Multi-source data fusion PM2.5 monitoring method based on neural network model

Abstract

The invention discloses a multi-source data fusion PM2.5 monitoring method based on a neural network model. The method specifically comprises the following steps: S1, performing super-resolution processing on remote sensing image data to obtain first image data; s2, error correction is carried out on pollutant output data of the obtained Slam model to obtain second image data; and S3, taking the first image data and the second image data as multi-source fusion data, and carrying out PM2.5 analysis and detection.

Description

Technical Field

[0001] The present invention relates to the technical field of environmental detection, and in particular to a method for monitoring PM2.5 using multi-source data fusion based on a neural network model. Background Art

[0002] In recent years, the national ambient air quality has maintained a long-term positive trend. Although the overall air quality in the country has improved, there are still challenges in improving the PM2.5 concentration in key areas. Due to factors such as high industrialization and dense population in these areas, the task of air pollution prevention and control is still arduous.

[0003] Ambient air quality monitoring stations are generally located and selected according to the principles of representativeness, comparability, integrity, foresight and stability. However, due to factors such as the cost of monitoring station equipment, operating costs, monitoring quality and monitoring range, many scholars and experts have adopted satellite remote sensing technology in recent years to carry out large-scale, continuous remote sensing monitoring of atmospheric pollution and achieved good monitoring results. "The spatiotemporal distribution characteristics of PM2.5 in the Beijing-Tianjin-Hebei region based on remote sensing data" studies the spatiotemporal distribution characteristics of PM2.5 in the Beijing-Tianjin-Hebei region, establishes a relationship model between AOD and PM2.5 concentration, and analyzes the spatiotemporal variation of PM2.5 concentration in the Beijing-Tianjin-Hebei region, providing a scientific basis for regional joint prevention and control of air pollution. "Exploration of the establishment of a model between PM2.5 concentration and atmospheric aerosol optical thickness based on remote sensing images - taking southern Hubei as an example" uses the MODIS remote sensing image inversion method to explore the relationship model between atmospheric aerosol optical thickness AOD and PM2.5 concentration in southern Hubei, obtains a cubic relationship model, and verifies that the model has a high degree of fit.

[0004] Satellite remote sensing technology has the characteristics of global coverage and high temporal resolution, and can realize continuous and real-time monitoring of PM2.5. By acquiring remote sensing images, we can timely understand the dynamic changes of key parameters such as atmospheric composition and pollutant concentration. At present, in PM2.5 remote sensing monitoring, satellite remote sensing data such as MODIS and Sentinel-5P are often used to carry out remote sensing inversion of aerosol optical thickness, and regression analysis models are constructed in combination with meteorological, ground-based and other multimodal data.

[0005] In practical applications, since satellite remote sensing data such as MODIS and Sentinel-5P are polar-orbiting satellite data, their coverage time resolution is mostly calculated on a daily basis. In addition, the regression analysis model of aerosol optical thickness and ground-based monitoring data is easily affected by weather and day and night, resulting in data loss, which leads to discontinuous and incomplete monitoring. Summary of the invention

[0006] In view of the technical problem in the prior art that remote sensing data is prone to defects, resulting in discontinuous and incomplete monitoring, the present invention proposes a method for monitoring PM2.5 by multi-source data fusion based on a neural network model. The resolution of multi-source remote sensing data is improved by using super-resolution reconstruction technology to ensure the consistency of spatial resolution of the training data set; the multi-source data errors are corrected by using the weighted extreme gradient boosting technology based on spatiotemporal information to improve the data accuracy of the training data set; the PM2.5 concentration is monitored using the multi-source and multi-modal data dynamic time warping MLP neural network technology to achieve hourly, continuous and complete remote sensing monitoring.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A method for monitoring PM2.5 by multi-source data fusion based on a neural network model specifically includes the following steps:

[0009] S1: Performing super-resolution processing on remote sensing image data to obtain first image data;

[0010] S2: performing error correction on the pollutant output data obtained from the Silam model to obtain the second image data;

[0011] S3: Use the first image data and the second impact data as multi-source fusion data to perform PM2.5 analysis and detection.

[0012] Preferably, the S1 includes:

[0013] S1-1: Obtaining remote sensing historical data from a server, and performing a first preprocessing on the remote sensing historical data to obtain input data; the first preprocessing includes matching processing, sampling processing, slicing processing, and data expansion processing;

[0014] S1-2: Input data to the existing ESRGAN model for training;

[0015] S1-3: Input the acquired real-time remote sensing image data into the trained ESRGAN model, thereby outputting the first image data.

[0016] Preferably, the S1-1 includes:

[0017] S1-1-1: Perform matching processing on the remote sensing historical data to obtain the first data.

