Lake and reservoir total phosphorus and total nitrogen concentration inversion method and system based on Sentinel-2 image data

Abstract

The invention discloses a lake and reservoir total phosphorus and total nitrogen concentration inversion method and system based on Sentinel-2 image data. The method comprises the following steps: S1, acquiring image data of a satellite Sentinel-2 sensor in a research area and extracting all wave band data; s2, constructing an inversion prediction model and performing model training by using the sample data set, wherein the inversion prediction model extracts and calculates a feature set required by total phosphorus prediction and a feature set required by total nitrogen prediction from all waveband data; s3, performing inversion operation on the basis of the feature set required by total phosphorus prediction by the inversion prediction model to obtain total phosphorus prediction concentration, and performing inversion operation on the basis of the feature set required by total nitrogen prediction by the inversion prediction model to obtain total nitrogen prediction concentration. According to the inversion prediction model, waveband data can be extracted from image data of a satellite Sentinel-2 sensor in a research area and subjected to inversion processing, and the total phosphorus prediction concentration and the total nitrogen prediction concentration of the large-area research area can be rapidly obtained.

Description

Technical Field

[0001] The present invention relates to the field of lake and reservoir water quality environmental monitoring, and in particular to a method and system for inverting total phosphorus and total nitrogen concentrations in lakes and reservoirs based on Sentinel-2 image data. Background Art

[0002] A large number of lakes are polluted by eutrophication to varying degrees. The concentrations of TN and TP in some inland lakes are constantly increasing, and there is a trend towards eutrophic lakes. Therefore, studying the changes in TN and TP concentrations in lakes and reservoirs and predicting their trends are of great significance to lake water quality monitoring and management. At present, the monitoring of total nitrogen TN and total phosphorus TP in lakes is mainly based on manual sampling monitoring, and the sampling method has a lot of limitations in terms of manpower and resources. Traditional monitoring methods cost a lot of manpower and time, and can only obtain point-like and non-continuous monitoring results, and are also severely restricted by traffic conditions. There is a physical and chemical relationship between total nitrogen TN, total phosphorus TP and water quality parameters. How to extract spectral feature information based on lake and reservoir water image data and construct a method that can invert and predict total nitrogen TN and total phosphorus TP concentrations is the current direction of technological development. It can also reduce the error caused by replacing the water quality concentration of surrounding waters with the water quality concentration of the sampling point to a certain extent, which is conducive to monitoring the water quality of large areas of water. Summary of the invention

[0003] The purpose of the present invention is to provide a method and system for inverting the total phosphorus and total nitrogen concentrations of lakes and reservoirs based on Sentinel-2 image data. The inversion prediction model uses a sample data set for model training. The inversion prediction model extracts band data from the image data of the Sentinel-2 sensor. Based on the model training of the sample data set, the inversion prediction model obtains the feature set required for total phosphorus prediction and the weight parameters of the total phosphorus concentration label, and obtains the feature set required for total nitrogen prediction and the weight parameters of the total nitrogen concentration label; the optimal inversion prediction model after model training is selected to invert the band data extracted from the image data of the satellite Sentinel-2 sensor in the study area, so as to obtain the predicted total phosphorus concentration and the predicted total nitrogen concentration.

[0004] The purpose of the present invention is achieved through the following technical solutions: A method for inverting total phosphorus and total nitrogen concentrations in lakes and reservoirs based on Sentinel-2 image data, the method comprising: S1. Collect image data from the satellite Sentinel-2 sensor in the study area and extract all band data; S2. Construct an inversion prediction model and use the sample data set to train the model. The inversion prediction model extracts and calculates the feature set required for total phosphorus prediction and the feature set required for total nitrogen prediction from all band data; S3. The inversion prediction model obtains the predicted total phosphorus concentration by inversion operation based on the feature set required for total phosphorus prediction, and the inversion prediction model obtains the predicted total nitrogen concentration by inversion operation based on the feature set required for total nitrogen prediction.

