Water reserve change inversion method based on particle swarm optimization optimization neural network

Abstract

The invention provides a water reserve change inversion method based on a particle swarm optimization optimization neural network, and the method mainly comprises the steps: carrying out the inversion of land water reserve change through a time-varying gravity field spherical harmonic coefficient provided by a GRACE / GRACE-FO satellite, and obtaining the underground water reserve change through an underground water reserve change calculation formula in combination with a hydrological model; selecting multi-source high-resolution grid data related to the change of the water reserves as the input of the neural network model, and taking a GRACE / GRACE-FO inversion result as the output; optimizing a weight and a threshold value of the BP neural network by adopting a particle swarm algorithm, and constructing a PSO-BP model for training to obtain an optimal network structure; inputting high-resolution multi-source data into the optimal model, carrying out inversion to obtain land water reserve change data with a spatial resolution of 0.25 degree * 0.25 degree and a continuous time sequence, and further calculating underground water reserve change with the same spatial resolution; according to the method, the spatial resolution and the time continuity of GRACE / GRACE-FO data for monitoring groundwater reserve changes in small and medium-scale areas are remarkably improved, high-precision data support can be provided for groundwater resource fine management and water safety evaluation, and the method has wide popularization and application values.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of earth science and information technology, specifically to gravity satellite remote sensing and hydrological information processing technology, and in particular to a method for improving the spatial resolution of gravity satellite inversion of water storage changes based on particle swarm optimization neural network. Background Technology

[0002] In recent years, numerous scholars both domestically and internationally have successfully used GRACE (Gravity Recovery and Climate Experiment) gravity satellite observation data to infer global-scale changes in terrestrial and groundwater storage. Validation studies, conducted using various hydrological models and groundwater monitoring well data, have confirmed the feasibility and effectiveness of this method in hydrological change monitoring. The widespread application of GRACE data provides crucial technical support for global water resource change monitoring and has played a positive role in supporting national water resource management, ecological environmental protection, and sustainable development strategic decision-making.

[0003] Due to the limited spatial resolution (approximately 1° × 1°) of GRACE and subsequent GRACE-FO (GRACE Follow On) satellite data, direct application to small, localized areas is difficult. Furthermore, observation interruptions and data gaps during satellite operation lead to discontinuities in the time-series information of water storage changes, further restricting its effectiveness in real-time dynamic monitoring and decision support. Therefore, improving the spatial resolution of GRACE satellite data and restoring the integrity of its time-series data has become a key research issue. To address these challenges, developing an inversion method combining multi-source remote sensing data with neural networks has significant scientific and practical value for monitoring regional-scale water storage changes and strengthening regional water resource management. Summary of the Invention

[0004] This invention aims to address the problems of low spatial resolution and discontinuous temporal series in the GRACE / GRACE-FO satellite inversion of terrestrial water storage changes. It proposes a method for terrestrial water storage change inversion that combines Particle Swarm Optimization (PSO) with Back Propagation Neural Network (BPNN) optimization, thereby improving the spatiotemporal adaptability and inversion accuracy of GRACE data in small and medium-scale regions, and further realizing dynamic calculation of high-resolution groundwater storage.

[0005] To achieve the above objectives, this invention proposes a water storage inversion method using a particle swarm optimization (PSO-BP) neural network. This method constructs a nonlinear mapping model between GRACE / GRACE-FO satellite data and topographic, high-resolution meteorological, and hydrological data to restore the continuity of the GRACE data time series and improve its spatial resolution. The method includes the following steps:

[0006] Step 1: Based on the time-varying gravity field spherical harmonic expansion data provided by the GRACE / GRACE-FO satellite, invert the changes in terrestrial water storage in the study area, and combine the known hydrological model data to calculate the changes in groundwater storage;

[0007] Step 1.1, the change in terrestrial water storage derived from GRACE / GRACE-FO time-varying gravity field data can be expressed as:

[0008]

