Remote sensing pollutant missing data acquisition method

Abstract

The invention discloses a remote sensing pollutant data acquisition method. Dividing window grid data corresponding to the remote sensing pollutant data into a missing data subset and a non-missing data subset, and randomly covering the remote sensing pollutant data in the non-missing data subset with a mask set corresponding to the missing data subset to generate a pseudo-missing data subset, generating a pre-training model data set according to the non-missing data subset and the pseudo-missing data subset, and training according to the pre-training model data set to obtain a codec model; and acquiring remote sensing pollutant missing data corresponding to the remote sensing pollutant data according to the pseudo missing data subset and the codec model. By constructing a real mask pre-training mechanism, a distribution mode of missing data can be efficiently identified to obtain missing remote sensing pollutant data. Manual intervention can be reduced, the repairing speed can be increased, the error rate can be reduced, and therefore the usability and reliability of remote sensing pollutant data are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing data restoration, and in particular relates to a method for acquiring missing data of remote sensing pollutants, a computer storable medium and an electronic device. Background Art

[0002] Remote sensing technology uses sensors carried by satellites, aircraft or other aircraft to collect information on electromagnetic radiation on the earth's surface. These sensors can detect electromagnetic radiation of different wavelengths, including visible light, infrared, ultraviolet, etc.

[0003] In terms of monitoring pollutants, remote sensing can obtain information on the distribution, concentration, and diffusion of pollutants in the atmosphere, water bodies, and soil. For example, for atmospheric pollutants, remote sensing can monitor the concentration and distribution of aerosols and assess the extent and scope of haze; for water bodies, it can detect changes in water quality parameters, such as chlorophyll concentration and suspended particulate matter, thereby reflecting the pollution status of water bodies; in terms of soil, it can monitor the content of pollutants such as heavy metals in the soil.

[0004] Remote sensing pollutant data refers to information and data related to pollutants obtained through remote sensing technology. Remote sensing pollutant data has the advantages of wide coverage, strong timeliness, and relatively low cost, and can provide an important basis for environmental monitoring, pollution control, and decision-making.

[0005] Although remote sensing pollutant data has a wide range of application value in fields such as meteorological monitoring, environmental assessment and agricultural decision-making, the accuracy of remote sensing pollutant data may be affected by many factors, such as weather conditions and sensor performance. In practical applications, remote sensing pollutant data are often missing due to sensor failures, observation conditions and other reasons. These missing data not only affect the integrity and accuracy of the data, but also limit the effectiveness and reliability of the data in subsequent analysis and application.

[0006] The inventors discovered during the process of implementing the present invention that there is no efficient and accurate method for repairing missing data of remote sensing pollutants in the prior art. Summary of the invention

[0007] In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides a method for acquiring missing data of remote sensing pollutants to solve the problem of missing data of remote sensing pollutants.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0009] A first aspect of the present invention provides a method for acquiring missing data of remote sensing pollutants.

[0010] A method for obtaining missing data of remote sensing pollutants, comprising the following steps:

[0011] Divide the window grid data corresponding to the remote sensing pollutant data into missing data subsets and non-missing data subsets;

[0012] Randomly masking the mask set corresponding to the missing data subset to the remote sensing pollutant data in the non-missing data subset to generate a pseudo-missing data subset, and generating a pre-trained model data set based on the non-missing data subset and the pseudo-missing data subset;

[0013] A codec model is obtained by training according to the pre-trained model data set;

[0014] Remote sensing pollutant missing data corresponding to the remote sensing pollutant data is acquired according to the pseudo missing data subset and the codec model.

[0015] According to one embodiment of the present invention, the window grid data corresponding to the remote sensing pollutant data is divided into missing data subsets and non-missing data subsets, specifically:

[0016] Merging the remote sensing pollutant data and auxiliary data corresponding to the remote sensing pollutant data into an original data set;

[0017] Dividing the original data set into the window grid data according to the specified length and width;

[0018] Dividing the window grid data into a plurality of subsets according to a specified window size;

[0019] According to whether each subset is missing remote sensing pollutant data, each subset is divided into the missing data subset and the non-missing data subset.

[0020] According to one embodiment of the present invention, the mask set corresponding to the missing data subset is randomly masked in the remote sensing pollutant data in the non-missing data subset to generate a pseudo-missing data subset, and a pre-trained model data set is generated according to the non-missing data subset and the pseudo-missing data subset, specifically:

[0021] Generating a mask set according to the extraction masks corresponding to the samples in the missing data subset;

[0022] Generate the pseudo missing data subset according to any extraction mask in the matrix corresponding to the mask set and any sample in the non-missing data set;

[0023] The data in the non-missing data subset and the data in the pseudo-missing data subset are respectively normalized to generate the pre-training model data set.

[0024] According to one embodiment of the present invention, the codec model is obtained by training the pre-trained model dataset, specifically:

[0025] Dividing the pre-trained model data set into a training set, a validation set, and a test set according to a preset ratio, and training the codec model using the training set;

[0026] Compare the output results of the test set obtained by the codec model with the corresponding samples in the validation set to obtain a loss function;

[0027] When the loss functions obtained multiple times meet a preset termination condition, the training of the codec model is terminated, and the codec model after the training is terminated is obtained.

[0028] According to an embodiment of the present invention, the preset termination condition is specifically:

[0029] The loss function obtained for a specified number of consecutive times does not decrease;

[0030] And / or, the loss function obtained for a specified number of consecutive times does not fluctuate;

[0031] and / or, reaching a predetermined minimum loss threshold;

[0032] and / or, reaching the maximum number of training rounds;

[0033] Among them, the loss function includes but is not limited to: mean square error, mean absolute error, mixed linear quadratic loss function, logarithmic hyperbolic cosine loss function and quantile loss function.

[0034] According to an embodiment of the present invention, after acquiring the remote sensing pollutant missing data corresponding to the remote sensing pollutant data according to the pseudo missing data subset and the codec model, the method further includes:

[0035] retaining missing data of remote sensing pollutants corresponding to the remote sensing pollutant data, and setting the remaining data in the remote sensing pollutant data to 0;

[0036] Retain the non-missing data in the pseudo-missing data subset, and set the missing data in the pseudo-missing data subset to 0;

[0037] Performing an element-by-element matrix addition operation on the remote sensing pollutant missing data corresponding to the remote sensing pollutant data and the non-missing data in the pseudo missing data subset, and splicing them according to longitude and latitude;

[0038] The spliced ​​data are used as the repair results of remote sensing pollutant data.

