CN115630686A - A method for recovering terrestrial water storage anomalies from satellite gravity data using machine learning - Google Patents

Abstract

The invention discloses a method for recovering land water reserve abnormality from satellite gravity data by machine learning, which comprises the following steps: s1: acquiring satellite data of a GRACE gravity satellite, land-water reserve data of a hydrological model and boundary and longitude and latitude information of a preset area, and setting boundary and longitude and latitude information of a background area according to a range in which gravity signals of the preset area are likely to leak; s2: calculating the smooth land water reserves abnormal of the GRACE gravity satellite inversion in the preset area and the background area; s3: acquiring lunar scale land-ground water reserve abnormal data of a global hydrological model, and carrying out leveling operation according to a research time period; s4: carrying out forward simulation on land and water reserve abnormal data of the globe, and acquiring smoothed land and water reserve abnormal data according to the boundary and longitude and latitude information of a background area; s5: constructing and training a neural network model; s6: and (3) inputting the land water reserves abnormal data after smoothing of the background area obtained in the step (S2) into a trained neural network model to recover the leakage signal.

Description

A method for recovering terrestrial water storage anomalies from satellite gravity data using machine learning

technical field

The present invention relates to the field of gravity satellite technology and hydrology, in particular, to the application field of inversion of land water storage anomalies by gravity satellites, and more specifically to a method for recovering land water storage anomalies from satellite gravity data using machine learning, that is, based on Method of Convolutional Neural Network Recovering Gravity Satellite Leaked Signals to Obtain Terrestrial Water Storage Anomalies.

Background technique

The sum of various forms of water storage on land is called terrestrial water storage, including vegetation canopy water, soil water content, surface water storage, snow water equivalent and groundwater storage. Terrestrial Water Storage Anomaly (TWSA) refers to the difference between the terrestrial water storage during the study period and the average terrestrial water storage over a period of time, which can reflect the increase or decrease of terrestrial water storage. The anomaly of terrestrial water storage is an important indicator to measure the status of regional water resources, and it is also an important part of the water and heat balance of the global terrestrial ecosystem. Quantitatively and accurately estimating the temporal and spatial changes of regional terrestrial water storage anomalies can provide strong support for sustainable management of water resources.

Commonly used land water storage monitoring methods include hydrometeorological observations and land surface process modeling. Traditional hydrometeorological observation methods are easily constrained by space-time distribution of space stations, complex topography, and human, financial, and material capabilities; land hydrological models such as the Global Land Data Assimilation System (GLDAS) and the Global Hydrological Model (hereinafter collectively referred to as Although hydrological models) can also simulate changes in terrestrial water storage near the surface, due to the limitations of the hydrological model itself and the absence of some hydrological elements (such as groundwater), it is often impossible to accurately estimate the regional TWSA. Since its launch in 2002, the gravity field recovery and climate experiment satellite GRACE (Gravity Recovery and Climate Experiment) has been widely used in TWSA inversion. However, the high-order coefficients of GRACE spherical harmonic products have large random errors and systematic errors, which need to be removed by low-pass filtering, resulting in signal leakage. Traditional signal leakage correction methods mainly use scale factor method, multiplication correction method, addition correction method and iterative correction method. However, the first two methods only rely on one hydrological model when they are applied, and are relatively dependent on the accuracy of the hydrological model. If an inaccurate hydrological model is used, it will often cause a large calculation deviation and cannot restore the leaked signal well. . In addition, the multiplication correction method assumes that the TWSA of the area is uniform, but the TWSA of the actual area is often difficult to meet this assumption. In the iterative correction method, the constrained iteration often needs to obtain the source range of the signal, while the unconstrained iteration has higher requirements on the uniformity of the TWSA in the inversion area, so it is easy to cause the spatial distribution of the signal to be too smooth and cause distortion. To sum up, there is a lack of a GRACE signal leakage recovery method that does not depend on a single hydrological model, has low dependence on prior signals, and can guarantee the spatial details of TWSA. The present invention aims at the above problems, and proposes a method for correcting GRACE leakage signals based on a convolutional neural network.

