GRACE downscaling method combining machine learning and dual water cycle model

Abstract

The present invention discloses a GRACE downscaling method combining machine learning and a binary water cycle model, and relates to the field of remote sensing technology. Compared with previous groundwater monitoring methods, the present invention solves the problems of large uncertainty and poor interpretability of existing GRACE downscaling methods; by combining GRACE and GLDAS data, a machine learning downscaling model based on the physical mechanism of binary water cycle is constructed, which can obtain long-term and high-spatial resolution groundwater reserve change information that is continuous in time and space, breaking the limitations of traditional ground monitoring in space and time, and can provide data support for the study of groundwater reserve changes in small and medium-scale areas where measured data are missing. In the process of machine learning model training, the physical mechanism of binary water cycle is introduced, and the machine learning model with physical mechanism is coupled. The attribution and interpretability of machine learning in solving geographical problems are increased, providing a feasible solution for the current combination of remote sensing science and artificial intelligence.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing technology, and in particular to a GRACE downscaling method combining machine learning and a binary water cycle model. Background Art

[0002] As a major component of water resources, groundwater plays a vital role in ecology, agriculture and domestic water use. According to statistics, about 1.5 billion people in the world get their drinking water from groundwater. Under the combined influence of climate change and human activities, the amount of groundwater resources has decreased significantly, causing a series of ecological and environmental problems such as groundwater level decline, ground subsidence, and land salinization. Traditional groundwater storage change monitoring generally uses ground monitoring wells for experiments. However, this method mainly relies on field measurements of groundwater levels at a single point, which is difficult to fully reflect the spatial and temporal variation characteristics of groundwater reserves, and is also limited by corresponding national policies, the number of stations, and instrument accuracy. The GRACE gravity satellite is a new means of monitoring terrestrial water storage changes, breaking the spatial and temporal limitations of traditional ground observations, and is widely used in groundwater storage change research. However, the current spatial resolution of GRACE products is up to 0.25°×0.25°, and the actual resolution is 1°×1°, which is difficult to support small and medium-scale regional groundwater research. Therefore, it is necessary to combine downscaling methods to gain a deeper understanding of the spatiotemporal distribution characteristics of groundwater storage changes.

[0003] The existing GRACE downscaling can be divided into dynamic downscaling and statistical downscaling. Dynamic downscaling is to assimilate GRACE data into the global climate model, but the accuracy of this method depends on the total error covariance matrix of GRACE and hydrological models, which not only increases uncertainty but also is relatively complex. Statistical downscaling achieves downscaling by establishing the relationship between input variables and target variables. Since the relationship between various hydrological elements and groundwater is nonlinear, machine learning models are widely used in downscaling research. However, machine learning models are not constrained by physical mechanisms, especially when targeting geographical problems, they have the disadvantages of poor attribution and interpretability. Coupling machine learning models with physical mechanisms is a key scientific issue in the current realization of the combination of remote sensing science and artificial intelligence. In summary, traditional groundwater monitoring methods are difficult to fully reflect the spatial and temporal variation characteristics of groundwater reserves. Although GRACE gravity satellite monitoring is not restricted by time and space, it is difficult to support the study of groundwater reserves in small and medium-scale regions.

[0004] In order to solve the above problems, the present invention proposes a GRACE downscaling method combining machine learning and a binary water cycle model. Summary of the invention

[0005] The purpose of the present invention is to propose a GRACE downscaling method combining machine learning and a binary water cycle model to solve the problems raised in the background technology:

[0006] The existing GRACE downscaling methods have problems such as large uncertainty and poor interpretability.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] The GRACE downscaling method combining machine learning and a dual water cycle model includes the following steps:

[0009] S1: Combine GRACE and GLDAS to obtain continuous groundwater storage change information in time and space;

[0010] S2: Based on the ERA5 and WaterGAP hydrological models, a dataset of key elements of the binary water cycle is constructed;

[0011] S3: Based on the water balance principle, a binary water cycle model is constructed;

[0012] S4: Under the framework of the geographic random forest algorithm, a geographic random forest downscaling model with a binary water cycle physical mechanism is constructed to obtain high-resolution groundwater storage change information in spatiotemporal and temporal continuity;

[0013] S5: Verify the downscaling results through measured groundwater level data.

[0014] Preferably, the GRACE is used to monitor changes in terrestrial water reserves, which include changes in soil water reserves, snow water reserves and groundwater reserves.