[0018] S1-1-2: resample the first data to 50m as a first-precision image, and then downsample the first-precision image to 200m as a second-precision image, to obtain a set of matching first-precision-second-precision image groups as the second data;

[0019] S1-1-3: Segment the first precision image in the second data into first slices of 96×96 pixels, segment the corresponding second precision image into second slices of 24×24 pixels, and combine the first slice and the second slice to obtain third data;

[0020] S1-1-4: Rotate the first slice and the second slice in the third data by 90°, 180°, and 270° around the geometric center of the slice, and use the slice data obtained after the rotation as input data.

[0021] Preferably, the S1-1-1 includes:

[0022] A: First, calculate the band mean of each data in the remote sensing historical data, match the band bandwidth and the band mean, and ensure the consistency of different satellite data sources in the same band;

[0023] B: Use the radiation brightness scale parameter and radiation brightness bias parameter to convert the digital quantization value into the radiation brightness value:

[0024] L=gain×DN+Bias (1)

[0025] In formula (1), L represents the apparent reflectance or radiance value; gain represents the radiance scale parameter; Bias represents the radiance bias parameter; DN represents the remote sensing image pixel brightness value;

[0026] C: Perform FLAASH atmospheric correction on the radiance value;

[0027] D: Select the near-infrared band, red band and green band in the historical remote sensing data after atmospheric correction as the first data.

[0028] Preferably, the S2 includes:

[0029] S2-1: performing a second preprocessing on the pollutant output data obtained from the Silam model, the second preprocessing including removing outliers and missing values;

[0030] S2-2: The pollutant output data of the Silam model after the second pretreatment is used as the influencing factor and the monitoring value of the Silam model is used as the result to construct the limiting gradient model:

[0031]

[0032] Actual value, represents the predicted value of the i-th sample of the model; Ω(f j ) represents the regularization term of the jth tree; t represents the total number of trees; p represents the total number of samples;

[0033] S2-3: Use the constructed extreme gradient boosting model to correct the existing Silam pollutant data to obtain the second image data.

[0034] Preferably, the S2 further includes:

[0035] S2-4: Constructing spatiotemporal geographic weighted models:

[0036]

[0037] In formula (3), y i represents the pollutant data of the i-th point in the Silam model, (μ i ,ν i ,t i ) represents the space-time coordinates of the i-th point, β0(μ i ,ν i ,t i )x ik represents the kth spatiotemporal regression parameter at the ith point, which can be regarded as a function of the geographical location and observation time, and is the random error of the spatiotemporal region of the ith point; β0(μ i ,ν i ,t i ) represents the pollutant data of the i-th point, ε i represents random error;

[0038] S2-5: Use the constructed spatiotemporal geographic weighted model to further correct Silam pollutant data at different times, spaces and geographical locations.

[0039] Preferably, S3 includes:

[0040] S3-1: taking the first image data and the second impact data as multi-source fusion data, performing dynamic time warping on the multi-source fusion data, and outputting convolution data;

[0041] S3-2: Perform CNN convolution on the convolution data and output the predicted data;

[0042] S3-3: Use the Transformer encoder to process the predicted data to obtain the output sequence, which is the prediction result of PM2.5.

[0043] Preferably, the S3-1 includes:

[0044] S3-1-1: Set the first time series X and the second time series Y to have lengths of m and n respectively, and create a (m+1)×(n+1) two-dimensional array, i.e., the original cumulative distance matrix; the elements in the two-dimensional array are all PM2.5 historical monitoring concentrations sorted by time series;

[0045] S3-1-2: Set the first row and first column of the original cumulative distance matrix corresponding to the distance between the empty sequence and another sequence to zero;

[0046] S3-1-3: Starting from the upper left corner of the original cumulative distance matrix, traverse each element and fill in the value before calculating the cumulative distance:

[0047] d(i,j)=d(X i ,Y j )+min[D(i-1,j),D(i,j-1),D(i-1,j-1)] (4)

[0048] In formula (4), d(i,j) represents the DTW distance from the first element to the i-th element of X and from the first element to the j-th element of Y; d(X i ,Y j ) represents the distance between the i-th element of X and the j-th element of Y; min[D(i-1,j),D(i,j-1),D(i-1,j-1)] means that when searching for the best path, the previous position of the current position (i,j) is the upper left, directly above, or to the left, and the direction with the smallest cumulative distance among these three directions is selected as the cumulative distance of the current position.

[0049] S3-1-4: Extract the elements whose cumulative distance is less than or equal to the preset threshold as convolution data.

