Water quality monitoring method based on multiple spectrums of unmanned aerial vehicle

Abstract

The invention discloses an unmanned aerial vehicle multispectral-based water quality monitoring method, which comprises the following steps of: constructing a two-layer learning model with a component learning layer and a meta learning layer, and after acquiring spectral characteristic parameters of a monitored water area, inverting the spectral characteristic parameters of the monitored water area by adopting the two-layer learning model to obtain water quality parameter prediction data; wherein the component learning layer is formed by stacking at least two base learning models of different types, each base learning model can independently carry out water quality parameter prediction based on spectral characteristic parameters to obtain a corresponding water quality parameter prediction result, and the meta learning layer is realized by adopting a meta learning algorithm. Performing weighted stacking on the water quality parameter prediction results obtained by the plurality of base learning models by using a weighted average method to obtain a final water quality parameter prediction value; the method has the advantage of high detection precision.

Description

Technical Field

[0001] The invention relates to a water quality monitoring method, and in particular to a water quality monitoring method based on unmanned aerial vehicle multispectrum. Background Art

[0002] The traditional water quality monitoring method mainly involves manual sampling at designated water locations to obtain water samples, which are then sent to laboratories for testing and analysis using laboratory instruments to obtain the required water quality indicators. Although the traditional water quality monitoring method has high accuracy, it takes a long time and can only provide data information at local points, and cannot fully reflect the spatial distribution of water quality.

[0003] In recent years, a water quality monitoring method based on multispectral drones has emerged. This water quality monitoring method uses drones equipped with multispectral sensors to obtain real-time spectral reflectance data of water bodies. Spectral reflectance data can reflect water quality parameters such as suspended matter concentration, chlorophyll content, dissolved organic matter, temperature, pH value and dissolved oxygen. Based on the spectral reflectance data, an inversion model is used to obtain the water quality parameters to be monitored.

[0004] Although the water quality monitoring method based on drone multispectral can realize rapid real-time monitoring of water quality in large areas of water, the inversion model used is a regression model of a single machine learning algorithm. Due to the low concentration of water quality parameters in river water, it has limitations in information mining capabilities and model generalization, making it difficult to fully utilize the potential of spectral reflectance data, resulting in low accuracy in water quality parameter detection. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a water quality monitoring method based on unmanned aerial vehicle multispectrum with high detection accuracy.

[0006] The technical solution adopted by the present invention to solve the above technical problems is: a water quality monitoring method based on unmanned aerial vehicle multispectrum, by constructing a two-layer learning model with a component learning layer and a meta-learning layer, after obtaining the spectral characteristic parameters of the monitored water area, the two-layer learning model is used to invert the spectral characteristic parameters of the monitored water area to obtain water quality parameter prediction data, wherein the component learning layer is formed by stacking at least two base learning models of different types, each base learning model can independently predict water quality parameters based on the spectral characteristic parameters to obtain corresponding water quality parameter prediction results, and the meta-learning layer is implemented by a meta-learning algorithm, and the water quality parameter prediction results obtained by multiple base learning models are weighted and stacked by a weighted average method to obtain the final water quality parameter prediction value.

[0007] The water quality parameter prediction result is either the prediction result of total nitrogen (TN), or the prediction result of total phosphorus (TP), or the prediction result of chlorophyll.

[0008] The meta-learning algorithm is the random forest (RF) algorithm.

[0009] The component learning layer is formed by stacking four base learning models of different types, and the four base learning models are MLR (multivariate linear regression), AdaBoost, GBDT (gradient boosted decision tree) and Catboost.