[0039] According to one embodiment of the present invention, the auxiliary data includes but is not limited to wind speed, wind direction, temperature, air pressure, altitude, longitude and latitude meteorological and geographic information data in the same area and at multiple times.

[0040] According to one embodiment of the present invention, the codec model is specifically a neural network model based on an encoder-decoder architecture, the encoder includes a three-layer two-dimensional convolutional neural network and a self-attention mechanism layer; the decoder includes a three-layer two-dimensional convolutional neural network.

[0041] The third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the method for repairing missing data of remote sensing pollutants based on mask pre-training as described in the first aspect of the present invention.

[0042] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the method for repairing missing data of remote sensing pollutants based on mask pre-training according to the first aspect of the present invention are implemented.

[0043] One or more of the above technical solutions have the following beneficial effects:

[0044] The window grid data corresponding to the remote sensing pollutant data is divided into missing data subsets and non-missing data subsets, and the mask set corresponding to the missing data subset is randomly masked on the remote sensing pollutant data in the non-missing data subset to generate a pseudo missing data subset, and a pre-trained model data set is generated according to the non-missing data subset and the pseudo missing data subset, and the codec model is trained according to the pre-trained model data set; the remote sensing pollutant missing data corresponding to the remote sensing pollutant data is obtained according to the pseudo missing data subset and the codec model. By constructing a real mask pre-training mechanism, the distribution pattern of missing data can be efficiently identified to obtain the missing remote sensing pollutant data. It can not only reduce manual intervention and improve the repair speed, but also reduce the error rate, thereby significantly improving the availability and reliability of remote sensing pollutant data.

[0045] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0047] Figure 1 A flowchart of a method for acquiring missing data of remote sensing pollutants provided in an embodiment of the present invention.

[0048] Figure 2 The overall architecture diagram of the codec model provided for a specific embodiment of the present invention.

[0049] Figure 3 A process diagram of training and repair provided for a specific embodiment of the present invention. DETAILED DESCRIPTION

[0050] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0051] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0052] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0053] The overall idea proposed by the present invention is: by dividing the window grid data corresponding to the remote sensing pollutant data into missing data subsets and non-missing data subsets; randomly masking the mask set corresponding to the missing data subset to the remote sensing pollutant data in the non-missing data subset to generate a pseudo missing data subset, and generating a pre-trained model data set according to the non-missing data subset and the pseudo missing data subset; training a codec model according to the pre-trained model data set; and obtaining the remote sensing pollutant missing data corresponding to the remote sensing pollutant data according to the pseudo missing data subset and the codec model. The technical problem that remote sensing pollutant data is often missing due to various reasons such as sensor failure and observation condition restrictions is solved.

[0054] Embodiment 1

[0055] S100, dividing the window grid data corresponding to the remote sensing pollutant data into a missing data subset and a non-missing data subset.

[0056] Before executing step S100, it is necessary to collect remote sensing pollutant data captured by remote sensing sensors. The remote sensing pollutant data is usually organized in the form of grids, and each grid represents a pollutant measurement value in a certain geographical area.

[0057] In the present invention, remote sensing pollutant data can be used as the original data set;

[0058] Furthermore, the present invention can also obtain auxiliary data of the same area and time as the remote sensing pollutant data, which can be meteorological parameters (such as wind speed, wind direction, temperature, air pressure) and geographical information (such as altitude, longitude and latitude). The auxiliary data also exists in the form of a grid, which is matched with the remote sensing pollutant data grid or appropriately resampled to ensure the spatial consistency of the two.

[0059] Integrate remote sensing pollutant data with auxiliary data to form an original dataset containing all information.

[0060] The above-mentioned original data set is the window grid data corresponding to the remote sensing pollutant data in step S100.

[0061] The sliding window method is used to divide the original data set into a series of sub-grids of the same size, and the size of each sub-grid is a preset length and width, such as 64 × 64 pixels. The sliding window moves with a certain step size (such as 1 grid) to ensure that there may be overlap or no overlap between adjacent sub-grids, depending on the specific application requirements.

[0062] According to whether each subgrid contains missing values, it is classified into missing data subset (DM) and non-missing data subset (DC). Missing values ​​may be due to sensor failure, cloud cover, data transmission problems, etc., resulting in data points not being recorded or invalid.

[0063] The missing data subset contains all the subgrid data with at least one grid cell containing a missing value. The non-missing data subset contains only those subgrid data with all grid cells having valid values.

[0064] In the specific implementation scenario of the present invention, the NO2 hourly data is combined with the auxiliary meteorological data to form an original data set. The data set at each moment will include the NO2 concentration and the corresponding meteorological parameters.

[0065] The sliding window method is used to divide the original data set into a series of window grid data D of the same length and width. Assuming the window size is 64×64 and the step size is 1, each window grid data d is obtained by sliding the original data set according to this window size and step size.

[0066] In this step, each window grid data d is checked to determine whether it contains missing values ​​of NO2. If the window grid data d contains at least one missing value, it will be classified as part of the missing data subset;

[0067] Conversely, if all NO2 values ​​in the window grid data d are valid, they will be classified as part of the non-missing data subset.

[0068] Assuming that the NO2 value of the window grid data di at the coordinate (x1, y1) of the i-th hour is missing, while the NO2 values ​​of all other grids are valid, then di will be classified into the missing data subset.

[0069] Similarly, assuming that the window grid data dj at the jth hour has no missing NO2 values ​​in all its grids, then dj will be classified into the non-missing data subset.

[0070] Two data subsets will be obtained: the missing data subset contains all window grid data with at least one missing value, and the non-missing data subset contains all window grid data without missing values.

[0071] It should be noted that, based on the above scheme, the size and shape of the window can be dynamically adjusted according to the distribution characteristics of remote sensing pollutant data. For example, a larger window can be used in areas where pollutants are densely distributed, and a smaller window can be used in areas where the distribution is sparse. This can make more efficient use of data resources and reduce the amount of calculation.

[0072] Furthermore, in other preferred embodiments, an intelligent screening mechanism can be introduced to give priority to the areas containing the most missing values, or to select the subgrids that need to be repaired the most based on some statistical indicators (such as coefficient of variation), so as to optimize the order of data repair and improve the overall repair efficiency. All of the above belong to the protection scope of the present invention.