[0050] Preferably, the S3-2 includes:

[0051] S3-2-1: At the initial stage of CNN, the data input layer performs the third preprocessing on the convolution data, which includes de-meaning and normalization operations;

[0052] S3-2-2: performing a convolution operation on the convolution data after the third preprocessing by using a plurality of convolution kernels;

[0053] S3-2-3: After the convolution calculation, the data will pass through the excitation layer, and the ReLU activation function will be used to perform nonlinear transformation on the output of the convolution layer;

[0054] S3-2-4: After the incentive layer, the data enters the pooling layer;

[0055] S3-2-5: Finally, after being processed by multiple convolutional and pooling layers, the data is flattened and input into the fully connected layer, and classified or recognized by the classifier, and finally the predicted data is output.

[0056] Preferably, the S3-3 includes:

[0057] S3-3-1: First, convert the input prediction data into a feature vector;

[0058] S3-3-2: The feature vector is sent to the encoder module for encoding to obtain encoded data;

[0059] S3-3-3: The encoded data is input into the decoder module for decoding to obtain decoded data;

[0060] S3-3-4: The decoded data is converted into the probability distribution of the monitoring concentration through the linear layer and the softmax function. According to the probability distribution, the monitoring concentration with the highest probability is selected as the prediction output of the current position;

[0061] S3-3-5: Repeat S3-3-1 to S3-3-4 until a complete output sequence is generated, i.e., the prediction result of PM2.5.

[0062] In summary, due to the adoption of the above technical solution, compared with the prior art, the present invention has at least the following beneficial effects:

[0063] This patent uses super-resolution reconstruction technology to improve the resolution of multi-source remote sensing data and ensure the consistency of spatial resolution of the training data set; uses weighted extreme gradient boosting technology based on spatiotemporal information to correct multi-source data errors and improve the data accuracy of the training data set; uses multi-source multi-modal data dynamic time warping MLP neural network technology to monitor PM2.5 concentration and achieve hourly, continuous and complete remote sensing monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 The figure is a flow chart of a method for monitoring PM2.5 using multi-source data fusion based on a neural network model according to an exemplary embodiment of the present invention.

[0065] Figure 2 FIG. 4 is a schematic diagram of a super-resolution processing flow according to an exemplary embodiment of the present invention.

[0066] Figure 3 Schematic diagram of a process of performing error correction on pollutant output data according to an exemplary embodiment of the present invention.

[0067] Figure 4 5 prediction process according to an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0068] The present invention is further described in detail below in conjunction with the examples and specific implementation methods. However, this should not be understood as the scope of the above subject matter of the present invention being limited to the following examples, and all technologies realized based on the content of the present invention belong to the scope of the present invention.

[0069] In the description of the present invention, it is necessary to understand that the terms "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0070] like Figure 1 As shown, the present invention provides a method for monitoring PM2.5 using multi-source data fusion based on a neural network model, which specifically includes the following steps:

[0071] S1: Perform super-resolution processing on the remote sensing image data to obtain first image data, thereby improving the resolution of the remote sensing image data and ensuring the consistency of the spatial resolution of the training data set.

[0072] In remote sensing monitoring, in order to ensure the temporal and spatial consistency of the data set, resampling methods are usually used, such as the nearest neighbor method, bilinear interpolation method, and cubic convolution interpolation method. The nearest neighbor method assigns the corresponding pixel of the output raster dataset by taking the nearest pixel value in the input raster dataset as the input value; the bilinear interpolation method calculates the new pixel value by weighted average of the pixel values ​​of the 4 nearest pixels (4 neighborhoods) in the input raster dataset, and assigns it to the corresponding pixel of the output raster dataset. The weight is determined by the distance between the center of each pixel in the 4 neighborhoods and the interpolation point; the cubic convolution interpolation method is based on the quintic polynomial interpolation method, which calculates the new pixel value by weighted average of the pixel values ​​of the 16 nearest pixels (16 neighborhoods) in the input raster dataset, and assigns it to the corresponding pixel of the output raster dataset. The weight is determined by the distance between the center point of each pixel in the 16 neighborhoods and the interpolation point. The above three methods simply copy the nearest pixel value or increase the resolution of raster data through interpolation, but the data accuracy is not actually improved.

[0073] In this embodiment, the sunflower 8 / 9 satellite is a geostationary meteorological observation satellite launched by the Earth Observation Research Center (EORC) of Japan in 2015 and 2016. The time resolution of the multispectral remote sensing data product based on the sunflower 8 / 9 satellite is 10 minutes, which has a high practical application value for carrying out regional PM2.5 short-term multi-frequency change monitoring. However, due to the sensor limitations of geostationary satellites and the limitations of data transmission, the spatial resolution of the acquired multispectral remote sensing images is at the kilometer level, which is difficult to meet the needs of high-precision analysis. In order to overcome this limitation, the super-resolution image reconstruction technology, which is widely used in the field of image reconstruction, is applied to the processing of sunflower 8 / 9 satellite remote sensing images to improve the details and spatial resolution of the images.