Contents of the invention

In order to solve the above problems, the present invention provides a method using machine learning to restore land water storage anomalies from satellite gravity data, which is based on convolutional neural networks, using WGHM (WaterGAP Global Hydrology Model, WaterGAP Global Hydrology model), Noah (Noah Land Surface model, Noah Land Surface Model), CLSM (Catchment Land Surface Model), Mosaic (Mosaic Land Surface Model, Mosaic Land Surface Model) and VIC (Variable Infiltration Capacity Macroscale Hydrologic Model) five hydrological models The terrestrial water storage anomaly obtained by the model imitates the GRACE leakage signal for model training, and the trained neural network model is used to correct the leakage signal during the processing of GRACE level-2 spherical harmonic coefficient products. The method of the present invention avoids the lack of prior knowledge when correcting signal leakage errors, does not depend on a single hydrological model, can consider the characteristics of spatial relationships by means of convolutional neural networks, and can accurately invert the spatial details of changes in land water storage, and expand The application space of gravity satellite technology in the inversion of terrestrial water storage changes and in the field of hydrology.

To achieve the above object, the present invention provides a method for recovering land water storage anomalies from satellite gravity data using machine learning, which includes:

Step S1: Obtain the satellite data of the GRACE gravity satellite, the land water storage data of the hydrological model, and the boundary and latitude and longitude information of the preset area, and set the boundary and latitude and longitude information of the background area according to the possible leakage range of the gravity signal of the preset area, where , the background area is the area outside the preset area that contains most of the leaked signals;

Step S2: According to the data obtained in step S1 and the latitude and longitude information of the preset area and the background area, calculate the smoothed land water storage anomaly retrieved by the GRACE gravity satellite in the preset area and the background area;

Step S3: Obtain the monthly scale terrestrial water storage anomaly data of the global hydrological model, and perform anomaly operations according to the research time period;

Step S4: Carry out forward modeling of global land water storage anomaly data, reduce errors in high-order spherical harmonic coefficients through low-pass filtering, and obtain smoothed land water storage anomaly data according to the boundary and latitude and longitude information of the background area;

Step S5: Construct and train the neural network model;

Step S6: Input the smoothed land water storage anomaly data retrieved by the GRACE gravity satellite in the background area obtained in step S2 into the trained neural network model to recover the leakage signal.

In an embodiment of the present invention, wherein the GRACE gravity satellite data is a level-2 spherical harmonic coefficient product, the land water storage information represented by it includes surface water, soil water and groundwater; the hydrological model obtained includes WGHM, Noah, CLSM There are five hydrological models, Mosaic and VIC. The terrestrial water storage data of the hydrological model include the sum of the snow water equivalent output results, soil water storage output results and groundwater storage output results simulated by the hydrological model.

In an embodiment of the present invention, the calculation of land water storage anomalies based on the satellite data of the GRACE gravity satellite in step S2 is to convert the high-order level-2 spherical harmonic data containing random errors and systematic errors into land after the background area is smoothed The process of anomalous water storage data is converted into smoothed land water storage anomalies based on the latitude and longitude information of the background area through the construction of the earth's gravity field model, low-pass filtering, destriping filtering, and anomaly operations.

In an embodiment of the present invention, wherein the low-pass filter adopts a Gaussian filter with a radius of 300km, and the destriping filter preferably adopts the P3M10 method, that is, the spherical harmonic coefficients of the first 10×10 orders remain unchanged, and a third-order polynomial is used Fit the spherical harmonic coefficients greater than or equal to the 10th order, the odd order and the even order are fitted separately, and the fitting value is deducted from the original spherical harmonic coefficients. The specific process includes:

Truncate the GRACE level-2 spherical harmonic coefficient products to a certain order of spherical harmonic coefficients;

Remove the 0th order item;

Replace item C20 with the results calculated from satellite laser ranging data;

Use the Technical Note TN-13 data released by JPL to replace the first-order term of the level-2 spherical harmonic coefficient product.

In an embodiment of the present invention, wherein, step S2 also includes the conversion between the spherical harmonic coefficient and the land water storage anomaly, which is the smoothed grid data of the GRACE gravity satellite level-2 spherical harmonic coefficient converted into the background area , including:

Step S201: Assuming that the mass redistribution of the earth that causes gravity anomalies is only concentrated on the earth's surface, mainly including changes in the atmosphere, oceans, ice sheets, and land water reserves, the density change corresponding to this mass redistribution is converted into The change of the geoid height caused by it, wherein the change includes the contribution of the direct gravitational attraction of the surface mass to the change of the geoid and the load deformation caused by the solid earth after the change of the surface load. The formula is expressed as formula (1),