[0005] In order to better implement the inversion method of total phosphorus and total nitrogen concentration in lakes and reservoirs of the present invention, the model training method of the inversion prediction model is as follows: A sample data set was constructed, in which data were stored in association with band data, total phosphorus concentration labels, and total nitrogen concentration labels. The inversion prediction model extracted the feature set required for total phosphorus prediction from the band data and the total phosphorus concentration labels for weight parameter training. The inversion prediction model extracted the feature set required for total nitrogen prediction from the band data and the total nitrogen concentration labels for weight parameter training.

[0006] Preferably, all band data in the image data of the Sentinel-2 sensor include band data of M bands, each band data includes a band name, a band number, and a band reflectivity; information gain and Pearson correlation coefficient processing are performed on the M bands to obtain a Pearson correlation matrix, and the first M1 bands with the lowest correlation are screened from the Pearson correlation matrix, M1<M; band combination processing is performed based on the screened M1 bands to obtain an initial feature set.

[0007] Preferably, the band combination processing method includes: A1, single band combination: use the single band data in M1 bands as the initial feature; A2, two-band combination: any two bands among the M1 bands are processed as follows: , ,in represents the band reflectivity of band i among the M1 bands, Indicates the band reflectivity of band j in M1 bands; the calculation result , , ,as the initial features; A3. Three-band combination: Perform the following operations on any three bands among the M1 bands and use the results as the initial features: ,in represents the band reflectivity of band k among M1 bands; The initial features obtained by methods A1 to A3 respectively constitute an initial feature set.

[0008] Preferably, the inversion prediction model uses the sample data set to perform recursive feature elimination and principal component analysis based on the initial feature set and performs initial feature dimensionality reduction on the initial feature set, thereby obtaining the feature set required for total phosphorus prediction and the feature set required for total nitrogen prediction.

[0009] Preferably, the inversion prediction model is constructed based on a neural network structure including a residual network, and the inversion prediction model adopts a determination coefficient when using a sample data set for model training. As a model evaluation indicator; , where n represents the total number of samples y, represents the sample value, Indicates the corresponding The predicted value of represents the mean of the sample y.

[0010] A total phosphorus and total nitrogen concentration inversion system for lakes and reservoirs based on Sentinel-2 image data comprises a data acquisition and extraction module, an inversion prediction model and a sample data set, wherein the data acquisition and extraction module is used to acquire image data of a satellite Sentinel-2 sensor in a study area and extract all band data; the data in the sample data set is stored in association with band data, total phosphorus concentration labels and total nitrogen concentration labels; the inversion prediction model extracts and calculates a feature set required for total phosphorus prediction and a feature set required for total nitrogen prediction from all band data; the inversion prediction model uses the sample data set for model training; the inversion prediction model performs an inversion operation based on the feature set required for total phosphorus prediction to obtain a total phosphorus predicted concentration; the inversion prediction model performs an inversion operation based on the feature set required for total nitrogen prediction to obtain a total nitrogen predicted concentration; the data acquisition and extraction module acquires image data of a satellite Sentinel-2 sensor in the study area at the prediction time and extracts all band data at the prediction time to input into the inversion prediction model; the inversion prediction model obtains a total phosphorus predicted concentration and a total nitrogen predicted concentration through an inversion operation.

[0011] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) The inversion prediction model of the present invention uses a sample data set for model training. The inversion prediction model extracts band data from the image data of the Sentinel-2 sensor. Based on the model training of the sample data set, the inversion prediction model obtains the feature set required for total phosphorus prediction and the weight parameters of the total phosphorus concentration label, and obtains the feature set required for total nitrogen prediction and the weight parameters of the total nitrogen concentration label; the optimal inversion prediction model after model training is selected to invert the band data extracted from the image data of the satellite Sentinel-2 sensor in the study area, so as to obtain the predicted total phosphorus concentration and the predicted total nitrogen concentration.