[0074] like Figure 2 As shown, S1 includes:

[0075] S1-1: Obtaining remote sensing historical data from a server, and performing a first preprocessing on the remote sensing historical data to obtain input data; the first preprocessing includes matching processing, sampling processing, slicing processing, and data expansion processing.

[0076] S1-1-1: In this embodiment, matching processing is performed on the remote sensing historical data to obtain the first data.

[0077] A: First, calculate the band mean of each data in the remote sensing historical data, match the band bandwidth and the band mean, and ensure the consistency of different satellite data sources in the same band.

[0078] B: Obtain the image scale parameters and bias parameters of the Sunflower 8 / 9 satellite and Sentinel-2 satellite in the historical remote sensing data; use the reflectivity gain parameter and reflectivity bias parameter to convert the digital quantization value of the remote sensing image data into the apparent reflectivity, and use the radiance scale parameter and radiance bias parameter to convert the digital quantization value into the radiance value. The formula is as follows:

[0079] L=gain×DN+Bias (1)

[0080] In formula (1), L represents the apparent reflectance or radiance value; gain represents the radiance scale parameter; Bias represents the radiance bias parameter; and DN represents the remote sensing image pixel brightness value.

[0081] C: Performing FLAASH atmospheric correction on the radiation brightness value is a prior art.

[0082] D: Select the near-infrared band, red band and green band from the atmospherically corrected remote sensing historical data (which itself consists of multiple channel bands) as the first data.

[0083] S1-1-2: Sampling the first data to obtain the second data.

[0084] First, the first data is resampled to 50m as the first precision (high precision) image, and then the high precision image is downsampled to 200m as the second precision (low precision) image, to obtain a set of matching first precision-second precision image groups as the second data.

[0085] S1-1-3: Slice the second data to obtain third data.

[0086] In this embodiment, the first precision image in the second data is segmented into first slices of 96×96 pixels, and the corresponding second precision image is segmented into second slices of 24×24 pixels. The first slice and the second slice are combined to obtain the third data.

[0087] S1-1-4: Perform data expansion processing on the third data to obtain input data.

[0088] In this embodiment, the first slice and the second slice in the third data are rotated 90°, 180°, and 270° around the geometric center of the slice (these three angles of rotation ensure that the square data remains a square after rotation, thus completing the symmetry and transformation processing), completing the symmetry and change processing, and performing data expansion (each rotated slice is retained, that is, the slice content is the same but the angle is different, and different angles are different input data in machine recognition) as input data for training the image super-resolution reconstruction model.

[0089] S1-2: Input the input data into the existing ESRGAN (enhanced super-resolution generative adversarial networks) model for training, and then input the acquired real-time remote sensing image data into the trained ESRGAN model, so as to output the first image data with high resolution.

[0090] In this embodiment, ESRGAN is an enhanced version of SRGAN. By introducing a dense residual block (RRDB), the performance of SRGAN is greatly improved, making the overall training more sufficient. According to the experience of parameter adjustment, the number of iterations of the ESRGAN model is set to 40,000, the initial learning rate lr = 0.0001, the learning rate is halved after every 10,000 iterations, and the batch size is set to 16.

[0091] In this embodiment, after the input data is input into the ESRGAN model, the generator performs low-precision to high-precision feature extraction of the image through a convolutional layer and a dense residual module (RRDB), and then further learns the mapping relationship between low precision (second precision) and high precision (first precision) through a second convolutional layer to extract a more abstract and advanced feature representation; the feature representation is used to convert the low-resolution image (second precision image) in the multi-source data into a super-resolution image (first precision image), i.e., the first image data, through an upsampling layer and two convolutional layers.

[0092] That is, in this embodiment, the remote sensing historical data is first analyzed to obtain input data (including a first precision-second precision image group), and then the input data is input into the ESRGAN model for training to obtain a mapping relationship from the second precision image to the first precision image. Finally, this mapping relationship is used to convert the low-resolution image (second precision image) in the real-time remote sensing image data into a super-resolution image (first precision image), that is, the first image data. This process is called super-resolution.

[0093] In this embodiment, the discriminator strategy is used to calculate the evaluation index, compare the similarity between the generated super-resolution image and the original high-resolution image, and then minimize the difference between the generated super-resolution image and the real high-resolution image by repeatedly adjusting the parameters of the generator and the discriminator.

