A method and system for remote sensing intelligent monitoring and early warning of river and lake eutrophication

Abstract

The present application belongs to the technical field of remote sensing image intelligent analysis, and particularly relates to a kind of lake eutrophication remote sensing intelligent monitoring and early warning method and system, method includes: the water quality data of the river lake to be monitored is collected, and pretreated;Based on the water quality data after pretreatment, using space-time matching algorithm, construct water quality parameter inversion sample library;Based on water quality parameter inversion sample library, using dynamic entropy weight method determines the weight of each water quality parameter, obtains comprehensive lake eutrophication index;Based on the physical constraint deep learning fusion model framework, construct lake eutrophication early warning model;Based on comprehensive lake eutrophication index and lake eutrophication early warning model, realize the real-time dynamic early warning to lake eutrophication degree.

Description

Technical Field

[0001] This invention belongs to the field of intelligent remote sensing image analysis technology, specifically relating to a remote sensing intelligent monitoring and early warning method and system for eutrophication of rivers and lakes. Background Technology

[0002] Eutrophication of rivers and lakes has become a global challenge for water environment management. Eutrophication not only damages aquatic ecosystems and triggers algal blooms, but also seriously threatens drinking water safety. Traditional methods for monitoring eutrophication, such as continuous monitoring at stations and on-site sampling and analysis, while accurately measuring local water quality parameters, have limitations including high manpower and material costs, limited monitoring range, and difficulty in comprehensively covering fragmented water bodies. These methods cannot promptly reflect the overall eutrophication status of large areas of rivers and lakes.

[0003] With the rapid development of satellite remote sensing technology, which features fast re-entry speed, wide monitoring range, and low cost, multi-source remote sensing collaborative observation technology can efficiently grasp the water quality status of rivers and lakes over a large area, revealing the spatiotemporal migration characteristics of water quality parameters. This has become an important component of the existing large-scale river and lake water quality monitoring system. Since the 1970s, remote sensing technology has been applied to river and lake water quality research. Early research focused on inverting water quality parameters through multispectral data band combinations. Subsequently, domestic and international scholars have conducted extensive research on chlorophyll a concentration inversion models based on different remote sensing data to assess the eutrophication status of rivers and lakes.

[0004] However, current remote sensing monitoring and early warning systems for eutrophication in rivers and lakes still have shortcomings. First, eutrophication largely depends on chlorophyll a concentration and lacks the utilization of other water quality parameters. Second, existing inversion models are mostly based on intelligent learning and lack the embedding of physical mechanisms, resulting in poor spatial applicability and consistency of the inversion models. Third, due to the lack of intelligent mining of eutrophication evolution characteristics, the accuracy of eutrophication early warning for rivers and lakes is not high, and it is impossible to realize the diffusion path of pollution sources.

[0005] Therefore, developing efficient and accurate remote sensing intelligent monitoring and early warning methods for river and lake eutrophication is of great significance for improving the level of river and lake environmental monitoring and management. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a remote sensing intelligent monitoring and early warning method and system for river and lake eutrophication. It aims to solve the problems of limited monitoring range, poor real-time performance, and high cost of traditional river and lake eutrophication monitoring methods, as well as the technical difficulties of existing deep learning methods lacking physical mechanism constraints and having insufficient interpretability of prediction results, so as to achieve real-time dynamic early warning of river and lake eutrophication levels.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] A remote sensing intelligent monitoring and early warning method for eutrophication of rivers and lakes, the method comprising:

[0009] Collect water quality data of the rivers and lakes to be monitored and perform preprocessing;

[0010] Based on the preprocessed water quality data, a water quality parameter inversion sample library is constructed using a spatiotemporal matching algorithm.

[0011] Based on the water quality parameter inversion sample library, the weight of each water quality parameter is determined by the dynamic entropy weight method to obtain the comprehensive river and lake eutrophication index.

[0012] A river and lake eutrophication early warning model is constructed based on a physical constraint deep learning fusion model framework.

[0013] Based on the comprehensive river and lake eutrophication index and the river and lake eutrophication early warning model, real-time dynamic early warning of the eutrophication level of rivers and lakes can be achieved.

[0014] Preferably, the water quality data of the rivers and lakes to be monitored includes: remote sensing image data of the rivers and lakes to be monitored, measured water quality data of the water quality monitoring stations, and meteorological and hydrological data of the area of ​​the water quality monitoring stations.

[0015] Preferably, based on the preprocessed water quality data, a spatiotemporal matching algorithm is used to construct a water quality parameter inversion sample library, including:

[0016] Based on the pre-processed water quality data, time-matching data that meets the preset time difference requirements is obtained;

[0017] Sub-cell spatial matching is performed on time-matched data to obtain a cell set;

[0018] Feature values ​​are extracted from pixel sets to construct a water quality parameter inversion sample library.