[0012] (2) Based on the differences in spectral characteristics of total nitrogen (TN) and total phosphorus (TP), the present invention proposes an inversion prediction model for water quality parameters based on multi-temporal Sentinel-2 images by selecting the band feature combination that reflects the best reaction rate. Compared with traditional monitoring methods that require huge manpower and time costs and are restricted by traffic conditions, and can only obtain point-like and non-continuous monitoring results, the present invention monitors the image data at a certain point in time through remote sensing technology, and performs inversion prediction of total phosphorus and total nitrogen concentrations through an inversion prediction model.

[0013] (3) The present invention constructs an inversion prediction model for inverting the total nitrogen and total phosphorus contents of key lakes and reservoirs. The key lakes and reservoirs are taken as the research area, and the contents of total nitrogen and total phosphorus are taken as the research objects. Different single-band, two-band and three-band combinations are used as features. The relative error between the predicted value and the measured value is small, and the concentrations of total nitrogen and total phosphorus in lakes and reservoirs can be quickly predicted over a large area. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a method flow chart of the method for inverting the total phosphorus and total nitrogen concentrations of lakes and reservoirs of the present invention; Figure 2 Schematic diagram of the principle of the method for inverting the total phosphorus and total nitrogen concentrations of lakes and reservoirs in the embodiment; Figure 3 It is a correlation matrix diagram of the Pearson correlation matrix in the embodiment; Figure 4 It is a rendering of the inversion prediction model in the total nitrogen concentration inversion in the embodiment; Figure 5 This is a diagram showing the effect of the inversion prediction model in the embodiment on the inversion of total phosphorus concentration. DETAILED DESCRIPTION

[0015] The present invention is further described in detail below in conjunction with embodiments: Example

[0016] like Figure 1 , Figure 2 As shown in FIG. 1 , a method for inverting total phosphorus and total nitrogen concentrations in lakes and reservoirs based on Sentinel-2 image data includes: S1. Collect image data from the Sentinel-2 satellite sensor in the study area and extract all band data. The Sentinel-2 satellite sensor includes two satellites, 2A and 2B. Select Sentinel-2A and Sentinel-2B image data with less cloud cover and better quality in the study area. Figure 2GEE (Google Earth Engine) is the Google Earth Engine platform, and S2-L2A represents the Sentinel-2 sensor. The band parameters of the Sentinel-2 sensor (including Sentinel-2A and Sentinel-2B image data) are shown in Table 1:

[0017] This embodiment takes a key lake as an example of the research object (i.e., the study area), collects data from three field water quality testing stations corresponding to the satellite Sentinel-2 sensor (the three field water quality testing stations correspond to the image data of the satellite Sentinel-2 sensor) and screens them (selecting data with less cloud cover and better quality), with a time span from January 7, 2020 to February 27, 2023.

[0018] S2. Construct an inversion prediction model and use the sample data set to train the model. The inversion prediction model extracts and calculates the feature set required for total phosphorus prediction and total nitrogen prediction from all band data. The model training method of the inversion prediction model is as follows: Construct a sample data set, in which data is stored in association with band data, total phosphorus concentration labels, and total nitrogen concentration labels. The inversion prediction model extracts the feature set required for total phosphorus prediction from the band data and the total phosphorus concentration label for weight parameter training, such as Figure 5 As shown in the figure, in the initial stage of training and testing of the inversion prediction model (i.e. the initial stage of model training), both the training inversion results and the test inversion results showed a good fit and were distributed around the total phosphorus concentration label (the true value of the total phosphorus concentration), with little difference from the total phosphorus concentration label. The inversion prediction model extracts the feature set required for total nitrogen prediction from the band data and the total nitrogen concentration label for weight parameter training, such as Figure 4 As shown, in the initial stages of training and testing of the inversion prediction model (that is, the initial stage of model training), both the training inversion results and the test inversion results showed a good fit, and were distributed around the total nitrogen concentration label (the true value of the total nitrogen concentration), with little difference from the total nitrogen concentration label.

