Method and equipment for constructing large-scale lake water salinity estimation model

Abstract

The invention discloses a method and equipment for constructing a large-scale lake water body salinity estimation model. Lake form data, lake geographical attribute data and hydrological characteristic data, climate characteristic data, land utilization ratio and soil component ratio in a basin where a lake is located are used as input characteristic variables, and actually measured lake salinity data are used as output; utilizing a machine learning XGBoost algorithm to construct a lake water body salinity estimation model and carrying out model training; models trained based on different training sets are constructed through random experiments, and the model with the optimal precision is selected as a lake water body salinity estimation model for unknown lake water body salinity estimation. According to the method disclosed by the invention, the unknown lake water salinity is estimated only by using the on-site measurement of the lake water salinity and the public acquisition of the related physical characteristic variables generated by the data source, so that the method has relatively high model precision; and the time and the economic cost for monitoring the water salinity of the lake which cannot be actually measured and investigated in an unknown remote area are greatly reduced.

Description

Technical Field

[0001] The present invention belongs to the field of lake hydrology and machine learning algorithms, and in particular relates to a method and device for constructing a large-scale lake water salinity estimation model. Background Art

[0002] Salinity is a basic attribute of natural water bodies and a basic parameter that describes the physical and chemical properties of water bodies. It plays an important role in the hydrology, water ecology, and water environment of lakes. It is very sensitive to the influence of geographical environment, climate factors, and human activities. Lake salinity has a profound impact on microorganisms and water resource management in the ecosystem, including the distribution and reproduction of aquatic organisms, the sustainable use of water resources, and the maintenance of ecological balance. In recent years, under the conditions of warming and wetter climate, the physical, chemical, and biological properties of many lakes have changed significantly. The area of ​​many salt lakes in the world is shrinking at an alarming rate, and the salinity of terminal lakes in the inland basins of the Qinghai-Tibet Plateau has generally decreased, resulting in a simplification of the structure of phytoplankton ecosystems. Therefore, monitoring the salinity of lake water bodies is crucial for ecological environmental protection and effective management of water resources.

[0003] In the early days, the monitoring process of lake water salinity was relatively cumbersome. Usually, water samples were collected, evaporated to dryness, and then the residual salt mass was measured to calculate the water salinity. At present, for lakes that can be reached by transportation, multi-parameter water quality monitoring instruments are mostly used to directly read the salinity of lake water. However, in some areas with harsh environmental conditions and relatively lack of accessibility, the monitoring of lake water salinity is still in a relatively difficult state. Although the development of remote sensing technology in recent years has provided a solution to this problem. Relevant studies have shown that in coastal and estuarine environments, colored soluble organic matter (CDOM) is correlated with salinity. Therefore, optical remote sensing data is used to indirectly invert the relationship between water surface reflectance and water salinity through CDOM as a bridge to connect the relationship between water surface reflectance and water salinity. In addition, the use of various prediction models to establish a mathematical relationship between water surface reflectance and water salinity can also effectively invert the surface salinity of lake water. However, the current research on inverting lake salinity using remote sensing means mainly remains in the qualitative research and exploration stage, and most of them are for salinity inversion of a single lake, and it is impossible to carry out regional large-scale research. Thanks to the availability of observational data and greatly improved computing power, various machine learning algorithms have been widely used in the field of water quality monitoring. Although most of these studies have focused on lake water quality aspects such as temperature, dissolved oxygen, chlorophyll and transparency, they have demonstrated the powerful ability of machine learning methods in estimating water quality parameters. However, the application of these machine learning algorithms to estimate lake salinity at a regional scale deserves further study.

[0004] In summary, lake salinity is a key factor affecting lake water environment and aquatic ecosystem, and has a significant impact on lake science, ecology, and water safety and management. Existing survey and research methods cannot conduct on-site surveys of many lakes due to harsh climate conditions and lack of accessible sites, resulting in the salinity of many lakes remaining unknown. Further, remote sensing technology is currently limited to estimating the salinity of a single lake, and it is still challenging to obtain lake water salinity over a large area. Summary of the invention

[0005] The first purpose of the present invention is to overcome the problems existing in the above-mentioned prior art and provide a method for estimating the salinity of large-scale lake water bodies.