[0073] In order to systematically identify and separate the missing parts in the data, thus providing a clear data basis for subsequent targeted data repair and model training, and enhancing the model's ability to handle data missing and prediction accuracy, the remote sensing pollutant data and the auxiliary data corresponding to the remote sensing pollutant data are merged into the original data set. The auxiliary data are meteorological data and geographic information data related to remote sensing pollutants.

[0074] Divide the original data set into window grid data according to the specified length and width;

[0075] Divide the window grid data into multiple subsets according to the specified window size;

[0076] Based on whether each subset is missing remote sensing pollutant data, each subset is divided into a missing data subset and a non-missing data subset; the missing data subset is the window grid data containing missing values, and the non-missing data subset is the window grid data without missing values.

[0077] By combining remote sensing pollutant data with auxiliary data (such as meteorological data and geographic information data), a raw data set containing all relevant information is formed. This is done to ensure that the model can fully take into account various factors that affect pollutant concentrations, such as wind speed, wind direction, temperature, humidity, topography, etc., which may affect the diffusion and deposition of pollutants.

[0078] The original data set is divided into a series of window grid data according to the predetermined length and width. Each window represents the data within a certain spatial range, which is helpful for local analysis and processing, and also facilitates subsequent model training because the model can focus on the characteristics of each small area.

[0079] Each window grid data is further divided into smaller subsets according to the specified window size. The purpose of this is to be able to analyze the data more carefully, and if there is missing data in a large window, it can be accurately located in a smaller subset, thereby avoiding the entire large window from being discarded.

[0080] Check whether the data in each subset is complete. If there is missing remote sensing pollutant data in a subset, it is classified as a missing data subset; conversely, if the data in the subset is complete, it is classified as a non-missing data subset.

[0081] S200, randomly masking the mask set corresponding to the missing data subset to the remote sensing pollutant data in the non-missing data subset to generate a pseudo-missing data subset, and generating a pre-training model data set based on the non-missing data subset and the pseudo-missing data subset.

[0082] A mask set (M) is extracted from the missing data subset by analyzing the missing pattern of each sample in the missing data subset. The mask set contains various missing patterns that represent the distribution of missing values ​​in remote sensing pollutant data.

[0083] The masks in the mask set are randomly applied to the remote sensing pollutant data in the non-missing data subset to generate a pseudo missing data subset (DM'). This step is achieved by matrix multiplication, that is, the mask matrix in the mask set is multiplied element by element with the data in the non-missing data subset.

[0084] In the mask matrix, 0 usually represents missing and 1 represents non-missing, so after multiplication, the original complete data will be set to missing values ​​at the corresponding positions.

[0085] In order to ensure the consistency and stability of the data, the non-missing data subset and the pseudo-missing data subset need to be normalized.

[0086] Normalization can scale the data in the solution of the present invention to the same scale, avoiding the situation where features with large numerical differences dominate the learning process of the model, including maximum and minimum value normalization, z-score standardization, etc.

[0087] The normalized non-missing data subset and pseudo-missing data subset are combined to construct a pre-training model dataset. This dataset will be used to train the encoder-decoder model, which can learn and understand the distribution pattern of pollutant data, especially the pattern of missing data, so that it can predict and repair the actual missing data in the subsequent stage.

[0088] In a specific implementation scenario of the present invention, a mask set is extracted from the missing data subset, and the mask set includes the position information of the missing values ​​in all samples.

[0089] Select a sample from the non-missing data subset and then randomly apply a mask from the mask set to generate a pseudo-missing data subset.

[0090] When the previously mentioned mask set is applied to this sample, a pseudo-missing data subset is obtained, in which some values ​​are marked as missing (which can be set to NaN or other missing value identifiers);

[0091] After obtaining the pseudo-missing data subset, the pseudo-missing data subset and the original non-missing data subset are normalized. Normalization is the process of scaling the data to a specific range (such as between 0 and 1), which helps the model learn. Assume that the maximum and minimum value normalization is used to transform the data in each data subset;

[0092] Then, the non-missing data subset and the pseudo-missing data subset are combined into a pre-trained model dataset for training the encoder-decoder model. The model will learn how to recover the missing values ​​from the pseudo-missing data subset to match the corresponding values ​​in the non-missing data subset.

[0093] It should be noted that, based on the above scheme, an algorithm can be designed to generate masks in specific patterns, such as masks based on connected component analysis, masks based on edge detection, etc. These masks can better simulate missing patterns in nature, such as cloud occlusion, terrain obstacles, etc.

[0094] Furthermore, in other preferred embodiments, the mask probability can be dynamically adjusted in the pre-training stage, that is, a lower mask probability is used in the early stage of training, and as the model training progresses, the mask probability is gradually increased to challenge the model and enable it to learn to recover data under more complex missing conditions.

[0095] Furthermore, in other preferred embodiments, when constructing a pre-training data set, the spatiotemporal characteristics of remote sensing data can be taken into consideration, and spatiotemporal consistency constraints can be added to ensure that masks at different time points in the same location maintain a certain degree of continuity, or masks at different locations at the same time point are correlated, which helps the model learn a more reasonable repair strategy. All of the above belong to the protection scope of the present invention.

[0096] In order to enhance the robustness and generalization ability of the model, and make the model more stable and accurate when facing incomplete data in practice, and to ensure the comprehensiveness and effectiveness of model training by simulating various possible missing patterns, a mask set is generated according to the extraction mask corresponding to each sample in the missing data subset, and the extraction mask is generated according to the distribution pattern of missing values ​​in remote sensing pollutant data;

[0097] Generate a pseudo missing data subset according to any extraction mask in the matrix corresponding to the mask set and any sample in the non-missing data set;

[0098] The data in the non-missing data subset and the data in the pseudo-missing data subset are normalized to generate a pre-training model dataset.

[0099] Based on the samples in the missing data subset, a series of masks are generated. These masks reflect the distribution pattern of missing values ​​in remote sensing pollutant data. The mask is a binary matrix in which 1 indicates that the data at the corresponding position is missing, and 0 indicates that the data exists.

[0100] All generated masks are collected together to form a mask set. This set contains a variety of missing modes, which can help the model learn performance under different missing conditions.