为l阶m次标准缔合勒让德函数，ρ_e表示地球平均密度(约为5517kg/m³)，k_l表示l阶负荷勒夫数，Δσ(θ，λ)为地表质量密度变化，a为地球平均半径(约6378km)；In the formula, θ represents the earth's co-latitude (0-180°), λ represents the earth's east longitude (-180°-180°), l and m represent the order and number of spherical harmonic expansion of the gravitational field, respectively, and the dimensionless coefficients ΔC _lm and ΔS _lm represents the spherical harmonic coefficient of the geoid change,

is the standard associated Legendre function of order m of order l, ρ _e represents the average density of the earth (about 5517kg/m ³ ), k _l represents the Love number of load of order l, and Δσ(θ, λ) is the change of surface mass density, a is the average radius of the earth (about 6378km);

Step S202: If the surface mass density change Δσ(θ, A) is also subjected to spherical harmonic expansion, then

In the formula, ρ _w is the density of water (1000kg/m ³ ). Compared with formula (1), it can be seen that ρ _w is often expressed in the form of Δσ/ρ _w due to the transformation of surface mass density, so we can get:

和

为地表密度变化的球谐系数；In the formula,

and

is the spherical harmonic coefficient of surface density change;

和

大地水准面变化的球谐系数ΔC_lm和ΔS_lm之间的函数关系为：Step S203: Obtain the spherical harmonic coefficient of surface density change according to formula (1) and formula (3)

and

The functional relationship between the spherical harmonic coefficients ΔC _lm and ΔS _lm of the geoid change is:

Step S204: Substituting formula (4) into formula (2) to get:

Among them, formula (5) is the basic formula for calculating the surface mass density change using the geoid spherical harmonic coefficient data released by the GRACE data center, that is, the basic conversion formula between the spherical harmonic coefficient and the anomaly of land water storage;

Step S205: When only the change of land water storage is considered, the change of land water storage is estimated according to the obtained spherical harmonic coefficient of geoid change provided by the GRACE satellite, combined with formula (1).

In an embodiment of the present invention, step S3 specifically includes:

Step S301: According to the global longitude and latitude, the soil water storage, surface water storage and other terrestrial water storage change components of the five hydrological models are respectively extracted by the following formula, so as to obtain the five terrestrial water storage change data:

TWS＝SMS+GWS+SWS+OS

In the formula, TWS is terrestrial water storage, SMS is soil water storage, GWS is groundwater storage, SWS is surface water storage, OS is other terrestrial water storage change components except SMS, GWS and SWS;

Step S302: Perform anomaly processing on the obtained land water storage data, that is, subtract the average value of land water storage in the research period from the value of each month, and obtain the abnormal data of land water storage in the background area simulated by the hydrological model.

In one embodiment of the present invention, wherein, the forward modeling in step S4 is to convert the abnormal land water storage data simulated by the hydrological model into spherical harmonic coefficients, refer to formula (3), and carry out the spherical harmonic coefficients with the gravity satellite spherical The data processing flow with similar harmonic coefficients, the specific processing flow includes:

Low-pass filtering adopts Gaussian filtering with a radius of 300km;

Remove the 0th order item;

The described GRACE level-2 spherical harmonic coefficient product is truncated to a certain order of spherical harmonic coefficient;

In step S4, the smoothed land water storage anomaly data obtained according to the boundary of the background area and the latitude and longitude information are converted according to formula (5).

In an embodiment of the present invention, wherein, step S5 specifically includes:

Step S501: Using the smoothed abnormal value of land water storage in the background area as the input data of the neural network model, and using the abnormal value of land water storage in the preset area as the target data of the neural network model to establish a neural network model;

Step S502: use the Python language to complete the training of the model.

Compared with the existing technology, the method of using machine learning to restore the anomaly of land water storage from satellite gravity data provided by the present invention can largely avoid the lack of prior knowledge when correcting signal leakage errors, and does not rely on a single hydrological The model can better reduce the impact of leakage errors, and can accurately invert the spatial details of land water storage anomalies.

Description of drawings

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 that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. Those skilled in the art can also obtain other drawings based on these drawings without creative work.

Fig. 1 is a method flowchart of an embodiment of the present invention;

Fig. 2 is the basic structural representation of the neural network model that an embodiment of the present invention adopts;

Fig. 3 is the performance comparison of the hydrological model in the test set of the Qinghai-Tibet Plateau in the convolutional neural network model in an embodiment of the present invention, wherein (a) represents the average value of the test set of the abnormal spherical harmonic expansion of the land water storage simulated by the hydrological model value; (b) represents the average value of the test set of land water storage anomalies simulated by the hydrological model; (c) represents the average value of the output results of the convolutional neural network; (d) represents the difference between (b) and (c);