[0094] The similarity calculation formula is as follows:

[0095]

[0096] In formula (2), PSNR represents the peak signal-to-noise ratio, which uses the sum of squared errors between the super-resolution image and the high-resolution image to represent the image quality. The higher the value, the smaller the image difference. V MSE represents mean square error; SSIM represents the degree of similarity, which is calculated by using the differences in image brightness, contrast, and structural similarity. The closer its value is to 1, the higher the similarity between the two images. a and b represent two images (for example, a represents a super-resolution image, and b represents a true high-resolution image). l(a,b), s(a,b), and c(a,b) represent the brightness, contrast, and structural similarity of the two images.

[0097] In this embodiment, the Silam model is an important tool for simulating the transmission and diffusion of atmospheric pollutants, but due to multiple influences such as data sources, model parameters and environmental factors, the simulation results often have certain errors. At present, data error correction mainly relies on traditional statistical methods or empirical formulas, which have limitations in correction accuracy and adaptability. Therefore, considering the influence of time and space dimensions, an efficient and accurate data error correction method is invented, which is of great significance for improving the data results of the Silam model.

[0098] S2: Error correction is performed on the pollutant output data obtained from the Silam model to obtain the second image data and improve the data accuracy of the training data set, such as Figure 3 shown.

[0099] S2-1: Perform a second preprocessing on the pollutant output data obtained from the Silam model. The second preprocessing includes removing outliers and missing values.

[0100] Among them, the pollutant output data of the Silam model include atmospheric pollutant concentrations and environmental parameters such as wind speed, wind direction, temperature, and rainfall.

[0101] S2-2: The pollutant output data of the Silam model after the second pretreatment is used as an influencing factor and the monitoring value of the Silam model is used as a result to construct a limiting gradient model.

[0102] A tree is a data structure that consists of a hierarchical set of n (n ≥ 0) finite nodes. The extreme gradient boosting model is an optimized distributed gradient boosting tree, and its objective function can be expressed as:

[0103] The actual value, represents the predicted value of the i-th sample of the model; Ω(f j ) represents the regularization term of the jth tree; t represents the total number of trees; p represents the total number of samples.

[0105] Use the greedy algorithm to divide the subtree (the maximum gain of the subtree is the ultimate pursuit of the extreme gradient boosting model, which can be understood as the extreme gradient boosting model is to find a way to maximize the parameter G) to obtain the maximum gain G, as shown in the following formula:

[0106]

[0107] In formula (4), G represents the maximum gain; G L , G R It represents the cumulative sum of the first-order partial derivatives of the samples contained in the leaf nodes L and R, which is a constant. L , H R It represents the cumulative sum of the second-order partial derivatives of the samples contained in the leaf nodes L and R, which is a constant. λ and γ are bias parameters. If the γ value is set too large, Gain is negative, indicating that the node is not split, because the tree structure after the split becomes worse.

[0108] S2-3: Use the constructed extreme gradient boosting model to correct the existing Silam pollutant data (S2-2 is correction, and bringing it into the model for processing is correction), supplement the abnormal values ​​and missing values ​​caused by weather and Silam model problems, and obtain the second image data.

[0109] In this embodiment, in order to further reduce the data errors caused by temporal and spatial geographical factors, it also includes:

[0110] S2-4: Constructing spatiotemporal geographic weighted models:

[0111]

[0112] In formula (5), yi represents the pollutant data of the i-th point in the Silam model, (μ i ,ν i ,t i ) represents the space-time coordinates of the i-th point, β0(μ i ,ν i ,t i )x ik represents the kth spatiotemporal regression parameter at the ith point, which can be regarded as a function of the geographical location and observation time, and is the random error of the spatiotemporal region of the ith point; β0(μ i ,ν i ,t i ) represents the pollutant data of the i-th point, ε i Represents random error.

[0113] S2-5: Use the constructed spatiotemporal geographic weighted model to further correct the Silam pollutant data at different times, spaces and geographic locations (S2-4 is correction, and bringing it into the model processing is correction) to reduce data errors caused by spatiotemporal geographic factors.

[0114] The weighting coefficient is dynamically adjusted according to the size of the model prediction error to ensure that the corrected data is closer to the true value.

[0115] S3: Use the first image data and the second impact data as fusion data to perform PM2.5 analysis and detection to achieve hourly, continuous and complete remote sensing monitoring, such as Figure 4 shown.

[0116] In this embodiment, the real-time remote sensing image data processed according to super resolution in S1 is used as the first image data output and the second image data output according to Silam pollutant data in S2 is used as fusion data to perform PM2.5 monitoring.

