Remote sensing monitoring and early warning method for concentration of chlorophyll in rivers and lakes

Abstract

The invention discloses a river and lake chlorophyll a concentration remote sensing monitoring and early warning method, which comprises the following steps: carrying out time sequence remote sensing data radiation normalization correction processing by utilizing multi-source sentinel No.2 and Landsat series remote sensing data; constructing a large-range and long-time-sequence river and lake chlorophyll a remote sensing sample library by combining water quality monitoring data continuously observed on the ground and a chlorophyll a characteristic wave band; further combining with a random forest algorithm to construct a river and lake chlorophyll a intelligent inversion model fused with convolutional features, and combining with a GEE satellite remote sensing data platform to realize large-range chlorophyll a remote sensing product generation; a time sequence chlorophyll a remote sensing data product is further combined, a water quality monitoring station is used as a truth value constraint, a long short-term memory network model with an improved step length is adopted, a river and lake chlorophyll a prediction product generation model is constructed, and river and lake water quality risk dynamic early warning is achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing image intelligent analysis, and in particular relates to a remote sensing monitoring and early warning method for chlorophyll a concentration in rivers and lakes. Background Art

[0002] Chlorophyll a concentration is an important factor in monitoring the safety of river and lake water quality, and is also a key content of water environment management. With the rapid development of cities and social economy, a large number of rivers and lakes have become eutrophic, leading to algae outbreaks and excessive chlorophyll a concentrations, which seriously threaten the safety of drinking water in rivers and lakes. Chlorophyll a in rivers and lakes is a key parameter for water quality monitoring stations. It is of great value for the eutrophication conversion of rivers and lakes and the carbon and nitrogen cycle, and is also an important factor in water pollution monitoring and management.

[0003] Traditional monitoring of chlorophyll a concentration in rivers and lakes is generally based on continuous monitoring at sites or field sampling and analysis. Although it can accurately measure chlorophyll a concentration, it consumes a lot of manpower, material resources, and financial resources, and can only obtain chlorophyll a concentration within the local range of the monitoring point. Remote sensing technology has the advantages of strong continuity, wide monitoring range, and relatively low cost. Using multispectral or hyperspectral remote sensing technology to monitor water quality can not only quickly and timely grasp the overall chlorophyll a concentration of rivers and lakes in a large area, improve monitoring efficiency, and reduce monitoring costs, but also reflect the temporal and spatial migration characteristics of chlorophyll a concentration that are difficult to reveal with conventional methods. Therefore, remote sensing, as an important part of the existing three-dimensional monitoring system of water quality parameters, is organically combined with traditional water quality site monitoring data, and has played an important role in the monitoring and early warning of chlorophyll a concentration in rivers and lakes.

[0004] Since the 1970s, remote sensing technology has been applied to the study of water quality in inland rivers and lakes. Zarco-Tejada et al. analyzed 72 bands of hyperspectral data CASI and found that the use of band R750 / R710 band combination can achieve better inversion accuracy of water quality parameter chlorophyll a concentration. Mattews et al. conducted research and development on the inversion of water quality parameters such as Chla, TSS, and CDOM in Lake Zeekoevlei in South Africa and found that the model established using empirical methods has higher inversion accuracy in water bodies with higher nutrient levels. Wen Jianguang et al. used semi-empirical methods to establish a remote sensing monitoring model for chlorophyll a in Taihu Lake water bodies and explored the feasibility of using hyperspectral and multispectral data to establish an inversion model.

[0005] At present, a large number of domestic and foreign scholars have conducted research on chlorophyll a concentration inversion models based on different types of remote sensing data, but there are the following shortcomings: (1) Chlorophyll a concentration inversion using Sentinel-2 and other data is mostly based on empirical models, and there is a lack of research on chlorophyll a concentration inversion models using multi-source satellite data coordinated by intelligent algorithms; (2) Most of the research is based on local areas, and there is a lack of remote sensing product production capabilities for large-scale chlorophyll a concentration continuous in time and space; (3) Chlorophyll a concentration inversion and monitoring are mostly carried out using remote sensing methods, and there is still a lack of products for chlorophyll a concentration early warning. Summary of the invention

[0006] In order to solve the above technical problems, the present invention proposes a remote sensing monitoring and early warning method for chlorophyll a concentration in rivers and lakes to achieve dynamic early warning of river and lake water quality risks.