Fig. 4 is the comparison of the 2003-2016 land water storage anomaly and other methods in the Qinghai-Tibet Plateau of the 2003-2016 GRACE inversion restored by the convolutional neural network in an embodiment of the present invention, wherein, (a), (b), (c) and (d) Respectively represent the trend value of GRACE inversion of land water storage anomalies restored by convolutional neural network, the trend value of scale factor method, the trend value of Mascon data officially released by CSR, the trend value of Mascon data officially released by JPL, (e) Time series representing four types of terrestrial water storage anomalies.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

In a specific embodiment, the technical method of the present invention can be realized through the following method steps. In addition, it should be pointed out that in this embodiment, the following steps are not in a strict logical sequence between the upper and lower steps, but only for the purpose of explaining the specific implementation process and main content of this solution in as much detail as possible. For example, in steps S2 to S4, except that the corresponding steps of the data that have a logical relationship between the front and back calculations have a front and back relationship in mathematical calculations, other parallel data calculations can be realized through parallel steps, not necessarily in a fixed The sequence of steps S1 to S6 below should not be construed as limiting the scope of protection of this solution.

Fig. 1 is the method flow chart of an embodiment of the present invention, as shown in Fig. 1, the present embodiment provides a kind of method utilizing machine learning to restore land water reserve anomaly from satellite gravity data, and it comprises:

Step S1: Obtain the satellite data of the GRACE gravity satellite, the land water storage data of the hydrological model, and the boundary and latitude and longitude information of the preset area, and set the boundary and latitude and longitude information of the background area according to the possible leakage range of the gravity signal of the preset area;

In this embodiment, the GRACE gravity satellite data obtained in step S1 includes the level-2 spherical harmonic coefficient products released by the Space Research Center (CSR) of the University of Texas in the United States, and also includes the Jet Propulsion Laboratory (JPL), Germany The level-2 spherical harmonic coefficient products released by Potsdam Geoscience Center (GFZ) and other institutions at home and abroad, the terrestrial water storage information represented by it includes surface water, soil water and groundwater; the obtained hydrological models include WGHM (WaterGAP global water Literature Model), Noah (Noah Land Surface Model), CLSM (Catchment Surface Model), Mosaic (Mosaic Land Surface Model) and VIC (Variable Infiltration Capability Macro Model), five hydrological models, the land water storage data of the hydrological model include the hydrological model The sum of the simulated snow water equivalent output, soil water storage output, and groundwater storage output. Among the hydrological models used in the present invention, the WGHM and CLSM hydrological models contain the output results of groundwater storage changes, while the other three do not.

Wherein, the background area is an area containing most leaked signals outside the preset area, and the range of the background area can be set to be 10° wider than the latitude and longitude of the preset area. For example, if the minimum longitude, maximum longitude, minimum latitude and maximum latitude of the preset area are 110°, 120°, 40° and 50° respectively, then the minimum longitude, maximum longitude, minimum latitude and maximum latitude of the background area are respectively 100°, 130°, 30° and 60°.

Step S2: According to the data obtained in step S1 and the latitude and longitude information of the preset area and the background area, calculate the smoothed land water storage anomaly retrieved by the GRACE gravity satellite in the preset area and the background area;

In this embodiment, step S2 calculates the anomaly of land water storage based on the satellite data of the GRACE gravity satellite, and specifically converts the high-order level-2 spherical harmonic data containing random errors and systematic errors into land water after the background area is smoothed. The process of anomalous reserves data can be converted into smoothed land water storage anomalies by constructing an earth gravity field model, performing low-pass filtering, destriping filtering, and anomaly operations, and according to the latitude and longitude information of the background area.

In this embodiment, the low-pass filtering can be preferably processed by Gaussian filtering with a radius of 300km, or other low-pass filtering can be used for processing; the destriping filtering preferably adopts the P3M10 method, that is, the first 10×10 order The spherical harmonic coefficients remain unchanged, and the third-order polynomial is used to fit the spherical harmonic coefficients greater than or equal to the 10th order. The odd and even orders are fitted separately, and the fitting value is deducted from the original spherical harmonic coefficients. The specific process includes:

Truncate the GRACE level-2 spherical harmonic coefficient products to a certain order of spherical harmonic coefficients, for example, truncate to 60th order;

Remove the 0th order item;

Replace item C20 with the results calculated from satellite laser ranging data;

Use the Technical Note TN-13 data released by JPL to replace the first-order term of the level-2 spherical harmonic coefficient product.