[0019] Preferably, sub-cell spatial matching is performed on the time-matched data to obtain a cell set, including:

[0020] Improved spectral angular distance SI(P) based on fused spectral weights i The calculation formula is:

[0021]

[0022] Among them, R j (P i ) represents pixel P i Reflectance in the j-th band; R j (P center ) represents the center pixel P of the buffer zone. center Reflectivity in the j-th band; ω j is the weighting coefficient for the j-th band, determined based on the band signal-to-noise ratio and information content; n is the number of bands.

[0023] Preferably, based on a water quality parameter inversion sample library, the weight of each water quality parameter is determined using the dynamic entropy weight method to obtain a comprehensive river and lake eutrophication index, including:

[0024] Based on the water quality parameter inversion sample library, a multi-level index system was established, secondary indicators were obtained and standardized.

[0025] Based on the standardized secondary indicators, the dynamic entropy weight method is used to obtain the comprehensive river and lake eutrophication index.

[0026] Preferably, the dynamic entropy weight method employs a dynamic weight update mechanism:

[0027] Calculate the basic entropy weight of the j-th secondary index according to the entropy weight calculation method.

[0028]

[0029] in, Let be the information entropy of the k-th secondary indicator at time step t, and n be the number of secondary indicators;

[0030] Time decay weight α (t) :

[0031]

[0032] Where t0 is the current time and τ is the decay time constant;

[0033] Dynamic fusion weight of the j-th secondary indicator

[0034]

[0035] Introducing a spatial heterogeneity weighting method, the spatial variation coefficient of the j-th secondary index...

[0036]

[0037] in, Let be the spatial standard deviation of the j-th secondary indicator. Let j be the spatial mean of the j-th secondary indicator;

[0038] Spatial heterogeneity adjustment factor β of the j-th secondary indicator j :

[0039]

[0040] Where γ is the adjustment parameter;

[0041] The final dynamically updated weights are:

[0042]

[0043] in, β is the dynamically updated weight for the j-th secondary indicator. k This is the spatial heterogeneity adjustment factor for the k-th secondary indicator. is the dynamic fusion weight of the k-th secondary indicator, and n is the number of indicators.

[0044] Preferably, the construction of the river and lake eutrophication early warning model based on the physical constraint deep learning fusion model framework includes:

[0045] A branched convolutional neural network was used to extract remote sensing spectral features and meteorological and hydrological features, respectively.

[0046] Based on the convection-diffusion reaction equation, a physically constrained TimesNet network model is constructed.

[0047] A physically constrained TimesNet network model was used to obtain the time series dependencies of remote sensing spectral features, meteorological and hydrological features, and various water quality parameters.

[0048] Construct a model constraint loss function based on physical constraint loss;

[0049] Based on time series dependencies and model constraint loss functions, an early warning model for river and lake eutrophication is obtained.

[0050] Preferably, based on physical constraint loss, a model constraint loss function is constructed, including:

[0051] Model constraint loss function

[0052]

[0053] Where λ is the adjustment coefficient. For data-driven loss functions, The physical constraint loss function;

[0054]

[0055] Where N is the total number of samples, and C i Let be the observed value and the predicted value of the i-th water quality parameter sample, respectively, and M be the number of collocation points used to verify the partial differential equation. For water quality parameter concentration, Water quality parameter concentration rate of change with time t For convection terms, For diffusion term, is the degradation term, k is the degradation rate constant, and S is the source-sink term.

[0056] The present invention also provides a remote sensing intelligent monitoring and early warning system for eutrophication of rivers and lakes. The system is used to implement the aforementioned method and includes: a data acquisition module, a sample library construction module, an index acquisition module, a model construction module, and a monitoring and early warning module.

[0057] The data acquisition module is used to collect water quality data of the rivers and lakes to be monitored and to perform preprocessing.

[0058] The sample library construction module is used to construct a water quality parameter inversion sample library based on preprocessed water quality data and using a spatiotemporal matching algorithm.

[0059] The index acquisition module is used to retrieve a sample library based on water quality parameters, determine the weight of each water quality parameter using the dynamic entropy weight method, and obtain the comprehensive river and lake eutrophication index.

[0060] The model building module is used to construct early warning models for river and lake eutrophication based on a physically constrained deep learning fusion model framework.

[0061] The monitoring and early warning module is used to achieve real-time monitoring and early warning of the eutrophication level of rivers and lakes based on the comprehensive river and lake eutrophication index and the river and lake eutrophication early warning model.

[0062] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0063] (1) By integrating multi-source data such as optical remote sensing, water quality monitoring, meteorology and hydrology through sub-pixel matching algorithm, a multi-dimensional water quality parameter spatiotemporal spectrum inversion sample library is established, which can achieve early warning 1-3 days in advance with an accuracy rate of over 85%, fundamentally changing the traditional "post-event monitoring" mode.

[0064] (2) The dynamic entropy weight method model is used to dynamically assign weights to each water quality parameter. Considering factors such as the decay of water quality parameter observation time and spatial heterogeneity, a stratified strategy is adopted to construct a comprehensive river and lake eutrophication index, which solves the problem of the difficulty in quantitatively quantifying heterogeneous river and lake eutrophication.