[0015] Preferably, the water balance equation for the change in terrestrial water reserves is expressed as:

[0016] TWSA=SMSA+SWEA+GWSA

[0017] Among them, TWSA, SMSA, SWEA and GWSA represent changes in terrestrial water storage, soil water storage, snow water storage and groundwater storage, respectively.

[0018] Preferably, ERA5 in S2 is used to provide monthly water volume and energy cycle related data with a resolution of 0.1°×0.1°; and the WaterGAP hydrological model is used to provide monthly groundwater extraction values ​​with a spatial resolution of 0.5°×0.5°.

[0019] Preferably, the binary water cycle key element data set constructed in S2 includes precipitation, evaporation, runoff, and groundwater extraction.

[0020] Preferably, the binary water cycle process in the binary water cycle model in S3 is as follows:

[0021] Global climate change affects the precipitation process by affecting atmospheric circulation and temperature field changes;

[0022] Precipitation provides a source of replenishment for surface water, and the water that falls to the surface is divided into four parts:

[0023] The first part infiltrates into the soil to form soil water, which infiltrates again to form groundwater, which is stored underground or in the social water cycle process;

[0024] The second part enters the social water cycle process, which enters the atmosphere through evaporation, forms precipitation again after cooling and condensation, or forms surface runoff through the drainage process;

[0025] The third part forms surface runoff, which also enters the atmosphere through evaporation and forms precipitation again after cooling and condensation;

[0026] The fourth part enters the atmosphere directly through evaporation, cools and condenses to form precipitation.

[0027] Preferably, the social water cycle process includes a water intake process, a water use process and a water discharge process.

[0028] Preferably, the groundwater recharge expression is:

[0029] R e =P-ET-RW

[0030] Among them, R e , P, ET, R, and W represent groundwater recharge, precipitation, evapotranspiration, runoff depth, and groundwater extraction, respectively.

[0031] Preferably, the construction of the geographic random forest downscaling model in S4 is as follows:

[0032] S4.1: Based on the geographic random forest model, the study area is divided according to the geographical characteristics of different regions in the study area;

[0033] S4.2: Unify the binary water cycle key element data set to a spatial resolution of 0.25°×0.25°, and input it into the geographic random forest model as an independent variable; use the groundwater storage change data with a spatial resolution of 0.25°×0.25° as the target variable of the geographic random forest model; select 80% of the data as sample data and 20% of the data as verification data, and construct the geographic random forest downscaling model according to the partitioning of S4.1; verify the simulation accuracy of the geographic random forest downscaling model through correlation coefficient, mean absolute error, root mean square error and Nash efficiency coefficient;

[0034] S4.3: Make a difference between the groundwater storage change data with a spatial resolution of 0.25°×0.25° obtained by the geographic random forest downscaling model simulation and the groundwater storage change data inverted by GRACE; use the Kriging method to convert the difference results with a spatial resolution of 0.25°×0.25° into the difference results with a spatial resolution of 0.1°×0.1°; use the difference results with a spatial resolution of 0.1°×0.1° as the error value simulated by the geographic random forest downscaling model;

[0035] S4.4: The binary water cycle key element dataset with a spatial resolution of 0.1°×0.1° is used as a prediction dataset and input into the geographic random forest downscaling model to obtain groundwater storage change data with a spatial resolution of 0.1°×0.1°; the error value in S4.2 is subtracted from the obtained result to weaken the error caused by the model simulation, and finally the downscaling of the groundwater storage change data is achieved.

[0036] Preferably, in S5, the water level data is processed by anomaly processing, the average water level depth in the study period is deducted, and then multiplied by the corresponding regional water supply degree to obtain the equivalent water column height of the water reserve change, so as to convert the groundwater level data into a groundwater reserve change value equivalent to that of gravity satellite monitoring:

[0037] ΔGWS 实测 =μ×ΔH

[0038] Among them, ΔGWS 实测 is the measured groundwater level data; μ is the dimensionless comprehensive water supply coefficient; ΔH is the groundwater level change.