[0006] In order to achieve the above technical objectives, the present invention adopts the following technical solutions: Large-scale lake water salinity estimation methods include: Using lake morphology data, lake geographic attribute data, hydrological characteristics data, climate characteristics data, land use ratio and soil composition ratio in the lake basin as input feature variables and measured lake salinity data as output, the machine learning XGBoost algorithm was used to build a lake water salinity estimation model and conduct model training. Through random experiments, models based on different training sets were constructed, and the model with the best accuracy was selected as the lake water salinity estimation model for estimating the salinity of unknown lake water.

[0007] As a preferred implementation, the lake morphology data includes lake area, water volume and average water depth; The geographic attribute data includes the lake's longitude, latitude, altitude, and distance around the lake x The average slope within the range, x is the preset threshold; The hydrological characteristic data include average annual runoff, lake recharge coefficient, river network density and water exchange time; The climate characteristic data include average annual rainfall, average annual evaporation, net rainfall and average annual temperature; The land use ratio includes grassland ratio, bare land ratio, urban ratio, forest land ratio and cultivated land ratio; The soil component ratio includes soil surface cation exchange capacity, total nitrogen content and pH value.

[0008] As a preferred implementation, the lake morphology data and geographic attribute data are obtained based on the HydroLAKES dataset, which is lake spatial distribution vector data.

[0009] As a preferred implementation, lake data with salinity ≥ 300 g / L are removed and used for model training.

[0010] As a preferred implementation, data screening is also performed based on the relationship between lake area, water volume, rainfall and evaporation and the salinity of the corresponding lake water body, including: Define the parameter E / SQ, where E is the average annual evaporation, S is the lake area, and Q is the lake water volume. If the E / SQ of a lake is greater than 10, the data of the lake will be discarded and not used for model training.

[0011] As a preferred embodiment, the method further comprises, before model training, calculating the correlation between the input feature variables and the corresponding lake salinity, and eliminating the feature variables with a correlation with salinity less than 0.1. 2 For lake basins, river network density is retained as one of the input characteristic variables.

[0012] As a preferred implementation, the acquired input characteristic variable data and the corresponding measured lake salinity data are used to construct different training sets through random experiments, and model training is performed using different training sets, and the model with the highest accuracy is selected as the lake salinity estimation model.

[0013] As a preferred implementation, the input characteristic variable data is normalized and used as model input data; the measured salinity data is logarithmically transformed and used as model output data.

[0014] As a preferred embodiment, based on the coefficient of determination (R 2 ), mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE) and relative root mean square error (RRMSE) five statistical indicators to evaluate the accuracy of the model, and select the best model as the estimation model of lake water salinity based on the model accuracy evaluation results. Furthermore, it also includes eliminating some overfitting models based on the accuracy evaluation results of the training set, and selecting the model with the best test set accuracy from the remaining models as the final model.

[0015] As a preferred embodiment, the method further includes adjusting parameters based on the model training process to determine the universal structural parameters of the XGBoost machine learning model, and performing random experiments using the determined universal structural parameters. Without considering the operation time, a universal model structural parameter is obtained, which tends to improve the model prediction accuracy and generalization ability as much as possible. Preferably, the adjustment parameters include the learning rate, the maximum depth of the tree, and the minimum leaf node sample weight.

[0016] As a preferred implementation, the sample data set is randomly divided into training data and test data in a ratio of 8:2, and a model based on different training sets is constructed through 1000 random experiments.

[0017] The second purpose of the present invention is to provide a method for estimating the salinity of large-scale lake water bodies, using the lake salinity estimation model constructed by the above-mentioned modeling method, obtaining relevant parameters of the lake to be estimated according to the input parameters required by the lake salinity estimation model, and inputting them into the lake salinity estimation model to obtain the lake salinity estimation value.