[0007] To achieve the above-mentioned object, the present invention provides a method for remote sensing monitoring and early warning of chlorophyll a concentration in rivers and lakes, comprising: obtaining multi-source remote sensing data of chlorophyll a concentration in rivers and lakes, performing radiation normalization processing on the multi-source remote sensing data, and obtaining a normalized data set;

[0008] Based on the normalized data set, a convolutional neural network is used to obtain the remote sensing characteristic bands of chlorophyll a of rivers and lakes, and a remote sensing characteristic set of chlorophyll a is constructed in combination with the radiation normalized characteristic bands;

[0009] Based on the chlorophyll a remote sensing feature set and combined with the ground-based continuous observation of water quality monitoring data, a large-scale, long-time series river and lake chlorophyll a remote sensing sample library is constructed, and the large-scale, long-time series river and lake chlorophyll a remote sensing sample library is used to train the river and lake chlorophyll a intelligent inversion model;

[0010] The chlorophyll a remote sensing feature set is input into the trained river and lake chlorophyll a intelligent inversion model, and combined with the GEE satellite remote sensing data platform, a large-scale, long-time series river and lake time-series chlorophyll a content map is obtained, and the chlorophyll a content is quantitatively inverted to generate a large-scale river and lake water body time-series chlorophyll a remote sensing product;

[0011] Based on the time-series chlorophyll a remote sensing products of the large-scale river and lake water bodies, taking the water quality monitoring stations as the true value constraints, a long short-term memory network model with improved step length is adopted to obtain a trained river and lake chlorophyll a prediction model to predict the future concentration of river and lake chlorophyll a;

[0012] Based on the time-series remote sensing monitoring of chlorophyll a in rivers and lakes and the future predicted concentration, early warnings are issued for the water quality of rivers and lakes over a large area.

[0013] According to the method for remote sensing monitoring and early warning of chlorophyll a concentration in rivers and lakes provided by the present invention, the normalized data set obtained includes:

[0014]

[0015] in, is the DN value of the radiation normalized image of the i-th band; S i is the DN value of the i-th band of the original target image; α i and β i are the gain coefficient and bias coefficient of the i-th band respectively.

[0016] According to the method for remote sensing monitoring and early warning of chlorophyll a concentration in rivers and lakes provided by the present invention, obtaining remote sensing characteristic bands of chlorophyll a in rivers and lakes includes:

[0017] Convolutional neural network is used to extract remote sensing features of chlorophyll a in rivers and lakes, including:

[0018] The convolution formula is as follows:

[0019]

[0020] The ReLU activation function formula is as follows:

[0021] f(z i,j )=max(0,z i,j ),

[0022] Among them, z i,j is the eigenvalue after convolution of the pixel at position (i, j); x is the radiation value of the pixel at position (i, j) in the normalized image; k is the convolution kernel.

[0023] According to the method for remote sensing monitoring and early warning of chlorophyll a concentration in rivers and lakes provided by the present invention, constructing a chlorophyll a remote sensing feature set includes:

[0024] Featureset = {I tf ,t=1,2,…,T; f=1,2,…,F},

[0025] Among them, t is the time series image; F is the number of features; f is the number of image bands.

[0026] According to the remote sensing monitoring and early warning method for river and lake chlorophyll a concentration provided by the present invention, the large-scale and long-time series river and lake chlorophyll a remote sensing sample library is:

[0027] dataset={(X tn ,Y tn )}, t=1,2…,T,n=1,2,…N,

[0028] X tn =[I t1 (i,j),I t2 (i,j),…,I tF (i,j)],

[0029] The dataset is a large-scale, long-time series river and lake chlorophyll a remote sensing sample library; T is the number of time-series remote sensing images; N is the number of samples; F is the number of features; X is the number of features. tn is the feature vector of the nth sample point at the tth moment; Y tn is the observed value of the nth sample point at the tth moment.

[0030] According to the method for remote sensing monitoring and early warning of river and lake chlorophyll a concentration provided by the present invention, the process of obtaining a trained river and lake chlorophyll a intelligent inversion model includes:

[0031] The continuous feature values ​​in the chlorophyll a remote sensing feature set are discretized and sorted by using a binary method to obtain a number of cut points of the first feature;

[0032] Perform segmentation according to the first feature to obtain a set of segmentation points;

[0033] Using a regression tree with a plurality of leaf nodes to divide the chlorophyll a remote sensing feature set into a plurality of subsets;

[0034] Selecting features and splitting points to obtain two feature space subsets that are split from the chlorophyll a remote sensing feature set;

[0035] Traversing all the features in the chlorophyll a remote sensing feature set to find the optimal segmentation feature and the optimal segmentation point;

[0036] Taking the average value of all targets in each feature space subset as the output value of the feature space subset;

[0037] Repeat the above six steps until the square error is minimized and a trained river and lake chlorophyll a intelligent inversion model is obtained.