[0010] Before inverting the spectral reflectance data, the two-layer learning model is trained first. The specific training process is as follows:

[0011] Step 1: Select any water area as the target area in the city where water quality monitoring is required, set multiple evenly distributed sampling points in the target area, and record the number of sampling points as N, which is set according to actual usage requirements;

[0012] Step 2, collecting water samples at each sampling point, storing the water samples in an incubator and then sending them to a laboratory for testing, to obtain measured data of the water quality parameters required to be measured at each sampling point, wherein the water quality parameters required to be measured are total nitrogen (Total Nitrogen) content concentration, total phosphorus (Total Phosphorus) content concentration or chlorophyll a (Chlorophyll-a) content concentration;

[0013] Step 3: Use a multispectral camera equipped with a DJI Phantom 4 multispectral drone to take aerial photos of the target area between 13:00 and 16:00 local time in clear and windless weather, and obtain drone images of each sampling point through the multispectral camera, where the drone images of each sampling point include images of five bands: blue (B), green (G), red (R), red edge (REG), and near infrared (NIR);

[0014] The operating altitude of the DJI Phantom 4 multispectral drone is 120-150m, the heading overlap is set to 80%, and the lateral overlap is set to 70%;

[0015] Step 4: Perform image stitching and radiation correction on the drone image of each sampling point to obtain the spectral reflectance data of the drone image of each sampling point, that is, the spectral characteristic parameters;

[0016] Step 5, the spectral reflectance data of the UAV image of each sampling point is processed using the band sum, band difference and band ratio of any two of the five bands of blue (B), green (G), red (R), red edge (REG) and near infrared (NIR) to obtain the spectral reflectance data of the UAV image under multiple dual-band combinations of each sampling point;

[0017] Step 6: Perform Pearson correlation analysis on the measured data of water quality parameters at each sampling point and the spectral reflectance data of the drone images under each dual-band combination of the sampling point to obtain the correlation between the spectral reflectance data of the drone images under each dual-band combination of each sampling point and the water quality parameters, and use the spectral reflectance data of the drone images under the dual-band combination with the greatest correlation with the water quality parameters among the spectral reflectance data of the drone images under all dual-band combinations of each sampling point as the spectral characteristic parameter of the sampling point;

[0018] Step 7: The spectral characteristic parameters of all sampling points and the measured data of water quality parameters of all sampling points are used to form a data set X. The spectral characteristic parameters of each sampling point and the measured data of water quality parameters constitute a sample of the data set X, X = {(x′1, y′1), (x′2, y′2), ..., (x′ M , y′ M )}, where M represents the number of samples in the data set X, (x′ t , y′ t ) represents the tth sample in the dataset X, x′ t is the spectral characteristic parameter of the tth sample in the data set X, y′ t is the measured data of water quality parameters of the t-th sample in the data set X, t = 1, 2, ..., M;

[0019] A random 70% sample of the data set is used to form the training set Y. If 70% of the data set is not an integer, it is rounded up and the remaining samples form the validation set Z; Y = {(x1, y1), (x2, y2), ..., (x N ,y N )},(x i ,y i ) represents the i-th sample in the training set Y, x i is the spectral feature parameter of the i-th sample in the training set Y, y i is the measured data of water quality parameters of the i-th sample in the training set Y, i = 1, 2, ..., N, N is the number of samples in the training set Y, and (x i ,y i ) is called the i-th observation;

[0020] Step 8: Use the training set Y to train the component learning layer. The specific training process is as follows:

[0021] S8.1. MLR (multivariate linear regression), AdaBoost, GBDT (gradient boosted decision tree) and Catboost are called model 1 to model 4 respectively. The model L is trained using the training set Y, where L = 1, 2, 3, 4. The specific process is as follows:

[0022] S8.1.1. Remove the i-th observation (x i ,y i ), use the remaining N-1 observations to train the model L, and then the model L outputs a predicted value, which is recorded as That is, the prediction result obtained by model L predicting the i-th observation value in the training set Y;

[0023] S8.1.2. Let the weight of model L be w L , initialize the weights w of model L with random values ​​or zero L The value of

[0024] S8.1.3. Construct by adjusting the weight w L The objective function of minimizing the sum of square errors of all observations in the training set Y is shown in formula (1):

[0025]

[0026] in, represents the predicted value of the component learning layer for the i-th observation in the training set Y, represents the square error of the i-th observation in the training set Y, is the objective function, argmin means finding the objective function The parameter value to be minimized, ∑ is the accumulation symbol;

