A deep neural network-based gas emission rate inversion method

Abstract

The application provides a SO2 gas emission rate inversion method based on a deep neural network. The method comprises the following steps: pre-processing original image data; obtaining an optical flow according to the pre-processed data; dividing the processed data into a training set and a debugging set; constructing a deep neural network model and optimizing the deep neural network model; and using the optimized deep neural network model to realize inversion of the emission rate. Compared with the prior art, the application has strong anti-interference ability, can greatly solve the problem of inversion of the plume emission rate at the image edge, and more importantly, realizes real-time, rapid and accurate inversion of the plume emission rate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of environmental protection monitoring, and particularly relates to a gas emission rate inversion method based on a deep neural network. BACKGROUND

[0002] Sulfur dioxide (SO2) as a toxic gas emitted by human and natural sources has different degrees of impact on social environment and economic nature and climate. For example, it can cause respiratory diseases in humans, inhibit or even destroy crop growth, and form acid rain. Therefore, it is necessary to study and monitor SO2, which helps to better assess the regional and global impact of SO2 emissions. SO2 ultraviolet imaging technology has a significant advantage in terms of time resolution and spatial resolution, and has successfully been applied to the remote sensing monitoring of SO2 from volcanoes, industrial chimneys and ship exhaust. SO2 ultraviolet imaging technology can obtain two-dimensional SO2 concentration images of plumes in real time, quickly and accurately, to analyze the spatial evolution of plume dynamics in real time. However, in engineering practice, more attention is often paid to the amount of SO2 emission. Therefore, the SO2 emission rate is an important parameter for better assessing the impact of SO2 emissions.

[0003] Currently, there are mainly three ways to convert SO2 concentration to SO2 emission rate. The first way is to combine SO2 concentration with wind speed to invert SO2 emission rate. This method takes wind speed as the movement speed of SO2 gas, which can only roughly estimate the size of SO2 emission rate and is not suitable for complex plume conditions. The second method is the cross-correlation method, which uses the correlation of SO2 concentration to calculate the average speed of SO2 gas in the same direction of movement, thereby inverting the SO2 emission rate of the plume cross section. This method is more accurate than the first inversion method, but cannot clearly show the real-time spatial changes of the plume. The third method is the optical flow method, which can convert the movement of spatial objects into the movement of pixels on the imaging plane of the sensor to accurately obtain real-time movement information of SO2 gas and thereby realize the conversion of SO2 concentration to emission rate, which is more valuable in assessing the impact of pollution areas. It is the current mainstream method for measuring gas emission rate. However, the limitation of the "small" movement assumption of optical flow hinders the development of the optical flow method, which makes the optical flow method not suitable for the edges of the image. In addition, the turbulence in the plume is the result of air thermodynamics and dynamics, which is usually manifested as a peak value and a continuous shift in direction of the optical flow vector, thereby rapidly reducing the accuracy of the emission rate. SUMMARY

[0004] In view of the defects in the prior art, the application provides a gas emission rate inversion method based on a deep neural network. The method can realize accurate inversion of SO2 gas emission rate, enhance anti-interference capability to turbulence, solve the problem of inversion of SO2 emission rate of edge plume section, and more importantly, realize real-time, rapid and accurate inversion of plume emission rate.

[0005] To achieve the above-mentioned application purposes, the technical solutions of the application are as follows:

[0006] A gas emission rate inversion method based on a deep neural network comprises the following steps:

[0007] In step S10, the following steps are included:

[0008] 1.1) Obtain the optical thickness image τ of the signal channel by using the sky background image I' and the plume image I of the SO2 ultraviolet camera signal channel A A0 A

[0009]

[0010] Obtain the optical thickness image τ of the reference channel by using the sky background image I' and the plume image I of the SO2 ultraviolet camera reference channel B B0 B

[0011]

[0012] Subtract the optical thickness image τ of the signal channel from the optical thickness image τ of the reference channel to obtain the optical thickness image τ of SO2 gas A B

[0013]

[0014] 1.2) Substitute the optical thickness image τ of SO2 gas into the calibration curve obtained by the spectrometer to calculate the SO2 concentration image.