[0019] All band data in the image data of the Sentinel-2 sensor include band data of M bands, and each band data includes a band name, a band sequence number, and a band reflectivity. In this embodiment, the M bands include 1-Coastalaerosol, 2-Blue, 3-Green, 4-Red, 5-Red Edge1, 6-Red Edge2, 7-Red Edge3, 8-NIR, 8a-RedEdge4, 9-Water Vapor, 10-SWIR–Cirrus, 11-SWIR1, and 12-SWIR2, a total of 13 bands.

[0020] The M bands are processed with information gain and Pearson correlation coefficient to obtain the Pearson correlation matrix, and the first M1 bands with the lowest correlation are selected from the Pearson correlation matrix, where M1<M. In one embodiment, the Sentinel-2 sensor selects 1-Coastal aerosol (B1 for short), 2-Blue (B2 for short), 3-Green (B3 for short), 4-Red (B4 for short), 5-Red Edge1 (B5 for short), 6-Red Edge2 (B6 for short), 7-Red Edge3 (B7 for short), 8-NIR (B8 for short), 8a-Red Edge4 (B8A for short), 9-Water Vapor (B9 for short), 11-SWIR1 (B111 for short), and 12-SWIR2 (B12 for short) for a total of 12 bands; in feature engineering, single-band combination and multi-band combination will generate a large number of features for use, some feature vectors carry rich information, some feature vectors carry overlapping information, and some feature vectors are irrelevant features. Too many single-band combinations will cause the dimensionality disaster problem and even reduce the accuracy of the model; due to the unclear spectral response mechanism of total nitrogen TN and total phosphorus TP, there is information redundancy in the band combination, so feature screening is required to exclude invalid / redundant features and select useful features as training data for the model. The methods used include information gain and Pearson correlation matrix, such as Figure 3 As shown, Figure 3The correlation matrix data of 12 bands are given. The data in the grid in the figure represent the correlation data of two bands (taking the grid in the lower left corner as an example, the correlation between band B1 and band B9 is 0.81). Set the correlation threshold (for example, the correlation threshold is 0.9) to divide the correlation matrix into two categories: greater than the correlation threshold and less than or equal to the correlation threshold. The ones greater than the correlation threshold are represented by red (the darker the red, the higher the correlation, which is the band that needs to be removed first), and the ones less than or equal to the correlation threshold are represented by blue (the darker the blue, the lower the correlation, which is the band that needs to be retained first). Select the first M1 bands with the lowest average correlation value from the Pearson correlation matrix (for example, if there are 6, you can choose B1, B11, B12, B2, B8 and B9).

[0021] Based on the selected M1 bands, band combination processing is performed to obtain an initial feature set. The band combination processing method includes: A1, single band combination: use the single band data in M1 bands as the initial feature; A2, two-band combination: any two bands among the M1 bands are processed as follows: , ,in represents the band reflectivity of band i among the M1 bands, Indicates the band reflectivity of band j in M1 bands; the calculation result , , as the initial features; A3. Three-band combination: Perform the following operations on any three bands among the M1 bands and use the results as the initial features: ,in represents the band reflectivity of band k among M1 bands; All the initial features obtained by methods A1 to A3 constitute an initial feature set.

[0022] In the key lakes used as the study area in this embodiment, the top M1 bands with the lowest correlation are B1, B11, B12, B2, B8 and B9. According to methods A1 to A3, the initial feature sets of total nitrogen TN and total phosphorus TP are obtained respectively.

[0023] The inversion prediction model uses a sample data set to perform recursive feature elimination processing and principal component analysis processing based on the initial feature set, and performs initial feature dimensionality reduction on the initial feature set. In the process of data processing and analysis, in order to efficiently and accurately construct a feature space with a specified dimension, the present invention adopts a strategy of combining principal component analysis PCA with recursive feature elimination (English full name Recursive Feature Elimination, referred to as REF) algorithm. First, a recursive feature elimination algorithm is introduced. By constructing a prediction model (the present invention preferably uses the SVR linear support vector machine regression algorithm), the importance of each feature is evaluated in each iteration, and those features that contribute the least to the model performance are gradually removed until the preset number of features is reached or the model performance is no longer significantly improved. Then, the initial feature set is preliminarily processed using principal component analysis technology, aiming to map high-dimensional data to low-dimensional space through linear transformation, while retaining the main information in the original data as much as possible, that is, the feature direction with the largest variance, further effectively reducing the dimension of the data and removing redundancy and noise.