[0027] S8.1.4. Use the gradient descent algorithm to adjust the weight w of the model L L The value of Minimize until the objective function Converge to a minimum value, and set the weight w of the model L at this time L The value of is taken as the weight w of model L L The final value of L ; Among them, when adjusting the weight w L In the process of selecting the value of , the constraints shown in formula (2) and formula (3) need to be satisfied:

[0028] w L ≥0 (2)

[0029]

[0030] S8.1.5. When the optimal weights of the four base learning models are obtained, the optimal weight of each base learning model is assigned to the base learning model, and the component learning layer training is completed;

[0031] Step 9: Train the meta-learning algorithm RF. The specific process is as follows:

[0032] S9.1. Based on the optimal weights of the four base learning models currently obtained, the predicted value of the i-th observation in the training set Y is calculated using formula (4):

[0033]

[0034] Among them, * represents the multiplication symbol;

[0035] S9.2. The currently obtained to As the input data of the meta-learning algorithm RF, the meta-learning algorithm RF is trained to obtain the trained meta-learning algorithm RF. At this time, the two-layer learning model completes one training and obtains the trained two-layer learning model;

[0036] Step 10, use the verification set Z to verify the prediction accuracy of the currently obtained trained two-layer learning model. If the prediction accuracy is not lower than the required accuracy, the verification is qualified, and the currently obtained trained two-layer learning model is the final trained two-layer learning model, which is used for the subsequent inversion of the spectral reflectance data. If the verification fails, return to step 7, re-obtain the training set Y and the verification set Z, and use steps 8 and 9 to train the currently obtained trained two-layer learning model again until the prediction accuracy of the trained two-layer learning model is verified to be qualified.

[0037] The spectral characteristic parameters of the monitored water area are obtained using the same method as steps 3 to 6.

[0038] Compared with the prior art, the advantage of the present invention is that multiple initial prediction results are obtained through the collaborative work of multiple base learning models, that is, the water quality parameter prediction results obtained by each base learning model. Since each base learning model will be trained through a training set before use, and the meta-learning layer is the final water quality parameter prediction result obtained based on the water quality parameter prediction results obtained by multiple base learning models, that is, the meta-learning layer learns from the stacked outputs of multiple base learning models and makes a final prediction, which not only increases the number of training samples of the meta-learning layer and improves the generalization ability of the meta-learning layer, but also makes full use of the unique advantages of each base learning model, so that the final water quality parameter prediction results have higher accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flow chart of the water quality monitoring method based on drone multispectrum of the present invention;

[0040] Figure 2 It is a structural schematic diagram of a two-layer learning model in the water quality monitoring method based on drone multispectral of the present invention;

[0041] Figure 3 This is a prediction chart for predicting chlorophyll water quality using the drone multi-spectral water quality monitoring method of the present invention. DETAILED DESCRIPTION

[0042] The present invention is further described in detail below with reference to the accompanying drawings.

[0043] Embodiment 1: Figure 1 A water quality monitoring method based on unmanned aerial vehicle multispectral is proposed. A two-layer learning model with a component learning layer and a meta-learning layer is constructed. After obtaining the spectral characteristic parameters of the monitored water area, the two-layer learning model is used to invert the spectral characteristic parameters of the monitored water area to obtain water quality parameter prediction data. The component learning layer is formed by stacking at least two base learning models of different types. Each base learning model can independently predict water quality parameters based on the spectral characteristic parameters to obtain corresponding water quality parameter prediction results. The meta-learning layer is implemented by a meta-learning algorithm. The water quality parameter prediction results obtained by multiple base learning models are weighted and stacked by a weighted average method to obtain the final water quality parameter prediction value.

[0044] In this embodiment, multiple initial prediction results are obtained through the collaborative work of multiple base learning models, that is, the water quality parameter prediction results obtained by each base learning model. Since each base learning model is trained with a training set before use, the meta-learning layer is the final water quality parameter prediction result obtained based on the water quality parameter prediction results obtained by multiple base learning models, that is, the meta-learning layer learns from the stacked outputs of multiple base learning models and makes a final prediction, which not only increases the number of training samples of the meta-learning layer and improves the generalization ability of the meta-learning layer, but also makes full use of the unique advantages of each base learning model, so that the final water quality parameter prediction results have higher accuracy.