[0018] As a preferred implementation, the lake is a lake in a plateau area, and the input parameters required for the lake salinity estimation model include: Geographic attribute data: lake longitude, latitude, altitude, and distance around the lake x The average slope within the range, x is the preset threshold; Lake morphological characteristic variables: lake area, water volume and average water depth; Hydrological characteristic data: average annual runoff, lake recharge coefficient and river network density; Climate characteristic data: average annual rainfall, average annual evaporation, net rainfall and average annual temperature; Land use percentage: grassland percentage, bare land percentage, urban percentage and cultivated land percentage; Soil composition ratio: cation exchange capacity of soil surface layer, total nitrogen content and pH value.

[0019] The third object of the present invention is to provide a computer device, which includes a memory, a processor and a lake salinity estimation model stored in the memory, characterized in that the lake salinity estimation model is constructed based on the above-mentioned modeling method; the processor obtains the input parameters of the model and performs model calculation on the lake salinity estimation model.

[0020] The present invention uses a machine learning algorithm to mine the intrinsic relationship between the water salinity data of a certain number of measured lake samples and their own lake morphology data and geographical attribute data as well as the climate data of the lake basin where the lake is located, and other physical characteristic variable data, and constructs a quantitative lake water salinity estimation model, thereby realizing the recording of unknown lake water salinity. The present invention has the following beneficial effects: (1) The lake water salinity estimation model construction method proposed in the present invention can obtain high model accuracy and significantly reduce the workload of field measurements. It provides a feasible technical means for lakes where field measurements cannot be carried out, and is particularly suitable for large-scale, multi-lake field measurements. (2) The algorithm of the present invention is simple to implement and can estimate the salinity of regional lake water bodies with high accuracy and efficiency, and can provide methodological support for regional lake surveys, lake water resources management, lake science-related research, etc. In addition, the invention can be extended to large-scale and even global-scale research, providing methodological support for large-scale lake water salinity assessment and water resources management under the background of global change. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings are not intended to be drawn to scale. In the accompanying drawings, each identical or nearly identical component shown in various figures may be represented by the same reference numeral. For the sake of clarity, not every component is labeled in each figure. Now, the various steps of the present invention will be described by studying a case and referring to the accompanying drawings, in which: Figure 1 (a) is the spatial distribution of the measured salinity data set of the lake water in the embodiment of the present invention; (b) is the statistics of the measured salinity sample set of the lake water in the embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the salt input that affects the salinity of lake water in the embodiment of the present invention. It mainly includes four aspects: climate factors, hydrological factors, human activities and geological and geographical factors.

[0023] Figure 3 It is a correlation diagram between the salinity of the sample lake water and related physical characteristic variables in the embodiment of the present invention.

[0024] Figure 4 It is a main flow chart of the inventive method in an embodiment of the present invention.

[0025] Figure 5 It is the accuracy evaluation result of the lake water salinity estimated based on the optimal model in the test set in the embodiment of the present invention. DETAILED DESCRIPTION

[0026] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and implementation examples. The following implementation examples are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0027] This embodiment takes the measured salinity data of 100 terminal lakes in the inland area of ​​the Qinghai-Tibet Plateau as an example to further describe the technical solution of the present invention. The experimental data selected in this experiment include the measured salinity data of terminal lakes in the Qinghai-Tibet Plateau, the lake morphology and geographical attribute data in the HydroLAKES dataset, the ERA5-Land climate data at the basin level, the SoilGrids global soil attribute data, the ESA-CCI land use data of the European Space Agency, and the basin hydrological attribute data provided in the paper Ongoing Drainage Reorganization Driven by Rapid Lake Growths on the Tibetan Plateau.

[0028] As an exemplary description, the implementation of the above method is specifically described below in conjunction with the accompanying drawings.

[0029] Step 1: Construct model training data set.

