Soil heavy metal accumulation space-time prediction algorithm based on CA-LSTM-RF under deep learning framework

Abstract

The invention discloses a CA-LSTM-RF soil heavy metal accumulation space-time prediction algorithm under a deep learning framework, and belongs to the technical field of soil heavy metal accumulation space-time prediction algorithms, and the algorithm comprises the steps: obtaining soil heavy metal content data of each sampling point in a plurality of historical periods in a target region and an environment variable data set affecting soil heavy metal accumulation; the method comprises the following steps: performing abnormal value detection, missing data supplementation, alignment processing, quantification processing, normalization processing and time sequence coding on data, performing cellular water flow direction judgment by improving a CA model and designing diffusion process simulation of adaptive time, and calculating heavy metal surface runoff diffusion by using a Manning formula. And taking the diffusion value of the heavy metal along with the surface runoff and the environment variable data as input, extracting a historical data relationship through LSTM, outputting a soil heavy metal accumulation prediction value at the current moment, and taking the output of the LSTM as input to construct an RF model to predict the spatial characteristics of the soil heavy metal.

Description

Technical Field

[0001] The present invention relates to a spatiotemporal prediction algorithm for soil heavy metal accumulation, and in particular to a spatiotemporal prediction algorithm for soil heavy metal accumulation using a CA-LSTM-RF deep learning framework, belonging to the technical field of spatiotemporal prediction algorithms for soil heavy metal accumulation. Background Art

[0002] Through searching, we found that: Patent A method for simulating and predicting the distribution of regional soil heavy metal pollution based on machine learning (CN118469060) proposed a spatial prediction method that couples Gaussian process autoregressive model machine learning and geostatistical spatial interpolation, which improves the prediction accuracy and precision of the spatial distribution of soil heavy metal pollution.

[0003] The patented method for spatial prediction of soil ammonium nitrogen pollution in rare earth mining areas proposes a hierarchical Bayesian three-dimensional surface mean modulus to predict soil ammonium nitrogen at different depths, reflecting the distribution differences in different strata. However, these methods all propose methods to improve the accuracy of spatial distribution from the perspective of considering spatial heterogeneity, and have not yet considered the accumulation of time scales. In summary, the existing models are insufficient in depicting complex spatial heterogeneity, making it difficult to explore the variable characteristics hidden in the time series and geographic space of soil heavy metal pollution accumulation, limiting the model's refined simulation of the soil heavy metal accumulation process and accurate prediction of future spatiotemporal distribution.

[0004] In recent years, deep learning has been able to discover temporal variation patterns of complex variables from time series and capture complex nonlinear relationships.

[0005] Patented method for watershed pollutant flux prediction based on LSTM-BP spatiotemporal combined model (CN111639748B) predicts watershed nitrogen flux using LSTM and BP neural networks;

[0006] Patent: A PM based on time series and deep learning framework 2.5 The estimation method uses spatiotemporal enhanced neural network (STENN) to estimate PM 2.5 Currently, only a few studies have used ensemble learning methods to predict the spatiotemporal accumulation of heavy metals in soil. However, the migration and diffusion of heavy metals in surface runoff has not been explored. To address this issue, we designed a CA-LSTM-RF spatiotemporal prediction algorithm for heavy metal accumulation in soil within a deep learning framework to address this technical issue. Summary of the Invention

[0007] The main purpose of this invention is to provide a CA-LSTM-RF soil heavy metal accumulation spatiotemporal prediction algorithm under a deep learning framework.