[0038] According to the method for remote sensing monitoring and early warning of river and lake chlorophyll a concentration provided by the present invention, the process of obtaining a trained river and lake chlorophyll a prediction model includes:

[0039] According to the time series chlorophyll a remote sensing data product, the time series chlorophyll a concentration of each observation point in the river and lake monitoring area is recorded as a set;

[0040] Extracting a training sample of length L from the set, and using the chlorophyll a concentrations at the first L-1 time points in the training sample as sample input, and using the chlorophyll a concentration at the last time point as a sample label;

[0041] The training samples are input into the long short-term memory network model to learn the driving factors of abnormal chlorophyll a quantity, and the back propagation algorithm is used to optimize the parameters of the long short-term memory network model to obtain a trained river and lake chlorophyll a prediction model.

[0042] According to the remote sensing monitoring and early warning method for chlorophyll a concentration in rivers and lakes provided by the present invention, a back propagation algorithm is used to train a long short-term memory network model, and the trained long short-term memory network model is obtained, which includes:

[0043] The output value of each LSTM neuron is obtained by forward propagation calculation method;

[0044] Calculate the error term of each LSTM neuron in two back-propagation directions: time and network level;

[0045] According to the error term of each LSTM neuron, the update gradient of each weight of the LSTM model is obtained;

[0046] Select a deep learning optimizer, set the minimum loss function as the optimization goal, give the sample segmentation length, number of network layers, number of hidden layer nodes, learning rate and number of training times, continuously update the weights of the long short-term memory network model, and obtain a trained long short-term memory network model.

[0047] Technical effect of the invention: The invention discloses a method for remote sensing monitoring and early warning of chlorophyll a concentration in rivers and lakes, constructs a large-scale river and lake chlorophyll a remote sensing intelligent inversion model, and breaks through the technical bottleneck that the original large-scale river and lake chlorophyll a remote sensing monitoring and early warning products are difficult to automatically produce; the invention adopts multi-source satellite remote sensing data to construct chlorophyll a remote sensing characteristic bands and sample libraries, and combines the random forest algorithm to construct an intelligent inversion model for chlorophyll a concentration in rivers and lakes; the invention further combines the time-series chlorophyll a remote sensing data product, and adopts a long short-term memory network algorithm based on an improved step size to construct a river and lake chlorophyll a prediction product generation model, so as to realize dynamic early warning of river and lake water quality risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0049] Figure 1 The present invention is a flowchart of a method for remote sensing monitoring and early warning of chlorophyll a concentration in rivers and lakes according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0051] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0052] like Figure 1 As shown, this embodiment provides a method for remote sensing monitoring and early warning of chlorophyll a concentration in rivers and lakes, including: obtaining multi-source remote sensing data of chlorophyll a concentration in rivers and lakes, performing radiation normalization processing on the multi-source remote sensing data, and obtaining a normalized data set;

[0053] Based on the normalized data set, a convolutional neural network is used to obtain the remote sensing characteristic bands of chlorophyll a of rivers and lakes, and a remote sensing characteristic set of chlorophyll a is constructed in combination with the radiation normalized characteristic bands;

[0054] Based on the chlorophyll a remote sensing feature set and combined with the ground-based continuous observation of water quality monitoring data, a large-scale, long-time series river and lake chlorophyll a remote sensing sample library is constructed, and the large-scale, long-time series river and lake chlorophyll a remote sensing sample library is used to train the river and lake chlorophyll a intelligent inversion model;

[0055] The chlorophyll a remote sensing feature set is input into the trained river and lake chlorophyll a intelligent inversion model, and combined with the GEE satellite remote sensing data platform, a large-scale, long-time series river and lake chlorophyll a content map is obtained, and the chlorophyll a content is quantitatively inverted to generate a large-scale river and lake water body time series chlorophyll a remote sensing product; the large range refers to the fact that the remote sensing imaging range is relatively large, and the chlorophyll a inversion of rivers and lakes in large areas can be achieved; the long time series refers to a 10-year long time series, with more than 30 monitoring results each year, forming a long-term monitoring result, and this time series result can be used to make predictions.

