured values, and use the Nash efficiency coefficient EF to evaluate the simulation accuracy for the comparison results;

[0016] Step S23: Obtain the water body spatial distribution pattern on a daily scale by extracting the daily water depth of all grids in the hydrodynamic model after evaluating the simulation accuracy, and assigning a value of 1 to the grids with a water depth > 0 and the rest with a value of 0.

[0017] Further, step S3 is specifically as follows:

[0018] Step S31: Read the reflectance data of the near-infrared band and the green band in the remote sensing image data, substitute the data of the corresponding bands into the water body normalized difference index calculation formula for calculation to obtain the water body normalized difference index value of each pixel in the study area. The calculation formula is as follows:

[0019]

[0020] In the formula, NDWI is the water body normalized difference index; ρ NIR is the reflectance of the near-infrared band; ρ Green is the reflectance of the green band;

[0021] Step S32: Extract the water body spatial distribution pattern according to the threshold division method; specifically:

[0022] In the calculated water body normalized difference index result data, set the threshold to 0. For the pixels with a water body normalized difference index value greater than 0, assign a value of 1 to represent water; for the pixels with a water body normalized difference index value less than or equal to 0, assign a value of 0 to represent non-water.

[0023] Further, the specific process of step S4 is as follows:

[0024] Step S41: Perform standardized preprocessing on the water body spatial distribution pattern data calculated from the obtained remote sensing image data and the water body spatial distribution pattern data obtained from the hydrodynamic model corresponding to the corresponding date;

[0025] For the remote sensing image data R ij , the standardization formula is μR represents the mean of R ij , and σR represents the standard deviation of R ij . For the hydrodynamic model data H ij , the standardization formula is μH is the mean of H ij , and σH is the standard deviation of H ij ; i = 1, 2,..., m; j = 1, 2,..., n; where m and n represent the length and width of the remote sensing image data and the hydrodynamic model data respectively; represents the standardized remote sensing image data; represents the standardized hydrodynamic model data;

[0026] Extract the timestamp information corresponding to the specific time period of the remote sensing image data, and perform encoding processing on the timestamp; convert the timestamp to a time feature, and set the encoded time feature as T, with a dimension of 1×k; specifically, it can be seen as follows:

[0027] Assume that the timestamps of the remote sensing image data are all t ij , in the date-time format, set the starting time t0, and calculate the time difference, which is expressed as:

[0028]

[0029] Among them, Δt ij is the time difference from the starting time t0 to the ending time t ij , and 1day represents the daily scale;

[0030] Use the sine-cosine encoding method to convert Δt ij into a time feature in the k-dimensional feature vector format, and the l-th dimensional time feature T ij,l is expressed as:

[0031]

[0032] Among them, T0 represents the total number of days in a specific time period;

[0033] Based on this, obtain the encoded time feature t ij ;

[0034] Step S42: Construct a convolutional neural network, input the synthetic data into the convolutional neural network, and the convolutional neural network outputs a probability map;

[0035] Step S43: Model training and spatial relationship fitting; all the paired data are divided into the training set, and the mean square error is used as the loss function, expressed as:

[0036]

[0037] In the formula, K is the number of training samples, k0 represents the index of the training sample, f(R * ij ,T) is the predicted output of the convolutional neural network for and T;

[0038] In the formula, K represents the number of training samples, f(R * ij ,T) represents the predicted output of the convolutional neural network for and T;

[0039] The Adam optimization algorithm is used to iteratively update the parameters of the convolutional neural network. During the training process, after each batch of training, the loss value on the validation set is calculated. When the training set loss value no longer decreases for several consecutive rounds or reaches the preset maximum number of training rounds, the training stops.

[0040] Furthermore, the specific process of step S5 is as follows:

[0041] Step S51: Data preparation and reshaping:

[0042] Obtain the input and output of the optimal convolutional neural network at T1 time points, with dimensions of m×n, and the feature dimension after timestamp encoding is k; reshape the input and output of the optimal convolutional neural network into three-dimensional tensors; specifically, stack the data at T1 time points along the feature dimension after timestamp encoding to obtain a tensor with a shape dimension of (T1, m×n + k); use the input of the optimal convolutional neural network in tensor format after dimension stacking as the input of the long short-term memory neural network, and use the output of the optimal convolutional neural network in tensor format after dimension stacking as the output of the long short-term memory neural network;

[0043] Step S52: Construct a long short-term memory neural network and train it based on the input and output in step S51;

[0044] Step S53: Model training and spatio-temporal relationship fitting; specifically, construct an MSE loss function and optimize the long short-term memory neural network by minimizing the MSE loss function; use the Adam optimization algorithm to iteratively update the parameters of the long short-term memory neural network to obtain the optimal long short-term memory neural network as the spatio-temporal relationship model.