[0101] A mask is randomly selected from the set of masks and then applied to a sample in the non-missing data subset. This means that the position data corresponding to 1 in the mask is replaced with a missing value in the sample, thus simulating the situation of missing data.

[0102] Repeat the above process until a sufficient number of pseudo-missing data samples are generated, which constitute a pseudo-missing data subset. This subset can help the model learn how to deal with missing values ​​that may be encountered in actual data.

[0103] In order to make the model converge better and improve training efficiency, it is necessary to normalize the data in the non-missing data subset and the pseudo-missing data subset. Normalization usually means scaling the data to a specific range, such as [0, 1] or [-1, 1], which can be achieved through methods such as minimum and maximum scaling and Z-score standardization.

[0104] The normalized non-missing data subset and the pseudo-missing data subset are combined to form a pre-trained model dataset. This dataset contains real data and simulated missing data, allowing the model to learn how to deal with various data integrity situations during training.

[0105] S300: Obtain a codec model by training according to a pre-trained model dataset.

[0106] The untrained encoder-decoder model receives as input a pre-training dataset consisting of a subset of non-missing data and a subset of pseudo-missing data, and tries to predict the data in the subset of non-missing data.

[0107] The prediction results are compared with the true values ​​of the non-missing data subset, and a loss function (such as RMSE loss) is calculated to measure the accuracy of the model prediction.

[0108] The loss function is back-propagated to the network, and the network weights are updated to minimize the prediction error. This process is repeated until the model loss on the validation set no longer decreases significantly, or the preset number of training rounds is reached. If the model loss on the validation set does not decrease for 10 consecutive rounds, or fluctuates up and down for more than 50 rounds, the training process will be terminated. This indicates that the model has converged and further training will not bring significant improvements. Once the model training is completed, it can be used to repair remote sensing pollutant data with missing values.

[0109] The pre-training dataset consists of a non-missing data subset and a pseudo-missing data subset, where the pseudo-missing data subset is generated by randomly masking different masks in the mask pattern set M onto the target pollutant data in the non-missing data subset.

[0110] During the training process, the output of the decoder is compared with the corresponding samples in the non-missing data subset to calculate the loss function (such as mean square error RMSE). The loss function measures the difference between the model output and the actual non-missing data, and the goal is to minimize this difference to improve the prediction accuracy of the model.

[0111] During the training process, the model loss on the validation set is continuously monitored. If the model loss does not decrease for 10 consecutive rounds, or fluctuates up and down for more than 50 rounds, this indicates that the model has converged or started to overfit. At this time, the training will be terminated and the current model weights will be saved as a stable model.

[0112] By training the encoder-decoder model, the goal is to learn how to reconstruct the missing parts from other parts (auxiliary data and non-missing parts) given that some parts of the remote sensing pollutant data are missing. This method can improve the generalization ability of the model and effectively perform data repair even in the face of previously unseen missing patterns, thereby significantly improving the availability and reliability of remote sensing pollutant data.

[0113] Among them, the encoder function: the encoder is responsible for converting the input remote sensing pollutant data into a compressed feature representation, usually called a "hidden state" or "latent variable". In this process, the encoder learns the intrinsic structure and pattern of the data so that it can capture the key features of the data. Composition: The encoder is usually composed of a multi-layer convolutional neural network (CNN), and each convolution layer can extract different levels of features of the input data. In the present invention, the encoder includes a three-layer two-dimensional convolutional neural network for extracting local spatial features of remote sensing pollutant data, and a SE (Squeeze-and-Excitation) attention mechanism layer for focusing on important areas and key information to better understand the distribution pattern of the input data.

[0114] SE attention mechanism layer functions:

[0115] Focus localization: Attention mechanisms enable the model to learn which input features are most important for predicting missing values. In remote sensing data, this may mean identifying which geographic locations or time points have the most correlation with other features (such as wind speed, wind direction, etc.).

[0116] Weight assignment: The attention mechanism assigns a weight to each part of the input, which reflects how much it contributes to the model output. This means that the model can rely more on features that are considered more important when reconstructing missing data.

[0117] Dynamic Adjustment: The attention mechanism is able to dynamically adjust its focus, which means that in different inputs or different inpainting tasks, it automatically identifies and adjusts which regions or features are most critical.

[0118] Composition: In the present invention, the attention mechanism is mainly reflected in the encoder part, specifically the SE (Squeeze-and-Excitation) attention mechanism layer, which is a form of self-attention mechanism used to enhance the feature selection ability of the model. The SE attention mechanism layer usually includes the following components:

[0119] Squeeze: This step compresses the high-dimensional feature map into a vector of fixed length, usually by performing global average pooling on the feature map to capture the information of the entire feature map.

[0120] Excitation: Based on the compressed vector, a nonlinear transformation is performed through the fully connected layers to generate the weights of each channel. These weights represent the importance of each feature channel.

[0121] Scale: Reassign the generated weights to each channel of the original feature map and combine the attention weights with the feature map through element-wise multiplication to highlight important features and suppress irrelevant features.

[0122] In the context of remote sensing pollutant data repair, the SE attention mechanism layer helps the model focus on areas that are critical to the distribution of pollutants, such as areas near pollutant emission sources or directly affected by wind direction and wind speed. This focus improves the accuracy and efficiency of the model in repairing missing data because it can better understand the distribution pattern of the input data and focus resources on the most critical information.

[0123] Decoder function: The task of the decoder is to reconstruct the original input data from the compressed feature representation produced by the encoder. The decoder learns how to go from the abstract feature space back to the specific pixel space to reconstruct the input data as accurately as possible.

[0124] Composition: The decoder is also composed of a multi-layer convolutional neural network to gradually reconstruct the details of the input data. In this case, the decoder is composed of three layers of two-dimensional convolutional neural networks to reconstruct the missing remote sensing pollutant data.

[0125] In a specific implementation scenario of the present invention, the original remote sensing pollutant data and auxiliary data (such as wind speed, wind direction, temperature, air pressure, etc.) are segmented into a series of grid data by a sliding window method. These data are divided into a missing data subset containing missing values ​​and a non-missing data subset that does not contain missing values.

[0126] The missing patterns are extracted from the missing data subset, and a mask set M is constructed. Then the masks in M ​​are randomly applied to the non-missing data in the non-missing data subset to generate a pseudo-missing dataset DM'. The pseudo-missing dataset and the non-missing data subset are normalized for efficient training in the network. The normalization operation ensures that the data is in a similar scale range, avoiding the gradient vanishing or exploding problem.