[0056] Based on the time-series chlorophyll a remote sensing products of the large-scale river and lake water bodies, taking the water quality monitoring stations as the true value constraints, a long short-term memory network model with improved step length is adopted to obtain a trained river and lake chlorophyll a prediction model to predict the future concentration of river and lake chlorophyll a;

[0057] Based on the time-series remote sensing monitoring of chlorophyll a in rivers and lakes and the future predicted concentration, early warnings are issued for the water quality of rivers and lakes over a large area.

[0058] Furthermore, the collection and preprocessing of time-series multi-source remote sensing image data

[0059] According to application requirements, collect as much long-term multi-source Sentinel-2 and hyperspectral remote sensing satellite data covering the study area as possible. After the image is acquired, it is preprocessed by atmospheric correction, radiation correction, geometric correction, etc. to build a time series multi-source image dataset within the study area.

[0060] Radiometric Normalization

[0061] Taking into account the inconsistency of radiation of remote sensing images from different sensors and different time series, the preprocessed time series image data is subjected to radiation normalization.

[0062] One image is taken as the reference image R, and the other images to be corrected are S. Based on the ideal radiator, a sample set is constructed and normalized using an analytical method.

[0063] Radiance Normalization Based on Conditional Probability Distribution

[0064] The radiation normalization formula is as follows:

[0065]

[0066] is the DN value of the normalized image of the ith band, S i is the DN value of the i-th band of the original target image, α i and β i They represent the gain and bias coefficient of the i-th band respectively.

[0067] According to the above steps, the time series images are processed to obtain a normalized data set of I = {I 1 ,I 2 ,…I T}, where T is the number of images.

[0068] Furthermore, the intelligent inversion model of river and lake chlorophyll a is constructed and products are generated, including:

[0069] (1) Convolutional neural network based feature band construction

[0070] A convolutional neural network was used to extract the remote sensing features of chlorophyll a in rivers and lakes. 32 groups of convolution kernels were used with a kernel size of 3, a step size of 1, and ReLU as the activation function to extract the remote sensing features of chlorophyll a.

[0071] The convolution formula is as follows:

[0072]

[0073] The ReLU activation function formula is as follows:

[0074] f(z i,j )=max(0,z i,j )

[0075] Among them, z i,j It represents the eigenvalue after convolution of the pixel at position (i, j), x is the radiation value of the pixel at position (i, j) in the normalized image, and k is the convolution kernel.

[0076] In order to retain the edge information of the image, the image is padded with zeros of width 1 before convolution extraction. The convolution kernel is traversed over the image with a step size of 1, and the ReLU activation function is used to obtain the feature map after convolution.

[0077] Based on the radiation normalized data set, according to the above method, the remote sensing characteristic bands of chlorophyll a concentration in rivers and lakes are obtained, and the chlorophyll a remote sensing feature set is constructed by combining the radiation normalized characteristic bands, which is expressed as:

[0078] Featureset = {I tf ,t=1,2,…,T; f=1,2,…,F}

[0079] Among them, t represents the time series image, F represents the number of features, and F is taken as 32+b, where b is the number of image bands.

[0080] (2) Training sample library

[0081] Combined with the time series chlorophyll a water quality parameter observation data, according to the geographical location of the observation data, the feature data of the corresponding position of the feature set obtained above is extracted to form a large-scale, long-time series river and lake chlorophyll a remote sensing sample library. The large-scale, long-time series river and lake chlorophyll a remote sensing sample library is expressed as follows:

[0082] dataset={(X tn ,Y tn )} t=1,2…,T,n=1,2,…N

[0083] X tn =[I t1 (i,j),I t2 (i,j),…,I tF (i,j)]

[0084] Among them, dataset means a large-scale, long-time series river and lake chlorophyll a remote sensing sample library; T is the number of time-series remote sensing images, N is the number of samples, and F is the number of features. tn represents the feature vector of the nth sample point at the tth moment, Y tn Represents the observation value of the nth sample point at the tth time.