[0065] (3) The physical process of nutrient migration and self-purification in water bodies is innovatively modeled as a differential equation and a deep learning loss function is embedded to construct a physically constrained TimesNet model, which significantly improves the interpretability and scientific rationality of the prediction results. Attached Figure Description

[0066] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 This is a schematic diagram of the process of the remote sensing intelligent monitoring and early warning method for eutrophication of rivers and lakes according to an embodiment of the present invention;

[0068] Figure 2 This is a schematic diagram of a remote sensing intelligent monitoring and early warning system module for eutrophication of rivers and lakes, according to an embodiment of the present invention. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0071] Example 1

[0072] like Figure 1 As shown, this invention provides a remote sensing intelligent monitoring and early warning method for river and lake eutrophication, comprising:

[0073] Collect water quality data of the rivers and lakes to be monitored and perform preprocessing;

[0074] Based on the preprocessed water quality data, a water quality parameter inversion sample library is constructed using a spatiotemporal matching algorithm.

[0075] Based on the water quality parameter inversion sample library, the weight of each water quality parameter is determined by the dynamic entropy weight method to obtain the comprehensive river and lake eutrophication index.

[0076] A river and lake eutrophication early warning model is constructed based on a physical constraint deep learning fusion model framework.

[0077] Based on the comprehensive river and lake eutrophication index and the river and lake eutrophication early warning model, real-time dynamic early warning of the eutrophication level of rivers and lakes can be achieved.

[0078] Furthermore, the specific implementation process of this invention is as follows:

[0079] Collect water quality data for the rivers and lakes to be monitored and perform preprocessing, including:

[0080] Collect water quality data for the rivers and lakes to be monitored:

[0081] (1) Remote sensing image data of the rivers and lakes to be monitored:

[0082] Collect multi-source satellite remote sensing image data covering the rivers and lakes to be monitored, including: Sentinel-2, Landsat series images (Landsat 8 and Landsat 9) and other domestically produced optical remote sensing data (such as Gaofen-1, environmental satellites, etc.).

[0083] (2) Actual water quality data from water quality monitoring stations:

[0084] Obtain hourly time-series water quality data from monitoring stations, including chlorophyll a concentration (Chla), total phosphorus concentration (TP), total nitrogen concentration (TN), water transparency (SD), permanganate concentration (CODMn), biomass, turbidity, dissolved oxygen, etc.

[0085] (3) Meteorological and hydrological data of the water quality monitoring station area:

[0086] Collect meteorological and hydrological data for the water quality monitoring station area, including meteorological data (rainfall, temperature, wind speed, wind direction) and hydrological data (water level, flow velocity, flow rate).

[0087] Preprocessing of collected water quality data from rivers and lakes to be monitored:

[0088] (1) Preprocessing of remote sensing image data, including:

[0089] The remote sensing image data were subjected to radiometric calibration, atmospheric correction, geometric correction, cloud detection and masking, and water body extraction in sequence. A set of sensitive spectral features for water quality parameters was then constructed, including original band reflectance, band combination and ratio, normalized vegetation index (NDVI), etc.

[0090] (2) Preprocessing of measured water quality data, including:

[0091] The measured water quality data may contain outliers, and different data have different dimensions and ranges. In order to ensure data quality, eliminate the influence of dimensions, and enable the data to be better used for model building and analysis, it is necessary to perform outlier detection and processing, as well as standardization and other preprocessing.

[0092] The modified Z-score method was used to detect outliers MZ in the measured water quality data. The calculation formula is as follows:

[0093]

[0094] Where, x i Let X be the measured water quality data to be tested, median(·) be the median function, and MAD be the absolute median deviation of the dataset. When MZ is greater than 3.5, it is judged as an outlier and corrected using front-to-back time-series interpolation.

[0095] Further standardization processing was performed on the measured water quality data after outlier monitoring. The calculation formula is as follows:

[0096]

[0097] Where, x norm Here, x represents the standardized measured water quality data, and x represents the measured water quality data after outlier monitoring. min and x max These represent the minimum and maximum values ​​of the measured water quality data after anomaly monitoring.

[0098] Through the above processing, a time-consistent measured water quality dataset is formed.

[0099] (3) Preprocessing of meteorological and hydrological data, including:

[0100] Due to their uneven spatial and temporal distribution, varying resolutions, and different dimensions, meteorological and hydrological data require preprocessing for data fusion and model computation. This preprocessing includes generating a unified grid data through spatiotemporal kriging interpolation, standardizing the dimensional differences, and resampling data at different resolutions to a resolution consistent with the main remote sensing impact data. Furthermore, the average values ​​of meteorological parameters (rainfall, temperature, wind speed, wind direction) and hydrological parameters (water level, flow velocity, flow rate) are extracted to form a meteorological and hydrological feature set.

[0101] To construct a water quality parameter inversion sample library and achieve accurate correspondence between remote sensing data, meteorological and hydrological data, and measured water quality data, thus providing reliable data support for subsequent water quality parameter inversion and model construction, a high-precision spatiotemporal matching algorithm is required first. Specifically, based on preprocessed water quality data, a spatiotemporal matching algorithm is used to construct a water quality parameter inversion sample library, including:

[0102] Based on the pre-processed water quality data, time-matching data that meets the preset time difference requirements is obtained;

[0103] Sub-cell spatial matching is performed on time-matched data to obtain a cell set;

[0104] Feature values ​​are extracted from pixel sets to construct a water quality parameter inversion sample library.