[0015] 1.3) Perform denoising processing on the SO2 concentration image to obtain a denoised image

[0016] The steps of step S20 include:

[0017] 2.1) Set the threshold value S of the SO2 gas concentration P , and screen the ROI region S of interest from the denoised image obtained in step 1)2.2) Calculate the SO2 emission rate of the ROI region S of interest2.3) Calculate the SO2 emission rate of the whole plume​​​​​​​ROI , select the denoised image greater than the threshold value S P , the ROI region S of interest ROI , extract the ROI region S of interest ROI , get the domain image

[0018] Then, from the domain image extracted to get the SO2 gas concentration S;

[0019] 2.2) Map the domain image to the interval

[0255] to get the interval image

[0020]

[0021] Where x, y represent the horizontal and vertical coordinates of the two-dimensional plane where the domain image and the interval image are located, S max is the maximum concentration value of SO2 gas concentration S in the domain image ;

[0022] Then, based on the constant value assumption of the interval image , the optical flow data V is calculated by the Farneback optical flow method,

[0023] The steps of step S30 include:

[0024] The modeling data set P is divided into training set P Train and debugging set P Test ,

[0025] Where the modeling data set P is N frames of domain images Based on the same position SO2 gas concentration data S and optical flow data V set, the training set P Train and debugging set P Test are N1, N2 frames, and N=N1+N2;

[0026] The steps of step S40 include:

[0027] 4.1) Define the input layer, hidden layer and output layer of the BP neural network topology structure, and connect the adjacent two layers in full connection mode,

[0028] Define the number of nodes of the input layer and the output layer, determine the number of layers of the hidden layer and the number of nodes of each layer,

[0029] Determine the transfer function between the input layer, the hidden layer and the output layer, and determine the training function between the hidden layers,

[0030] Setting momentum factor J, training times K, learning rate F, training target minimum error U;

[0031] 4.2) Defining training set P Train forward concentration set and lag light flow set Defining debugging set P Test forward concentration set and lag light flow set

[0032] Wherein, the forward concentration set Train of training set P is the SO2 gas concentration data of the first plume section along the overall motion direction of the plume in each frame image of N1 frames The lag light flow set Train of training set P is the SO2 gas concentration data of the second plume section along the overall motion direction of the plume in each frame image of N1 frames The light flow data at i , 0≤i≤N1;

[0033] The forward concentration set Test of debugging set P is the SO2 gas concentration data of the first plume section along the overall motion direction of the plume in each frame image of N2 frames The lag light flow set Test of debugging set P is the SO2 gas concentration data of the second plume section along the overall motion direction of the plume in each frame image of N2 frames The light flow data at j , 0≤j≤N2;

[0034] 4.3) Calculating the weighted time T Train , T Test ' and the weighted temperature C i , C j ' of training set P i and debugging set P j respectively,

[0035]

[0036]

[0037]

[0038]

[0039]

[0040] where T i is the time of the i-th image in the training set P Train where T j is the time of the j-th image in the debugging set P Test where T N is the time of the k-th image in the modeling data set P i where T Train is the atmospheric temperature of the i-th image in the training set P j where T Test is the atmospheric temperature of the j-th image in the debugging set P k where T i is the atmospheric temperature of the k-th image in the modeling data set P, 0 < k ≤ N, where T

[0041] where D is the pixel distance between the two (between the first and the second) plumes cross sections, and G i is the variable factor of the i-th image, and G j is the variable factor of the j-th image.

[0042] G i = [T i , C i , D]

[0043] G j = [T j , C j , D]

[0044] where G Train is the variable factor set of the training set P where G Test is the variable factor of the debugging set P where G is the variable factor set of the N1-th image in the training set P Train where G i is the variable factor set of the N2-th image in the debugging set P Test ;

[0045] 4.4) The training set P j is input into the BP neural network model, and the lagging light flow set is output, so that a mapping relationship is established between the forward concentration set and the lagging light flow set , that is, the initial BP neural network model DNN0 is obtained.

[0046] 4.5) The debugging set P Train is input into the BP neural network model, and the lagging light flow set is output, so that a mapping relationship is established between the forward concentration set and the lagging light flow set , that is, the initial BP neural network model DNN0 is obtained.