[0085] (3) Intelligent remote sensing inversion model and product generation of chlorophyll a in rivers and lakes

[0086] The continuous feature values ​​in the training sample set X are discretized using the binary method. j1 ,x j2 ,…,x jn}, that is, the eigenvalues ​​of all training samples when j = 1 are arranged in ascending order from small to large, denoted as {F 1 ,F 2,…,F n}, get feature x 1 The n-1 split points:

[0087]

[0088] in represents the eigenvalue of the i-th training sample, Represents feature x 1 A set of n-1 split points;

[0089] According to the feature x 1 Perform segmentation to obtain a set S of n-1 segmentation points X ={s i |1≤i≤n-1};

[0090] Use a regression tree with K leaf nodes to divide the training sample set X into K subsets R k , 1≤k≤K;

[0091] Select the jth feature x j and the i-th split point s i , we get two feature space subsets divided in the training sample set X:

[0092] R 1 (j,s)={x|x j ≤s i},R 2 (j,s)={x|x j >s i}

[0093] Where x represents the sample point;

[0094] Traverse all features in the training sample set X and find the optimal segmentation feature x j And the optimal split point s:

[0095]

[0096] Among them, x (i) is the jth eigenvalue of the i-th sample point, y (i) is the target value corresponding to the jth feature of the i-th sample point; c 1 To select the split feature x j When the split point s is reached, the feature space subset R 1 The average value of all corresponding targets; c 2 To select the split feature x j When the split point s is reached, the feature space subset R 2 The average value of all corresponding targets;

[0097] The average value of all targets in each feature space subset is taken as the output value c of the feature space subset k :

[0098] c k =mean(y|x∈R k )

[0099] Where mean(·) represents the mean value function;

[0100] Repeat the above steps until the square error is minimized and the trained river and lake chlorophyll a intelligent inversion model is obtained:

[0101]

[0102] in represents the estimated value of the kth sample space subset, K represents the number of regression tree child nodes, I(x∈R k ) is an exponential when x∈R k When I(x∈R k )=1, otherwise I(x∈R k )=0.

[0103] Based on the above trained inversion model, the obtained multi-feature set Featureset is used as input, and combined with the GEE remote sensing big data platform, a large-scale river and lake time series chlorophyll a content map {Cha 1 , Cha 2 ,…,Cha T}, to achieve quantitative inversion of chlorophyll a content and obtain time-series chlorophyll a remote sensing data products.

[0104] Chlorophyll a prediction product generation and water quality risk warning for rivers and lakes, including:

[0105] (1) Constructing an LSTM chlorophyll a prediction model with improved step size

[0106] According to the time series chlorophyll a remote sensing data product, the time series chlorophyll a concentration of each observation point in the river and lake monitoring area is recorded as a set D m ={d 1 , d 2 , …, d m}.

[0107] From the set D m Extract the training sample D of length L from train ={d m-L , d m-L+1 , …, d m-1}, and the chlorophyll a concentrations at the first L-1 time points in the training sample are used as sample input, and the chlorophyll a concentration at the last time point is used as the sample label.

[0108] The training samples are input into the LSTM network model to learn the driving factors of abnormal chlorophyll a content, and the back propagation algorithm is used to adjust the LSTM network model parameters to obtain a trained prediction model.

[0109] In the embodiment of the present invention, the BPTT algorithm is used to train the LSTM network model, including the following steps:

[0110] The output value O of each LSTM neuron is obtained by the forward propagation calculation method. t :

[0111] O t =LSTM forward (X t , H t-1 , C t-1 )

[0112] Among them, LSTM forward (·) represents the forward propagation calculation function, X t represents the input of the current LSTM neuron, H t-1 and C t-1 Represent the hidden state and cell state of the previous LSTM neuron respectively. t , H t-1 and C t-1 They are all sets, and the specific calculation formula of the elements is as follows:

[0113] i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i )

[0114] f t =σ(W xf x t +W hf h t-1 +W cf c t-1 +b f )

[0115] c t =f t c t-1 +i t tanh(W xc x t +W hc h t-1 +b c )

[0116] o t =σ(W xo x t +W ho h t-1 +W co c t +b o )

[0117] h t =o t tanh(c t )

[0118] Among them, x t represents the sample time series data input at time t, h t represents the hidden state output of the corresponding neuron at time t, and each neuron performs data input, information forgetting, cell state update, and hidden state output in sequence. t 、f t 、c t , o t They represent the input gate, forget gate, cell state and output gate at the tth moment respectively; W and b are the corresponding weight coefficient matrix and bias term respectively; σ and tanh are the sigmoid and hyperbolic tangent activation functions respectively.

[0119] The error term of each LSTM neuron is calculated in two back-propagation directions: time and network level.

