Basin nutrition load estimation method based on remote sensing data and machine learning

Abstract

The invention discloses a drainage basin nutrition load estimation method based on remote sensing data and machine learning. The method comprises the following steps: constructing target variables of a model, constructing characteristic variables of the model, selecting a combination of the characteristic variables by a search algorithm, establishing the model, and estimating an area to be measured by using the constructed model. According to the method, a GIS analysis tool, data acquisition on a GEE platform and a machine learning model are integrated together by using a python language. According to the method, feature variables are obtained through remote sensing data, drainage basin nutrition load estimation of a data-deficient area is achieved by building a machine learning model, and a scientific basis is provided for drainage basin management and non-point source pollution abatement.

Description

Technical Field

[0001] The present invention relates to the technical field of watershed nutrient load estimation, and in particular to a watershed nutrient load estimation method based on remote sensing data and machine learning. Background Art

[0002] Non-point source nutrient loads (such as total nitrogen and total phosphorus) have a significant impact on water quality at the basin scale. However, it is affected by hydrological processes and land use patterns, and accurate estimation of non-point source nutrient loads has always been a challenge. Physical-based models such as SWAT or material balance methods usually require a large amount of data, resulting in a very complicated calculation process and are not suitable for areas with insufficient data.

[0003] Machine learning models based on remote sensing data are an effective tool. First, data-driven machine learning can provide a model between prediction results and characteristic parameters based on a large amount of statistical inference, and shows good universality and predictive ability in solving nonlinear problems. Secondly, remote sensing data directly or indirectly provide information related to human activities, such as land use type and further available urbanization rate, percentage of crop area and urban area, and terrain data related to nutrient output such as slope aspect and slope. Reasonable use of remote sensing data and machine learning provides a scientific basis for predicting the nutrient load of a watershed. Summary of the invention

[0004] In response to the problems existing in the prior art, the present invention provides a method for estimating the nutrient load of a watershed based on remote sensing data and machine learning, and establishes a machine learning model that uses remote sensing and existing environmental data for transfer learning to estimate the nutrient load of the area to be tested, thereby solving the problem of estimating the nutrient load of a watershed in areas where data is lacking.

[0005] To achieve the above object, the present invention provides the following technical solutions: A method for estimating nutrient load in a watershed based on remote sensing data and machine learning, including Step 1: Construct the target variable of the model Using spatial analysis tools, through the processing and analysis of DEM data, we selected sub-basins with relatively complete data; obtained the nutrient load data of the sub-basin, and used the data as the target variable of the machine learning model;

[0006] Step 2: Construct the characteristic variables of the model Use multi-source remote sensing data to collect and extract appropriate characteristic variables, including land use type, vegetation index, rainfall data, terrain characteristics, and population density;

[0007] Step 3: Search algorithm to select the combination of feature variables Use the search algorithm to determine at least one characteristic variable combination with a fitting degree of more than 80% with the target variable, and the maximum number of characteristic variable combinations is determined by the characteristic vector weight method;

[0008] Step 4: Build the model Use the machine learning model to build models using the feature variable combinations in step 3;

[0009] Step 5: Use the constructed model to estimate the area to be measured Multi-source remote sensing data are used to collect and extract characteristic variables of the target area, and the constructed model is used to estimate the nutrient load of other areas or the same area at different time periods.

[0010] A further improvement plan is to integrate spatial analysis tools, data acquisition on the GEE platform, and machine learning models through the Python language.

[0011] A further improvement plan is that the specific implementation of step one includes: (1) Select sub-basins: Use GIS analysis tools to process and analyze DEM data and select areas with longer time coverage and higher spatial resolution to determine the basin and its sub-basins; (2) Estimation of nutrient load value: Use physical model or material balance method to estimate the nutrient load value of the sub-basin. The data involved mainly come from county and city statistical data; (3) Data cleaning and processing: The estimated nutrient load data are cleaned and processed to ensure the accuracy and completeness of the data. The processed data will be used as the target variable of the machine learning model.

[0012] A further improvement is that in step 2, the feature variables The land use type includes the use of satellite image data or domestic and foreign global land cover databases to calculate the urbanization rate, the percentage of urban area, the percentage of cultivated land area, and the percentage of forest area; The vegetation index includes spectral indexes NDVI, EVI, NDBI, and NDWI calculated from remote sensing data; The rainfall data is obtained by interpolation using the measured data of the global climate database or the basin meteorological monitoring station; The terrain features include slope and aspect, calculated by digital elevation model (DEM), derived from SRTM or ASTERGDEM; The population density is obtained by using the satellite night light data via the GEE platform.

[0013] A further improvement is to set the number of characteristic variables in the combination in step 3 to no more than 6, in order to limit the complexity of the model and the availability of later data.