[0046] 4.6) The modeling data set P TestForward concentration set and variable factor set Input an initial backpropagation (BP) neural network model DNN0, and the initial BP neural network model DNN0 outputs a debug set P. Test Output optical flow set

[0047] 4.6) Calculate the output optical flow set SO2 emission rate Φ′ in the j-th frame of the image j ,

[0048]

[0049] Where f is the camera focal length, and m and M are the positions of the cross-section of the second plume. Let be the normal vector of the plume cross section. For output optical flow set The output optical flow data of the j-th frame of the image, d p Let Δs be the distance between the camera and the target plume source, and Δs be the integration step size.

[0050] With debug set P Test The hysteresis optical flow collection As the reference optical flow, calculate the debugging set P. Test Lag optical flow collection The baseline SO2 emission rate Φ of the j-th frame image j ,

[0051]

[0052] 4.7) Calculate the debug set P Test SO2 emission rate Φ′ in each of the N2 frames of the image j Compared with the baseline SO2 emission rate Φ j The absolute value of the deviation between them, E j If the absolute value of the deviation in each frame is E', j If all values ​​are less than the set error threshold E, then the deep BP neural network model DNN2 is obtained.

[0053] 4.8) Conversely, the absolute value of the deviation E j Each frame whose error threshold E is greater than or equal to is used as feedback input to the initial BP neural network model DNN0, and steps 4.5)-4.7) of S40 are repeated until the SO2 emission rate Φ′ of each frame in N2 frames is obtained. j Compared with the baseline SO2 emission rate Φ j The absolute value of the deviation between them, E j If all values ​​are less than the set error threshold E, then the deep BP neural network model DNN2 is obtained.

[0054] In step S50, the concentration set to be measured and the corresponding variable factor set G are input into the deep BP neural network model DNN2, and the output optical flow set is obtained.Then, the accurate SO2 emission rate Φ' is obtained by substituting the model SO2 emission rate Φ' in step S40 into the inversion formula. j

[0055] Compared with the traditional method, the advantages of the present application are:

[0056] 1. In the actual pollution gas monitoring process, the camera often takes the whole plume source, where the plume may exist turbulence, thereby causing the calculation accuracy of the optical flow algorithm to decrease rapidly. The present application enhances the anti-interference ability of the method by incorporating a deep neural network, and can effectively suppress the rapid change of turbulence.

[0057] 2. The traditional optical flow algorithm does not meet the "small" motion assumption near the edge of the image, resulting in the loss of correlation between adjacent frames of the optical flow image, and thus is not conducive to the accurate inversion of the edge SO2 emission rate. The proposal of the deep neural network optical flow method can well solve this problem. The present application has strong applicability and can realize the inversion of the emission rate of the edge plume cross section, greatly solving the problem of distortion of the edge emission rate inversion of the traditional optical flow algorithm.

[0058] 3. The present application combines the high temporal resolution and high spatial resolution of the SO2 ultraviolet camera, the pixel-level accurate detection of object motion information of the optical flow algorithm, and the technical advantages of high-speed optimization solution of the neural network, and can realize the real-time, rapid and accurate inversion of the plume emission rate. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 is a comparison chart of the SO2 average emission rate and the error of the average emission rate inverted by the optical flow method and the deep neural network of the present application (columns 1243-1343).

[0060] Figure 2 is a comparison chart of the operation time of each frame of SO2 gas optical flow data obtained by the optical flow method and the deep neural network method of the present application (columns 1243-1343).

[0061] Figure 3 is an example of the selection of the plume cross section of the present application (marking the plume cross section).

[0062] Figure 4 is an example of the selection of the plume cross section of the present application (marking the edge plume cross section). Figure 3

[0063] Figure 5 is a comparison chart of the SO2 average emission rate and the error of the average emission rate inverted by the optical flow method and the deep neural network of the present application (columns 1243-1343).​Figure 3 Contrast plot of optical flow of examples. DETAILED DESCRIPTION

[0064] A deep neural network-based SO2 gas emission rate inversion method will be further explained and described below in combination with the accompanying drawings.