[0045] Furthermore, the processing flow of the long short-term memory neural network in step S52 is as follows:

[0046] f t = σ(W f · [h t-1 , x t + b f );

[0047] i t = σ(W i · [h t-1 , x t + b i );

[0048]

[0049] o t = σ(W o · [h t-1 , x t + b o );

[0050] h t = o t * tanh(C t );

[0051] Among them, represents the candidate cell state, C t represents the updated cell state, that is, the cell state at time step t, C t-1 represents the cell state at time step t - 1, h t represents the hidden state, W f , W i , W C and W o all represent the preset weight matrices, b f , b i , b C and b o all represent the preset weight parameters, σ represents the SigMoid activation function, tanh represents the hyperbolic tangent activation function, h t represents the hidden state, h t-1 represents the hidden state at the previous moment, x t represents the input of the long short - term memory neural network, f t represents the activation value of the forget gate, i t represents the activation value of the input gate, o t represents the activation value of the output gate, · represents matrix multiplication, * represents element - wise multiplication;

[0052] The first fully - connected layer performs a fully - connected operation on the hidden state, and the processing process of the fully - connected layer is expressed as:

[0053] F1 = σ(W1h t + b2);

[0054] Wherein, F1 represents the output of the fully connected layer, and W1 and b2 respectively represent the weights and biases of the fully connected layer;

[0055] The second fully connected layer performs a fully connected operation on the output of the first fully connected layer and performs dimensionality conversion to obtain the output of the second fully connected layer as the output of the long short-term memory network.

[0056] The beneficial effects of the present invention are as follows:

[0057] Traditional methods for obtaining wetland water body information often have limitations. For example, although relying solely on a single hydrodynamic model can simulate some dynamic changes of the water body, it has deficiencies in accurately matching with actual observation data and obtaining high-frequency information; while simply relying on remote sensing images is difficult to deeply analyze the physical mechanisms and long-term trends behind water body changes. A method for obtaining wetland water body information by deeply integrating a hydrodynamic model and remote sensing images according to the present invention innovatively deeply integrates the hydrodynamic model and remote sensing images through a deep learning method, giving full play to the advantages of the hydrodynamic model in depicting the physical process of the water body and providing continuous time series data, as well as the advantages of remote sensing images in obtaining high-resolution spatial distribution information. Compared with the prior art, the present invention can accurately obtain the spatial distribution pattern of wetland water bodies at different time scales. Whether it is the daily water body change trend provided by the hydrodynamic model or the accurate water body boundary information reflected by remote sensing images at a specific time period, they can be effectively integrated and optimized. This not only greatly improves the accuracy and integrity of wetland water body information acquisition, but also provides a more reliable data basis for the research, management and protection of wetland ecosystems, fills the technical gap in deeply integrating two data sources to obtain comprehensive wetland water body information, and strongly promotes the further development of research and practice in related wetland fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0059] Figure 1 It is a flowchart of a method for obtaining wetland water body information by deeply integrating a hydrodynamic model and remote sensing images. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0060] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without making creative efforts shall fall within the protection scope of the present invention.

[0061] Referring to Figure 1 , a method for obtaining wetland water body information based on a deeply integrated hydrodynamic model and remote sensing images, the method comprising:

[0062] Step S1: Obtain remote sensing image data, hydrological data, topographic data, and meteorological data through investigating the research area;

[0063] Step S2: Construct a hydrodynamic model based on the hydrological data, topographic data, and meteorological data, extract the daily water depth data of all grids in the hydrodynamic model, and further obtain the daily water body spatial distribution pattern;

[0064] Step S3: Calculate the water body normalized index according to the remote sensing image data, and extract the daily water body spatial distribution pattern within a specific time period through the threshold division method;