[0127] The convolutional attention mechanism neural network based on the encoder-decoder architecture is used. The encoder contains three layers of two-dimensional convolutional neural network and SE (Squeeze-and-Excitation) attention mechanism layer, and the decoder also consists of three layers of two-dimensional convolutional neural network. The encoder is used to extract the features of the input data, the SE attention mechanism helps the network focus on the most important features, and the decoder is responsible for reconstructing the complete data based on the output of the encoder.

[0128] The pre-training dataset (non-missing data subset and pseudo-missing data subset) is divided into training set, validation set and test set, and the ratio can be 8:1:1. The training set is used to train the model, the validation set is used to adjust the model parameters to prevent overfitting, and the test set is used to evaluate the generalization ability of the model.

[0129] During the training process, the model receives the training set as input and tries to predict the data in the non-missing data subset. The prediction results are compared with the true values ​​of the non-missing data subset, and the loss function (such as RMSE loss) is calculated to measure the accuracy of the model's prediction. The loss function is back-propagated to the network and the network weights are updated to minimize the prediction error.

[0130] This process is repeated until the model's loss on the validation set no longer decreases significantly, or the preset number of training rounds is reached. If the model's loss on the validation set does not decrease for 10 consecutive rounds, or fluctuates up and down for more than 50 rounds, the training process will be terminated. This indicates that the model has converged and further training will not bring significant improvements. Once the model training is completed, it can be used to repair remote sensing pollutant data with missing values.

[0131] It should be noted that random transformations such as rotation, scaling, flipping, etc. can be applied to the training data based on the above scheme to increase the robustness and generalization ability of the model. Gaussian noise or other types of noise can also be added to the input data to help the model learn the ability to make effective predictions under noisy conditions.

[0132] Furthermore, in other preferred embodiments, Transformer can also be used: Transformer architecture can be introduced to process sequence data using self-attention mechanism, which may be more suitable for processing long-distance dependencies. Remote sensing images and other types of data (such as meteorological data, geographic information data) can also be combined as input to design a multimodal encoder-decoder model to improve prediction accuracy.

[0133] Furthermore, in other preferred embodiments, the prediction results can be outputted simultaneously at multiple levels of the decoder, and the losses can be calculated separately, which is called deep supervision. This can not only speed up the training speed, but also enable the middle layer of the model to better learn useful features. All of the above belong to the protection scope of the present invention.

[0134] In order to enable the model to learn the deep representation and structure of the data, so as to show better performance in tasks such as data reconstruction, feature extraction and noise reduction, and improve the model's prediction accuracy and ability to process complex data, the pre-trained model dataset is divided into training set, validation set and test set according to a preset ratio, and the encoder-decoder model is trained through the training set;

[0135] Compare the output of the test set through the encoder-decoder model with the corresponding samples in the validation set to obtain the loss function;

[0136] When the loss functions obtained multiple times meet the preset termination condition, the training of the codec model is terminated, and the codec model after the termination of the training is obtained.

[0137] The pre-trained model dataset is divided into three subsets: training set, validation set, and test set. Generally, the ratio of the dataset can be determined according to the specific task and data volume. For example, the common ratios are 70% training set, 15% validation set, and 15% test set, or 80% training set, 10% validation set, and 10% test set. The ratio of the training set, validation set, and test set in this embodiment is 8:1:1.

[0138] The training set is used for actual training of the model; the validation set is used to adjust the model's hyperparameters and evaluate the model's performance on unseen data; the test set is used to finally evaluate the generalization ability of the model. The test set is not used in the entire training process until the model training is completed.

[0139] The encoder-decoder model is an unsupervised learning model consisting of an encoder and a decoder. The encoder compresses the input data into a low-dimensional feature vector, while the decoder attempts to reconstruct the original data from these feature vectors.

[0140] During the training phase, the model tries to minimize the reconstruction error, which is the difference between the decoder output and the original input. The goal of model training is to optimize a loss function, usually the reconstruction error. For encoder-decoder models, the loss function may be mean squared error (MSE), cross entropy, or other metrics, depending on the data type and the specific design of the model.

[0141] At the end of each training cycle, the model uses the validation set to evaluate its performance in the current state. The loss function is calculated by comparing the model's output of the validation set samples with the actual samples.

[0142] The training process stops when the model's loss function on the validation set meets the preset termination criteria. This may include but is not limited to reaching a predetermined number of iterations, the loss function no longer significantly decreases or begins to increase (a sign of overfitting), or reaching a set loss threshold.

[0143] Once the training is terminated, the trained encoder-decoder model is obtained. The model at this point is considered to be the best version on the given dataset and can be used for further tasks such as feature extraction, anomaly detection, or data dimensionality reduction. Through this process, the encoder-decoder model is able to learn the intrinsic structure and characteristics of the data, and can effectively perform feature learning and data reconstruction even in the absence of labels.

[0144] S400, acquiring remote sensing pollutant missing data corresponding to the remote sensing pollutant data according to the pseudo missing data subset and the codec model.

[0145] Extract a mask set from the missing data subset; the mask set contains the location information of missing values ​​in all samples; select a sample from the non-missing data subset, and then randomly apply a mask from the mask set to generate a pseudo-missing data subset; when the previously mentioned mask is applied to this sample, a pseudo-missing data subset is obtained, in which some values ​​are marked as missing (can be set to NaN or other missing value identifiers);

[0146] Input the remote sensing pollutant data obtained from the remote sensing sensor into the trained codec model to obtain the output result;

[0147] According to the pseudo missing data subset and the output results obtained by the encoder-decoder model, the remote sensing pollutant missing data corresponding to the remote sensing pollutant data are obtained.

[0148] In a specific implementation scenario of the present invention, assume that there is a remote sensing image with a size of 256x256 pixels, and each pixel represents a PM2.5 concentration value. A 16x16 pixel area in the image is randomly selected, and all pixel values ​​in the area are set to NaN (Not a Number), which represents a pseudo-missing data subset. The convolutional autoencoder model is trained using images containing missing data to allow the model to learn how to infer the value of the missing area from the surrounding pixel values. When a new remote sensing image is received in which a 16x16 pixel area is missing due to cloud cover, this image is input into the trained model. The model outputs a complete image in which the originally missing 16x16 pixel area is filled with estimated PM2.5 concentration values. These estimates are based on the relationship between the surrounding pixel values ​​and the missing area learned by the model during the training phase.