[0065] The overall process mainly includes five steps: data preprocessing, optical flow acquisition, data classification, model construction and optimization, and emission rate inversion. It should be noted that:

[0066] Step S10, data preprocessing;

[0067] Specifically, the sky background image I' of the signal channel of the SO2 ultraviolet camera A and the plume image I A0 obtain the optical thickness image τ A of the signal channel,

[0068]

[0069] The sky background image I' of the reference channel of the SO2 ultraviolet camera B and the plume image I B0 obtain the optical thickness image τ B of the reference channel,

[0070]

[0071] The optical thickness image τ A of the signal channel and the optical thickness image τ B of the reference channel are subtracted to obtain the optical thickness image of SO2 gas

[0072]

[0073] The optical thickness image of SO2 gas is substituted into the calibration curve obtained by the spectrometer to calculate the SO2 concentration image.

[0074] A 3x3 median filter is used to denoise the SO2 concentration image to obtain a denoised image

[0075] Step S20, optical flow acquisition, the specific flow chart is shown in Figure 3 .

[0076] Specifically, the threshold S P of SO2 gas concentration is set to 2.5x10 15 molecules / cm 2 , and greater than the threshold S Ppart of the ROI region S of interest ROI extracting the ROI region S of interest ROI obtaining the intra-domain image mapping the intra-domain image to the interval image in the interval

[0255]

[0077]

[0078] wherein x, y respectively represent the horizontal and vertical coordinates of the two-dimensional plane in which the intra-domain image and the interval image are located, S max is the maximum concentration value in the intra-domain image ;

[0079] Based on the assumption that the numerical value of the interval image is constant, the Farneback algorithm is called by Matlab R2021b software to calculate the optical flow V.

[0080] Step S30, data classification.

[0081] Specifically, the modeling data set P is the intra-domain image Based on the set of SO2 gas concentration data S and optical flow data V at the same position, P = (S, V) 9000 , the training set P Train and the debugging set P Test have 7000, 2000 frames, respectively.

[0082] Step S40, model construction and optimization.

[0083] 4.1) Define the input layer, hidden layer and output layer of the BP neural network topology structure, and connect the adjacent two layers in a full connection mode.

[0084] Define the number of nodes in the input layer and the output layer as 501, determine the number of layers of the hidden layer as 10 and the number of nodes in each layer [3, 5, 5, 5, 5, 5, 6, 7, 8, 10] and the training function trainlm,

[0085] Determine the transfer function between the input layer, the hidden layer and the output layer, wherein the transfer function from the input layer to the 9th layer of the hidden layer is the tansig function, and the transfer function from the 10th layer of the hidden layer to the output layer is the purelin function.

[0086] Set the momentum factor J = 0.95, the training times K = 1000, the learning rate F = 0.01, and the training target minimum error U = 0.01.

[0087] 4.2) Define the training set P Trainforward concentration set of P and a set of lag optical flow of P

[0088] wherein the forward concentration set of P Train is the SO2 gas concentration data of the first plume section along the plume overall motion direction of each frame image in the training set P Train the set of lag optical flow of P is the SO2 gas concentration data of the second plume section along the plume overall motion direction of each frame image in the training set P

[0089] wherein the forward concentration set of P Test and a set of lag optical flow of P

[0090] wherein the forward concentration set of P Test is the SO2 gas concentration data of the first plume section along the plume overall motion direction of each frame image in the training set P Test is the SO2 gas concentration data of the second plume section along the plume overall motion direction of each frame image in the training set P

[0091] 4.3) calculating the weighted time T Train , T j and the weighted temperature C i , C j of the training set P Test and the debugging set P i respectively,

[0092]

[0093]

[0094]

[0095]

[0096]

[0097] wherein T i is the training set P Train ​​​​​​​​​​The time of the i-th frame, T j For debug set P Test The time of the j-th frame, T 9000 To model the time of the 9000th frame image in dataset P, C i For training set P Train The atmospheric temperature, C, of ​​the i-th frame image. j For debug set P Test The atmospheric temperature, C, of ​​the j-th frame image. k To model the atmospheric temperature of the k-th frame image in dataset P, where 0 < k ≤ 9000, The average atmospheric temperature is the value of 9000 frames of images.