[0008] The purpose of the present invention can be achieved by adopting the following technical solutions:

[0009] A CA-LSTM-RF deep learning framework based spatiotemporal prediction algorithm for soil heavy metal accumulation includes the following steps:

[0010] Step 1: Obtain soil heavy metal content data and environmental variable datasets affecting soil heavy metal accumulation at each sampling point in the target area over multiple historical periods, perform outlier detection, missing data supplementation, alignment processing, quantification processing, normalization processing, and time series coding on the data;

[0011] Step 2: By improving the CA model, designing a time-adaptive diffusion process simulation, determining the cellular water flow direction, and using the Manning formula to calculate the surface runoff diffusion of heavy metals;

[0012] Step 3: Take the diffusion value of heavy metals with surface runoff and environmental variable data as input, extract the historical data relationship through LSTM, and output the predicted value of soil heavy metal accumulation at the current moment;

[0013] Step 4: Use the output of LSTM as input to build a RF model to predict the spatial characteristics of soil heavy metals;

[0014] Step 5: Use Keras of TensorFlow to build a model. Input the diffusion values ​​of environmental variables and CA model simulation into the LSTM time simulator, and input their predicted values ​​into the RF spatial simulator. Output the predicted values ​​of soil heavy metals at the grid scale, and perform model performance evaluation.

[0015] Preferably, the environmental variable data set in step 1 includes historical enterprise information data, soil property data, meteorological climate data, high-resolution remote sensing image data and socioeconomic data.

[0016] Preferably, the outlier test in the data preprocessing of step 1 adopts the 3σ principle, the missing data is supplemented by bilinear interpolation or spline interpolation, and the quantitative processing includes using the kernel density method to calculate the enterprise density, the population to area ratio to calculate the population density, the road length to area ratio to calculate the road density, and the Euclidean distance to calculate the distance between the sample point and the enterprise and residential area.

[0017] Preferably, in the CA model of step 2, the cellular water flow direction is determined by calculating the average water level of the central cell and the neighboring cells, eliminating the neighboring cells with water levels greater than the average, and determining the water flow direction cell;

[0018] The diffusion calculation of heavy metal surface runoff uses the Manning formula, taking into account the influence of the Reynolds coefficient conversion factor, hydraulic radius, permeability coefficient and the mass of infiltrating pollutants on the hydraulic radius, and calculates the time step of runoff from the central cell to the neighboring cells and the transfer equation of heavy metals on the cells.

[0019] Preferably, in step 3, the input layer of the LSTM time simulator is the diffusion value of heavy metals with surface runoff and historical data of environmental variables, and the output layer is the heavy metal content in the soil at time t. The information flow is controlled by the forget gate, input gate, memory unit state and output gate. The specific formula is as follows:

[0020] f t =σ[w f (h t-1 ,x t )+b f );

[0021] i t =σ[w i (h t-1 ,x t )+b i );

[0022] c t =f t ×c t-1 +i t ×tanh[w c (h t-1 ,x t )]+b c );

[0023] o t =σ[w o (h t-1 ,x t )+b o );

[0024] h t =o t ×tanh(c t );

[0025] Among them, ft,i t , c t and o t are the vectors of forget gate, input gate, memory cell state and output gate respectively;

[0026] w f , w i , w c and w o are weights respectively;

[0027] bf, b i , b c and b o are the bias vectors of the forget gate, input gate, memory unit state, and output gate respectively;

[0028] x tis the input vector at time t;

[0029] h t-1 is the output vector at time t-1;

[0030] h t is the output vector at time t;

[0031] c t-1 is the memory unit at time t-1;

[0032] c t is the memory unit at time t;

[0033] σ is the Sigmoid activation function and tanh is the hyperbolic tangent activation function.

[0034] Preferably, the input of the RF spatial simulator in step 4 is the output of the LSTM model at time t and the environmental feature vector affecting the accumulation of heavy metals in the soil, and the prediction formula is:

[0035]

[0036] Where K is the number of trees, f k (x) is the prediction result of the k-th tree;

[0037] Hyperparameter settings: Estimators: 125;

[0038] Max_depth:18;

[0039] Max_features: 17;

[0040] Min_samples_leaf:2;

[0041] Min_samples_split:2.

[0042] Preferably, in step 5, the model performance evaluation adopts ten-fold cross validation and time cross validation, and the accuracy evaluation index is the determination coefficient R 2 , root mean square error RMSE and mean absolute error MAE, the specific formulas are as follows:

[0043]

[0044] Where m is the total number of samples;

[0045] and y i It represents the cumulative amount of heavy metal pollution predicted by the model and the actual monitored cumulative amount of heavy metal pollution.