[0045] Embodiment 2: This embodiment is basically the same as Embodiment 1, except that: in this embodiment, the water quality parameter prediction result is either the prediction result of total nitrogen (TN), or the prediction result of total phosphorus (TP), or the prediction result of chlorophyll.

[0046] In this embodiment, the meta-learning algorithm is a random forest (RF) algorithm.

[0047] In this embodiment, the component learning layer is formed by stacking four base learning models of different types, such as Figure 2 ,The four base learning models are MLR (Multiple Linear Regression), AdaBoost, GBDT (Gradient Boosted Decision Tree) and Catboost.

[0048] Embodiment 3: This embodiment is basically the same as Embodiment 2, except that: in this embodiment, before inverting the spectral reflectance data, the two-layer learning model is first trained. The specific training process is:

[0049] Step 1: Select any water area as the target area in the city where water quality monitoring is required, set multiple evenly distributed sampling points in the target area, and record the number of sampling points as N, which is set according to actual usage requirements;

[0050] Step 2, collecting water samples at each sampling point, storing the water samples in an incubator and then sending them to a laboratory for testing, to obtain measured data of the water quality parameters required to be measured at each sampling point, wherein the water quality parameters required to be measured are total nitrogen (Total Nitrogen) content concentration, total phosphorus (Total Phosphorus) content concentration or chlorophyll a (Chlorophyll-a) content concentration;

[0051] Step 3: Use a multispectral camera equipped with a DJI Phantom 4 multispectral drone to take aerial photos of the target area between 13:00 and 16:00 local time in clear and windless weather, and obtain drone images of each sampling point through the multispectral camera, where the drone images of each sampling point include images of five bands: blue (B), green (G), red (R), red edge (REG), and near infrared (NIR);

[0052] The operating altitude of the DJI Phantom 4 multispectral drone is 120-150m, the heading overlap is set to 80%, and the lateral overlap is set to 70%;

[0053] Step 4: Perform image stitching and radiation correction on the drone image of each sampling point to obtain the spectral reflectance data of the drone image of each sampling point, that is, the spectral characteristic parameters;

[0054] Step 5, the spectral reflectance data of the UAV image of each sampling point is processed using the band sum, band difference and band ratio of any two of the five bands of blue (B), green (G), red (R), red edge (REG) and near infrared (NIR) to obtain the spectral reflectance data of the UAV image under multiple dual-band combinations of each sampling point;

[0055] Step 6: Perform Pearson correlation analysis on the measured data of water quality parameters at each sampling point and the spectral reflectance data of the drone images under each dual-band combination of the sampling point (refer to the Pearson correlation analysis in Chapter 3, 3.2.1 of the paper "Research on Inversion of River Water Quality Parameters Based on UAV Remote Sensing Technology") to obtain the correlation between the spectral reflectance data of the drone images under each dual-band combination of each sampling point and the water quality parameters, and use the spectral reflectance data of the drone images under the dual-band combination with the greatest correlation with the water quality parameters among the spectral reflectance data of the drone images under all dual-band combinations of each sampling point as the spectral characteristic parameter of the sampling point;

[0056] Step 7: The spectral characteristic parameters of all sampling points and the measured data of water quality parameters of all sampling points are used to form a data set X. The spectral characteristic parameters of each sampling point and the measured data of water quality parameters constitute a sample of the data set X, X = {(x′1, y′1), (x′2, y′2), ..., (x′ M , y′ M )}, where M represents the number of samples in the data set X, (x′ t , y′ t ) represents the tth sample in the dataset X, x′ t is the spectral characteristic parameter of the tth sample in the data set X, y′ t is the measured data of water quality parameters of the t-th sample in the data set X, t = 1, 2, ..., M;