[0014] A further improvement is that the fourth step of the machine learning model uses the random forest classification method. The specific steps are as follows: (1) Dataset division: The dataset is divided into a training set and a validation set in a ratio of 7:3; (2) Model training: Use the training set data to build a random forest model and set relevant parameters for model training; (3) Model evaluation: Use validation set data to evaluate the predictive performance of the model; (4) Parameter optimization: Optimize model parameters through grid search or random search to improve model performance.

[0015] A further improvement plan is that step five specifically includes the following: (1) Data preparation: Collect and process the characteristic variable data of the area to be monitored to ensure consistency with the model training data; (2) Model prediction: Use the trained machine learning model to predict the nutrient load of the monitored area. Beneficial Effects

[0016] This method constructs a model for predicting nutrient load in a watershed, and it is crucial to select appropriate characteristic variables. It is not easy to obtain data due to the lack of data in the region. The use of remote sensing data to calculate and obtain breaks the lack of regional data. At the same time, the combination of characteristic variables must have a high degree of fit with the target variable, so that the estimated nutrient load value in the watershed can be accurate. Therefore, the combination screening and determination of characteristic variables is also critical. The present invention realizes the estimation of nutrient load in watersheds in areas lacking data by constructing a machine learning model and calculating using remote sensing data, and provides a scientific basis for watershed management and non-point source pollution control. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is the TN spatial distribution map of the Huaihe River Basin (Anhui section) predicted by the model of the present invention (2020); Figure 2 It is a relationship diagram between the TN predicted by the model of the present invention and the measured value (including the verification data of Suzhou City, Bengbu City and Chuzhou City). DETAILED DESCRIPTION

[0018] Based on the estimation of the total nitrogen value of the nutrient load in the Huaihe River Basin (Anhui section) predicted by the model of the present invention, the spatial analysis tool, data acquisition on the GEE platform and the machine learning model are integrated together through the Python language. The steps are as follows: Constructing the target variable of the model

[0019] (1) Select a sub-basin: Use GIS analysis tools to process and analyze DEM data to determine the basin and its sub-basins, and select areas with longer data coverage and higher spatial resolution (Suzhou, Bengbu, and Chuzhou in the Anhui section of the Huaihe River basin); obtain the nutrient load data of the sub-basin, and use the data as the target variable of the machine learning model; (2) Estimate the nutrient load value: Physical models or material balance methods can be used to estimate the nutrient load value of the sub-basin. In this operation, the nutrient load is calculated based on the material balance method and the output coefficient, taking total nitrogen (TN) as an example. The total nitrogen input here mainly considers the use of nitrogen fertilizers, nitrogen emissions from livestock waste, and nitrogen emissions from human sewage and waste. The data involved mainly come from county and city statistical data; (3) Data cleaning and processing: The estimated nutrient load data is cleaned and processed to ensure the accuracy and completeness of the data. The processed data will be used as the target variable of the machine learning model.

[0020] Step 2: Construct the characteristic variables of the model Extract appropriate feature variables from multi-source remote sensing data collected by GEE platform, including Land use type: It can be obtained by classification using satellite image data such as Landsat, MODIS, and Sentinel-2, or by using existing domestic or global land cover databases. The choice depends on the spatial resolution of the target variable to determine its accuracy. Based on the land use type data, the urbanization rate, the percentage of urban area, the percentage of cultivated land area, and the percentage of forest area are calculated.

[0021] Vegetation index: Remote sensing data mainly reflects the surface and vegetation conditions. It is impossible to directly extract indicators related to livestock N emissions and farmland fertilization, but the spectral index extracted from remote sensing data can be used for indirect speculation. For example, vegetation indices such as NDVI and EVI can be used to monitor the vegetation coverage and health of crops and pastures. For this reason, indices such as NDVI, EVI, NDBI, and NDWI were selected during this implementation.

[0022] Precipitation data: You can interpolate the measured data from the basin's meteorological monitoring stations, or you can use a global climate database (such as WorldClim).

[0023] Topographic features: including slope and aspect, calculated by digital elevation model (DEM), derived from SRTM or ASTERGDEM.

[0024] Population density: The GEE platform obtains night light data from satellites.

[0025] The image data of all feature variables and target variables are converted into discrete data according to a 3×3 km grid to facilitate subsequent modeling.

[0026] Step 3: Search algorithm to select the combination of feature variables The search algorithm is used to determine at least one characteristic variable combination with a fit of more than 80% with the target variable. The maximum number of characteristic variable combinations is determined by the characteristic vector weight method. In order to limit the complexity of the model and the accessibility of later data, the algorithm is set to allow no more than 6 features in the combination. The OOB (Out-of-Bag) precision verification technology is used to resample the original data to form a bootstrap sample, which does not rely on additional verification sets or cross-validation. Finally, three or four most characteristic variable combinations are selected.