It should be noted that when the above-mentioned filtering and truncation processing is performed on the GRACE gravity satellite level-2 spherical harmonic data, while weakening the high-order errors, it will also cause loss to the real signal. In this embodiment, in step S6 restore it. The Gaussian filter used in this embodiment is essentially a spatially averaged weighting function, and the smoothing of the signal is achieved by weighting the averaged space within a certain radius range. The destriping filter used in this embodiment can remove the correlation between the spherical harmonic coefficients, but it will also cause distortion to the real signal; an important characteristic of the order of the spherical harmonic coefficients is to reflect the level of data spatial resolution , theoretically spherical harmonics can be expanded to infinite order, showing more spatial details, but higher order will bring more errors.

In this embodiment, step S2 also includes the conversion between spherical harmonic coefficients and land water storage anomalies, which is to convert the GRACE gravity satellite level-2 spherical harmonic coefficients into the smoothed grid data of the background area, specifically include:

Step S201: Assuming that the redistribution of the earth's mass that causes gravity anomalies is only concentrated on the earth's surface (within 10-15 km), mainly including changes in the atmosphere, oceans, ice sheets, and land water reserves, etc., the density corresponding to this mass redistribution The change is converted into the change of the geoid height caused by the radial integral, wherein the change includes the contribution of the direct gravitational attraction of the surface mass to the change of the geoid and the load deformation of the solid earth caused by the change of the surface load. The formula The expression is as formula (1),

式中，θ表示地球余纬(0～180°)，λ表示地球东经(-180°～180°)，l和m分别表示重力场球谐展开的阶数和次数，无量纲系数ΔC_lm和ΔS_lm表示大地水准面变化球谐系数，

为l阶m次标准缔合勒让德函数，ρ_e表示地球平均密度(约为5517kg/m³)，k_l表示l阶负荷勒夫数，Δσ(θ，λ)为地表质量密度变化，a为地球平均半径(约6378km)；

In the formula, θ represents the earth's co-latitude (0-180°), λ represents the earth's east longitude (-180°-180°), l and m represent the order and number of spherical harmonic expansion of the gravitational field, respectively, and the dimensionless coefficients ΔC _lm and ΔS _lm represents the spherical harmonic coefficient of the geoid change,

is the standard associated Legendre function of order m of order l, ρ _e represents the average density of the earth (about 5517kg/m ³ ), k _l represents the Love number of load of order l, and Δσ(θ, λ) is the change of surface mass density, a is the average radius of the earth (about 6378km);

Step S202: If the surface mass density change Δσ(θ, λ) is also subjected to spherical harmonic expansion, then

In the formula, ρ _w is the density of water (1000kg/m ³ ). Compared with formula (1), it can be seen that ρ _w is often expressed in the form of Δσ/ρ _w due to the transformation of surface mass density, so we can get:

和

为地表密度变化的球谐系数；In the formula,

and

is the spherical harmonic coefficient of surface density change;

和

大地水准面变化的球谐系数ΔC_lm和ΔS_lm之间的函数关系为：Step S203: Obtain the spherical harmonic coefficient of surface density change according to formula (1) and formula (3)

and

The functional relationship between the spherical harmonic coefficients ΔC _lm and ΔS _lm of the geoid change is:

Step S204: Substituting formula (4) into formula (2) to get:

Among them, formula (5) is the basic formula for calculating the surface mass density change using the geoid spherical harmonic coefficient data released by the GRACE data center, that is, the basic conversion formula between the spherical harmonic coefficient and the anomaly of land water storage;

In hydrological research, the focus is usually on the change of terrestrial water storage. Therefore, based on the "thin layer" assumption (the assumption in step S201), with the help of relevant model simulation data, the part of the disturbance caused by the mass redistribution of the atmosphere and ocean can be simulated The disturbance signal (or geoid change) is subtracted from the original GRACE signal, and the remaining disturbance signal (or geoid change) is mainly caused by the change of terrestrial water storage. On the premise that only the change of land water storage is considered, the change of land water storage (or equivalent water height change ^Δ σ/ρ _w ) can be estimated by using the spherical harmonic coefficient of geoid change provided by the GRACE satellite. .

Step S205: When only the change of land water storage is considered, the change of land water storage is estimated according to the obtained spherical harmonic coefficient of geoid change provided by the GRACE satellite, combined with formula (1).