[0046] Beneficial technical effects of the present invention:

[0047] The present invention provides a CA-LSTM-RF spatiotemporal prediction algorithm for soil heavy metal accumulation under a deep learning framework, which introduces cellular automata into time series and machine learning models for the spatiotemporal prediction of soil heavy metal accumulation. It can capture the spatial details of the migration and distribution of soil heavy metals at a microscale, and is expected to make up for the shortcomings of traditional models and more accurately depict the accumulation trend of soil heavy metal pollution. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a schematic diagram of a method flow chart of a preferred embodiment of a CA-LSTM-RF soil heavy metal accumulation spatiotemporal prediction algorithm under a deep learning framework according to the present invention;

[0049] Figure 2 This is a diagram of the LSTM unit structure of a preferred embodiment of a CA-LSTM-RF soil heavy metal accumulation spatiotemporal prediction algorithm under a deep learning framework according to the present invention;

[0050] Figure 3-6 A scatter plot of the model predicted values ​​and measured values ​​is shown in FIG1 , which is a preferred embodiment of the soil heavy metal accumulation spatiotemporal prediction algorithm of CA-LSTM-RF under a deep learning framework of the present invention. Figure 3-6 It can be seen that the R of the CA-LTSM-RF combination model 2 The range is between 0.82-0.90, and the RMSE range is between 0.33-8.93. DETAILED DESCRIPTION

[0051] In order to make the technical solution of the present invention more clear and specific to those skilled in the art, the present invention is further described in detail below with reference to embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0052] The purpose of the present invention is to solve the problems of the prior art and provide a CA-LSTM-RF spatiotemporal prediction model to achieve spatiotemporal prediction of soil heavy metal accumulation at a regional scale;

[0053] The specific technical methods adopted in the present invention are as follows:

[0054] S1: Data collection and preprocessing;

[0055] S11: Obtain soil heavy metal content data for each sampling point within the target area over multiple historical periods, and obtain relevant environmental variable datasets that influence soil heavy metal accumulation, including historical enterprise information data, soil property data, meteorological and climate data, high-resolution remote sensing imagery data, and socioeconomic data. Enterprise information data includes enterprise latitude and longitude coordinates; soil physical and chemical property data includes soil organic matter, soil pH, soil type, soil erodibility, and soil cation exchange capacity; meteorological and climate data includes temperature, rainfall, wind speed, and direction; remote sensing data includes elevation, slope, terrain roughness, normalized difference vegetation index, and river flow; socioeconomic data includes land use type, population density, road density, residential areas, and GDP.

[0056] S12: Preprocess enterprise data, soil properties, meteorological climate, remote sensing data, and socioeconomic data:

[0057] ①Outlier test: in Python, we use algorithms such as the 3σ principle to identify and process the data, that is, to remove abnormal data that is outside the range of 3 times the standard deviation;

[0058] ②Supplement missing data by using bilinear interpolation, spline interpolation and other methods;

[0059] ③ Data alignment processing: pre-process and resample the data to ensure that all data are at the same spatiotemporal resolution. The present invention selects a spatiotemporal resolution of 100×100m;

[0060] S13: Quantify the data:

[0061] ① In ArcGIS 12.0, the kernel density method was used to perform kernel density analysis on the spatial locations of enterprises (Formula 1) to obtain the enterprise density value;

[0062] ② Calculate population density using the population to area ratio using the field calculator;

[0063] ③ Calculate road density by the ratio of road length to area;

[0064] ④ Calculate the distance between the sample points and enterprises and residential areas using Euclidean distance.

[0065]

[0066] In the formula, n is the total number of data points, h is the bandwidth, K(x) is the kernel function, n is the number of dimensions, and x is the number of dimensions. i and y i are the coordinates of two points in space.