[0105] The specific steps are as follows:

[0106] (1) Time matching: Select remote sensing images and meteorological and hydrological data that are close to the time of acquisition of water quality measurement data, with a time difference of no more than ±24 hours. When multiple images meet the conditions, the image with the closest time and the least cloud cover shall be selected first.

[0107] (2) Sub-pixel spatial matching: Considering the spatial heterogeneity of water body optical properties, a sub-pixel level spatial matching method is designed:

[0108] Smatch ={P i ∈W|d(P i ,P sample )≤R buffer and SI(P i )≥T SI};

[0109] Among them, S match W is the set of matching pixels; P is the set of water body pixels; i For the sub-pixel to be matched; P sample d(P) represents the sampling point location; i ,P sample R is the Euclidean distance between two points; buffer The buffer zone radius (usually set to 30m-100m); SI(P) i ) represents the improved spectral angular distance; T SI This is the similarity threshold (usually set to 2.5).

[0110] Improved spectral angular distance SI(P) based on fused spectral weights i The calculation formula is:

[0111]

[0112] Among them, R j (P i ) represents pixel P i Reflectance in the j-th band; R j (P center ) represents the center pixel P of the buffer zone. center Reflectivity in the j-th band; ω j is the weighting coefficient for the j-th band, determined based on the band signal-to-noise ratio and information content; n is the number of bands.

[0113] (4) Feature extraction: Calculate the weighted average of the matched pixel set as the final matched feature. The calculation formula is as follows:

[0114]

[0115] Among them, R j w is the eigenvalue of the j-th feature; i The weight of the i-th pixel can be set as an inverse ratio to the distance from the center point: w i =1 / d(P i ,P sample ); m is the number of matched pixels.

[0116] Using the above matching methods, spatiotemporal matching of various water quality parameters, remote sensing spectral features, and basic environmental features is achieved, thus constructing a multidimensional spatiotemporal spectral inversion sample library for water quality parameters.

[0117] D = {(SPEC)} i,t ,ENV i,t WQY i,t |i = 1, 2, ..., N};

[0118] SPEC i,t Indicates matching to the spectral feature set, ENV i,t This indicates a match to the meteorological and hydrological feature set, WQY i,t This indicates the matching of monitoring data from various water quality parameter stations, including chlorophyll a concentration (Chla), total phosphorus concentration (TP), total nitrogen concentration (TN), water transparency (SD), permanganate concentration (CODMn), biomass, turbidity, dissolved oxygen, etc., where N is the total number of samples and t is the sample time label. Furthermore, based on the water quality parameter inversion sample library, the dynamic entropy weight method is used to determine the weight of each water quality parameter, obtaining the comprehensive river and lake eutrophication index, including:

[0119] Based on the water quality parameter inversion sample library, a multi-level index system was established, secondary indicators were obtained and standardized.

[0120] Based on the standardized secondary indicators, the dynamic entropy weight method is used to obtain the comprehensive river and lake eutrophication index.

[0121] Specifically, it includes:

[0122] (1) Based on the water quality parameter inversion sample library, a multi-level index system was established, secondary indicators were obtained and standardized:

[0123] Based on the constructed water quality parameter inversion sample library, the dynamic entropy weight method is used to determine the weight of each water quality parameter, thereby constructing a river and lake eutrophication index. A multi-level index system is established based on the water quality parameter inversion sample library.

[0124] Primary indicator X:

[0125] X = {X1, X2, X3, X4, X5};

[0126] Among them, X1 is the nutrient salt index (total nitrogen (TN), total phosphorus (TP), ammonia nitrogen (NH3N)), X2 is the biological index (chlorophyll a (Chla), biomass), X3 is the transparency index (SD, turbidity), X4 is the dissolved oxygen index (DO, BOD5, COD), and X5 is the pH index.

[0127] Secondary indicators:

[0128] X1 = {TN, TP, NH3N}

[0129] X2 = {Chla, Biomass}

[0130] X3 = {SD, Turbidity};

[0131] X4 = {DO, BOD5, COD}

[0132] X5 = {pH}

[0133] Further standardization of the secondary indicators: Indicator standardization:

[0134]

[0135] Where, x ij y represents the original value of the j-th secondary indicator for the i-th sample; ij These are the standardized secondary indicator values.

[0136] (2) Based on the standardized secondary indicators, the dynamic entropy weight method was used to obtain the comprehensive river and lake eutrophication index:

[0137] 1) Time series entropy weight calculation:

[0138] The matrix Y of secondary index values ​​at time step t (t) :

[0139]

[0140] Where m is the number of samples and n is the number of secondary indicators. This is the secondary index value for time step t.