[0030] First, in ArcGIS spatial analysis software, obtain the lake’s own morphological data and geographic attribute data in the HydroLAKES dataset and spatially connect them with the known measured lake water salinity data. After the connection is successful, check the corresponding data, remove outliers, and fill in empty values. Then, in the watershed corresponding to the lake, spatially connect the calculated climate characteristic variables, hydrological characteristic variables, land use ratio, and soil attribute ratio with the previously processed data in ArcGIS. After the connection is successful, process them, remove outliers, and fill in missing values. Finally, build a complete model training dataset containing lake water salinity data and related physical characteristic variables. The spatial distribution of the collected measured terminal lake water salinity data is shown in the figure. Figure 1 As shown in (a), the sample data statistics are as follows Figure 1 As shown in (b), it can be seen that the sample set fully reflects the differences in morphology, size, spatial distribution, water salinity, etc., and has good research representativeness.

[0031] Step 2: Processing and screening of target variables and feature variables First, target variables are processed. First, ultra-high salinity lakes with salinity higher than 300 g / L are eliminated. The number of such lakes is small, and using them for model training may make it difficult to build an effective generalization model for such special lakes. Second, the salinity of lake water is comprehensively judged based on lake area, water volume, rainfall and evaporation data to determine whether the salinity of lake water meets the corresponding relationship. For example, some lakes have a large area, more rainfall, less evaporation, but their salinity is very high, which means that the selected physical characteristic variables may not be sufficient to measure the changes in the salinity of these lake waters. Therefore, we define the parameter E / SQ (E is the average annual evaporation, S is the lake area, and Q is the lake water volume). If E / SQ is greater than 10, these lakes are excluded from the construction of the model.

[0032] Secondly, the relevant physical characteristic variables were processed. The Pearson correlation coefficient was used to calculate the correlation between the relevant physical characteristic variables and the salinity of the lake water body, and the relevant physical characteristic variables with a correlation of less than 0.1 with the lake water body salinity were eliminated to reduce variable redundancy, such as Figure 3 It is worth noting that we retain the special variable of river network density as a proxy for smaller basins (usually less than 1000 km 2The annual average runoff of the lake basin unit) is supplemented by the hydrological characteristics. Since the spatial resolution of ERA5-Land is relatively coarse (0.1° × 0.1°, the horizontal resolution is about 10 km), it cannot fully reflect the annual average runoff in small basins. Finally, the area, water volume, average water depth, altitude, longitude and latitude of the lake, the annual average rainfall, annual average evapotranspiration, annual average runoff, annual average temperature, net rainfall, lake recharge coefficient and river network density in the basin where the lake is located, as well as the pH, total nitrogen content, cation exchange capacity of the soil surface in the basin where the lake is located, and the proportion of grassland, cultivated land, urban area and bare land in land use were selected as the final model input.

[0033] Finally, the input feature variables are normalized and the target variables are logarithmically transformed to reduce the error in model training caused by excessive differences in data distribution, thereby increasing the generalization ability of the model.

[0034] Step 3: The construction process of the lake water salinity estimation method based on XGBoost is as follows Figure 4 As shown in the figure, with the help of Python's XGBoost package, the method of determining the structural parameters of the XGBoost machine learning model is to adjust the parameters based on the model training process, including the learning rate, the maximum depth of the tree, and the minimum leaf node sample weight. Without considering the operation time, a universal model structural parameter is obtained, which tends to improve the model prediction accuracy and generalization ability as much as possible.

[0035] In this embodiment, the model parameters obtained by training through the above-mentioned method principle are: the learning rate is 0.1, the maximum depth of the tree is 2 layers, and the minimum leaf node sample weight is 1.

[0036] Step 4: Use the universal model structure parameters determined in step 3 above, randomly split the sample data set into training and test data according to the ratio of 8:2, and build a model based on different training sets through 1000 random experiments.

[0037] Step 5: Based on the test set, use the coefficient of determination (R 2 ), mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE) and relative root mean square error (RRMSE) were used to evaluate the performance of the 1000 test models. First, some overfitting models were eliminated based on the accuracy evaluation results of the training set, and the final model was selected from the remaining models with the best accuracy on the test set.