[0149] It should be noted that more complex architectures can be considered based on the above solutions, such as variational autoencoders (VAE), generative adversarial networks (GANs) or U-Net, which perform better in image reconstruction tasks.

[0150] Furthermore, in other preferred embodiments, other types of remote sensing data (such as infrared images, synthetic aperture radar SAR images) or ground monitoring station data can also be combined to enhance the prediction ability of the model using multimodal information. All of the above belong to the protection scope of the present invention.

[0151] In a preferred embodiment of the present application, the preset termination condition is specifically:

[0152] The loss function obtained for a specified number of consecutive times does not decrease;

[0153] And / or, the loss function obtained for a specified number of consecutive times does not fluctuate;

[0154] and / or, reaching a predetermined minimum loss threshold;

[0155] and / or, reaching the maximum number of training rounds;

[0156] Among them, loss functions include but are not limited to: mean square error, mean absolute error, mixed linear quadratic loss function, log hyperbolic cosine loss function and quantile loss function.

[0157] During the training process, the model's loss function (such as mean square error, mean absolute error, etc.) is calculated after each training round. If the loss function no longer decreases for several consecutive times (preset number of 10 times), it may mean that the model has reached its learning limit and continued training will not bring significant performance improvements.

[0158] If the value of the loss function fluctuates within a very small range without an obvious downward trend in several consecutive training rounds (the preset number of rounds is 50), this may also indicate that the model has converged or fallen into a local minimum, and further training may not bring substantial improvement.

[0159] Before training, you can set a very low loss threshold, which can be set to 0.001. When the model's loss function is lower than this threshold, the model is considered to have achieved sufficient performance and no further training is required. This is suitable for situations where there are clear requirements for model performance.

[0160] To avoid indefinite training, you can pre-set a maximum number of training rounds, which can be set to 200 rounds. Regardless of how the model's loss function changes, when this number of rounds is reached, the training will automatically stop. This is a safety measure to prevent the model from training for too long.

[0161] Mean Squared Error (MSE): One of the most commonly used loss functions, suitable for regression problems, calculating the average of the squares of the differences between the predicted values ​​and the actual values.

[0162] Mean Absolute Error (MAE): Calculates the average of the absolute values ​​of the differences between the predicted values ​​and the actual values. It is less sensitive to outliers than MSE.

[0163] Hybrid linear quadratic loss function: It combines the advantages of linear loss and quadratic loss, and can use different loss function forms in different error ranges.

[0164] Log hyperbolic cosine loss function: A loss function used for regression tasks that can handle problems when the predicted value is significantly different from the actual value and is suitable for data with a large error range.

[0165] Quantile loss function: It is applicable to predicting a certain quantile of a distribution. For example, the quantile loss function can be used when predicting the median. It is particularly suitable for dealing with extreme values ​​and outliers.

[0166] By setting the above termination conditions, the model training process can be effectively controlled to ensure that optimal or near-optimal performance is achieved within a reasonable time and resources, avoiding unnecessary computational waste and overfitting.

[0167] In a preferred embodiment of the present application, after acquiring the remote sensing pollutant missing data corresponding to the remote sensing pollutant data according to the pseudo missing data subset and the codec model, it also includes:

[0168] The missing data of remote sensing pollutants corresponding to the remote sensing pollutant data are retained, and the remaining data in the remote sensing pollutant data are set to 0;

[0169] Keep the non-missing data in the pseudo-missing data subset and set the missing data in the pseudo-missing data subset to 0;

[0170] Perform element-by-element matrix addition operations on the non-missing data in the remote sensing pollutant missing data and pseudo-missing data subsets corresponding to the remote sensing pollutant data, and splice them according to longitude and latitude;

[0171] The spliced ​​data are used as the repair results of remote sensing pollutant data.

[0172] The original remote sensing pollutant data is processed. In this step, only the data values ​​that were originally missing are retained, and all other non-missing data values ​​are set to 0. The purpose of this is to clearly distinguish which are the missing values ​​that need to be supplemented by the model and which are the valid values ​​in the original data in the subsequent steps.

[0173] A similar process is performed on the pseudo-missing data subset. Here, the non-missing data is retained, and the missing data previously marked by the mask is set to 0. In this way, a subset containing only non-missing data can be obtained, in which the positions of the missing values ​​are cleared to facilitate subsequent integration.

[0174] In this step, the missing data of remote sensing pollutants predicted by the model and the non-missing data retained in the pseudo-missing data subset are added element by element in the matrix. This is actually to merge the missing values ​​predicted by the model with the non-missing values ​​in the original data to obtain a complete data set, in which each item is either a value predicted by the model or a value that originally existed in the original data.

[0175] In order to ensure the integrity and correctness of the data set, the data needs to be spliced ​​according to geographic coordinates (longitude and latitude). This is because remote sensing data is usually grid data based on geographic coordinates, and each grid point corresponds to a specific location on the earth's surface. By splicing according to longitude and latitude, the continuity and consistency of the data in space can be guaranteed, thereby obtaining the correct repair results.

[0176] In a preferred embodiment of the present application, the auxiliary data includes, but is not limited to, wind speed, wind direction, temperature, air pressure, altitude, longitude and latitude meteorological and geographic information data in the same area and at multiple times.

[0177] Wind conditions have a direct impact on the spread of pollutants. Strong winds may accelerate the spread of pollutants, while light winds may cause pollutants to accumulate in local areas. Wind direction determines the main direction of movement of pollutants.

[0178] Temperature changes affect the structure of the atmosphere, which in turn affects the vertical distribution of pollutants. For example, under temperature inversion conditions, pollutants may be trapped near the surface and are difficult to disperse.

[0179] Changes in air pressure can affect atmospheric stability, which in turn affects the diffusion patterns of pollutants.

[0180] Atmospheric conditions can vary significantly at different altitudes, and altitude can affect the concentration and distribution of pollutants.

[0181] Precise geolocation information helps identify the sources of pollutants at specific locations, as well as trends in pollutants over time and space.