[0141] Information entropy of the j-th secondary indicator at time step t

[0142]

[0143] Among them, the probability of the j-th secondary indicator at time step t

[0144]

[0145] Where ε is a very small positive number (usually taken as 10). -6 ).

[0146] 2) Dynamic weight update mechanism:

[0147] Calculate the basic entropy weight of the j-th secondary index according to the entropy weight calculation method.

[0148]

[0149] in, Let be the information entropy of the k-th secondary indicator at time step t.

[0150] Time decay weight α (t) :

[0151]

[0152] Where t0 is the current time and τ is the decay time constant;

[0153] Dynamic fusion weight of the j-th secondary indicator

[0154]

[0155] To further ensure the robustness of the calculation results, considering the spatial heterogeneity of each index, a spatial heterogeneity weighting adjustment method is introduced, where the spatial variation coefficient of the j-th secondary index is...

[0156]

[0157] in, Let be the spatial standard deviation of the j-th secondary indicator. Let j be the spatial mean of the j-th secondary indicator;

[0158] Spatial heterogeneity adjustment factor β of the j-th secondary indicator j :

[0159]

[0160] Where γ is the adjustment parameter;

[0161] The final dynamically updated weights are:

[0162]

[0163] in, β is the dynamically updated weight for the j-th secondary indicator. k This is the spatial heterogeneity adjustment factor for the k-th secondary indicator. is the dynamic fusion weight of the kth secondary indicator.

[0164] 3) Calculation method for comprehensive river and lake eutrophication index:

[0165] For the secondary indicators, the basic eutrophication index (EI) is calculated separately. basic :

[0166]

[0167] Further calculate the weight W of the primary indicator k :

[0168]

[0169] Among them, G k This is the set of secondary indicators contained in the k-th primary indicator.

[0170] Primary indicator score S k :

[0171]

[0172] The comprehensive eutrophication index of rivers and lakes is calculated as follows:

[0173]

[0174] Based on the above calculations, the comprehensive eutrophication index of rivers and lakes at each station can be calculated, which serves as the true value for the eutrophication early warning model of rivers and lakes, and is used to train the model parameters.

[0175] Furthermore, to achieve dynamic early warning of river and lake eutrophication, the physical processes of nutrient migration and self-purification in river and lake water bodies are modeled as differential equations. A physical constraint-based deep learning fusion model framework is designed, and an improved TimesNet network is used to mine the time-series nonlinear relationship between river and lake eutrophication and meteorological and hydrological scenario characteristics. A guided constraint loss function is constructed through physical mechanism equations. That is, a river and lake eutrophication early warning model is constructed based on the physical constraint-based deep learning fusion model framework, including:

[0176] A branched convolutional neural network was used to extract remote sensing spectral features and meteorological and hydrological features, respectively.

[0177] Based on the convection-diffusion reaction equation, a physically constrained TimesNet network model is constructed.

[0178] A physically constrained TimesNet network model was used to obtain the time series dependencies of remote sensing spectral features, meteorological and hydrological features, and various water quality parameters.

[0179] Construct a model constraint loss function based on physical constraint loss;

[0180] Based on time series dependencies and model constraint loss functions, an early warning model for river and lake eutrophication is obtained.

[0181] Specifically, it includes:

[0182] (1) Spatial feature extraction based on branched convolutional neural networks:

[0183] To achieve the extraction of multi-source remote sensing spectral features and meteorological and hydrological features, a convolutional neural network is used to extract these features. The input layer contains multi-source remote sensing spectral features and meteorological and hydrological features, with dimensions of H×W×C, where H and W are the spatial dimensions and X is the number of features. For both remote sensing imagery and meteorological and hydrological data, a convolutional neural network with a two-branch structure is used to extract remote sensing spectral features and meteorological and hydrological features respectively, and finally the extracted features are concatenated.

[0184] Each branch has the same convolutional structure, and the convolutional layers use a multi-layer convolutional structure. Each layer has a 3×3 convolutional kernel with a stride of 1 and the ReLU activation function. The forward propagation calculation formula is as follows:

[0185] Z (l) =W (l) *A (l-1) +b (l) ;

[0186] A (l) =ReLU(Z) (l) );

[0187] P (l) =MaxPool(A (l) );

[0188] Among them, Z (l) W is the output of the convolution of the l-th layer; (l) and b (l) These are the kernel weights and biases, respectively; A (l-1) A is the activation value of the (l-1)th layer; (l) P is the activation value of the l-th layer; (l) This is the output after pooling.

[0189] The remote sensing spectral features extracted by the first branch are designated as F1, and the meteorological and hydrological features extracted by the second branch are designated as F2. Connecting F1 and F2 yields the final feature map P. (L) The connection operation Concat is as follows:

[0190] P (L) =Concat(F1,F2);

[0191] Convert the feature map extracted by convolution into a one-dimensional feature vector:

[0192] F cnn =Flatten(P (L) );

[0193] Among them, F cnn These are the flattened feature vectors.