[0065] Step S4: Construct a convolutional neural network, standardize the remote sensing image data and hydrodynamic model data corresponding to the time points within a specific time period, and obtain the encoded time features within the specific time period. Combine the standardized hydrodynamic model data and the encoded features to obtain the synthetic data as the input of the convolutional neural network. Combine the standardized hydrodynamic model data and the encoded features to obtain the label as the output of the convolutional neural network. Input the synthetic data into the convolutional neural network. The convolutional neural network outputs a probability map, and the value of each pixel in the probability map represents the probability that the pixel is a water body. Construct an MSE loss function and optimize the convolutional neural network by minimizing the MSE loss function. Obtain the input and corresponding output of the optimal convolutional neural network;

[0066] Step S5: Construct a long short-term memory neural network, stack the dimensions of the input and corresponding output of the optimal convolutional neural network, use the stacked input of the optimal convolutional neural network as the input of the long short-term memory neural network, and use the stacked output of the optimal convolutional neural network as the output of the long short-term memory neural network. Construct an MSE loss function and optimize the long short-term memory neural network by minimizing the MSE loss function. Obtain the optimal long short-term memory neural network as the spatio-temporal relationship model;

[0067] Step S6: Use the daily water body spatial distribution pattern and the corresponding timestamp information obtained in real time by the hydrodynamic model as independent variables, and input them into the spatio-temporal relationship model constructed in Step S5 to obtain the accurate daily-scale water body spatial distribution pattern.

[0068] Further, the specific process of step S1 through investigation and research of the area is as follows:

[0069] Select a specific lake area as the research area, and obtain hydrological, topographical, and meteorological data at a daily scale through area surveys. The hydrological data includes the observed values of the flow rates of the inflowing and outflowing rivers of the lake and the water levels at stations within the lake area. The topographical data includes the digital elevation model of the lake area. The meteorological data includes the precipitation and wind speed in the lake area; obtain the moderate-resolution imaging spectrometer surface reflectance data of the lake area during a specific period, that is, remote sensing image data, which covers the near-infrared band and the short-infrared band.

[0070] Further, the specific process of constructing a hydrodynamic model based on the hydrological data, topographical data, and meteorological data in step S2, extracting the daily water depth data of all grids in the hydrodynamic model, and then obtaining the daily high-frequency water body spatial distribution pattern is as follows:

[0071] Step S21: Perform grid division operations on the research area containing hydrological data, topographical data, and meteorological data, complete the interpolation of the grids based on the topographical data, respectively set the upper and lower boundaries of the research area, clarify the roughness parameters of the hydrodynamic model, construct a hydrodynamic model based on the hydrological data and meteorological data, and obtain the water depths of all grids in the hydrodynamic model during a specific period;

[0072] Step S22: Extract the water depths of the grids in the hydrodynamic model that are in the same positions as the stations in the research area as the simulated values, and use the measured water depths of the corresponding stations in the research area as the measured values. Compare the simulated values and the measured values, and use the Nash efficiency coefficient EF to evaluate the simulation accuracy for the comparison results;

[0073] Step S23: Obtain the daily-scale water body spatial distribution pattern by extracting the daily water depths of all grids in the hydrodynamic model after evaluating the simulation accuracy, and assigning a value of 1 to the grids with a water depth > 0 and the rest with a value of 0.

[0074] Further, step S3 is specifically as follows:

[0075] Step S31: Read the reflectance data of the near-infrared band and the green band in the remote sensing image data, substitute the data of the corresponding bands into the water body normalized difference index calculation formula for calculation, and obtain the water body normalized difference index value of each pixel in the research area. The calculation formula is as follows:

[0076]

[0077] In the formula, NDWI is the water body normalized difference index; ρ NIR is the reflectance of the near-infrared band; ρ Green is the reflectance of the green band;

[0078] Step S32: Extract the water body spatial distribution pattern according to the threshold division method; specifically:

[0079] In the calculated water body normalized index result data, set the threshold to 0. For pixels with a water body normalized index value greater than 0, assign them a value of 1, indicating water bodies; for pixels with a water body normalized index value less than or equal to 0, assign them a value of 0, indicating non-water bodies.