Step S3: Obtain the monthly scale terrestrial water storage anomaly data of the global hydrological model, and perform anomaly operation according to the research time period; the hydrological model in this embodiment includes five hydrological models including WGHM, Noah, CLSM, Mosaic and VIC, and , other embodiments may also include other hydrological models released by other institutions at home and abroad that can obtain water storage data such as regional soil water storage and snow water equivalent;

In this embodiment, step S3 specifically includes:

Step S301: According to the global longitude and latitude, the soil water storage, surface water storage and other terrestrial water storage change components of the five hydrological models are respectively extracted by the following formula, so as to obtain the five terrestrial water storage change data:

TWS＝SMS+GWS+SWS+OS

In the formula, TWS is terrestrial water storage, SMS is soil water storage, GWS is groundwater storage, SWS is surface water storage, OS is other terrestrial water storage change components except SMS, GWS and SWS;

Step S302: Perform anomaly processing on the obtained land water storage data, that is, subtract the average value of land water storage in the research period from the value of each month, and obtain the abnormal data of land water storage in the background area simulated by the hydrological model.

Among them, it should be noted that: since the resolutions of the five hydrological models are different, as a more preferred implementation, in this embodiment, all the above five hydrological models can be converted into 1°×1° resolution. Among them, WGHM and CLSM of the five hydrological models contain the output results of groundwater storage changes, while the other three do not.

Step S4: Carry out forward modeling of global land water storage anomaly data, reduce errors in high-order spherical harmonic coefficients through low-pass filtering, and obtain smoothed land water storage anomaly data according to the boundary and latitude and longitude information of the background area;

In this embodiment, wherein, the forward modeling in step S4 is to convert the abnormal data of land water storage simulated by the hydrological model into spherical harmonic coefficients, and formula (3) can be referred to, and the spherical harmonic coefficients are compared with the gravity satellite spherical harmonic coefficients. The data processing flow with similar coefficients, the specific processing flow includes:

Low-pass filtering adopts Gaussian filtering with a radius of 300km;

Remove the 0th order item;

Truncating the GRACE level-2 spherical harmonic coefficient products to spherical harmonic coefficients of a certain order, such as truncating to 60th order;

In step S4, the smoothed land water storage anomaly data obtained according to the boundary of the background area and the latitude and longitude information are converted according to formula (5).

The land water storage of the hydrological model is calculated as a smoothed land water storage anomaly. Specifically, the land water storage of the hydrological model is expanded spherical harmonically, and the smoothed This operation is also called forward modeling.

Step S5: Construct and train the neural network model;

Fig. 2 is a schematic diagram of the basic structure of a neural network model used in an embodiment of the present invention. As shown in Fig. 2, the neural network used in this embodiment is preferably a convolutional neural network model, and other application networks can also be used in other embodiments Model, the platform for constructing the neural network model is Tensorflow 2.2.0 and Keras 2.4.3 packages of Python 3.8, the learning rate of the model is 0.01, the batch size is 32, the number of training iterations (epoch) is 300, and the discarding rate (dropout) is 0.1, the optimizer uses a stochastic gradient descent optimizer (SGD), and the loss function is MSE.

In this embodiment, wherein, step S5 specifically includes:

Step S501: Using the smoothed abnormal value of land water storage in the background area as the input data of the neural network model, and using the abnormal value of land water storage in the preset area as the target data of the neural network model to establish a neural network model;

Specifically, the abnormal land water storage data of the five hydrological models in the preset area in step S3 and the smoothed land water storage abnormal data in the background area by the five hydrological models in step S4 through forward modeling can be used as model training, the former As the target data of the neural network model, the latter serves as the input data of the neural network model. 80% of the input data and target data of the five hydrological models used for convolutional neural network model training can be extracted for model training, and the other 20% can be used for model verification. The present invention does not strictly limit the training data and verification data It can be adjusted appropriately according to the needs.

Step S502: use the Python language to complete the training of the model.

In order to confirm the recovery effect of the neural network model for the leaked signal, 20% of the data randomly extracted were evaluated by the neural network model, and the evaluation results are as shown in Figure 3, and Figure 3 is the hydrological model in an embodiment of the present invention on the Qinghai-Tibet Plateau The performance comparison of the test set in the convolutional neural network model, where (a) represents the average value of the spherical harmonic expansion of the test set of the abnormal land water storage simulated by the hydrological model; (b) represents the anomaly of the land water storage simulated by the hydrological model The average value of the test set; (c) represents the average value of the convolutional neural network output; (d) represents the difference between (b) and (c).

In this example, the input data used in the training of the model is the smoothed land water storage anomaly in the background area of the hydrological model, and the target data is the abnormal value of the land water storage in the preset area of the hydrological model, so the convolution of the training is completed The neural network model can be used to restore the smoothed land water storage anomaly in the background area to the land water storage anomaly in the preset area.