[0194] (2) Physical process modeling and differential equation construction:

[0195] To address the modeling needs of nutrient migration and self-purification processes in water bodies, a physically constrained TimesNet model is proposed. This model innovatively embeds the convection-diffusion reaction equation into the TimesNet loss function, achieving high-precision spatiotemporal prediction through multi-period decomposition and synergistic optimization based on physical laws. The migration and self-purification processes of pollutants in water bodies can be described by the following convection-diffusion reaction equation:

[0196]

[0197] in, Let c be the rate of change of pollutant concentration over time t. This is the convection term, where u is the water flow velocity, and the direction is downstream. Let D be the diffusion term, -kc be the degradation term, k be the degradation rate constant, and S(x,t) be the source-sink term, where S>0 is the source and S<0 is the sink. The diffusion coefficient D and degradation rate k are used as trainable parameters, and the output is generated through a TimesNet deep learning model.

[0198] D = TimesNet(X) input );

[0199] k = TimesNet(X) input );

[0200] Among them, X input The input values ​​include temporal, spatial, and meteorological / hydrological data (such as temperature and flow rate). In TimesNet, local spatiotemporal features and long-range dependencies are extracted using the following temporal convolution and temporal attention modules, respectively.

[0201] (3) Time series feature extraction based on improved TimesNet:

[0202] TimesNet is used to mine the time-series dependencies of remote sensing spectral features, meteorological and hydrological features, and various water quality parameters. The time series of the above data can be regarded as a complex time pattern formed by the superposition of multiple periods. Two-dimensional tensor decomposition is used to capture the periodicity of the time series, while forcing the model to output the temporal variation characteristics of the diffusion coefficient D and degradation rate k, satisfying the differential equations of the physical model, improving the physical rationality of long-term predictions, and achieving multi-period-physical co-modeling. TimesBlock converts the one-dimensional time series into a two-dimensional space and uses the parameter-efficient Inception module for modeling. The specific process includes the following steps:

[0203] ① Convert one-dimensional changes into two-dimensional changes

[0204] For a time series of length T containing c recorded variables (remote sensing spectral features extracted by CNN, diffusion coefficient D, and degradation rate k, etc.), the original one-dimensional structure obtained by spatial feature extraction based on a branched convolutional neural network is as follows: First, the time series is periodically decomposed in the frequency domain using Fast Fourier Transform (FFT), as follows:

[0205] A = Avg(Amp(FFT(X)) 1D )));

[0206]

[0207] Here, FFT(·) and Amp(·) represent Fast Fourier Transform and amplitude calculation, respectively. The calculated amplitude for each frequency is represented by Avg(·), which is averaged across c dimensions. The top n amplitude values ​​are selected to obtain the unnormalized amplitude {A}. f1 ,…,A fn The most significant frequencies of {f1,…,f} n} and the corresponding period lengths {p1,…,p n Based on the selected frequencies {f1,…,f} k} and the corresponding period lengths {p1,…,p k The one-dimensional time series X is expressed by the following formula. 1D Remodeled into multiple two-dimensional tensors

[0208]

[0209] Padding(·) is used to extend the time series with zeros along the time dimension, making it suitable for... Where p i and f i These represent the number of rows and columns of the transformed two-dimensional tensor, respectively.

[0210] ②The TimesBlock module captures two-dimensional changes

[0211] The TimesBlock module is constructed using a residual approach. Specifically, it is used for one-dimensional input time series of length T, including spectral features, diffusion coefficient D, and degradation rate k. Firstly, through the embedding layer Project the original input onto deep features For the l-th layer of TimesNet, the input is Its processing can be formalized as follows:

[0212]

[0213] For the processed features, Period(·) is used to capture the two-dimensional changes over time, and softmax is used for adaptive aggregation at different periods to predict the concentration of water quality parameters.

[0214] ③ Capturing two-dimensional changes in time

[0215] Using Period(·) as a deep feature The period length is estimated, and based on the estimated period length, the one-dimensional time series is transformed into a two-dimensional space, obtaining a set of two-dimensional tensors. Information representation is then extracted from these tensors using the parameter-efficient Inception module. This process is formalized as follows:

[0216]

[0217] in, This is the i-th transformed two-dimensional tensor. After transformation, the two-dimensional tensor is processed using Inception(·). Then, the learned two-dimensional representation is... Convert back to one-dimensional space Perform aggregation, using Trunc(·) to fill the padded length (p i ×f i The sequence is truncated to its original length T.

[0218] ④ Adaptive Aggregation

[0219] Fusion of k distinct one-dimensional representations Amplitude is calculated using the softmax function. The relative importance of the frequency and period reflected Therefore, the importance of each transformed two-dimensional tensor can be represented in one-dimensional form based on amplitude aggregation as follows:

[0220]

[0221] The degree of eutrophication is predicted by using the aggregated features and the mean squared error (MSE) as the prediction accuracy measure.