[0080] Furthermore, the specific process of step S4 is as follows:

[0081] Step S41: Perform standardized preprocessing on the water body spatial distribution pattern data calculated from the obtained remote sensing image data and the water body spatial distribution pattern data obtained from the hydrodynamic model for the corresponding date;

[0082] For the remote sensing image data R ij , the standardization formula is μR represents the mean of R ij , and σR represents the standard deviation of R ij . For the hydrodynamic model data H ij , the standardization formula is μH is the mean of H ij , and σH is the standard deviation of H ij ; i = 1, 2,..., m; j = 1, 2,..., n; where m and n represent the length and width of the remote sensing image data and the hydrodynamic model data respectively; represents the standardized remote sensing image data; represents the standardized hydrodynamic model data;

[0083] Extract the timestamp information corresponding to a specific time period of the remote sensing image data, and perform encoding processing on the timestamp; convert the timestamp to a time feature, and set the encoded time feature as T, with a dimension of 1×k; specifically as follows:

[0084] Assume that the timestamps of the remote sensing image data are all t ij , in the date-time format, set the starting time t0, and calculate the time difference, expressed as:

[0085]

[0086] Among them, Δt ij is the time difference from the starting time t0 to the ending time t ij , and 1day represents the daily scale;

[0087] Use the sine-cosine encoding method to convert Δt ij into a time feature in the k-dimensional feature vector format. The l-th dimensional time feature T ij,l is expressed as:

[0088]

[0089] Among them, T0 represents the total number of days in a specific time period;

[0090] Based on this, the encoded time feature t is obtained ij ;

[0091] Step S42: Construct a convolutional neural network, input the synthetic data into the convolutional neural network, and the convolutional neural network outputs a probability map;

[0092] Step S43: Model training and spatial relationship fitting; The paired data are all divided into the training set, and the mean square error is used as the loss function, expressed as:

[0093]

[0094] In the formula, K is the number of training samples, k0 represents the index of the training sample, and f(R * ij , T) is the prediction output of the convolutional neural network for and T;

[0095] The Adam optimization algorithm is used to iteratively update the parameters of the convolutional neural network. During the training process, after each training batch, the loss value on the validation set is calculated. When the training set loss value no longer decreases for several consecutive rounds or reaches the preset maximum number of training rounds, the training stops.

[0096] Furthermore, the specific process of step S5 is as follows:

[0097] Step S51: Data preparation and reshaping:

[0098] Obtain the input and output of the optimal convolutional neural network at T1 time points, with dimensions of m×n, and the dimension of the timestamp-encoded feature is k; reshape the input and output of the optimal convolutional neural network into a three-dimensional tensor; specifically, stack the data at T1 time points along the dimension of the timestamp-encoded feature to obtain a tensor with a shape dimension of (T1, m×n + k); use the input of the optimal convolutional neural network in the tensor format after dimension stacking as the input of the long short-term memory neural network, and use the output of the optimal convolutional neural network in the tensor format after dimension stacking as the output of the long short-term memory neural network;

[0099] Step S52: Construct a long short-term memory neural network and train it based on the input and output in step S51;

[0100] Step S53: Model training and spatio-temporal relationship fitting; specifically, construct the MSE loss function and optimize the long short-term memory neural network by minimizing the MSE loss function; use the Adam optimization algorithm to iteratively update the parameters of the long short-term memory neural network to obtain the optimal long short-term memory neural network as the spatio-temporal relationship model.

[0101] Further, the specific process of step S5 is as follows:

[0102] Step S51: Data preparation and reshaping:

[0103] Obtain the input and output of the optimal convolutional neural network at T1 time points, with dimensions of m×n, and the feature dimension after timestamp encoding is k; reshape the input and output of the optimal convolutional neural network into a three-dimensional tensor; specifically, stack the data at T1 time points along the feature dimension after timestamp encoding to obtain a tensor with a shape dimension of (T1, m×n + k); use the tensor format input of the optimal convolutional neural network after dimension stacking as the input of the long short-term memory neural network, and use the tensor format output of the optimal convolutional neural network after dimension stacking as the output of the long short-term memory neural network;

[0104] Step S52: Construct a long short-term memory neural network and train it based on the input and output in step S51;

[0105] Step S53: Model training and spatio-temporal relationship fitting; specifically, construct the MSE loss function and optimize the long short-term memory neural network by minimizing the MSE loss function; use the Adam optimization algorithm to iteratively update the parameters of the long short-term memory neural network to obtain the optimal long short-term memory neural network as the spatio-temporal relationship model.