Step S6: Input the smoothed land water storage anomaly data retrieved by the GRACE gravity satellite in the background area obtained in step S2 into the trained neural network model to recover the leakage signal.

By inputting the smoothed land water storage anomaly data in the background area inverted by GRACE into the trained convolutional neural network model, the land water storage anomaly data after the leakage signal is restored can be obtained. Fig. 4 is the comparison of the 2003-2016 land water storage anomaly and other methods in the Qinghai-Tibet Plateau of the GRACE inversion restored by the convolutional neural network in an embodiment of the present invention, as shown in Fig. 4, (a), (b), (c) and (d) respectively represent the trend value of land water storage anomalies retrieved by GRACE restored by convolutional neural network, the trend value of scale factor method, the trend value of Mascon data officially released by CSR, and the Mascon data officially released by JPL The trend value of , (e) represents the time series of four types of terrestrial water storage anomalies.

The method for recovering abnormal land water storage from satellite gravity data using machine learning provided by the present invention is a method for recovering the leakage signal of abnormal land water storage based on GRACE gravity satellite level-2 spherical harmonic data. The model data is processed in the same way as the GRACE spherical harmonic data, and the convolutional neural network model is used to obtain a model that can be used for leak signal recovery. Compared with the existing technology, the method of the present invention can avoid the lack of prior knowledge when correcting the signal leakage error, does not rely on a single hydrological model, can better reduce the influence of leakage error, and can accurately invert the anomaly of land water storage space details.

Those skilled in the art can understand that the accompanying drawing is only a schematic diagram of an embodiment, and the modules or processes in the accompanying drawing are not necessarily necessary for implementing the present invention.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it can still be Modifications are made to the technical solutions described in the foregoing embodiments, or equivalent replacements are made to some of the technical features; these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (8)

1. A method for recovering terrestrial water reserve anomalies from satellite gravity data using machine learning, comprising:

step S1: acquiring satellite data of a GRACE gravity satellite, land-water reserve data of a hydrological model and boundary and longitude and latitude information of a preset area, and setting boundary and longitude and latitude information of a background area according to a range in which gravity signals of the preset area are likely to leak, wherein the background area is an area which is out of the range of the preset area and contains most of leaked signals;

step S2: according to the data obtained in the step S1 and longitude and latitude information of the preset area and the background area, calculating the abnormal smooth land water reserves inverted by GRACE gravity satellites in the preset area and the background area;

and step S3: acquiring month scale land water reserve abnormal data of a global hydrological model, and performing leveling operation according to a research time period;

and step S4: forward modeling is carried out on land water reserve abnormal data of the whole world, errors in high-order spherical harmonic coefficients are reduced through low-pass filtering, and smoothed land water reserve abnormal data are obtained according to the boundary and longitude and latitude information of a background area;

step S5: constructing and training a neural network model;

step S6: and (3) inputting the smoothed land water storage quantity abnormal data obtained by the background region GRACE gravity satellite inversion in the step (S2) into a trained neural network model to recover the leakage signal.

2. The method for recovering terrestrial water reserves abnormality from satellite gravity data using machine learning as claimed in claim 1 wherein GRACE gravity satellite data is level-2 spherical harmonic coefficient product, which represents terrestrial water reserves information including surface water, soil water and ground water; the acquired hydrological models comprise five hydrological models of WGHM, noah, CLSM, mosaic and VIC, and land water reserve data of the hydrological models comprise the sum of snow water equivalent output results, soil water reserve output results and underground water reserve output results simulated by the hydrological models.

3. The method for recovering terrestrial water reservoir anomaly from satellite gravity data by using machine learning according to claim 2, wherein the step S2 of calculating the terrestrial water reservoir anomaly from the satellite data of the GRACE gravity satellite is a process of converting level-2 spherical harmonic data containing random errors and system errors at a high order into terrestrial water reservoir anomaly data with a smoothed background region by constructing an earth gravity field model, and converting the terrestrial water reservoir anomaly data into a smoothed late terrestrial water reservoir anomaly value according to latitude and longitude information of the background region through low pass filtering, band-removing filtering and distance leveling operations.

4. The method for recovering terrestrial water reserve abnormality from satellite gravity data by machine learning according to claim 3, wherein the low pass filtering is gaussian filtering with a radius of 300km, the de-banding filtering is preferably P3M10, that is, the first 10 x 10 order spherical harmonic coefficients are kept unchanged, a 3 rd order polynomial is used to fit the spherical harmonic coefficients with a value greater than or equal to 10 th order, the odd and even orders are fitted separately, and the fitted value is subtracted from the original spherical harmonic coefficients, and the method comprises:

cutting off a GRACE level-2 spherical harmonic coefficient product to a spherical harmonic coefficient of a certain order;

removing the 0 order term;

replacing the C20 item by adopting the result of the satellite laser ranging data calculation;

the Technical Note TN-13 data published by JPL was used to replace the 1 st order term of the level-2 spherical harmonic coefficient product.