[0009] In the formula, Equivalent water height represents changes in terrestrial water storage. As the geocentric coplanarity, Longitude of the Earth's core The average radius of the Earth The average density of the Earth ( 5517kg / m 3 ), The density of water ( 1000kg / m 3 N represents the maximum order of the gravity field model. Let be the order and degree of the spherical harmonic function, respectively. To fully normalize the associated Legendre function, for Order load Love number, and This represents the variation of the fully normalized spherical harmonic coefficients in the time-varying gravity field model. This is the Gaussian smoothing kernel function;

[0010] Step 1.2: The changes in terrestrial water storage retrieved by the GRACE satellite reflect the comprehensive changes in multiple hydrological elements, mainly including soil water, snow water equivalent, and groundwater storage. Therefore, the changes in groundwater storage can be expressed as:

[0011]

[0012] In the formula, For changes in groundwater storage, The change in terrestrial water storage retrieved from the GRACE satellite data in step 1.1 , These represent soil water and snow water equivalents, respectively, which can be obtained through a hydrological model.

[0013] Step 2: Analyze and collect relevant factors affecting changes in terrestrial water storage in the study area. After downsampling the relevant data, use them as input variables for the model. The changes in terrestrial water storage obtained by GRACE inversion are the output variables of the model.

[0014] Step 3: Use the particle swarm optimization algorithm to optimize the weights and thresholds in the BP neural network model for global optimization, construct the PSO-BP neural network, and carry out model training based on the input and output data in Step 1 and Step 2 to obtain the optimal network structure and parameter configuration within the study area.

[0015] Step 3.1: The BP neural network consists of an input layer, a hidden layer, and an output layer. The preprocessed input features in step 2 are first received by the input layer and then passed to the hidden layer through weighted connections. In the hidden layer, neurons use activation functions to perform nonlinear transformations on the input signals and then pass them to the next layer. Finally, the network prediction value is generated in the output layer.

[0016] Step 3.2: In the particle swarm optimization algorithm, each particle in the swarm is encoded to represent all values ​​and thresholds in the BP neural network. Each particle is evaluated by constructing a fitness function, its fitness value is calculated, and the velocity and position of the particles are dynamically updated based on the information of the individual historical best and the global best.

[0017] Step 3.3: After iterative optimization, the particle with the best fitness is finally obtained, and its corresponding parameters are the optimal weights and thresholds of the network, thus realizing the global optimization of the BP neural network structure.

[0018] Step 4: Input the relevant factors that were not processed in Step 2 into the PSO-BP neural network model in Step 3 to obtain the land water storage changes with high spatial resolution and complete time series in the target area. Combine the soil water and snow water equivalent data of the hydrological model in Step 1 to further obtain the groundwater storage changes in the target area, and realize the high-precision inversion of groundwater dynamics in small and medium-scale areas.

[0019] The present invention provides a method for optimizing a backpropagation neural network based on particle swarm optimization (PSO) algorithm. This method integrates multiple sources of factors related to changes in terrestrial water storage, enhances the training model's ability to perceive regional water storage changes, and effectively avoids the problem of traditional backpropagation neural networks easily getting trapped in local optima by using PSO algorithm to globally optimize the weights and thresholds of the neural network. This allows for the construction of an empirical model between changes in terrestrial water storage and related factors. This method can be used to recover missing GRACE / GRACE-FO data and significantly improve the spatial resolution of water storage change inversion in small and medium-scale regions. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the time-varying gravity field filtering effect according to an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of the error surface according to an embodiment of the present invention.

[0023] Figure 4 This is a comparison of the spatial resolution of GRACE-TWS and PSOBP-TWS in the study area according to an embodiment of the present invention.

[0024] Figure 5 This is a time series comparison chart of GRACE-TWS and PSOBP-TWS in the study area according to an embodiment of the present invention.

[0025] Figure 6 This is a correlation diagram of the study area GRACE-TWS and PSOBP-TWS in an embodiment of the present invention.

[0026] Figure 7 This is a comparison of the spatial resolution of the study area using GRACE-GWS and PSOBP-GWS in an embodiment of the present invention.

[0027] Figure 8 This is a time series comparison chart of GRACE-GWS and PSOBP-GWS in the study area according to an embodiment of the present invention.