[0120] The mean square error is selected as the error statistical indicator, and the loss function loss of the training process is defined as:

[0121]

[0122] where Y pred It represents the set of prediction results of the LSTM network model, Y represents the corresponding true concentration, and m represents the number of observation points.

[0123] The updated gradient of each weight of the LSTM network model is obtained according to the error term of each LSTM neuron.

[0124] The ADAM algorithm is selected as the optimizer, and the minimum loss function is set as the optimization goal. Given the sample segmentation length L, the number of network layers K, the number of hidden layer nodes S, the learning rate l, and the number of training epochs, the weights of the LSTM network model are continuously updated to obtain a trained prediction model.

[0125] The trained prediction model is used to predict the future concentration of chlorophyll a in rivers and lakes.

[0126] In the embodiment of the present invention, a set of length L is intercepted from the time series concentration of each observation point as a prediction sample, which is the same as the model training sample. The chlorophyll a concentration of the first L-1 time points in the prediction sample is used as the sample input, and the chlorophyll a concentration of the Lth time point is used as the sample label. By continuously adjusting the hyperparameters, the prediction model can obtain the expected effect in the training and verification process at the same time, and finally obtain a model for predicting the amount of chlorophyll a at future moments. The standardized prediction samples of all observation points are input, and the output result can be expressed as:

[0127] Y pred ={y 1 ,y 2 , …, y m}

[0128] By dynamically adjusting the LSTM network model, the nonlinear correlation between chlorophyll a concentrations in different time series can be fully learned, and the future concentration of chlorophyll a in rivers and lakes can be predicted.

[0129] (2) Water quality safety warning for large-scale rivers and lakes

[0130] Combined with the time-series remote sensing monitoring of chlorophyll a in rivers and lakes and future predicted concentrations, in accordance with the national river and lake water quality prediction standards and the graded early warning strategy, early warnings are issued for the water quality safety of rivers and lakes over a large area according to high risk, medium risk and low risk, and water quality safety risk areas are delineated to support large-scale, efficient and precise river and lake water quality supervision.

[0131] The present invention discloses a method for remote sensing monitoring and early warning of chlorophyll a concentration in rivers and lakes, constructs a large-scale river and lake chlorophyll a remote sensing intelligent inversion model, and breaks through the technical bottleneck that the original large-scale river and lake chlorophyll a remote sensing monitoring and early warning products are difficult to automatically produce; the present invention adopts multi-source satellite remote sensing data to construct chlorophyll a remote sensing characteristic bands and sample libraries, and combines with a random forest algorithm to construct an intelligent inversion model for chlorophyll a concentration in rivers and lakes; the present invention further combines time-series chlorophyll a remote sensing data products, adopts a long short-term memory network algorithm based on an improved step size, constructs a river and lake chlorophyll a prediction product generation model, and realizes dynamic early warning of river and lake water quality risks.

[0132] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for remote sensing monitoring and early warning of chlorophyll a concentration in rivers and lakes, characterized in that: include: Acquiring multi-source remote sensing data of chlorophyll a concentration in rivers and lakes, performing radiometric normalization on the multi-source remote sensing data, and obtaining a normalized data set; Based on the normalized data set, a convolutional neural network is used to obtain the chlorophyll a remote sensing characteristic bands of rivers and lakes, and a chlorophyll a remote sensing feature set is constructed by combining the radiation normalized characteristic bands; Based on the chlorophyll a remote sensing feature set and combined with ground-based continuous observation water quality monitoring data, a large-scale, long-term river and lake chlorophyll a remote sensing sample library is constructed. The large-scale, long-term river and lake chlorophyll a remote sensing sample library is used to train the river and lake chlorophyll a intelligent inversion model; The chlorophyll a remote sensing feature set is input into the trained river and lake chlorophyll a intelligent inversion model, and combined with the GEE satellite remote sensing data platform, a large-scale, long-term river and lake time-series chlorophyll a content map is obtained, and the chlorophyll a content is quantitatively inverted to generate a large-scale river and lake water body time-series chlorophyll a remote sensing product; Based on the time-series chlorophyll-a remote sensing products of the large-scale river and lake water bodies, with water quality monitoring stations as the true value constraint, a long short-term memory network model with an improved step size is used to obtain a trained river and lake chlorophyll-a prediction model to predict the future concentration of chlorophyll-a in rivers and lakes; Based on the time-series remote sensing monitoring of chlorophyll a in rivers and lakes and the future predicted concentration, early warnings are issued for the water quality of rivers and lakes over a large area.