[0182] These auxiliary data are used in various ways in remote sensing pollutant data processing, such as:

[0183] Spatiotemporal analysis: Combining meteorological data at multiple times can analyze the changing patterns of pollutants under different temporal and spatial conditions, helping to identify pollution sources and diffusion paths.

[0184] Model correction: Use auxiliary data to correct the atmospheric transport model and improve the accuracy of the model's prediction of pollutant concentrations.

[0185] Data fusion: remote sensing data is integrated with meteorological and pollutant data from ground monitoring stations to obtain a more detailed pollutant distribution map.

[0186] Anomaly detection: Auxiliary data can help identify outliers in remote sensing data, such as sudden increases in pollutant concentrations caused by special meteorological events.

[0187] In a preferred embodiment of the present application, the codec model is specifically a neural network model based on an encoder-decoder architecture, wherein the encoder includes a three-layer two-dimensional convolutional neural network and a self-attention mechanism layer; the decoder includes a three-layer two-dimensional convolutional neural network.

[0188] 3D Convolutional Neural Networks Convolutional Layer: The encoder contains three layers of 2D Convolutional Neural Networks (2D CNN). Convolutional layers capture local features such as edges and textures by sliding small filters (kernels) over the input data, which is particularly effective when processing images and spatial data.

[0189] The first convolutional layer is usually used to capture the most basic features.

[0190] The second and third layers learn progressively more abstract and higher-level feature representations.

[0191] A self-attention mechanism layer is added to the encoder to enhance the model's ability to focus on key information in the input data.

[0192] The self-attention mechanism allows the model to assign different weights based on the relationship between different parts of the input, thereby allocating resources more intelligently to the most important features when processing sequence or spatial data.

[0193] SE (Squeeze-and-Exci tat ion) network: Perform a global average pooling operation on a given feature map, compress it into a vector, then calculate the weight of each channel through a fully connected layer, and finally reapply these weights to the corresponding channels of the feature map to emphasize important features and suppress unimportant features.

[0194] Decoder: Deconvolution layer: The decoder also contains three layers of 2D convolutional neural network, but its working principle is opposite to that of the encoder. The purpose of the decoder is to reconstruct or generate a complete image or data set based on the feature representation produced by the encoder. Typically, the first layer of the decoder receives the feature map output by the encoder and then gradually restores it to the size of the original input through deconvolution operations. Subsequent layers continue to deconvolve and upsample until the output is a reconstructed data of the same size as the input.

[0195] Overall Architecture: End-to-End Training: The entire encoder-decoder model can be trained end-to-end as a whole, which means that the model can directly learn the mapping from input data to output data without the need for manual design of feature extraction or reconstruction processes.

[0196] Loss function: During training, the model optimizes its parameters by minimizing the difference between the reconstructed data and the original data. Commonly used loss functions include mean square error (MSE), cross entropy loss, etc. The specific choice depends on the specific task requirements.

[0197] Embodiment 2

[0198] This embodiment discloses a method for acquiring missing data of remote sensing pollutants.

[0199] This example takes TROPOMI pollutant data restoration as an example to collect NO in Shandong from 2020 to 2021 provided by TROPOMI. 2 hourly data, and obtain auxiliary data for the same area and time period provided by ERA5.

[0200] like Figure 1 As shown, a method for obtaining missing data of remote sensing pollutants includes the following steps:

[0201] S1. Construct the acquired remote sensing pollutant data and auxiliary data into an original data set; specifically:

[0202] The remote sensing pollutant data with missing values ​​obtained in this example are NO in Shandong area from 2020 to 2021 provided by TROPOMI. 2 Hourly data For the data of the i-th hour, the grid size is L = 190, W = 110, that is,

[0203]

[0204] Where L represents the length of the grid, W represents the width of the grid, and R represents the real number space.

[0205] The auxiliary data obtained in this embodiment without missing values ​​include: 10mU wind, 10mV wind, 2m relative humidity, 2m temperature hourly four kinds of auxiliary data in the same area and time period provided by ERA5 products, forming an auxiliary meteorological data matrix A. For the data of the i-th hour, the grid size is L=190, W=110, that is,

[0206] A i ∈R L×W×4

[0207] No 2 The hourly data and four auxiliary meteorological data are combined to form the original data set O, namely:

[0208]

[0209] O={O 1 ,O 2 ,...,O T}∈R T×L×W×5 .

[0210] By introducing meteorological and geographic information data such as wind speed, wind direction, temperature, air pressure, altitude, longitude and latitude in the same area as auxiliary data and increasing input information, the model can capture the spatial correlation between pollutant values ​​and meteorological values, analyze the impact of topography on pollutant distribution, and improve the model's fitting ability and repair accuracy.

[0211] S2, divide the original data set into missing data set and non-missing data set according to whether it contains missing values;

[0212] Specifically, the original data set O is divided into window pollutant grid data D with a length of 64 and a width of 64, that is, by the formula:

[0213] d=f(64,64,1,O)

[0214] d is the output sample, (64×64) is the window size, 1 is the step size, O is the original data set; f is the method of sliding window interception of data, according to whether there is NO 2 Missing, divided into missing data set D M and the non-missing dataset D C ,in:

[0215] D M ∈R N×64×64×5

[0216] D C ∈R K×64×64×5

[0217] N and K represent the size of the data set, and M and C are used to distinguish the superscripts of the missing data (Miss) and complete data (Complete) data sets, respectively.

[0218] S3. Extract missing dataset D M The missing patterns in , construct the mask pattern set M;

[0219] Among them, for D M Sample d in m , and its mask extraction formula is:

[0220]

[0221] where ∈R 64×64×5 , represents window data with missing values, x, y represent d m The horizontal and vertical coordinates of any grid in

[0222] The matrix representation of the mask pattern set M is:

[0223] M={m 1 ,m 2 ,...,m n}∈R N×64×64×5 .

[0224] S4, masking different mask patterns in the mask pattern set onto the non-missing data set to construct a pseudo-missing data set, and normalizing the non-missing data set and the pseudo-missing data set to construct a pre-training data set;

[0225] Specifically: for the non-missing data set D C Any sample d in c , and perform matrix multiplication with any mask pattern m in the mask pattern set M to obtain D C Dataset D after random masking C’ , the formula is:

[0226] d c′ =m×d c

[0227]

[0228] where d c Indicates any NO with a size of 64*64*5 and no missing values 2 Window Data

[0229] Right D C , D C’ The data d∈R 64×64×5 The normalization operation performed in this embodiment adopts the maximum and minimum value normalization method, and the formula is:

[0230]

[0231] where X∈R 64×64 is any kind of grid data, Y is the normalized data, x and y represent the horizontal and vertical coordinates of each data in the grid, and max and min represent the maximum and minimum values, respectively.