[0028] Figure 9 This is a correlation diagram of the study area GRACE-GWS and PSOBP-GWS in an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0030] Figure 1 This is a flowchart illustrating an embodiment of the present invention. The specific implementation steps are as follows:

[0031] Step 1: Using the GRACE / GRACE-FO time-varying gravity field spherical harmonic expansion data released by the Center for Space Research (CSR) at the University of Texas, the terrestrial water storage changes in the area under study are inverted. This data serves as... Figure 1The output variables in the model framework shown are used to train the PSO-BP neural network model.

[0032] Step 1.1: In the original GRACE / GRACE-FO spherical harmonic data, the first-order terms are 0, and the C20 and C30 terms have low accuracy. They can be replaced by external data to improve the inversion accuracy. The Glacial Isostatic Adjustment (GIA) process will have a certain impact on the GRACE inversion of terrestrial water storage changes. Existing GIA models can be selected to eliminate its impact.

[0033] Step 1.2, the spherical harmonic coefficients provided by the GRACE / GRACE-FO satellite contain correlation errors and high-frequency errors. It is difficult to eliminate both types of errors simultaneously using a single filtering method. As shown in Figure (2), the P3M8 decorrelation filter can be combined with a Gaussian filter with a radius of 300km to form a combined filter to suppress the errors in the spherical harmonic coefficients.

[0034] Step 1.3: In step 1.2, filtering to eliminate errors will weaken the actual signal to some extent. Forward modeling can be used to continuously adjust the difference between the analog signal and the observed value through multiple iterations to correct the leakage error.

[0035] Step 1.4: Divide the study area into grids with a spatial resolution of 1°×1°. Based on latitude and longitude, extract all grid center points within the coverage area using a regular grid method, which serve as the spatial units for data extraction and model construction.

[0036] Step 2: To construct the input variable dataset for training the PSO-BP neural network model, the driving factors closely related to changes in terrestrial water storage are extracted by fusing multiple sources of meteorological, hydrological, vegetation, and topographic surface elements.

[0037] Step 2.1: Based on the hydrophysical mechanism of changes in terrestrial water storage in the area under study, the relevant influencing factors selected include: topographic change, surface soil water, land temperature, normalized vegetation index (NDVI), terrestrial evapotranspiration, and terrestrial precipitation. Among them, topographic change data are obtained from DEM data provided by SRTM, surface soil water, terrestrial evapotranspiration and precipitation are obtained from GLDAS hydrological model, and land temperature and NDVI data are provided by MODIS products.

[0038] Step 2.2: Downsample the input data from Step 2.1 according to the grid point range of the area to be studied in Step 1.4 to ensure that the input data corresponds to the GRACE output data in spatial dimension.

[0039] Step 2.3: Ensure that the input data and output data are registered in the time dimension, remove time points in the output dataset where GRACE data is missing, and ensure one-to-one correspondence of training samples in the spatiotemporal dimension;

[0040] Step 3: After completing the preprocessing of input and output data in Steps 1 and 2, construct a particle swarm optimization BP neural network model, reasonably set network structure parameters and optimization strategies, establish a nonlinear mapping relationship between changes in terrestrial water storage in the study area and related factors, and obtain the optimal PSO-BP neural network model between input and output variables through model training, so as to achieve accurate modeling and inversion of changes in terrestrial water storage at the regional scale.

[0041] Step 3.1, as follows Figure 3 As shown, traditional BP neural networks rely on backpropagation and gradient descent algorithms for network training, which have good function approximation capabilities. However, they are prone to getting trapped in local minima when facing complex nonlinear systems, which limits the model's performance. To improve the model's stability and global optimization capabilities, an intelligent optimization algorithm with global search advantages is introduced. Compared with the traditional gradient descent algorithm, the particle swarm optimization (PSO) algorithm has the characteristics of fewer parameters, ease of implementation, fast convergence speed, and excellent global search capabilities. It can effectively avoid BP neural networks getting trapped in local optima and improve the model's training stability and final inversion accuracy.

[0042] Step 3.2: In order to avoid the impact on training accuracy that may be caused by the inconsistency of the magnitude, units, etc. of the various types of data in the input and output ends, it is necessary to normalize all the data participating in the training before the neural network is trained to ensure that the response range of each layer of neurons in the network is consistent during the training process, thereby improving the convergence speed and training stability.