[0106] Further, the processing flow of the long short-term memory neural network in step S52 is as follows:

[0107] f t = σ(W f ·[h t-1 , x t +b f );

[0108] i t = σ(W i ·[h t-1 , x t +b i );

[0109]

[0110] o t = σ(W o ·[h t-1 , x t+b o );

[0111] h t =o t *tanh(C t );

[0112] Wherein, represents the candidate cell state, C t represents the updated cell state, that is, the cell state at time step t, C t-1 represents the cell state at time step t-1, h t represents the hidden state, W f , W i , W C and W o all represent preset weight matrices, b f , b i , b C and b o all represent preset weight parameters, σ represents the SigMoid activation function, tanh represents the hyperbolic tangent activation function, h t represents the hidden state, h t-1 represents the hidden state at the previous moment, x t represents the input of the long short-term memory neural network f t represents the activation value of the forget gate, i t represents the activation value of the input gate, o t represents the activation value of the output gate, · represents matrix multiplication, * represents element-wise multiplication;

[0113] The first fully connected layer performs a fully connected operation on the hidden state, and the processing process of the fully connected layer is expressed as:

[0114] F1 = σ(W1h t +b2);

[0115] In the formula, F1 represents the output of the fully connected layer, and W1 and b2 represent the weight and bias of the fully connected layer respectively;

[0116] The second fully connected layer performs a fully connected operation on the output of the first fully connected layer and performs dimensional conversion to obtain the output of the second fully connected layer as the output of the long short-term memory network.

[0117] In this article, specific examples are used to elaborate on the principles and implementation methods of the present invention. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A method for obtaining wetland water body information by deeply integrating a hydrodynamic model and remote sensing images, characterized in that, It includes the following steps: Step S1: Obtain remote sensing image data, hydrological data, topographic data, and meteorological data through investigating the research area; Step S2: Build a hydrodynamic model based on the hydrological data, topographic data, and meteorological data, extract the daily water depth data of all grids in the hydrodynamic model, and then obtain the daily water body spatial distribution pattern; Step S3: Calculate the water body normalized index according to the remote sensing image data, and extract the daily water body spatial distribution pattern within a specific time period through the threshold division method; Step S4: Build a convolutional neural network, standardize the remote sensing image data and hydrodynamic model data corresponding to the time points within a specific time period, and obtain the encoded time features within the specific time period. Combine the standardized hydrodynamic model data and the encoded features to obtain the synthetic data used as the input of the convolutional neural network. Combine the standardized hydrodynamic model data and the encoded features to obtain the label used as the output of the convolutional neural network. Input the synthetic data into the convolutional neural network, and the convolutional neural network outputs a probability map. The value of each pixel in the probability map represents the probability that the pixel is a water body. Build an MSE loss function and optimize the convolutional neural network by minimizing the MSE loss function; obtain the input and corresponding output of the optimal convolutional neural network; Step S5: Build a long short-term memory neural network, stack the dimensions of the input and corresponding output of the optimal convolutional neural network, use the stacked input of the optimal convolutional neural network as the input of the long short-term memory neural network, and use the stacked output of the optimal convolutional neural network as the output of the long short-term memory neural network. Build an MSE loss function and optimize the long short-term memory neural network by minimizing the MSE loss function; obtain the optimal long short-term memory neural network as the spatio-temporal relationship model; Step S6: Use the daily water body spatial distribution pattern and the corresponding timestamp information obtained in real time by the hydrodynamic model as independent variables, and input them into the spatio-temporal relationship model built in Step S5 to obtain the accurate daily-scale water body spatial distribution pattern.

2. The method for obtaining wetland water body information by deeply integrating a hydrodynamic model and remote sensing images according to claim 1, wherein The specific process of "through investigating the research area" in Step S1 is as follows: Select a specific lake area as the research area, obtain hydrological, topographic, and meteorological data at the daily scale through area investigation. The hydrological data includes the observed values of the flow rates of the inflowing and outflowing rivers and the water levels of the stations within the lake area. The topographic data includes the digital elevation model of the lake area. The meteorological data includes the precipitation and wind speed in the lake area; obtain the moderate resolution imaging spectrometer surface reflectance data of the lake area within a specific time period, that is, the remote sensing image data, and the remote sensing image data covers the near-infrared band and the short-wave infrared band.