[0038] The accuracy of lake water salinity estimated based on the optimal model in this example is as follows: Figure 5 The results show that the XGBoost optimal model trained based on the test set has a high accuracy (R 2=0.90, MAE=8.11 g / L, MAPE=36.40%, RMSE=12.51 g / L, RRMSE=36.96 g / L). In summary, the present invention can ensure that the constructed lake water salinity estimation model has high accuracy under the premise of greatly reducing the workload of field measured data, and has great application potential in large-scale lake water salinity modeling and prediction. It is expected to provide a feasible solution for lake water salinity estimation in remote and data-scarce areas, and promote regional lake water resources management.

Claims

1. A method for constructing a large-scale lake water salinity estimation model, characterized in that: include: Using lake morphology data, lake geographic attribute data, hydrological characteristics data, climate characteristics data, land use ratio and soil composition ratio in the lake basin as input feature variables and measured lake salinity data as output, the machine learning XGBoost algorithm was used to build a lake water salinity estimation model and conduct model training. Through random experiments, models based on different training sets were constructed, and the model with the best accuracy was selected as the lake water salinity estimation model for estimating the salinity of unknown lake water.

2. The method according to claim 1, characterized in that: The lake morphological data include lake area, water volume and average water depth; The geographic attribute data includes the lake's longitude, latitude, altitude, and distance around the lake x The average slope within the range, x is the preset threshold; The hydrological characteristic data include average annual runoff, lake recharge coefficient, river network density and water exchange time; The climate characteristic data include average annual rainfall, average annual evaporation, net rainfall and average annual temperature; The land use ratio includes grassland ratio, bare land ratio, urban ratio, forest land ratio and cultivated land ratio; The soil component ratio includes soil surface cation exchange capacity, total nitrogen content and pH value.

3. The method according to claim 1, characterized in that Lake data with salinity ≥ 300 g / L were excluded and used for model training.

4. The method according to claim 1, characterized in that: It also includes data screening based on the relationship between lake area, water volume, rainfall and evaporation and the salinity of the corresponding lake water body, including: Define the parameter E / SQ, where E is the average annual evaporation, S is the lake area, and Q is the lake water volume. If the E / SQ of a lake is greater than 10, the data of the lake will be discarded and not used for model training.

5. The method according to claim 1, characterized in that: It also includes calculating the correlation between the input feature variables and the corresponding lake salinity before model training, and eliminating feature variables with a correlation with salinity of less than 0.

1.

6. The method according to claim 1, characterized in that The obtained input characteristic variable data and the corresponding measured lake salinity data are used to construct different training sets through random experiments. Different training sets are used for model training, and the model with the highest accuracy is selected as the lake salinity estimation model.

7. The method according to claim 1, characterized in that The input characteristic variable data are normalized and used as model input data; the measured salinity data are logarithmically transformed and used as model output data.

8. A method for estimating salinity of large-scale lake water, characterized in that: A lake salinity estimation model is constructed using the modeling method described in any one of claims 1 to 7, and relevant parameters of the lake to be estimated are obtained according to the input parameters required by the lake salinity estimation model, and the parameters are input into the lake salinity estimation model to obtain the lake salinity estimation value.

9. The method according to claim 8, characterized in that: The lake is a lake in the plateau area, and the input parameters required for the lake salinity estimation model include: Geographic attribute data: lake longitude, latitude, altitude, and distance around the lake x The average slope within the range, x is the preset threshold; Lake morphological characteristic variables: lake area, water volume and average water depth; Hydrological characteristic data: average annual runoff, lake recharge coefficient and river network density; Climate characteristic data: average annual rainfall, average annual evaporation, net rainfall and average annual temperature; Land use percentage: grassland percentage, bare land percentage, urban percentage and cultivated land percentage; Soil composition ratio: cation exchange capacity of soil surface layer, total nitrogen content and pH value.

10. A computer device comprising a memory, a processor and a lake salinity estimation model stored in the memory, characterized in that: The lake salinity estimation model is constructed based on the method described in any one of claims 1 to 7; the processor obtains the input parameters of the model and performs model calculation on the lake salinity estimation model.

Images