[0098] The variable factors G for the i-th and j-th frames are set by combining the pixel spacing D of the cross-sections of two adjacent plumes (between the first and second plumes). i G j ,

[0099] G i =[T i C i [D]

[0100] G j =[T j C j [D]

[0101] Define training set P Train The set of variable factors G 7000 Debug set P Test The set of variable factors G 2000 Variable factor set G 7000 For training set P Train Variable factors G of 7000 frames of images i The set of variables, G 2000 For debug set P Test Variable factors G of 2000 frames of images j A set;

[0102] 4.4) The training set P Train Forward concentration set and variable factor set G 7000 As input, hysteresis optical flow set Substituting the output into the BP neural network model makes the forward concentration set and delayed optical flow collection Establish a mapping relationship between them to obtain the initial BP neural network model DNN0.

[0103] 4.5) Debug set P Test Forward concentration set and variable factor set G 2000Input an initial backpropagation (BP) neural network model DNN0, and the initial BP neural network model DNN0 outputs a debug set P. Test Output optical flow set

[0104] 4.6) Calculate the output optical flow set SO2 emission rate Φ′ in the j-th frame of the image j ,

[0105]

[0106] Where f is the camera focal length and f = 25mm, m and M are the positions of the second plume cross-section, and n is the normal vector of the plume cross-section. For output optical flow set The output optical flow data of the j-th frame of the image, d p Let d be the distance between the camera and the target plume source. p =10.3km, Δs is the integration step size and Δs = 4.65μm.

[0107] With debug set P Test The hysteresis optical flow collection As the reference optical flow, calculate the debugging set P. Test Lag optical flow collection The baseline SO2 emission rate Φ of the j-th frame image j ,

[0108]

[0109] 4.7) Calculate the debug set P Test SO2 emission rate Φ′ in each of the 2000 frames of images j Compared with the baseline SO2 emission rate Φ j The absolute value of the deviation between them, E j If the absolute value of the deviation in each frame is E', j All of them are less than the set error threshold E, where E = 1 kg / s, thus obtaining the deep BP neural network model DNN2;

[0110] 4.8) Conversely, the absolute value of the deviation E j Each frame with an SO2 emission rate greater than or equal to 1 kg / s is used as feedback input to the initial BP neural network model DNN0. Steps 4.5)-4.7) of S40 are repeated until the SO2 emission rate Φ′ of each frame in 2000 frames is obtained. j Compared with the baseline SO2 emission rate Φ j The absolute value of the deviation between them, E j The values ​​are all less than 1 kg / s, thus obtaining the deep backpropagation neural network model DNN2;

[0111] Step S50, the emission rate inversion.

[0112] The concentration set to be measured is input into the deep BP neural network model DNN2, and the output optical flow set is obtained and the corresponding variable factor set G, that is, the output optical flow set Then, the accurate SO2 emission rate Φ' is obtained by substituting the model SO2 emission rate Φ' in step S40 into the inversion formula. j

[0113] Case analysis

[0114] MatlabR2021b software is used for programming simulation, and the inversion results of the deep neural network model DNN2 and the Farneback optical flow method are compared and verified.

[0115] The simulation results are shown in Figure 1 The SO2 gas velocity changes obtained by the deep neural network model DNN2 and the Farneback optical flow method of the application are very good in time sequence, which shows that the calculation accuracy of the deep neural network model DNN2 of the application is high, and also shows that the application has the advantages of accurately simulating and analyzing the complex motion state of the gas plume. At the same time, the real-time processing of the method of the application to the ultraviolet camera data can be much better than the optical flow method. The simulation results are shown in Figure 2 The time required for the Farneback optical flow method is longer, and it is slightly insufficient in real-time processing of data obtained by the ultraviolet camera with high time resolution, while the deep neural network model of the application significantly reduces the calculation amount and shortens the time consumption, which can greatly meet the processing demand of real-time data, and has the advantages of real-time, fast and accurate compared with the Farneback optical flow method.