[0057] A random 70% sample of the data set is used to form the training set Y. If 70% of the data set is not an integer, it is rounded up and the remaining samples form the validation set Z; Y = {(x1, y1), (x2, y2), ..., (x N ,y N )},(x i ,y i ) represents the i-th sample in the training set Y, x i is the spectral feature parameter of the i-th sample in the training set Y, y i is the measured data of water quality parameters of the i-th sample in the training set Y, i = 1, 2, ..., N, N is the number of samples in the training set Y, and (x i ,y i ) is called the i-th observation;

[0058] Step 8: Use the training set Y to train the component learning layer. The specific training process is as follows:

[0059] S8.1. MLR (multivariate linear regression), AdaBoost, GBDT (gradient boosted decision tree) and Catboost are called model 1 to model 4 respectively. The model L is trained using the training set Y, where L = 1, 2, 3, 4. The specific process is as follows:

[0060] S8.1.1. Remove the i-th observation (x i ,y i ), use the remaining N-1 observations to train the model L, and then the model L outputs a predicted value, which is recorded as That is, the prediction result obtained by model L predicting the i-th observation value in the training set Y;

[0061] S8.1.2. Let the weight of model L be w L , initialize the weights w of model L with random values ​​or zero L The value of

[0062] S8.1.3. Construct by adjusting the weight w L The objective function of minimizing the sum of square errors of all observations in the training set Y is shown in formula (1):

[0063]

[0064] in, represents the predicted value of the component learning layer for the i-th observation in the training set Y, represents the square error of the i-th observation in the training set Y, is the objective function, arg min means finding the objective function The parameter value to be minimized, ∑ is the accumulation symbol;

[0065] S8.1.4. Use the gradient descent algorithm to adjust the weight w of the model L L The value of Minimize until the objective function Converge to a minimum value, and set the weight w of the model L at this time L The value of is taken as the weight w of model L L The final value of L ; Among them, when adjusting the weight w L In the process of selecting the value of , the constraints shown in formula (2) and formula (3) need to be satisfied:

[0066] w L ≥0 (2)

[0067]

[0068] S8.1.5. When the optimal weights of the four base learning models are obtained, the optimal weight of each base learning model is assigned to the base learning model, and the component learning layer training is completed;

[0069] Step 9: Train the meta-learning algorithm RF. The specific process is as follows:

[0070] S9.1. Based on the optimal weights of the four base learning models currently obtained, the predicted value of the i-th observation in the training set Y is calculated using formula (4):

[0071]

[0072] Among them, * represents the multiplication symbol;

[0073] S9.2. The currently obtained to As the input data of the meta-learning algorithm RF, the meta-learning algorithm RF is trained to obtain the trained meta-learning algorithm RF. At this time, the two-layer learning model completes one training and obtains the trained two-layer learning model;

[0074] Step 10, use the verification set Z to verify the prediction accuracy of the currently obtained trained two-layer learning model. If the prediction accuracy is not lower than the required accuracy, the verification is qualified, and the currently obtained trained two-layer learning model is the final trained two-layer learning model, which is used for the subsequent inversion of the spectral reflectance data. If the verification fails, return to step 7, re-obtain the training set Y and the verification set Z, and use steps 8 and 9 to train the currently obtained trained two-layer learning model again until the prediction accuracy of the trained two-layer learning model is verified to be qualified.

[0075] In this embodiment, the spectral characteristic parameters of the monitored water area are obtained using the same method as steps 3 to 6.

[0076] In this embodiment, the Nanxintang River in Ningbo City is taken as the target water area. When the water quality remote sensing inversion is performed on the target water area, the collection of measured data of water quality parameters is carried out simultaneously with the collection of drone images. In the Nanxintang River in Ningbo City, we set 5 evenly distributed sampling points. Each sampling point collected measured data of water quality parameters 9 times. Water samples were collected 0.5m below the water surface, and 1000mL were collected at each sampling point. The water quality parameter chlorophyll (Chl-a) was collected, and a total of 45 measured data of water quality parameters (i.e., 45 chlorophyll sample data) were collected. During the detection process, we followed national standards and industry standards. The maximum, minimum and average values ​​of the 45 chlorophyll sample data collected in the Nanxintang River in Ningbo City are statistically shown in Table 1:

[0077] Table 1 Water quality data statistics

[0078]

[0079] The water quality monitoring method based on multi-spectral unmanned aerial vehicle of the present invention was used to monitor the chlorophyll of Nanxintang River in Ningbo. The monitoring results are as follows: Figure 3 14 chlorophyll sample data were randomly selected from the 45 chlorophyll sample data collected as prediction data, and then the inversion progress was verified. The inversion results were judged by observing the relative error value of the fitting degree between the prediction data and the 1:1 line. Figure 3 It can be seen that the prediction results are relatively good.

Claims

1. A water quality monitoring method based on drone multispectral, characterized in that By constructing a two-layer learning model with a component learning layer and a meta-learning layer, after obtaining the spectral characteristic parameters of the monitored water area, the two-layer learning model is used to invert the spectral characteristic parameters of the monitored water area to obtain water quality parameter prediction data, wherein the component learning layer is formed by stacking at least two base learning models of different types, each base learning model can independently predict water quality parameters based on the spectral characteristic parameters to obtain corresponding water quality parameter prediction results, and the meta-learning layer is implemented using a meta-learning algorithm, and the water quality parameter prediction results obtained by multiple base learning models are weighted stacked by using a weighted average method to obtain a final water quality parameter prediction value.

2. The water quality monitoring method based on drone multispectral according to claim 1 is characterized in that The water quality parameter prediction result is either the prediction result of total nitrogen (TN), or the prediction result of total phosphorus (TP), or the prediction result of chlorophyll.

3. The water quality monitoring method based on drone multispectral according to claim 1 is characterized in that The meta-learning algorithm is the random forest (RF) algorithm.

4. The water quality monitoring method based on drone multispectral according to claim 1 is characterized in that The component learning layer is formed by stacking four base learning models of different types, and the four base learning models are MLR (multivariate linear regression), AdaBoost, GBDT (gradient boosted decision tree) and Catboost.

5. The water quality monitoring method based on drone multispectral according to claim 4 is characterized in that Before inverting the spectral reflectance data, the two-layer learning model is trained first. The specific training process is as follows: Step 1: Select any water area as the target area in the city where water quality monitoring is required, set multiple evenly distributed sampling points in the target area, and record the number of sampling points as N, which is set according to actual usage requirements; Step 2, collecting water samples at each sampling point, storing the water samples in an incubator and then sending them to a laboratory for testing, to obtain measured data of the water quality parameters required to be measured at each sampling point, wherein the water quality parameters required to be measured are total nitrogen (Total Nitrogen) content concentration, total phosphorus (Total Phosphorus) content concentration or chlorophyll a (Chlorophyll-a) content concentration; Step 3: Use a multispectral camera equipped with a DJI Phantom 4 multispectral drone to take aerial photos of the target area between 13:00 and 16:00 local time in clear and windless weather, and obtain drone images of each sampling point through the multispectral camera, where the drone images of each sampling point include images of five bands: blue (B), green (G), red (R), red edge (REG), and near infrared (NIR); The operating altitude of the DJI Phantom 4 multispectral drone is 120-150m, the heading overlap is set to 80%, and the lateral overlap is set to 70%; Step 4: Perform image stitching and radiation correction on the drone image of each sampling point to obtain the spectral reflectance data of the drone image of each sampling point, that is, the spectral characteristic parameters; Step 5, the spectral reflectance data of the UAV image of each sampling point is processed using the band sum, band difference and band ratio of any two of the five bands of blue (B), green (G), red (R), red edge (REG) and near infrared (NIR) to obtain the spectral reflectance data of the UAV image under multiple dual-band combinations of each sampling point; Step 6: Perform Pearson correlation analysis on the measured data of water quality parameters at each sampling point and the spectral reflectance data of the drone images under each dual-band combination of the sampling point to obtain the correlation between the spectral reflectance data of the drone images under each dual-band combination of each sampling point and the water quality parameters, and use the spectral reflectance data of the drone images under the dual-band combination with the greatest correlation with the water quality parameters among the spectral reflectance data of the drone images under all dual-band combinations of each sampling point as the spectral characteristic parameter of the sampling point; Step 7: The spectral characteristic parameters of all sampling points and the measured data of water quality parameters of all sampling points are used to form a data set X. The spectral characteristic parameters of each sampling point and the measured data of water quality parameters constitute a sample of the data set X, X = {(x′1, y′1), (x′2, y′2), ..., (x′ M , y′ M )}, where M represents the number of samples in the data set X, (x′ t , y′ t ) represents the tth sample in the dataset X, x′ t is the spectral characteristic parameter of the tth sample in the data set X, y′ t is the measured data of water quality parameters of the t-th sample in the data set X, t = 1, 2, ..., M; A random 70% sample of the data set is used to form the training set Y. If 70% of the data set is not an integer, it is rounded up and the remaining samples form the validation set Z; Y = {(x1, y1), (x2, y2), ..., (x N ,y N )},(x i ,y i ) represents the i-th sample in the training set Y, x i is the spectral feature parameter of the i-th sample in the training set Y, y i is the measured data of water quality parameters of the i-th sample in the training set Y, i = 1, 2, ..., N, N is the number of samples in the training set Y, and (x i ,y i ) is called the i-th observation; Step 8: Use the training set Y to train the component learning layer. The specific training process is as follows: S8.