[0027] Step 4: Build the model The machine learning model in this implementation uses the random forest classification method (RF). As a general classification and regression method, random forest has a mature module. The RF model is established by using the combination of feature variables in step three. (1) Dataset division: The dataset is divided into a training set and a validation set in a ratio of 7:3; (2) Model training: Use the training set data to build a random forest model and set relevant parameters (such as the number of trees, maximum depth, etc.) for model training; (3) Model evaluation: Use validation set data to evaluate the prediction performance of the model. Common evaluation indicators include R², RMSE, etc. (4) Parameter optimization: Optimize model parameters through grid search or random search to improve model performance.

[0028] See Figure 1 As shown in the figure, the model predicted TN values ​​of 30% of the validation sets in the sub-basins of Suzhou, Bengbu and Chuzhou, and the fit degree with the measured values ​​reached more than 86%.

[0029] Step 5: Use the constructed model to estimate the area to be measured Multi-source remote sensing data was used to collect and extract characteristic variables of the target area (Anhui section of the Huaihe River Basin), and the constructed model was used to estimate the nutrient load of other areas in the same time period. The specific steps are as follows: (1) Data preparation: Collect and process the characteristic variable data of the area to be monitored to ensure consistency with the model training data; (2) Model prediction: Use the trained machine learning model to predict the nutrient load of the monitored area.

[0030] The TN spatial distribution map of the Huaihe River Basin (Anhui section) predicted by the model of the present invention (2020) is as follows: Figure 2 shown.

[0031] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A method for estimating nutrient load in a watershed based on remote sensing data and machine learning, characterized by: include Step 1: Construct the target variable of the model Using spatial analysis tools, through the processing and analysis of DEM data, we selected sub-basins with relatively complete data; obtained the nutrient load data of the sub-basin, and used the data as the target variable of the machine learning model; Step 2: Construct the characteristic variables of the model Use multi-source remote sensing data to collect and extract appropriate characteristic variables, including land use type, vegetation index, rainfall data, terrain characteristics, and population density; Step 3: Search algorithm to select the combination of feature variables Use the search algorithm to determine at least one characteristic variable combination with a fitting degree of more than 80% with the target variable, and the maximum number of characteristic variable combinations is determined by the characteristic vector weight method; Step 4: Build the model Use the machine learning model to build models using the feature variable combinations in step 3; Step 5: Use the constructed model to estimate the area to be measured Multi-source remote sensing data are used to collect and extract characteristic variables of the target area, and the constructed model is used to estimate the nutrient load of other areas or the same area at different time periods.

2. The method for estimating nutrient load of a watershed based on remote sensing data and machine learning according to claim 1, characterized in that: Spatial analysis tools, data acquisition on the GEE platform, and machine learning models are integrated through the Python language.

3. The method for estimating nutrient load of a watershed based on remote sensing data and machine learning according to claim 1, characterized in that: The specific implementation of step one includes: (1) Select sub-basins: Use GIS analysis tools to process and analyze DEM data and select areas with longer time coverage and higher spatial resolution to determine the basin and its sub-basins; (2) Estimation of nutrient load value: Use physical model or material balance method to estimate the nutrient load value of the sub-basin. The data involved mainly come from county and city statistical data. (3) Data cleaning and processing: The estimated nutrient load data is cleaned and processed to ensure the accuracy and completeness of the data. The processed data will be used as the target variable of the machine learning model.

4. The method for estimating nutrient load of a watershed based on remote sensing data and machine learning according to claim 1, characterized in that: Step 2: Feature variables The land use type includes the use of satellite image data or domestic and foreign global land cover databases to calculate the urbanization rate, the percentage of urban area, the percentage of cultivated land area, and the percentage of forest area; The vegetation index includes spectral indexes NDVI, EVI, NDBI, and NDWI calculated from remote sensing data; The rainfall data is obtained by interpolation using the measured data of the global climate database or the basin meteorological monitoring station; The terrain features include slope and aspect, calculated by digital elevation model (DEM), derived from SRTM or ASTERGDEM; The population density is obtained by using the satellite night light data via the GEE platform.

5. The method for estimating nutrient load of a watershed based on remote sensing data and machine learning according to claim 1, characterized in that: In step 3, the number of characteristic variables in the combination is set to no more than 6.

6. The method for estimating nutrient load of a watershed based on remote sensing data and machine learning according to claim 1, characterized in that: Step 4: The machine learning model uses the random forest classification method. The specific steps are as follows: (1) Dataset division: The dataset is divided into a training set and a validation set in a ratio of 7:3; (2) Model training: Use the training set data to build a random forest model and set relevant parameters for model training; (3) Model evaluation: Use validation set data to evaluate the predictive performance of the model; (4) Parameter optimization: Optimize model parameters through grid search or random search to improve model performance.

7. The method for estimating nutrient load of a watershed based on remote sensing data and machine learning according to claim 1, characterized in that: Step 5 specifically includes the following: (1) Data preparation: Collect and process the characteristic variable data of the area to be monitored to ensure consistency with the model training data; (2) Model prediction: Use the trained machine learning model to predict the nutrient load of the monitored area.

Images