[0039] Compared with the prior art, the present invention provides a GRACE downscaling method combining machine learning and a binary water cycle model, which has the following beneficial effects:

[0040] The present invention combines GRACE and GLDAS data to construct a machine learning downscaling model based on the physical mechanism of binary water cycle, which can obtain long-term and high-spatial resolution groundwater reserve change information that is continuous in time and space, breaking the limitations of traditional ground monitoring in space and time, and can provide data support for the study of groundwater reserve changes in small and medium-scale areas where measured data are missing. In the process of machine learning model training, the physical mechanism of binary water cycle is introduced, and the machine learning model with physical mechanism is coupled. The attribution and interpretability of machine learning in solving geographical problems are increased, providing a feasible solution for the current combination of remote sensing science and artificial intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1This is a flow chart of the method mentioned in Example 1 of the present invention;

[0042] Figure 2 This is a technical flow chart of extracting groundwater reserves changes by combining GRACE and GLDAS mentioned in Example 1 of the present invention;

[0043] Figure 3 This is a schematic diagram of the binary water cycle mentioned in Example 1 of the present invention;

[0044] Figure 4 This is a schematic diagram of the geographic random forest model mentioned in Example 1 of the present invention;

[0045] Figure 5 A flow chart for building the geographic random forest downscaling model mentioned in Example 1 of the present invention. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0047] The present invention combines GRACE and GLDAS data to construct a machine learning downscaling model based on the physical mechanism of binary water cycle, which can obtain long-term and high-spatial resolution groundwater reserve change information that is continuous in time and space, breaking the limitations of traditional ground monitoring in space and time, and can provide data support for the study of groundwater reserve changes in small and medium-scale areas where measured data are missing. In the process of machine learning model training, the physical mechanism of binary water cycle is introduced, and the machine learning model with physical mechanism is coupled. The attribution and interpretability of machine learning in solving geographical problems are increased, providing a feasible solution for the current combination of remote sensing science and artificial intelligence. Specifically include the following contents.

[0048] Embodiment 1:

[0049] See also Figure 1-5 The present invention combines machine learning with the GRACE downscaling method of the binary water cycle model, including:

[0050] S1: Combine GRACE and GLDAS to obtain the temporal and spatial continuous groundwater storage change information; the details are as follows:

[0051] GRACE monitors changes in terrestrial water storage, which include changes in soil water storage, snow water storage, vegetation canopy water storage, surface water storage, and groundwater storage. Changes in groundwater storage can be obtained by deducting other water components from changes in terrestrial water storage. my country's arid areas are widely distributed, and the impact of surface water on the overall water storage is relatively small. Therefore, surface water was not included in the overall research scope during the research process. In addition, when processing GLDAS (Global Land Data Assimilation System) data, it was found that changes in vegetation canopy water storage were very weak, and it had almost no effect on the overall trend of changes in terrestrial water storage. Therefore, changes in vegetation canopy water storage were ignored. In summary, the balance equation for changes in terrestrial water storage can be expressed as:

[0052] TWSA=SMSA+SWEA+GWSA

[0053] Among them, TWSA, SMSA, SWEA, and GWSA represent the change in terrestrial water storage, soil water storage, snow water storage, and groundwater storage, respectively. The method for calculating the change in groundwater storage is based on the above formula, referring to Figure 2 The terrestrial water storage change data are extracted from GRACE, and the soil water storage change and snow water storage change are simulated by the GLDAS hydrological model.

[0054] S2: Based on the ERA5 and WaterGAP hydrological models, a dataset of key elements of the binary water cycle is constructed; the details are as follows:

[0055] ERA5 is the fifth generation of atmospheric reanalysis dataset released by ECMWF (European Centre for Medium-Range Weather Forecasts), which provides monthly water and energy cycle data with a resolution of 0.1°×0.1° since 1950. The WaterGAP hydrological model was developed by the University of Kassel and the University of Frankfurt. The water use model included in the hydrological model can provide monthly groundwater extraction values ​​with a spatial resolution of 0.5°×0.5° from 1901 to 2016. This step can use ERA5 and WaterGAP data to construct a binary water cycle key element dataset.

[0056] S3: Based on the water balance principle, a binary water cycle model is constructed; the details are as follows:

[0057] Under the background of global climate change and intensified human activities, the water cycle process in the natural state of the river basin is constantly undergoing profound changes. With the influence of human beings, the traditional natural water cycle is gradually forming a system with deep coupling between nature and human society in the region. The regional water cycle has undergone significant changes in its circulation path and characteristics, and has gradually evolved into a "nature-society" dual water cycle model. The schematic diagram of the dual water cycle is shown in the figure below. Figure 3 shown.