1. MLR (multivariate linear regression), AdaBoost, GBDT (gradient boosted decision tree) and Catboost are called model 1 to model 4 respectively. The model L is trained using the training set Y, where L = 1, 2, 3, 4. The specific process is as follows: S8.1.

1. Remove the i-th observation (x i ,y i ), use the remaining N-1 observations to train the model L, and then the model L outputs a predicted value, which is recorded as That is, the prediction result obtained by model L predicting the i-th observation value in the training set Y; S8.1.

2. Let the weight of model L be w L , initialize the weights w of model L with random values ​​or zero L The value of S8.1.

3. Construct by adjusting the weight w L The objective function of minimizing the sum of square errors of all observations in the training set Y is shown in formula (1): in, represents the predicted value of the component learning layer for the i-th observation in the training set Y, represents the square error of the i-th observation in the training set Y, is the objective function, arg min means finding the objective function The parameter value to be minimized, ∑ is the accumulation symbol; S8.1.

4. Use the gradient descent algorithm to adjust the weight w of the model L L The value of Minimize until the objective function Converge to a minimum value, and set the weight w of the model L at this time L The value of is taken as the weight w of model L L The final value of L ; Among them, when adjusting the weight w L In the process of selecting the value of , the constraints shown in formula (2) and formula (3) need to be satisfied: w L ≥0 (2) S8.1.

5. When the optimal weights of the four base learning models are obtained, the optimal weight of each base learning model is assigned to the base learning model, and the component learning layer training is completed; Step 9: Train the meta-learning algorithm RF. The specific process is as follows: S9.

1. Based on the optimal weights of the four base learning models currently obtained, the predicted value of the i-th observation in the training set Y is calculated using formula (4): Among them, * represents the multiplication symbol; S9.

2. The currently obtained to As the input data of the meta-learning algorithm RF, the meta-learning algorithm RF is trained to obtain the trained meta-learning algorithm RF. At this time, the two-layer learning model completes one training and obtains the trained two-layer learning model; Step 10, use the verification set Z to verify the prediction accuracy of the currently obtained trained two-layer learning model. If the prediction accuracy is not lower than the required accuracy, the verification is qualified, and the currently obtained trained two-layer learning model is the final trained two-layer learning model, which is used for the subsequent inversion of the spectral reflectance data. If the verification fails, return to step 7, re-obtain the training set Y and the verification set Z, and use steps 8 and 9 to train the currently obtained trained two-layer learning model again until the prediction accuracy of the trained two-layer learning model is verified to be qualified.

6. The water quality monitoring method based on drone multispectral according to claim 5 is characterized in that The spectral characteristic parameters of the monitored water area are obtained using the same method as steps 3 to 6.

Images