2. The method for remote sensing monitoring and early warning of chlorophyll a concentration in rivers and lakes according to claim 1, characterized in that: The normalized data sets include: in, is the DN value of the radiometric normalized image of the i-th band; S i is the DN value of the i-th band of the original target image; α i and β i are the gain coefficient and bias coefficient of the i-th band respectively.

3. The method for remote sensing monitoring and early warning of chlorophyll a concentration in rivers and lakes according to claim 1, characterized in that: The remote sensing characteristic bands of chlorophyll a in rivers and lakes include: Convolutional neural networks are used to extract remote sensing features of chlorophyll a in rivers and lakes, including: The convolution formula is as follows: The ReLU activation function formula is as follows: f(z i，j )max(0,z i，j ), Among them, z i,j is the eigenvalue after convolution of the pixel at position (i, j); x is the radiation value of the pixel at position (i, j) in the normalized image; k is the convolution kernel.

4. The method for remote sensing monitoring and early warning of chlorophyll a concentration in rivers and lakes according to claim 1, characterized in that: Constructing the chlorophyll a remote sensing feature set includes: Featureset={I tf ,t=1,2,...,T;f=1,2,...,F}, Among them, t is the time series image; F is the number of features; R is the number of image bands.

5. The method for remote sensing monitoring and early warning of chlorophyll a concentration in rivers and lakes according to claim 1, characterized in that: The large-scale, long-time series river and lake chlorophyll-a remote sensing sample library is: dataset={(X tn ,Y tn )},t=1,2...,T,n=1,2,...N, X tn =[I t1 (i,j),I t2 (i,j),...,I tF (i,j)], The dataset is a large-scale, long-term river and lake chlorophyll a remote sensing sample library; T is the number of time-series remote sensing images; N is the number of samples; F is the number of features; X tn is the feature vector of the nth sample point at the tth moment; Y tn is the observation value of the nth sample point at the tth moment.

6. The method for remote sensing monitoring and early warning of chlorophyll a concentration in rivers and lakes according to claim 1, characterized in that: The process of obtaining a trained river and lake chlorophyll-a intelligent inversion model includes: The continuous feature values ​​in the chlorophyll a remote sensing feature set are discretized and sorted using a dichotomy method to obtain a number of cut points of the first feature; Perform segmentation according to the first feature to obtain a set of segmentation points; Using a regression tree with a plurality of leaf nodes to divide the chlorophyll a remote sensing feature set into a plurality of subsets; Selecting features and splitting points to obtain two feature space subsets of the chlorophyll a remote sensing feature set; Traversing all features in the chlorophyll a remote sensing feature set to find the optimal segmentation feature and the optimal segmentation point; Taking the average value of all targets in each feature space subset as the output value of the feature space subset; Repeat the above six steps until the square error is minimized and a trained river and lake chlorophyll-a intelligent inversion model is obtained.

7. The method for remote sensing monitoring and early warning of chlorophyll a concentration in rivers and lakes according to claim 1, characterized in that: The process of obtaining a trained river and lake chlorophyll-a prediction model includes: According to the time series chlorophyll a remote sensing data product, the time series chlorophyll a concentration of each observation point in the river and lake monitoring area is recorded as a set; Extracting a training sample of length L from the set, and using the chlorophyll a concentrations of the first L-1 time points in the training sample as sample input, and using the chlorophyll a concentration of the last time point as the sample label; The training samples are input into the long short-term memory network model to learn the driving factors of abnormal chlorophyll a content, and the back propagation algorithm is used to optimize the parameters of the long short-term memory network model to obtain a trained river and lake chlorophyll a prediction model.

8. The method for remote sensing monitoring and early warning of chlorophyll a concentration in rivers and lakes according to claim 7, characterized in that: The back propagation algorithm is used to train the long short-term memory network model. The trained long short-term memory network model includes: The output value of each long short-term memory network neuron is obtained through the forward propagation calculation method; Calculate the error term of each LSTM neuron in two back-propagation directions: time and network level; According to the error term of each LSTM neuron, the updated gradient of each weight of the LSTM model is obtained; Select a deep learning optimizer, set the minimum loss function as the optimization goal, give the sample segmentation length, number of network layers, number of hidden layer nodes, learning rate and number of training times, continuously update the weights of the long short-term memory network model, and obtain a trained long short-term memory network model.

Images