[0058] Depend on Figure 3 It can be seen that global climate change affects the precipitation process by affecting atmospheric circulation and temperature. Precipitation is an important source of surface water and groundwater recharge. Part of the water that falls to the ground flows on the surface to form surface runoff, and the other part forms groundwater runoff through infiltration. Water bodies on land form water vapor through evaporation from the ground and water surface and transpiration from vegetation, which is carried to the sky above the land by air currents, cooled and condensed, and finally forms precipitation again. In addition, each link in the natural water cycle will affect the amount of natural recharge of groundwater. Humans obtain water resources through the exploitation of surface water and groundwater, and use them in agriculture, industry and various aspects of life, and finally discharge wastewater in the form of runoff. In addition to the conventional water extraction process, the implementation of inter-basin water transfer projects has introduced water resources outside the basin into the regional water cycle, alleviated the water resource pressure in the region, and increased surface runoff. The social water cycle is integrated into and participated in all links of the natural water cycle. The over-exploitation of groundwater generated by it is the main factor leading to the continuous decline of groundwater levels. Therefore, the recharge of groundwater can be expressed as:

[0059] R e =P-ET-RW

[0060] Among them, R e , P, ET, R, and W represent groundwater recharge, precipitation, evapotranspiration, runoff depth, and groundwater extraction, respectively.

[0061] S4: Under the framework of the geographic random forest algorithm, a geographic random forest downscaling model with a binary water cycle physical mechanism is constructed to obtain high-resolution groundwater storage change information in spatiotemporal continuity; the details are as follows:

[0062] The random forest model is a self-service sampling technique that randomly extracts samples and features, builds multiple unrelated decision trees through the bagging algorithm, and finally votes on the modeling results of the decision trees to obtain the final prediction result. The geographic random forest is a machine learning algorithm that takes into account both geographic weighting and random forest ideas. It can be divided according to the geographic characteristics of different small areas in the study area and perform random forest modeling separately. The specific process of the geographic random forest model is as follows Figure 4 The geographic random forest model is suitable for solving spatial nonlinear problems. The algorithm is easy to implement, the simulation results have spatial local effects, and the accuracy is high. It has been widely used in hydrological simulation and prediction, image recognition and other research.

[0063] Reference Figure 5 ,The construction of the geographic random forest downscaling model is mainly divided into the following four steps:

[0064] Step 1: Based on the geographic random forest model, the study area is divided according to the geographical characteristics of different areas in the study area, so as to construct the model in each small area later.

[0065] Step 2: Unify the binary water cycle key element data set to a spatial resolution of 0.25°×0.25°, and input it as an independent variable into the geographic random forest model; use the groundwater storage change data with a spatial resolution of 0.25°×0.25° (the actual spatial resolution is 1°×1°) as the target variable of the geographic random forest model; select 80% of the data as sample data and 20% of the data as verification data, and construct the geographic random forest downscaling model based on the partition in step 1; verify the model simulation accuracy through correlation coefficient, mean absolute error, root mean square error and Nash efficiency coefficient.

[0066] Step 3: Make a difference between the groundwater storage change data (simulated data) with a spatial resolution of 0.25°×0.25° obtained by the geographic random forest downscaling model simulation and the groundwater storage change data (original data) inverted by GRACE; use the Kriging method to convert the difference results with a spatial resolution of 0.25°×0.25° into a difference result with a spatial resolution of 0.1°×0.1°; and use the difference result with a spatial resolution of 0.1°×0.1° as the error value simulated by the geographic random forest downscaling model.

[0067] Step 4: The binary water cycle key element dataset with a spatial resolution of 0.1°×0.1° is used as the prediction dataset and input into the geographic random forest downscaling model to obtain groundwater storage change data with a spatial resolution of 0.1°×0.1°. The error value obtained in step 2 is subtracted from the obtained result to reduce the error caused by the model simulation, and finally the groundwater storage change data is downscaled.

[0068] S5: Verify the downscaling results by measuring groundwater level data. The details are as follows:

[0069] The available groundwater level data are usually point features or line features in the form of contour lines. In order to convert the groundwater level data into groundwater storage changes equivalent to those monitored by gravity satellites, it is necessary to first perform anomaly processing on the water level data, deduct the average water level depth during the study period, and then multiply it by the corresponding regional water supply to obtain the equivalent water column height of the water storage change. The calculation formula is as follows:

[0070] ΔGWS 实测 =μ×ΔH

[0071] Among them, ΔGWS 实测is the measured groundwater level data; μ is the dimensionless comprehensive water supply coefficient, which represents the volume of water released per unit area of ​​the aquifer under the action of gravity when the groundwater level of the aquifer decreases by one unit; ΔH is the change in groundwater level.