[0067] S14: Since the data of environmental variables that affect the accumulation of heavy metals in soil have large dimensional differences and large variations in numerical ranges, in order to avoid the impact of dimensions and ranges on model training, convergence speed and performance, linear normalization processing (Formula 2) is used before inputting the model to map the data to the interval [0,1]:

[0068]

[0069] In the formula, x ′ is the standardized data, x is the original data, x min and x max are the minimum and maximum values ​​of the original data respectively. In the output layer, it is converted into the soil heavy metal content through the inverse operation of formula (2).

[0070] S15: Arrange these multi-source data in chronological order and use Python to encode them into a time series format, that is, divide the data window into fixed time steps, and the data in each window constitutes a sample to obtain time series data.

[0071] S2: Simulation of heavy metal surface runoff diffusion based on cellular automata (CA);

[0072] CA is a discrete spatiotemporal dynamics model that abstracts the study area into cells in a cellular space. It can describe complex geographic conditions by extrapolating local information to the whole, and has applications in simulating surface water pollution diffusion and flooding. This paper improves on CA by designing an adaptive time-based diffusion simulation based on the time differences in diffusion between adjacent cells and the decay of pollution diffusion rate with mass. In the CA model, the direction of water flow within the cells is determined and the runoff diffusion of heavy metals is calculated. Furthermore, a minimum difference algorithm combined with the Manning formula is used to simulate the diffusion of heavy metals in soil with runoff.

[0073] S21: Determination of the direction of cellular water flow;

[0074] The CA model is based on the principle that dynamic systems tend to develop in the direction of equilibrium. There is a height difference between the central cell and the eight surrounding neighboring cells. Heavy metals will flow from cells with higher elevations to cells with lower elevations with runoff, and balance is achieved by reducing the elevation difference between each grid.

[0075] ① Calculate the average water level of the central cell and the eight neighboring cells:

[0076]

[0077] In the formula, h0 is the water level of the central cell (m), h i is the water level (m) of the central cell in its neighborhood that has not been removed, and m is the number of neighboring cells that have not been removed.

[0078] ② Calculate the average water level of the remaining neighborhood cells and the central cell, and continue to remove neighborhood cells whose water level is greater than the average.

[0079] ③ Repeat the above steps until the water levels of the remaining neighborhood cells are lower than the average value.

[0080] ④ The remaining neighboring cells are the cells in which the water flows in a single time step. The water flow from the central cell is distributed so that the central cell and the remaining neighboring cells have the same water level.

[0081] S22: Calculation of heavy metal surface runoff diffusion;

[0082] Since the water level difference, roughness and slope in local areas are different, the runoff velocity is also different. Therefore, under the condition of uniform time step of CA model, the cell velocity is calculated using Manning formula (4).

[0083]

[0084] In the formula: k is the conversion constant, which is 1; v is the flow rate (m / s); R h is the hydraulic radius, which can be replaced by the average depth of the fluid according to experience; s is the slope, which is the ratio of the height difference between two points to the distance; n is the roughness (roughness coefficient);

[0085] The calculation formula of the target fluid flow rate is obtained using the Reynolds conversion factor:

[0086]

[0087] In the formula: f re is the Reynolds number conversion factor, is the Reynolds number of the conversion target fluid; is the Reynolds number of water. If the target fluid is water, then f re is 1.

[0088] The hydraulic radius in formula (6) can be replaced by the average water depth. Assuming that the distribution of heavy metals in each grid unit is uniform, the average height of heavy metals in each grid, i.e., the hydraulic radius, can be calculated based on the diffusion mass of heavy metals at the next moment and the size of the grid. The calculation formula is as follows (7):

[0089]

[0090] In the formula, M is the mass of heavy metal in the current grid; ρ is the density of the fluid; and s is the ground area of ​​the grid.

[0091] In addition, for general soil, the permeability coefficient cannot be ignored. The permeability coefficient of water is shown in formula (8).