[0116] At the same time, the gas emission rate inversion method based on the deep neural network of the application overcomes the problems of turbulent flow effect and emission rate inversion distortion of the edge plume section. Further analysis is made with Figure 3 the number of columns selected in the turbulent plume section being 1055, 1075, and the number of rows being from 70-570 (corresponding to the values of m, M), and the comparison of the optical flow of one frame obtained by the optical flow method and the deep neural network is shown in Figure 5 The SO2 gas velocity obtained by the optical flow method is affected by the turbulent flow, and there is an obvious speed peak and direction deviation, while the deep neural network model DNN2 of the application greatly weakens the abnormal influence of the turbulent flow on the inversion result due to the optimized solution, and has higher inversion accuracy and anti-interference compared with the Farneback optical flow method.

[0117] Figure 4 ​​The column numbers of the edge plume section are 1323, 1343, and the row numbers are from 70 to 570 (corresponding to the values of m and M). The SO2 emission rates and the corresponding error of the 1243-1343 columns of the 200 frames of the optical flow method and the deep neural network inversion are compared as shown in Figure 1 As shown, the closer to the image edge (i.e. after the 1315th column), the error of the average SO2 emission rate inverted by the Farneback optical flow method is significantly larger, and the error at the edge is as high as 90%, which is at the image edge, since the motion information of the pixels of the previous frame image used for inversion is more difficult to match the motion information corresponding to the pixels of the current frame image, thus causing a large error in the optical flow inversion; however, under the double influence of the low contrast of the image and the plume turbulence effect, the deep neural network model DNN2 of the present application still maintains high accuracy, and the inversion error is not more than 10%. It can be seen that the deep neural network model DNN2 of the present application can realize accurate inversion of the emission rate of the edge plume section.

[0118] In summary, the gas emission rate inversion method based on the deep neural network of the present application greatly overcomes the influence of the plume turbulence effect and the pixels at the image edge that do not meet the "small" motion assumption, has strong anti-interference ability, can effectively suppress the rapid change of turbulence, and greatly solves the problem of emission rate inversion distortion of the edge plume section. The present application realizes real-time, rapid and accurate inversion of the plume emission rate.