[0117] The use of the first image data and the second image data as fusion data will increase the amount of data. The increase in the amount of data can provide more information and patterns for the network to learn, so that it can more accurately capture the relationship between input and output and improve the generalization ability of the model. However, when a large amount of data is used to train the neural network, its computational efficiency is low, and the imbalance of the data set leads to a large difference between the output result and the actual result, which reduces the accuracy of PM2.5 monitoring.

[0118] S3-1: Dynamic time warping is performed on the fused data to effectively extract similar features from the data set. Data with large feature differences will not be used as training sets, and convolution data will be output.

[0119] The present invention mainly uses the time period to compare with the historical data set in the dynamic time warping module, and selects a data set that is more in line with the current change law, so as to quickly and accurately improve the training efficiency.

[0120] S3-1-1: Initialize the time matrix: set the first time series X and the second time series Y to lengths of m and n respectively, and create a (m+1)×(n+1) two-dimensional array, i.e. the original cumulative distance matrix; the elements in the two-dimensional array are all PM2.5 historical monitoring concentrations sorted according to the time series.

[0121] S3-1-2: Set boundary conditions: The first row and the first column of the original cumulative distance matrix correspond to the distance between the empty sequence and another sequence, which are usually set to zero or some initial value.

[0122] S3-1-3: Fill the original cumulative distance matrix: Starting from the upper left corner of the original cumulative distance matrix, traverse each element to fill in the value and then calculate the cumulative distance.

[0123] Each element in the original cumulative distance matrix is ​​empty and will have a value after filling. The value of each element is calculated based on the values ​​of the three adjacent elements in the upper left corner, left and top, and the distance between the current two time series points. The minimum value of the sum of the three adjacent elements and the current distance is selected as the cumulative distance of the current position.

[0124] d(i,j)=d(X i ,Y j )+min[D(i-1,j),D(i,j-1),D(i-1,j-1)] (6)

[0125] In formula (6), d(i,j) represents the DTW (Dynamic Time Warping) distance from the first element to the i-th element of X and from the first element to the j-th element of Y; d(X i ,Y j ) represents the distance between the i-th element of X and the j-th element of Y; min[D(i-1,j),D(i,j-1),D(i-1,j-1)] means that when searching for the best path, the previous position of the current position (i,j) can be the upper left, directly above, or to the left, and the one with the smallest cumulative distance among these three directions is selected as the cumulative distance of the current position.

[0126] S3-1-4: Extract the elements whose cumulative distance is less than or equal to the preset threshold as convolution data, that is, remove the elements whose cumulative distance is greater than the preset threshold, and select a data set that is more in line with the current change law, so as to quickly and accurately improve the training efficiency

[0127] For example, the current time is 9:00 on October 14, 2024, and the time series X is from 9:00 on October 14 to 9:00 on October 13. This method will search for the past time series Y and calculate the cumulative distance through the cumulative distance matrix. The smaller the cumulative distance, the more similar the features are. In this way, the data set can be filtered out and provided for CNN training.

[0128] S3-2: Perform CNN (Convolutional Neural Networks) convolution on the convolution data and output the predicted data. This network has the characteristic that the number of channels doubles and the sequence size is halved every time it passes through the convolution pooling layer. The entire network uses the same convolution kernel size and maximum pooling size.

[0129] S3-2-1: At the initial stage of CNN, the data input layer will perform the third preprocessing on the convolution data. The third preprocessing includes operations such as de-meaning and normalization to improve the training effect and the generalization ability of the model.

[0130] S3-2-2: This layer is the core of CNN. It uses multiple convolution kernels to perform convolution operations on the convolution data after the third preprocessing. The convolution kernel slides on the convolution data and performs linear product summation of the local area to extract local features in the image. This process can capture the spatial structure and feature information in the image.

[0131] S3-2-3: After the convolution calculation, the data will pass through the excitation layer, and use activation functions such as ReLU to perform nonlinear transformation on the output of the convolution layer to enhance the expressive power of the model.

[0132] S3-2-4: After the excitation layer, the data enters the pooling layer. The pooling operation is used to reduce the dimension of the data and the amount of calculation while retaining important features and improving the robustness of the model.

[0133] S3-2-5: Finally, after multiple convolution and pooling layers, the data is flattened and input to the fully connected layer. The fully connected layer is responsible for integrating the extracted features, classifying or identifying them through the classifier, and finally outputting the predicted data.

[0134] S3-3: Using the Transformer encoder, only the encoder layer is used in this case. Neither the sequence length nor the number of dimensions is changed. By processing the predicted data, the output sequence, i.e., the prediction result of PM2.5, is obtained.