[0043] Step 3.3, in the BP neural network for the first... neurons in the first layer, The output of each neuron is:

[0044]

[0045] in, For the first layer to the first Layer weights For the first Layer The activation value of each neuron. For the first Layer The bias of each neuron, and the activation value of the neuron through the activation function. The calculation yields the following formula:

[0046]

[0047] In the hidden layer, the hyperbolic tangent function (tansig) is chosen as the activation function, and its expression is:

[0048]

[0049] Step 3.4: Set particle encoding and population initialization. Encode all values ​​and biases of the neural network as particle vectors in floating-point format. Each particle represents a complete parameter configuration of a network structure. Let the population size be N, and the dimension of each particle be D, the total number of network parameters. Initialize the particle positions. With speed The range is:

[0050]

[0051] Step 3.5: Map the particle positions to neural network parameters, evaluate the output error on the training set as the fitness function, and select mean squared error (MSE) as the objective function.

[0052]

[0053] in, For the network's predicted output, This represents the actual output, where n is the number of samples.

[0054] Step 3.6: Dynamically adjust the particle search direction using the PSO update formula, as follows:

[0055]

[0056]

[0057] in, For the individual's optimal position, It is the globally optimal position. and It is a learning factor. and It is a random number between [0,1]. It is the position update step size;

[0058] Step 3.7 decodes the particle position vectors with optimal fitness into neural network weights and bias matrices, and fine-tunes the neural network using the Levenberg-Marquardt (LM) algorithm to improve accuracy. The loss function is:

[0059] in, Where is the number of samples in the test set. This represents the activation value of the unique neuron in the output layer. The desired output is data on changes in terrestrial water storage.

[0060] Step 3.8: Divide the dataset into training set, validation set and test set according to 70%:15%:15%, and set the maximum number of iterations to 1000.

[0061] Step 4: Select the PSO-BP neural network model with the best training effect in Step 3, use the high-resolution input data in Step 2 as the model input, and invert to obtain the land water storage change results of the study area with the same resolution. On this basis, combined with the soil water content and snow water equivalent data provided by the GLDAS hydrological model, the distribution of groundwater storage change under the same spatial resolution is further calculated.

[0062] In step 4.1, the spatial resolutions of the various related influencing factors selected in step 2 differ. In this embodiment, the spatial resolution of the GLDAS hydrological model (0.25°×0.25°) is uniformly followed, and spatial downsampling is performed on the input data with a resolution higher than 0.25°×0.25° to ensure that all input data remain consistent on the spatial grid.

[0063] Step 4.2: After completing the spatial consistency processing of all input data in Step 4.1, input data with a resolution of 0.25°×0.25° is input into the optimal PSO-BP neural network model selected in Step 3 to invert and obtain the terrestrial water storage change (PSOBP-TWS) data with a spatial resolution of 0.25°×0.25° in the study area.

[0064] Step 4.3: Compare and analyze the original GRACE-TWS and PSOBP-TWS, such as... Figure 4 As shown, PSOBP-TWS improves the spatial resolution of terrestrial water storage changes in the study area from 1°×1° to 0.25°×0.25°, significantly enhancing the ability to express spatial details in terrestrial water storage monitoring. Figure 5 As shown, this model effectively fills in the missing data in certain time periods of the GRACE data, improving the completeness of the time series. Figure 6 As shown, the Pearson correlation coefficient between PSOBP-TWS and GRACE-TWS is as high as 0.99, which verifies the accuracy and stability of the constructed model while maintaining the consistency of the original data.

[0065] Step 5: Based on the high-resolution terrestrial water storage change (PSOBP-TWS) obtained by inversion from the PSO-BP neural network model in Step 4, combined with the soil water and snow water equivalent data provided in the GLDAS hydrological model, the groundwater storage change (PSOBP-GWS) with the same spatial resolution as PSOBP-TWS is obtained.

[0066] Step 5.1, based on the formula for changes in groundwater storage The solution obtained a PSOBP-GWS with a spatial resolution of 0.25°×0.25° and continuous time in the study area, as shown in Figure (7). The spatial resolution of groundwater storage change in the study area was improved to 0.25°×0.25°, which significantly enhanced the ability to identify and monitor groundwater anomalies in small and medium-scale areas.