[0092]

[0093] In the formula: K is the permeability coefficient; p is the soil permeability, which is related to porosity; ρ is the fluid density; η is the viscosity coefficient, and g is the acceleration of gravity.

[0094] Since the factors affecting the permeability coefficient are mainly fluid density and viscosity coefficient, the Reynolds number ratio f can be used re Perform the conversion, and the conversion formula (9) is:

[0095] K T =f re K (9);

[0096] Where: K T is the fluid permeability coefficient; K is the water permeability coefficient. Considering the mass loss of pollutants penetrating downward, the mass of pollutants penetrating downward is calculated based on the permeability coefficient of the pollutants and the first evolution of the downward infiltration time Δt. Formula (7) is improved to obtain the hydraulic radius (Formula 10):

[0097]

[0098] Due to the differences in concentration gradients and slopes in different neighborhood directions of a cell during one evolution process, the flow velocity is different and the diffusion time is different. The time step Δt of runoff from the central cell to the neighboring cells is xy The calculation formula is:

[0099]

[0100] In the formula, Δt xy is the time step of heavy metal diffusion in the (x, y) direction; l is the distance between the central cell and the neighboring cells (m).

[0101] Therefore, the transfer (migration) equation of heavy metals on the cell can be expressed as formula (12):

[0102]

[0103] In the formula, S i,j is the original value of heavy metal surface runoff; v i,j is the diffusion speed of heavy metals in the i and j directions; Δt i,j is the time step.

[0104] S3: Time simulator based on long short-term memory network (LSTM);

[0105] LSTM can effectively process time-dependent data, utilizing more complex internal processing units to handle the storage and updating of contextual information. By introducing memory cells and gating units, it enables the preservation of historical information and long-term states, as well as the control of information flow. This paper uses an LSTM-based time series method to extract relationships between historical data, using environmental variables affecting soil heavy metal accumulation and the diffusion of heavy metals with surface runoff as input, and outputting the predicted soil heavy metal accumulation at the current moment.

[0106] S31: input layer;

[0107] The diffusion value of heavy metals obtained in S2 with surface runoff and other variables including enterprise data, soil properties, meteorological climate, remote sensing data and socio-economic historical data as spatiotemporal series data (x t , x t-1 ,…x t-n ) Input the input layer of the LSTM time simulator.

[0108] S32: output layer;

[0109] The output layer is the soil heavy metal content at time t (h t ).

[0110] Specifically, the LSTM memory unit structure at time t is shown in Figure 2 The LSTM input parameters include the current moment input variable x t , the hidden layer state variable h at the previous moment t-1 and the memory unit state variable c at the previous moment t-1 ; In the LSTM unit: the model passes through the forget gate f in turn t , input gate i t and output gate o t ; The parameters of LSTM output include the current output variable h t and the current memory unit state variable c t Therefore, in the prediction of soil heavy metal accumulation, the input gate is mainly used to control the impact of new observations on the current prediction value; while the output gate controls the influence of past trends.

[0111] More specifically: the forget gate f t Determine how much information to retain from time t-1 (Formula 13); Input gate i t Determine how much information to store starting from the current time t (Formula 14); the memory cell state c t Used to update the current unit state (Formula 15), part of it is to discard some long-term memory information through the forget gate, and the other part is to add the new information at the current moment through the input gate; the output gate ( t) determines the current cell state c t How much information should be output (Formula 16), the output value is h t is the cumulative amount of heavy metal pollution in soil at time t (Formula 17).

[0112] f t =σ[w f (h t-1 ,x t )+b f ) (13);

[0113] i t =σ[w i (h t-1 ,x t )+b i ) (14);

[0114] c t =f t ×c t-1 +i t ×tanh[w c (h t-1 ,x t )]+b c ) (15);

[0115] o t =σ[w o (h t-1 ,x t )+b o ) (16);

[0116] h t =o t ×tanh(c t ) (17);

[0117] Among them, f t ,i t , c t and o t are the vectors of forget gate, input gate, memory unit state and output gate respectively, w f , w i , w c and w o are weights, b f , b i , b c and b o are the forget gate, output gate, memory unit state and bias vector of the output gate respectively, σ is the Sigmoid activation function and tanh is the hyperbolic tangent activation function.