[0024] The ultimate goal is to construct a feature space that not only meets the specified dimensional requirements but also has a high amount of information and predictive efficiency, providing strong support for subsequent model training and predictive analysis.

[0025] After the initial feature set is processed by recursive feature elimination and principal component analysis, the feature set required for total phosphorus prediction and the feature set required for total nitrogen prediction are obtained. In the key lakes used as the study area in this embodiment, after the initial feature set is processed by recursive feature elimination and principal component analysis, the feature set required for total nitrogen TN prediction includes: , , Examples of feature sets required for total phosphorus TP prediction include: , , .

[0026] If the key lake used as an example in this embodiment is Dongqian Lake, taking the Sentinel-2 satellite image data obtained on June 14, 2023 as an example, Dongqian Lake is taken as a specific research area, and the concentration distribution of total nitrogen TN, a key parameter of water quality in the area, is inverted. The concentration distribution of total phosphorus TP, a key parameter of water quality in the area, is inverted.

[0027] In some embodiments, the inversion prediction model is constructed based on a neural network structure including a residual network, and the inversion prediction model uses a determination coefficient when training the model using a sample data set. As the model evaluation index (the present invention selects the determination coefficient The inversion prediction model with a larger value is regarded as the optimal model, and the model parameters in the inversion prediction model are regarded as the optimal parameters); , where n represents the total number of samples y, represents the sample value, Indicates the corresponding The predicted value of It represents the average value of sample y. The inversion prediction model with a larger value is the most preferred post-training model, with a greater goodness of fit, a higher degree of explanation of the independent variable on the dependent variable, and a denser number of observation points near the regression line.

[0028] S3. The inversion prediction model obtains the predicted total phosphorus concentration by inversion operation based on the feature set required for total phosphorus prediction, and the inversion prediction model obtains the predicted total nitrogen concentration by inversion operation based on the feature set required for total nitrogen prediction.

[0029] A total phosphorus and total nitrogen concentration inversion system for lakes and reservoirs based on Sentinel-2 image data comprises a data acquisition and extraction module, an inversion prediction model and a sample data set, wherein the data acquisition and extraction module is used to acquire image data of a satellite Sentinel-2 sensor in a study area and extract all band data; the data in the sample data set is stored in association with band data, total phosphorus concentration labels and total nitrogen concentration labels; the inversion prediction model extracts and calculates a feature set required for total phosphorus prediction and a feature set required for total nitrogen prediction from all band data; the inversion prediction model uses the sample data set for model training; the inversion prediction model performs an inversion operation based on the feature set required for total phosphorus prediction to obtain a total phosphorus predicted concentration; the inversion prediction model performs an inversion operation based on the feature set required for total nitrogen prediction to obtain a total nitrogen predicted concentration; the data acquisition and extraction module acquires image data of a satellite Sentinel-2 sensor in the study area at the prediction time and extracts all band data at the prediction time to input into the inversion prediction model; the inversion prediction model obtains a total phosphorus predicted concentration and a total nitrogen predicted concentration through an inversion operation.

[0030] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for inverting total phosphorus and total nitrogen concentrations in lakes and reservoirs based on Sentinel-2 image data, characterized by: The methods include: S1. Collect image data from the satellite Sentinel-2 sensor in the study area and extract all band data; S2. Construct an inversion prediction model and use the sample data set to train the model. The inversion prediction model extracts and calculates the feature set required for total phosphorus prediction and the feature set required for total nitrogen prediction from all band data; S3. The inversion prediction model obtains the predicted total phosphorus concentration by inversion operation based on the feature set required for total phosphorus prediction, and the inversion prediction model obtains the predicted total nitrogen concentration by inversion operation based on the feature set required for total nitrogen prediction.