[0119] The above only describes the preferred embodiments of the present application and is not intended to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for gas emission rate inversion based on deep neural networks, characterized in that The method comprises the following steps: Step S10 1.1) Sky background image I' through the SO2 UV camera signal channel A and plume image I A0 Compute optical thickness image τ of signal channel A B and plume image I B0 Compute optical thickness image τ of reference channel B ,​ Optical thickness image τ of the signal channel A Optical thickness image τ of the reference channel B Difference processing to obtain optical thickness image τ of SO2 gas 1.2) The optical thickness image of SO2 gas is calculated The SO2 concentration image is calculated by substituting the calibration curve obtained by the spectrometer, 1.3) denoising the SO2 concentration image to obtain a denoised image Step S20 2.1) setting a threshold value S of SO2 gas concentration P From the denoised image The part greater than the threshold value S P is the ROI region S of interest ROI , obtaining the in-domain image Then, the SO2 gas concentration S is extracted from the in-domain image ; 2.2) mapping the intra-domain image to the interval [0255] to get the interval image wherein x, y represent horizontal and vertical coordinates of the domain image and the interval image respectively, S max is the maximum concentration value in the domain image Then, based on the assumption that the value of the interval image is constant, the optical flow data V is calculated by the Farneback optical flow method. Step S30 The modeling dataset P is divided into a training set P Train and a debugging set P Test , Wherein, the modeling data set P is N frames of in-domain images Based on the set of SO2 gas concentration data S and optical flow data V of the same position, the training set P Train And the debugging set P Test N1, N2 frames, and N=N1+N2; The steps of step S40 comprise: 4.1) defining the input layer, the hidden layer and the output layer of the BP neural network topology, and connecting the adjacent two layers in a full connection mode, defining the number of nodes of the input layer and the output layer, determining the number of layers of the hidden layer and the number of nodes of each layer, determining the transfer function between the input layer, the hidden layer and the output layer, and determining the training function between the hidden layers, setting the momentum factor J, the training number K, the learning rate F, and the training target minimum error U, 4.2) Defining the training set P Train the forward concentration set and the lag optical flow set Defining the debug set P Test the forward concentration set and the lag optical flow set The training set P Train The forward concentration set The SO2 gas concentration data of the first plume section along the overall motion direction of the plume in each frame image in the N1 frames The training set P Train The lag optical flow set The SO2 gas concentration data of the second plume section along the overall motion direction of the plume in each frame image in the N1 frames The optical flow data at the position of the first plume section The set, 0≤i≤N1; debug set P Test forward concentration set SO2 gas concentration data for a first plume section along the overall plume motion direction for each of the N2 frames of images set, debug set P Test lagging optical flow set SO2 gas concentration data for a second plume section along the overall plume motion direction for each of the N2 frames of images optical flow data at set, 0 < j < N2; 4.3) Calculate the training set P respectively. Train and debug set P Test Weighted time T i ′、T j ′ and weighted temperature C i C j ′, where T i is the training set P Train is the time of the i-th image, T j is the debug set P Test is the time of the j-th image, T N is the time of the k-th image of the modeling dataset P, 0 < k ≤ N, i is the training set P Train is the atmospheric temperature of the i-th image, C j is the debug set P Test is the atmospheric temperature of the j-th image, C k is the atmospheric temperature of the k-th image of the modeling dataset P, 0 < k ≤ N, is the mean of the atmospheric temperature of the N images, A variable factor G of the i-th, j-th frame is set by the pixel interval D between the smoke plume cross sections i , G j , G i = [T i , C i , D] G j = [T j , C j , D] a set of variable factors G for the N1 frames of images of the training set P Train a set of variable factors G for the N1 frames of images of the training set P a set of variable factors G for the N1 frames of images of the training set P Test a set of variable factors G for the N1 frames of images of the training set P a set of variable factors G for the N1 frames of images of the training set P a set of variable factors G for the N1 frames of images of the training set P Train a set of variable factors G for the N1 frames of images of the training set P i a set of variable factors G for the N1 frames of images of the training set P a set of variable factors G for the N1 frames of images of the training set P Test a set of variable factors G for the N1 frames of images of the training set P j a set of variable factors G for the N1 frames of images of the training set P 4.4) Train the set P Train Forward concentration set And variable factor set As input, lag light flow set As output, substitute BP neural network model, so that the forward concentration set And lag light flow set Between the mapping relationship, that is, the initial BP neural network model DNN0; 4.5) the set of debug sets P Test of forward concentration sets and variable factor sets input an initial BP neural network model DNN0, the initial BP neural network model DNN0 feedback output the set of output optical flow sets P Test of debug sets 4.6) From the set of output optical flow Compute the tuning set P Test SO2emission rate Φ' of the jth image j , where f is the camera focal length, m, M are the positions of the second plume cross section, is the normal vector of the plume cross section, d p is the distance between the camera and the target plume source, Δs is the integration step, is the output optical flow set is the output optical flow data of the jth frame of images in the set With debug set P Test The hysteresis optical flow collection As the reference optical flow, calculate the debugging set P. Test Lag optical flow collection The baseline SO2 emission rate Φ in the j-th frame of the image j , 4.7) calculating the set P of adjustments Test the absolute value E' of the deviation of the SO2 emission rate Φ' of each of the N2 frame images from the reference SO2 emission rate Φ j j j if the absolute value E' of the deviation of each of the N2 frame images from the reference SO2 emission rate Φ j is smaller than a set error threshold E, then the depth BP neural network model DNN2 is obtained.​​ 4.8) Conversely, the absolute value E' of the deviation between the SO2 emission rate Φ' of each frame of the N2 frames of images and the reference SO2 emission rate Φ j equal to or greater than the error threshold value E is input into the initial BP neural network model DNN0 as feedback, and steps 4.5)-4.7) of step S40 are repeated until the absolute value E' of the deviation between the SO2 emission rate Φ' of each frame of the N2 frames of images and the reference SO2 emission rate Φ j is less than the set error threshold value E, i.e. the deep BP neural network model DNN2 is obtained. j j is less than the set error threshold value E, i.e. the deep BP neural network model DNN2 is obtained.​ Step S50 Input the set of concentrations to be tested and its corresponding set of variable factors G into the deep BP neural network model DNN2, i.e. output the set of optical flow Then calculate the emission rate by the SO2 emission rate inversion formula.

Images