[0135] S3-3-1: First, the input prediction data is converted into feature vectors, which capture the spatiotemporal context information.

[0136] S3-3-2: The feature vector is sent to the encoder module for encoding to obtain encoded data.

[0137] The encoder module usually contains multiple encoders. Each encoder performs operations such as multi-head self-attention mechanism and feedforward neural network. The multi-head self-attention mechanism allows the model to capture the correlation of input sequences in different representation subspaces, enhancing the model's understanding ability. The output of the encoder is a matrix containing the deep features of the input sequence.

[0138] S3-3-3: The encoded data is input into the decoder module for decoding to obtain decoded data.

[0139] The decoder module also contains multiple decoders. Each decoder also uses a multi-head self-attention mechanism and a feed-forward neural network, but also adds a masked multi-head self-attention layer to ensure that future words are not taken into account when predicting the current position word. The decoder combines the output of the encoder with its own input to gradually generate an output sequence.

[0140] S3-3-4: The decoded data is converted into a probability distribution of monitored concentrations through a linear layer and a softmax function. Based on this probability distribution, the monitored concentration with the highest probability is selected as the predicted output for the current location. This process is repeated in each step of sequence generation until a complete output sequence is generated.

[0141] S3-3-5: During the training process, the cross entropy loss function is used to measure the gap between the model prediction and the actual target. The gradient is calculated through the back-propagation algorithm, and the model parameters are updated using the optimizer to minimize the loss function. After multiple rounds of iterative training, the model gradually learns the mapping relationship from input to output and improves its ability to generate correct output sequences.

[0142] In this embodiment, through loop iteration, the sample data set and accuracy most suitable for this prediction are calculated for MLP neural network PM2.5 remote sensing monitoring.

[0143] Those skilled in the art will appreciate that the above-mentioned embodiments are specific examples for implementing the present invention, and in actual applications, various changes may be made thereto in form and detail without departing from the spirit and scope of the present invention.

Claims

1. A method for monitoring PM2.5 using multi-source data fusion based on a neural network model, characterized in that: The specific steps include: S1: Performing super-resolution processing on remote sensing image data to obtain first image data; S2: performing error correction on the pollutant output data obtained from the Silam model to obtain the second image data; S3: Use the first image data and the second impact data as multi-source fusion data to perform PM2.5 analysis and detection.

2. A method for monitoring PM2.5 using multi-source data fusion based on a neural network model as claimed in claim 1, characterized in that: The S1 includes: S1-1: Acquire historical remote sensing data, and perform a first preprocessing on the historical remote sensing data to obtain input data; the first preprocessing includes matching processing, sampling processing, slicing processing, and data expansion processing; S1-2: Input data to the existing ESRGAN model for training; S1-3: Input the acquired real-time remote sensing image data into the trained ESRGAN model, thereby outputting the first image data.

3. A method for monitoring PM2.5 using multi-source data fusion based on a neural network model as claimed in claim 2, characterized in that: The S1-1 includes: S1-1-1: Perform matching processing on the remote sensing historical data to obtain the first data. S1-1-2: resample the first data to 50m as a first-precision image, and then downsample the first-precision image to 200m as a second-precision image, to obtain a set of matching first-precision-second-precision image groups as the second data; S1-1-3: Segment the first precision image in the second data into first slices of 96×96 pixels, segment the corresponding second precision image into second slices of 24×24 pixels, and combine the first slice and the second slice to obtain third data; S1-1-4: Rotate the first slice and the second slice in the third data by 90°, 180°, and 270° around the geometric center of the slice, and use the slice data obtained after the rotation as input data.

4. A method for monitoring PM2.5 using multi-source data fusion based on a neural network model as claimed in claim 3, characterized in that: The S1-1-1 includes: A: First, calculate the band mean of each data in the remote sensing historical data, match the band bandwidth and the band mean, and ensure the consistency of different satellite data sources in the same band; B: Use the radiation brightness scale parameter and radiation brightness bias parameter to convert the digital quantization value into the radiation brightness value: L=gain×DN+Bias (1) In formula (1), L represents the apparent reflectance or radiance value; gain represents the radiance scale parameter; Bias represents the radiance bias parameter; DN represents the remote sensing image pixel brightness value; C: Perform FLAASH atmospheric correction on the radiance value; D: Select the near-infrared band, red band and green band in the historical remote sensing data after atmospheric correction as the first data.