[0067] Step 5.2, compare and analyze the groundwater storage change inverted by PSOBP-GWS and the groundwater storage change inverted by the traditional GRACE+GLDAS method (GRACE-GWS). The results are shown in Figure (8). The interannual variation trends of the two are -11.45 mm / a and -11.28 mm / a, respectively, with a difference of no more than 0.17 mm / a. The Pearson correlation coefficient between the two is 0.98, as shown in Figure (9). The above results show that PSOBP-GWS can maintain the long-term variation trend of GRACE-GWS and can also provide high-precision supplementary measurements during the period when GRACE is missing. It can give full play to the advantages of GRACE satellite in long-term monitoring of groundwater changes.

[0068] The above description is only a preferred embodiment of the present invention. For those skilled in the art, there will be changes in the specific implementation and application scope based on the ideas of the present invention. The content of this specification should not be construed as a limitation of the present invention.

Claims

1. A water storage change inversion method based on a particle swarm algorithm optimized neural network, characterized in that, Comprise the following steps: Step 1, the original land water storage change data is obtained by using GRACE / GRACE-FO satellite gravity field spherical harmonic expansion data inversion, and combined with multi-source ground information, an input variable set related to land water storage change is constructed, and all data are uniformly registered and normalized in space and time dimensions to ensure the spatio-temporal consistency of neural network training; Step 2, the BP neural network structure is constructed, the initial weight and bias parameters are globally optimized by using particle swarm optimization algorithm, and the network is further fine-tuned by combining Levenberg-Marquardt gradient algorithm to improve the convergence and precision of model training, and finally the nonlinear mapping relationship between input factors and land water storage change is established; Step 3, the input factor data with uniform spatial resolution of 0.25°x0.25° is input into the trained PSO-BP neural network to obtain high-resolution land water storage change, and further combined with soil water and snow water equivalent information in hydrological model to calculate high-resolution groundwater storage change in small-scale area.

2. The method of claim 1, wherein, Step 1 specifically comprises the following sub-steps: Step 1.1, the spherical harmonic expansion data provided by GRACE / GRACE-FO satellite needs to replace the first-order term with insufficient accuracy, C20 and C30, and the leakage error is corrected by forward modeling method through de-correlation filtering and Gaussian filtering to suppress error; Step 1.2, the input data set is constructed including terrain data, ground soil water data, precipitation, evapotranspiration, land temperature and vegetation normalized index data, wherein the terrain data is provided by SRTM, and other data is derived from GLDAS and MODIS products, and all input data are uniformly resampled to a spatial resolution of 0.25°x0.25°.

3. The method of claim 1, wherein, Step 2 specifically comprises the following sub-steps: Step 2.1, the BP neural network in step 2 includes an input layer, two hidden layers and an output layer, the activation function uses hyperbolic tangent function, the input node is 6, the output node is 1, and the hidden layer is 25-10; Step 2.2, the neural network weight and bias parameters are coded as particles in floating point form, the population size, maximum iteration number, learning factor and inertia weight are set, and the fitness function is set as mean square error (MSE); Step 2.3, the global optimal parameters are obtained by iteratively updating the particle position and velocity, and the neural network is fine-tuned by combining Levenberg-Marquardt algorithm.

4. The method of claim 1 wherein, Step 3 specifically comprises the following sub-steps: Step 3.1, all input factor data are resampled and aligned according to the uniform spatial grid, and the input variable structure of the training model is kept consistent; Step 3.2, the normalized input data is input into the trained PSO-BP neural network to obtain high-resolution land water storage change data under the corresponding grid; Step 3.3, combined with the soil water and snow water equivalent data provided by the hydrological model, the high-resolution groundwater storage change is calculated by inverting the groundwater storage change formula.

5. The method of claim 1, wherein: The inversion result has a spatial resolution of 0.25°*0.25° and a time resolution of a month, can provide continuous land water storage and groundwater storage estimation results in a missing period of original GRACE data, and effectively improves the space-time adaptability of GRACE data in water resource monitoring in a small and medium scale region.

Images