[0118] S4: Spatial simulator based on random forest (RF);

[0119] By utilizing the nonlinear relationship of random forest, the output of LSTM is used as the input of random forest to construct a spatial feature based on RF to predict soil heavy metals (Formula 18).

[0120]

[0121] Where K is the number of trees and fk(X) is the prediction result of the kth tree.

[0122] X combined =[h tLSTM , X spatial ] (19);

[0123] X LSTM is the output of the LSTM model at time t, X spatial is the environmental characteristic vector that affects the accumulation of heavy metals in soil.

[0124] RF hyperparameter settings: Estimators: 125; Max_depth: 18; Max_features: 17; Min_samples_leaf: 2; Min_samples_split: 2.

[0125] S5: Building an LSTM-RF spatiotemporal combination model under the framework of deep learning

[0126] S51: LSTM-RF spatiotemporal combination model;

[0127] A spatiotemporal LSTM-RF model was constructed using the Keras framework of the deep learning framework TensorFlow. First, environmental variables influencing soil heavy metal accumulation and regional-scale soil heavy metal diffusion with surface runoff simulated using the CA model were input into the LSTM-based temporal simulator component to simulate changes in soil heavy metal accumulation at time t. Second, the predicted values ​​from the temporal simulator were input into the RF-based spatial simulator. Finally, the output layer obtained the predicted soil heavy metal values ​​at the current grid level.

[0128] S52 model performance evaluation;

[0129] ① Stability test uses sample-based cross-validation and time cross-validation.

[0130] Ten-fold cross-validation was used to randomly divide the dataset into 10 parts, 9 of which were used to build and train the model, and the remaining 1 was used for model verification. The verification was repeated alternately in parallel until all years were verified.

[0131] Time cross-validation was used, and data from any one year was selected for model validation. Data from other years were used for simulation fitting, and validation was performed alternately in parallel until all years were validated.

[0132] ② The accuracy evaluation was performed using the coefficient of determination (R 2 ), root mean square error (RMSE), and mean absolute error (MAE).

[0133] R2, RMSE, and MAE are calculated using equations (20), (21), and (22).

[0134]

[0135] Where m is the total number of samples, and y i It represents the cumulative amount of heavy metal pollution predicted by the model and the actual monitored cumulative amount of heavy metal pollution.

[0136] The LTSM-RF model was trained using the training set, and its performance was evaluated using the test set. The prediction results are shown in Table 1. The model accuracy is higher than that of the traditional inverse distance weighted interpolation method (Table 2).

[0137] Table 1 Prediction results of soil heavy metals based on CA-LTSM-RF;

[0138]

[0139] Table 2 Prediction results of soil heavy metals based on inverse distance weighted interpolation;

[0140]

[0141]

[0142] The above is only a further embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes based on the technical solutions and concepts of the present invention within the scope disclosed by the present invention, which fall within the scope of protection of the present invention.

Claims

1. A CA-LSTM-RF deep learning framework for spatiotemporal prediction of soil heavy metal accumulation, characterized by: The steps include: Obtain soil heavy metal content data and environmental variable datasets affecting soil heavy metal accumulation at each sampling point in the target area over multiple historical periods. Perform outlier testing, missing data supplementation, alignment processing, quantification processing, normalization processing, and time series coding on the data. By improving the CA model, designing a time-adaptive diffusion process simulation, determining the cellular water flow direction, and using the Manning formula to calculate the diffusion of heavy metal surface runoff; The diffusion value of heavy metals with surface runoff and environmental variable data are used as input, and the historical data relationship is extracted through LSTM to output the predicted value of soil heavy metal accumulation at the current moment; The output of LSTM was used as input to construct the RF model to predict the spatial characteristics of soil heavy metals; The model was constructed using Keras of TensorFlow. The diffusion values ​​simulated by environmental variables and the CA model were input into the LSTM time simulator, and their predicted values ​​were input into the RF spatial simulator. The predicted values ​​of soil heavy metals at the grid scale were output, and the model performance was evaluated.