[0222] (4) Design of physical constraint loss function

[0223] Total loss function of river and lake eutrophication early warning model Data-driven loss function With physical constraint loss function composition:

[0224]

[0225] Where λ is the adjustment coefficient. For data-driven loss functions, The physical constraint loss function;

[0226]

[0227] Where N is the total number of samples, and C i Let be the observed value and the predicted value of the i-th water quality parameter sample, respectively, and M be the number of collocation points used to verify the partial differential equation. For water quality parameter concentration, Water quality parameter concentration rate of change with time t This refers to the convection term (concentration changes caused by fluid movement). This refers to the diffusion term (the change in water quality parameter concentration due to molecular diffusion). is the degradation term, k is the degradation rate constant, and S is the source-sink term.

[0228] Furthermore, based on the comprehensive river and lake eutrophication index and the river and lake eutrophication early warning model, real-time dynamic early warning of the eutrophication level of rivers and lakes is achieved, including:

[0229] To better adapt the river and lake eutrophication early warning model to data characteristics and improve its accuracy, a comprehensive river and lake eutrophication index was used to train and optimize the model. A multi-task weighted loss function was employed, taking into account the prediction accuracy of each parameter.

[0230] L total =∑ k∈{Chla,TP,TN,SD,CODMn} λ k L k ;

[0231]

[0232] Among them, L k The predicted loss for the eutrophication index of rivers and lakes is determined by a data-driven loss function. With physical constraint loss function Composition; λ k The corresponding weighting coefficients can be automatically adjusted based on task uncertainty.

[0233]

[0234] Where, σ k The uncertainty parameters for task k can be learned through backpropagation; The mean squared error loss for task k includes a data-driven loss function and a physical constraint loss function:

[0235]

[0236] The Adam optimizer is further used to train the model. By dynamically adjusting the learning rate, setting an early stopping mechanism and L2 regularization, the model parameters are optimized to avoid overfitting and improve generalization ability.

[0237] Finally, the model performance was evaluated using indicators such as the coefficient of determination and root mean square error, enabling dynamic early warning of eutrophication indices in multiple rivers and lakes. The coefficient of determination (R²) 2 The calculation methods for the root mean square error (RMSE) are as follows:

[0238]

[0239] Among them, y i These are actual observations. It is the average of the actual observed values. This is the model's predicted value. R 2 The value of RMSE ranges from 0 to 1. The larger the value, the stronger the correlation between the predicted value and the observed value, that is, the higher the model accuracy. The smaller the RMSE, the smaller the error between the predicted value and the true value, and the higher the model accuracy.

[0240] In summary, this invention proposes an intelligent monitoring and early warning method for river and lake eutrophication based on physical mechanism-constrained deep learning. First, a spatiotemporal matching algorithm is used to perform sub-pixel matching between multi-source optical remote sensing data and time-series water quality monitoring data (including chlorophyll a concentration (Chla), total phosphorus concentration (TP), total nitrogen concentration (TN), water transparency (SD), permanganate concentration (CODMn), biomass, turbidity, dissolved oxygen, etc.). This is then integrated with meteorological and hydrological data such as rainfall, temperature, and water level to construct a multi-dimensional spatiotemporal spectrum inversion sample library for water quality parameters. A dynamic entropy weighting model is used to dynamically assign weights to each water quality parameter, considering factors such as observation time decay and spatial heterogeneity, to construct a comprehensive river and lake eutrophication index. The physical processes of nutrient migration and self-purification in river and lake water bodies are modeled as differential equations. A physical constraint-based deep learning fusion model framework is designed, and an improved TimesNet network is used to mine the time-series nonlinear relationship between river and lake eutrophication and meteorological and hydrological scenario characteristics. A guided constraint loss function is constructed through physical mechanism equations to achieve real-time dynamic early warning of river and lake eutrophication levels.

[0241] Example 2

[0242] like Figure 2 As shown, based on the same inventive concept, the present invention also provides a remote sensing intelligent monitoring and early warning system for river and lake eutrophication, used to implement the methods described in the foregoing embodiments. The system includes: a data acquisition module, a sample library construction module, an index acquisition module, a model construction module, and a monitoring and early warning module.

[0243] The data acquisition module is used to collect water quality data of the rivers and lakes to be monitored and to perform preprocessing.

[0244] The sample library construction module is used to construct a water quality parameter inversion sample library based on preprocessed water quality data and using a spatiotemporal matching algorithm.

[0245] The index acquisition module is used to retrieve a sample library based on water quality parameters, determine the weight of each water quality parameter using the dynamic entropy weight method, and obtain the comprehensive river and lake eutrophication index.

[0246] The model building module is used to construct early warning models for river and lake eutrophication based on a physically constrained deep learning fusion model framework.

[0247] The monitoring and early warning module is used to achieve real-time monitoring and early warning of the eutrophication level of rivers and lakes based on the comprehensive river and lake eutrophication index and the river and lake eutrophication early warning model.