2. The method for inverting total phosphorus and total nitrogen concentrations in lakes and reservoirs based on Sentinel-2 image data according to claim 1 is characterized in that: The model training method of the inversion prediction model is as follows: A sample data set was constructed, in which data were stored in association with band data, total phosphorus concentration labels, and total nitrogen concentration labels. The inversion prediction model extracted the feature set required for total phosphorus prediction from the band data and the total phosphorus concentration labels for weight parameter training. The inversion prediction model extracted the feature set required for total nitrogen prediction from the band data and the total nitrogen concentration labels for weight parameter training.

3. The method for inverting total phosphorus and total nitrogen concentrations in lakes and reservoirs based on Sentinel-2 image data according to claim 1 is characterized in that: All band data in the image data of Sentinel-2 sensor include band data of M bands, each of which contains band name, band number and band reflectance; information gain and Pearson correlation coefficient are processed for M bands to obtain Pearson correlation matrix, and the first M1 bands with the lowest correlation are selected from the Pearson correlation matrix, where M1<M; band combination processing is performed based on the selected M1 bands to obtain the initial feature set.

4. The method for inverting total phosphorus and total nitrogen concentrations in lakes and reservoirs based on Sentinel-2 image data according to claim 3 is characterized in that: Band combination processing methods include: A1, single band combination: use the single band data in M1 bands as the initial feature; A2, two-band combination: any two bands among the M1 bands are processed as follows: , ,in represents the band reflectivity of band i among the M1 bands, Indicates the band reflectivity of band j in M1 bands; the calculation result , , as the initial features; A3. Three-band combination: Perform the following operations on any three bands among the M1 bands and use the results as the initial features: ,in represents the band reflectivity of band k among M1 bands; The initial features obtained by methods A1 to A3 respectively constitute an initial feature set.

5. The method for inverting total phosphorus and total nitrogen concentrations in lakes and reservoirs based on Sentinel-2 image data according to claim 3 is characterized in that: The inversion prediction model uses the sample data set to perform recursive feature elimination and principal component analysis based on the initial feature set, and then performs initial feature dimensionality reduction on the initial feature set to obtain the feature set required for total phosphorus prediction and the feature set required for total nitrogen prediction.

6. The method for inverting total phosphorus and total nitrogen concentrations in lakes and reservoirs based on Sentinel-2 image data according to claim 1, characterized in that: The inversion prediction model is constructed based on a neural network structure including a residual network. The inversion prediction model uses a determination coefficient when training the model using a sample data set. As a model evaluation indicator; , where n represents the total number of samples y, represents the sample value, Indicates the corresponding The predicted value of represents the mean of the sample y.

7. A lake and reservoir total phosphorus and total nitrogen concentration inversion system based on Sentinel-2 image data, characterized by: The invention comprises a data acquisition and extraction module, an inversion prediction model and a sample data set. The data acquisition and extraction module is used to acquire the image data of the satellite Sentinel-2 sensor in the study area and extract all the band data. The data in the sample data set is stored in association with the band data, the total phosphorus concentration label and the total nitrogen concentration label. The inversion prediction model extracts and calculates the feature set required for total phosphorus prediction and the feature set required for total nitrogen prediction from all the band data. The inversion prediction model uses the sample data set for model training. The inversion prediction model performs an inversion operation based on the feature set required for total phosphorus prediction to obtain the total phosphorus predicted concentration. The inversion prediction model performs an inversion operation based on the feature set required for total nitrogen prediction to obtain the total nitrogen predicted concentration. The data acquisition and extraction module acquires the image data of the satellite Sentinel-2 sensor in the study area at the prediction time and extracts all the band data at the prediction time to input into the inversion prediction model. The inversion prediction model obtains the total phosphorus predicted concentration and the total nitrogen predicted concentration through the inversion operation.

Images