5. The method for recovering terrestrial water reserves anomalies from satellite gravity data by machine learning as claimed in claim 4, wherein step S2 further comprises the conversion between spherical harmonics and terrestrial water reserves anomalies, which is the smoothed grid data of the level-2 spherical harmonics of the GRACE gravity satellite into background regions, specifically comprising:

step S201: assuming that the earth mass weight distribution causing gravity anomaly is only concentrated on the earth surface layer, mainly comprising the changes of atmosphere, sea, ice cover and land water reserves, converting the density change corresponding to the mass weight distribution into the caused ground level surface height change by using radial integration, wherein the change comprises two parts of contribution of ground mass direct gravity attraction to the ground level surface change and load deformation caused by solid earth after the ground surface load change, and the formula is expressed as formula (1),

in the formula, theta represents earth complementary latitude, lambda represents earth east longitude, l and m respectively represent the order and the number of times of gravity field spherical harmonic expansion, and a dimensionless coefficient delta C _lm And Δ S _lm Representing the spherical harmonic coefficient of the change of the geohorizon,

is a standard associated Legendre function of order l m, ρ _e Representing the mean density of the earth, k _l Expressing the load lux number of the order l, wherein delta sigma (theta, lambda) is the change of the mass density of the earth surface, and a is the average radius of the earth;

step S202: if the change of the earth surface mass density delta sigma (theta, lambda) is also subjected to spherical harmonic expansion, then

In the formula, ρ _w Rho is the density of water, and is known from the formula (1) _w The delta sigma/rho is often adopted for surface mass density conversion _w Is expressed in terms of, thus yielding:

in the formula (I), the compound is shown in the specification,

and

spherical harmonic coefficients for surface density variations;

step S203: obtaining spherical harmonic coefficient of surface density variation according to formula (1) and formula (3)

And

spherical harmonic coefficient Delta C of large ground level surface change _lm And Δ S _lm The functional relationship between the two is as follows:

step S204: substituting formula (4) into formula (2) yields:

wherein, the formula (5) is a basic formula for calculating the change of the earth surface mass density by using the terrestrial level spherical harmonic coefficient data published by the GRACE data center, namely a basic conversion formula between the spherical harmonic coefficient and the land water reserve abnormality;

step S205: when only the land water reserve change is considered, the land water reserve change is estimated according to the acquired spherical harmonic coefficient of the large ground level change provided by the GRACE satellite in combination with the formula (1).

6. The method of claim 5, wherein the step S3 comprises:

step S301: according to the global longitude and latitude, the soil water reserves, the surface water reserves and other land water reserve change components of the five hydrological models are respectively extracted through the following formula, so that five land water reserve change data are obtained:

TWS＝SMS+GWS+SWS+OS

wherein TWS is land water reserve, SMS is soil water reserve, GWS is groundwater reserve, SWS is surface water reserve, OS is other land water reserve change component except SMS, GWS and SWS;

step S302: and (4) carrying out distance processing on the obtained land water reserves data, namely deducting the average land water reserves of the research time period from the value of each month to obtain land water reserve abnormal data of the background area simulated by the hydrological model.

7. The method for recovering terrestrial water reserve abnormality from satellite gravity data by machine learning according to claim 6, wherein the forward modeling in step S4 is to convert the terrestrial water reserve abnormality data simulated by the hydrological model into spherical harmonic coefficients, refer to equation (3), and perform a data processing procedure similar to the spherical harmonic coefficients of the gravity satellite on the spherical harmonic coefficients, and the specific processing procedure includes:

the low-pass filtering adopts Gaussian filtering with the radius of 300 km;

removing the 0 order term;

cutting off the GRACE level-2 spherical harmonic coefficient product to a spherical harmonic coefficient of a certain order;

in step S4, land and water storage quantity abnormal data after smoothing is obtained according to the boundary and longitude and latitude information of the background area and is converted according to the formula (5).

8. The method of claim 7, wherein the step S5 comprises:

step S501: taking the land water reserve abnormal value after smoothing of the background area as input data of the neural network model, taking the land water reserve abnormal value of the preset area as target data of the neural network model, and establishing the neural network model;

step S502: the training of the model is done using Python language.

Images