2. The CA-LSTM-RF deep learning framework-based spatiotemporal prediction algorithm for soil heavy metal accumulation according to claim 1 is characterized by: The environmental variable data set includes historical enterprise information data, soil property data, meteorological climate data, high-resolution remote sensing image data and socio-economic data.

3. The CA-LSTM-RF deep learning framework-based spatiotemporal prediction algorithm for soil heavy metal accumulation according to claim 1 is characterized by: In the data preprocessing, the 3σ principle is used for outlier testing, and bilinear interpolation or spline interpolation is used to supplement missing data. Quantitative processing includes calculating enterprise density using the kernel density method, population density using the population-to-area ratio, road density using the road length-to-area ratio, and the distance between sample points and enterprises and residential areas using the Euclidean distance method.

4. The CA-LSTM-RF deep learning framework-based spatiotemporal prediction algorithm for soil heavy metal accumulation according to claim 1 is characterized by: In the CA model, the cellular water flow direction is determined by calculating the average water level of the central cell and the neighboring cells, eliminating the neighboring cells with water levels greater than the average, and determining the water flow direction cell; The diffusion calculation of heavy metal surface runoff uses the Manning formula, taking into account the influence of the Reynolds coefficient conversion factor, hydraulic radius, permeability coefficient and the mass of infiltrating pollutants on the hydraulic radius, and calculates the time step of runoff from the central cell to the neighboring cells and the transfer equation of heavy metals on the cells.

5. The CA-LSTM-RF deep learning framework-based spatiotemporal prediction algorithm for soil heavy metal accumulation according to claim 1 is characterized by: The input layer of the LSTM time simulator is the diffusion value of heavy metals with surface runoff and historical data of environmental variables. The output layer is the heavy metal content in the soil at time t. The information flow is controlled by the forget gate, input gate, memory unit state and output gate. The specific formula is as follows: f t =σ[w f (h t-1 ,x t )+b f ); i t =σ[w i (h t-1 ,x t )+b i ); c t =f t ×c t-1 +i t ×tanh[w c (h t-1 ,x t )]+b c ); the t =σ[w o (h t-1 ,x t )+b o ); h t =o t ×tanh(c t ); Among them, f t ,i t , c t and o t are the vectors of forget gate, input gate, memory cell state and output gate respectively; w f , w i , w c and w o are weights respectively; b f , b i , b c and b o are the bias vectors of the forget gate, input gate, memory unit state, and output gate respectively; x t is the input vector at time t; h t-1 is the output vector at time t-1; h t is the output vector at time t; c t-1 is the memory unit at time t-1; c t is the memory unit at time t; σ is the Sigmoid activation function and tanh is the hyperbolic tangent activation function.

6. The CA-LSTM-RF deep learning framework-based spatiotemporal prediction algorithm for soil heavy metal accumulation according to claim 1 is characterized by: The input of the RF spatial simulator is the output of the LSTM model at time t and the environmental feature vector that affects the accumulation of heavy metals in soil. The prediction formula is: Where K is the number of trees, f k (X) is the prediction result of the k-th tree; Hyperparameter settings: Estimators: 125; Max_depth:18; Max_features: 17; Min_samples_leaf:2; Min_samples_split:

2.

7. The CA-LSTM-RF deep learning framework-based spatiotemporal prediction algorithm for soil heavy metal accumulation according to claim 6 is characterized by: The performance evaluation of the model was performed using ten-fold cross validation and time cross validation, and the accuracy evaluation index was the determination coefficient R 2 , root mean square error RMSE and mean absolute error MAE, the specific formulas are as follows: Where m is the total number of samples; and y i It represents the cumulative amount of heavy metal pollution predicted by the model and the actual monitored cumulative amount of heavy metal pollution.

Images