[0072] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. The GRACE downscaling method combining machine learning and a binary water cycle model is characterized by: The steps include: S1: Combine GRACE and GLDAS to obtain continuous groundwater storage change information in time and space; S2: Based on the ERA5 and WaterGAP hydrological models, a data set of key elements of the binary water cycle is constructed, which specifically includes: precipitation, evaporation, runoff, and groundwater extraction; S3: Based on the water balance principle, a binary water cycle model is constructed; S4: Under the framework of the geographic random forest algorithm, a geographic random forest downscaling model with a binary water cycle physical mechanism is constructed to obtain high-resolution groundwater storage change information in spatiotemporal continuity. The specific contents include the following: S4.1: Based on the geographic random forest model, the study area is divided into different zones according to the geographical characteristics of different regions in the study area; S4.2: Unify the binary water cycle key element data set to a spatial resolution of 0.25°×0.25° and input it into the geographic random forest model as an independent variable; use the groundwater storage change data with a spatial resolution of 0.25°×0.25° as the target variable of the geographic random forest model; select 80% of the data as sample data and 20% of the data as verification data, and construct the geographic random forest downscaling model based on the partition in S4.1; verify the simulation accuracy of the geographic random forest downscaling model through correlation coefficient, mean absolute error, root mean square error and Nash efficiency coefficient; S4.3: Make a difference between the groundwater storage change data with a spatial resolution of 0.25°×0.25° obtained by the geographic random forest downscaling model simulation and the groundwater storage change data inverted by GRACE; use the Kriging method to convert the difference results with a spatial resolution of 0.25°×0.25° into the difference results with a spatial resolution of 0.1°×0.1°; use the difference results with a spatial resolution of 0.1°×0.1° as the error value simulated by the geographic random forest downscaling model; S4.4: The binary water cycle key element data set with a spatial resolution of 0.1°×0.1° is used as a prediction data set and input into the geographic random forest downscaling model to obtain groundwater storage change data with a spatial resolution of 0.1°×0.1°; the error value in S4.2 is subtracted from the obtained result to weaken the error caused by the model simulation, and finally the groundwater storage change data is downscaled; S5: Verify the downscaling results through measured groundwater level data; specifically include the following: The water level data was processed by anomaly, the average water level depth during the study period was deducted, and then multiplied by the corresponding regional water supply degree to obtain the equivalent water column height of the water storage change. In this way, the groundwater level data was converted into a groundwater storage change value equivalent to that of gravity satellite monitoring: in, It is the measured groundwater level data; is the dimensionless comprehensive water supply coefficient; is the groundwater level change.

2. The GRACE downscaling method of combining machine learning and a binary water cycle model according to claim 1, characterized in that: The GRACE is used to monitor changes in terrestrial water reserves, which include changes in soil water reserves, snow water reserves, and groundwater reserves.

3. The GRACE downscaling method for combining machine learning and a binary water cycle model according to claim 2, characterized in that: The water balance equation for the change in terrestrial water storage is expressed as: in, , , , They represent changes in land water storage, soil water storage, snow water storage and groundwater storage respectively.

4. The GRACE downscaling method for combining machine learning and a binary water cycle model according to claim 1, characterized in that: The ERA5 in S2 is used to provide monthly water volume and energy cycle related data with a resolution of 0.1°×0.1°; the WaterGAP hydrological model is used to provide monthly groundwater extraction values ​​with a spatial resolution of 0.5°×0.5°.

5. The GRACE downscaling method for combining machine learning and a binary water cycle model according to claim 1, characterized in that: The water cycle process in the binary water cycle model in S3 is as follows: Global climate change affects the precipitation process by affecting atmospheric circulation and temperature field changes; Precipitation provides replenishment for surface water and groundwater. The water that falls to the surface is divided into four parts: The first part infiltrates into the soil to form soil water, which infiltrates again to form groundwater, which is stored underground or enters the social water cycle; The second part enters the social water cycle process, which enters the atmosphere through evaporation, forms precipitation again after cooling and condensation, or forms surface runoff through the drainage process; The third part forms surface runoff, which also enters the atmosphere through evaporation and forms precipitation again after cooling and condensation; The fourth part enters the atmosphere directly through evaporation, cools and condenses to form precipitation.

6. The GRACE downscaling method of combining machine learning and a binary water cycle model according to claim 5, characterized in that: The social water cycle process includes water intake process, water use process and water discharge process.

7. The GRACE downscaling method of combining machine learning and a binary water cycle model according to claim 5, characterized in that: The groundwater recharge expression is: in, , , , , They represent groundwater recharge, precipitation, evapotranspiration, runoff depth and groundwater extraction respectively.

Images