3. The method for obtaining wetland water body information by deeply integrating a hydrodynamic model and remote sensing images according to claim 2, characterized in that, The specific process of "build a hydrodynamic model based on the hydrological data, topographic data, and meteorological data, extract the daily water depth data of all grids in the hydrodynamic model, and then obtain the daily water body spatial distribution pattern" in Step S2 is as follows: Step S21: Perform grid division operations on the study area containing hydrological data, topographic data, and meteorological data. Complete the interpolation of the grids based on the topographic data, set the upper and lower boundaries of the study area respectively, clarify the roughness parameters of the hydrodynamic model, construct a hydrodynamic model based on the hydrological data and meteorological data, and obtain the water depths of all grids in the hydrodynamic model during a specific period; Step S22: Extract the water depths of the grids in the hydrodynamic model that have the same positions as the stations in the study area as the simulated values, and use the measured water depths of the corresponding stations in the study area as the measured values. Compare the simulated values and the measured values, and use the Nash efficiency coefficient EF to evaluate the simulation accuracy for the comparison results; Step S23: By extracting the daily water depths of all grids in the hydrodynamic model after evaluating the simulation accuracy, and assigning a value of 1 to the grids with water depth > 0 and the rest with a value of 0, obtain the water body spatial distribution pattern at the daily scale.

4. A method for obtaining wetland water body information by deeply integrating a hydrodynamic model and remote sensing images according to claim 3, characterized in that Step S3 is specifically as follows: Step S31: Read the reflectance data of the near-infrared band and the green band in the remote sensing image data, substitute the data of the corresponding bands into the water body normalization index calculation formula for calculation, and obtain the water body normalization index value of each pixel in the study area. The calculation formula is as follows: In the formula, NDWI is the Normalized Difference Water Index; ρ NIR is the reflectance in the near-infrared band; ρ Green is the reflectance in the green band; Step S32: Extract the water body spatial distribution pattern according to the threshold division method; specifically: In the calculated water body normalization index result data, set the threshold to 0. For the pixels with a water body normalization index value greater than 0, assign a value of 1 to represent the water body; For the pixels with a water body normalization index value less than or equal to 0, assign a value of 0 to represent non-water body.

5. A method for obtaining wetland water body information by deeply integrating a hydrodynamic model and remote sensing images according to claim 4, characterized in that The specific process of Step S4 is as follows: Step S41: Perform standardized preprocessing on the water body spatial distribution pattern data calculated from the obtained remote sensing image data and the water body spatial distribution pattern data obtained from the hydrodynamic model corresponding to the corresponding date; For remote sensing image data R ij , the normalization formula is where μR represents the mean of R ij , and σR represents the standard deviation of R ij ; for hydrodynamic model data H ij , the normalization formula is where μH is the mean of H ij , and σH is the standard deviation of H ij ; i = 1, 2,..., m; j = 1, 2,..., n; where m and n represent the length and width of the remote sensing image data and the hydrodynamic model data respectively; represents the normalized remote sensing image data; represents the normalized hydrodynamic model data; Extract the timestamp information of the remote sensing image data corresponding to a specific period, and perform encoding processing on the timestamp; convert the timestamp into a time feature, and set the encoded time feature as T, with a dimension of 1×k; specifically as follows: Assume that the timestamps of the remote sensing image data are all \(t\). ij , in the date-time format. Set the starting time \(t_0\) and calculate the time difference, which is expressed as: where, Δt ij is the time difference from the start time t0 to the end time t ij and 1day represents the daily scale; Convert Δt using the sine-cosine coding method ij into a time feature in the format of a k-dimensional feature vector. The time feature T of the l-th dimension ij,l is expressed as: Among them, T0 represents the total number of days in a specific period; Based on this, the encoded time feature t is obtained ij ; Step S42: Construct a convolutional neural network, input the synthetic data into the convolutional neural network, and the convolutional neural network outputs a probability map; Step S43: Model training and spatial relationship fitting; all the paired data are used as the training set, and the mean square error is used as the loss function, which is expressed as: where K is the number of training samples, k0 represents the index of the training sample, and f(R * ij , T) is the predicted output of the convolutional neural network for and T; Use the Adam optimization algorithm to iteratively update the parameters of the convolutional neural network. During the training process, after each batch of training, calculate the loss value on the validation set. When the training set loss value no longer decreases for several consecutive rounds or reaches the preset maximum number of training rounds, stop the training.