5. A method for monitoring PM2.5 using multi-source data fusion based on a neural network model as claimed in claim 3, characterized in that: The S2 includes: S2-1: performing a second preprocessing on the pollutant output data obtained from the Silam model, the second preprocessing including removing outliers and missing values; S2-2: The pollutant output data of the Silam model after the second pretreatment is used as the influencing factor and the monitoring value of the Silam model is used as the result to construct the limiting gradient model: Actual value, represents the predicted value of the i-th sample of the model; Ω(f j ) represents the regularization term of the jth tree; t represents the total number of trees; p represents the total number of samples; S2-3: Use the constructed extreme gradient boosting model to correct the existing Silam pollutant data to obtain the second image data.

6. A method for monitoring PM2.5 using multi-source data fusion based on a neural network model as claimed in claim 5, characterized in that: The S2 further includes: S2-4: Constructing spatiotemporal geographic weighted models: In formula (3), y i represents the pollutant data of the i-th point in the Silam model, (μ i ,ν i ,t i ) represents the space-time coordinates of the i-th point, β0(μ i ,ν i ,t i )x ik represents the kth spatiotemporal regression parameter at the ith point, which can be regarded as a function of the geographical location and observation time, and is the random error of the spatiotemporal region of the ith point; β0(μ i ,ν i ,t i ) represents the pollutant data of the i-th point, ε i represents random error; S2-5: Use the constructed spatiotemporal geographic weighted model to further correct Silam pollutant data at different times, spaces and geographical locations.

7. A method for monitoring PM2.5 using multi-source data fusion based on a neural network model as claimed in claim 1, characterized in that: The S3 includes: S3-1: taking the first image data and the second impact data as multi-source fusion data, performing dynamic time warping on the multi-source fusion data, and outputting convolution data; S3-2: Perform CNN convolution on the convolution data and output the predicted data; S3-3: Use the Transformer encoder to process the predicted data to obtain the output sequence, which is the prediction result of PM2.

5.

8. A method for monitoring PM2.5 using multi-source data fusion based on a neural network model as claimed in claim 7, characterized in that: The S3-1 includes: S3-1-1: Set the first time series X and the second time series Y to have lengths of m and n respectively, and create a (m+1)×(n+1) two-dimensional array, i.e., the original cumulative distance matrix; the elements in the two-dimensional array are all PM2.5 historical monitoring concentrations sorted by time series; S3-1-2: Set the first row and first column of the original cumulative distance matrix corresponding to the distance between the empty sequence and another sequence to zero; S3-1-3: Starting from the upper left corner of the original cumulative distance matrix, traverse each element and fill in the value before calculating the cumulative distance: d(i,j)=d(X i ,Y j )+min[D(i-1,j),D(i,j-1),D(i-1,j-1)] (4) In formula (4), d(i,j) represents the DTW distance from the first element to the i-th element of X and from the first element to the j-th element of Y; d(X i ,Y j ) represents the distance between the i-th element of X and the j-th element of Y; min[D(i-1,j),D(i,j-1),D(i-1,j-1)] means that when searching for the best path, the previous position of the current position (i,j) is the upper left, directly above, or to the left, and the direction with the smallest cumulative distance among these three directions is selected as the cumulative distance of the current position. S3-1-4: Extract the elements whose cumulative distance is less than or equal to the preset threshold as convolution data.

9. A method for monitoring PM2.5 using multi-source data fusion based on a neural network model as claimed in claim 7, characterized in that: The S3-2 includes: S3-2-1: At the initial stage of CNN, the data input layer performs the third preprocessing on the convolution data, which includes de-meaning and normalization operations; S3-2-2: performing a convolution operation on the convolution data after the third preprocessing by using a plurality of convolution kernels; S3-2-3: After the convolution calculation, the data will pass through the excitation layer, and the ReLU activation function will be used to perform nonlinear transformation on the output of the convolution layer; S3-2-4: After the incentive layer, the data enters the pooling layer; S3-2-5: Finally, after being processed by multiple convolutional and pooling layers, the data is flattened and input into the fully connected layer, and classified or recognized by the classifier, and finally the predicted data is output.

10. The method for monitoring PM2.5 by multi-source data fusion based on a neural network model as claimed in claim 7, characterized in that: The S3-3 includes: S3-3-1: First, convert the input prediction data into a feature vector; S3-3-2: The feature vector is sent to the encoder module for encoding to obtain encoded data; S3-3-3: The encoded data is input into the decoder module for decoding to obtain decoded data; S3-3-4: The decoded data is converted into a probability distribution of monitoring concentration through a linear layer and a softmax function. According to the probability distribution, the monitoring concentration with the highest probability is selected as the predicted output for the current location. S3-3-5: Repeat S3-3-1 to S3-3-4 until a complete output sequence is generated, i.e., the prediction result of PM2.5.

Images