[0248] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A remote sensing intelligent monitoring and early warning method for eutrophication of rivers and lakes, characterized in that, The method includes: Collect water quality data of the rivers and lakes to be monitored and perform preprocessing; Based on the preprocessed water quality data, a water quality parameter inversion sample library is constructed using a spatiotemporal matching algorithm. Based on the water quality parameter inversion sample library, the weight of each water quality parameter is determined by the dynamic entropy weight method to obtain the comprehensive river and lake eutrophication index. A river and lake eutrophication early warning model is constructed based on a physical constraint deep learning fusion model framework. Based on the comprehensive river and lake eutrophication index and the river and lake eutrophication early warning model, real-time dynamic early warning of the degree of river and lake eutrophication can be achieved. Based on the preprocessed water quality data, a spatiotemporal matching algorithm was used to construct a water quality parameter inversion sample library, including: Based on the pre-processed water quality data, time-matching data that meets the preset time difference requirements is obtained; Sub-cell spatial matching is performed on time-matched data to obtain a cell set; Feature values ​​are extracted from pixel sets to construct a water quality parameter inversion sample library; Sub-cell spatial matching is performed on the temporal matching data to obtain a cell set, including: Improved spectral angular distance with fused spectral weights SI ( P i The calculation formula is: ； in, For pixels In the Reflectivity of each band; ) is the center pixel of the buffer zone In the Reflectivity of each band; For the first The weighting coefficients for each band are determined based on the band signal-to-noise ratio and information content. Number of bands; Based on a water quality parameter inversion sample library, the weights of each water quality parameter are determined using the dynamic entropy weight method to obtain a comprehensive river and lake eutrophication index, including: Based on the water quality parameter inversion sample library, a multi-level index system was established, secondary indicators were obtained and standardized. Based on the standardized secondary indicators, the dynamic entropy weight method is used to obtain the comprehensive river and lake eutrophication index. The dynamic entropy weight method employs a dynamic weight update mechanism: According to the entropy weight calculation method, calculate the first... j Basic entropy weight of each secondary indicator : ； in, For the first k Each secondary indicator at the time step t Information entropy n The number of secondary indicators; Time decay weight : ； in, For the current moment, The decay time constant; No. j Dynamic fusion weights of secondary indicators : ； Introducing a spatial heterogeneity weight adjustment method, the first j Spatial variation coefficient of each secondary indicator : ； in, For the first j Spatial standard deviation of each secondary indicator For the first j Spatial mean of each secondary indicator; No. j Spatial heterogeneity adjustment factor for each secondary indicator : ； in, To adjust the parameters; The final dynamically updated weights are: ； in, For the first j The dynamic update weights of each secondary indicator For the first k Spatial heterogeneity adjustment factor for each secondary indicator For the first k Dynamic fusion weights of each secondary indicator.

2. The remote sensing intelligent monitoring and early warning method for eutrophication of rivers and lakes according to claim 1, characterized in that, The water quality data of the rivers and lakes to be monitored includes: remote sensing image data of the rivers and lakes to be monitored, measured water quality data of water quality monitoring stations, and meteorological and hydrological data of the water quality monitoring station area.

3. The remote sensing intelligent monitoring and early warning method for eutrophication of rivers and lakes according to claim 1, characterized in that, The aforementioned physical constraint-based deep learning fusion model framework is used to construct an early warning model for river and lake eutrophication, including: A branched convolutional neural network was used to extract remote sensing spectral features and meteorological and hydrological features, respectively. Based on the convection-diffusion reaction equation, a physically constrained TimesNet network model is constructed. A physically constrained TimesNet network model was used to obtain the time series dependencies of remote sensing spectral features, meteorological and hydrological features, and various water quality parameters. Construct a model constraint loss function based on physical constraint loss; Based on time series dependencies and model constraint loss functions, an early warning model for river and lake eutrophication is obtained.

4. The remote sensing intelligent monitoring and early warning method for eutrophication of rivers and lakes according to claim 3, characterized in that, Based on physical constraint loss, a model constraint loss function is constructed, including: Model constraint loss function : ； in, For adjustment coefficients, For data-driven loss functions, The physical constraint loss function; ； ； in, The total number of samples, and The first i Observed and predicted values ​​of water quality parameters for each sample. M The number of collocation points used to verify the partial differential equation. For water quality parameter concentration, Water quality parameter concentration Over time rate of change, For convection terms, For diffusion term, For degradation items, The degradation rate constant is S For source and sink items.

5. A remote sensing intelligent monitoring and early warning system for eutrophication of rivers and lakes, the system being used to implement the method described in any one of claims 1-4, characterized in that, The system includes: a data acquisition module, a sample library construction module, an index acquisition module, a model construction module, and a monitoring and early warning module; The data acquisition module is used to collect water quality data of the rivers and lakes to be monitored and to perform preprocessing. The sample library construction module is used to construct a water quality parameter inversion sample library based on preprocessed water quality data and using a spatiotemporal matching algorithm. The index acquisition module is used to retrieve a sample library based on water quality parameters, determine the weight of each water quality parameter using the dynamic entropy weight method, and obtain the comprehensive river and lake eutrophication index. The model building module is used to construct early warning models for river and lake eutrophication based on a physically constrained deep learning fusion model framework. The monitoring and early warning module is used to achieve real-time monitoring and early warning of the eutrophication level of rivers and lakes based on the comprehensive river and lake eutrophication index and the river and lake eutrophication early warning model.

Images