6. The method for obtaining wetland water body information by deeply integrating a hydrodynamic model and remote sensing images according to claim 5, characterized in that, The specific process of Step S5 is as follows: Step S51: Data preparation and reshaping: Obtain the input and output of the optimal convolutional neural network at T1 time points, with dimensions of m×n, and the feature dimension after timestamp encoding is k; reshape the input and output of the optimal convolutional neural network into three-dimensional tensors; specifically, stack the data at T1 time points along the feature dimension after timestamp encoding to obtain a tensor with a shape dimension of (T1, m×n + k); use the input of the optimal convolutional neural network in tensor format after dimension stacking as the input of the long short-term memory neural network, and use the output of the optimal convolutional neural network in tensor format after dimension stacking as the output of the long short-term memory neural network; Step S52: Construct a long short-term memory neural network and train it based on the input and output in Step S51; Step S53: Model training and spatio-temporal relationship fitting; specifically, construct an MSE loss function and optimize the long short-term memory neural network by minimizing the MSE loss function; use the Adam optimization algorithm to iteratively update the parameters of the long short-term memory neural network to obtain the optimal long short-term memory neural network as the spatio-temporal relationship model.

7. A method for obtaining wetland water body information by deeply fusing a hydrodynamic model and remote sensing images according to claim 5, characterized in that, Step S42 is specifically as follows: Step S421: The convolutional layer processes the synthetic data. Convolutional layer 1 uses a 3×3 convolutional kernel, with a stride of 1 and an output channel number of 16. Its calculation process is: Where O1(x, y) is the output of convolutional layer 1, w1(a, b) is the convolutional kernel weight, a represents the X-axis offset, b represents the Y-axis offset, b1 represents the bias term of the convolutional layer, and (x, y) represents the position of the pixel of the synthetic data at (x, y); Step S422: The activation function layer uses the ReLU activation function to perform a non-linear transformation on the output of convolutional layer 1, expressed as: A1(x, y) = max(0, O1(x, y)); where A1(x, y) represents the output of the activation function layer at position (x, y); max(·) represents the ReLU activation function; Step S423: The pooling layer uses a 2×2 max pooling operation with a stride of 2 to perform pooling on the output of the activation function layer, expressed as: In the formula, R pool represents a pooling region centered at the position (x, y) with a size of i0×j0. Both i0 and j0 represent preset parameters, and P1(x, y) represents the output of the pooling layer; Step S424: Process P1(x, y) according to the processing flow in Steps S421 to S423 to obtain P1(x, y), which is also the second output of the pooling layer; Step S425: The fully connected layer flattens the second output of the pooling layer. The fully connected layer has N neurons, and the calculation method is: where F l represents the output of the l-th neuron in the fully connected layer, that is, the probability map, and w l (x,y) represents the connection weight of the fully connected layer at the position (x,y), and b l represents the bias term of the fully connected layer, M represents the total number of sampling points of P2(x,y) in the X-axis direction, and N represents the total number of sampling points in the Y-axis direction.

8. A method for obtaining wetland water body information by deeply fusing a hydrodynamic model and remote sensing images according to claim 6, characterized in that, The processing flow of the long short-term memory neural network in Step S52 is: f t = σ(W f · [h t-1 , x t + b f ); i t = σ(W i · [h t-1 , x t + b i ); o t = σ(W o · [h t-1 , x t + b o ); h t = o t *tanh(C t ) Among them, represents the candidate cell state, C t represents the updated cell state, that is, the cell state at time step t, C t-1 represents the cell state at time step t - 1, h t represents the hidden state, W f 、W i 、W C and W o all represent the preset weight matrices, b f 、b i 、b C and b o all represent the preset weight parameters, σ represents the SigMoid activation function, tanh represents the hyperbolic tangent activation function, h t represents the hidden state, h t-1 represents the hidden state at the previous moment, x t represents the input of the long short-term memory neural network, f t represents the activation value of the forget gate, i t represents the activation value of the input gate, o t represents the activation value of the output gate, · represents matrix multiplication, * represents element-wise multiplication; The first fully connected layer performs a fully connected operation on the hidden state. The processing process of the fully connected layer is expressed as: F1 = σ(W1h t + b2); Where F1 represents the output of the fully connected layer, and W1 and b2 represent the weight and bias of the fully connected layer respectively; The second fully connected layer performs a fully connected operation on the output of the first fully connected layer and conducts dimensional transformation, obtaining the output of the second fully connected layer as the output of the long short-term memory network.

Images