[0232] By building a true mask pre-training mechanism, the distribution pattern of missing data can be efficiently identified and fill-in values ​​can be automatically generated.

[0233] S5. Train the model using the pre-trained model dataset.

[0234] The model is as follows Figure 2 The convolutional attention mechanism neural network based on the encoder-decoder architecture shown in FIG. 1 is composed of a three-layer two-dimensional convolutional neural network and an SE attention mechanism layer. The output end of the three-layer two-dimensional convolutional neural network of the encoder is connected to the SE attention mechanism layer, and the decoder is composed of a three-layer two-dimensional convolutional neural network. C , D C’ Divide the training set, validation set, and test set into a ratio of 8:1:1; C’ The samples are passed into the input of the model, and the output of the model is compared with D C The loss function is calculated by comparing the corresponding samples in the example. The loss function L adopted in this embodiment is RMSE loss, and its calculation formula is:

[0235]

[0236] in Represents the true value of the covered grid with coordinates (x, y), (x,y) is its repair value, and S represents the number of grids covered by the sample.

[0237] For the model obtained in S5, this embodiment terminates the training and obtains a stable model when the loss of the model on the validation set does not decrease for 10 consecutive rounds, or fluctuates up and down for more than 50 rounds.

[0238] The advantage of RMSE as a loss function is that it takes into account both the size and direction of the prediction error. And because its calculation process involves squaring, it gives more penalties to predictions with larger errors. By minimizing this loss function, that is, finding the model parameters that minimize the mean square error, we can guide model learning and gradually improve the prediction ability.

[0239] By extracting the real missing pattern and applying it to the pre-trained model, the distribution characteristics of missing data can be simulated more accurately, the generalization ability of the model can be improved, and filling values ​​that are more in line with the actual situation can be generated, so that it can adapt to more scenarios; compared with the traditional random batch masking mode, the model constructed by the present invention is better capable of solving the missing repair problem in real scenarios, which makes the repaired remote sensing pollutant data better support subsequent analysis and application.

[0240] The present invention adopts a codec architecture model, combined with a convolutional neural network (CNN) and an attention mechanism, to achieve the restoration of missing data of remote sensing pollutants. CNN can extract local spatial features of input data, and can better focus on the relationship between multi-source data in the real missing mode. The attention mechanism focuses on important areas and key information, and better understands the distribution pattern of input data. The encoder-decoder architecture model realizes end-to-end learning, automatically learns data features and reconstructs missing data. The three work together to effectively capture the data correlation and spatial correlation between pollutant data and auxiliary data, and improve the restoration accuracy and reliability.

[0241] S6, such as Figure 3 As shown in Figure 2, the remote sensing pollutant data and auxiliary data with missing values ​​are input into the trained model to obtain the repaired pollutant data D R , R stands for Repai r, used to distinguish superscripts M, C, and C', and the missing data set D M Splicing, inverse normalization processing, and obtain the final data repair result.

[0242] Among them, for the repaired pollutant data D in S6 R With missing data set D M The steps of splicing are:

[0243] For any single sample d∈D M , retain the pollutant values ​​of the grids that are not missing, and set the grids of the missing parts to 0. The formula is as follows:

[0244]

[0245] The repair result d for sample d R , only the repair results of the missing positions of the original data are retained, and the data of the non-missing positions are set to 0. The formula is as follows:

[0246]

[0247] For sample d, repair result d R Perform element-by-element matrix addition to obtain the final repair result d F , the formula is as follows:

[0248] dF =d+d R .

[0249] Here, F stands for Final and is used to distinguish superscripts R, M, C and C'.

[0250] Furthermore, the repaired data is spliced ​​according to longitude and latitude, and inverse normalization is performed to obtain the final data repair result.

[0251] Embodiment 3

[0252] The purpose of this embodiment is to provide a computer-readable storage medium.

[0253] A computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps in the method for acquiring missing remote sensing pollutant data as described in Example 1 of the present disclosure.

[0254] Embodiment 4

[0255] The purpose of this embodiment is to provide an electronic device.

[0256] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the method for acquiring missing data of remote sensing pollutants of embodiment 1 of the present disclosure are implemented.

[0257] The steps involved in the apparatuses of the above embodiments 2, 3 and 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium, any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method of the present invention.

[0258] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0259] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A method for obtaining missing data of remote sensing pollutants, characterized in that: The following steps are involved: Divide the window grid data corresponding to the remote sensing pollutant data into missing data subsets and non-missing data subsets; Randomly masking the mask set corresponding to the missing data subset to the remote sensing pollutant data in the non-missing data subset to generate a pseudo-missing data subset, and generating a pre-trained model data set based on the non-missing data subset and the pseudo-missing data subset; A codec model is obtained by training according to the pre-trained model data set; Remote sensing pollutant missing data corresponding to the remote sensing pollutant data is acquired according to the pseudo missing data subset and the codec model.

2. The method for acquiring missing data of remote sensing pollutants according to claim 1, characterized in that: The window grid data corresponding to the remote sensing pollutant data are divided into missing data subsets and non-missing data subsets, specifically: Merging the remote sensing pollutant data and auxiliary data corresponding to the remote sensing pollutant data into an original data set; Dividing the original data set into the window grid data according to the specified length and width; Dividing the window grid data into a plurality of subsets according to a specified window size; According to whether each subset is missing remote sensing pollutant data, each subset is divided into the missing data subset and the non-missing data subset.

3. The method for acquiring missing data of remote sensing pollutants according to claim 1, characterized in that: The mask set corresponding to the missing data subset is randomly masked on the remote sensing pollutant data in the non-missing data subset to generate a pseudo-missing data subset, and a pre-trained model data set is generated according to the non-missing data subset and the pseudo-missing data subset, specifically: Generating a mask set according to the extraction masks corresponding to the samples in the missing data subset; Generate the pseudo missing data subset according to any extraction mask in the matrix corresponding to the mask set and any sample in the non-missing data set; The data in the non-missing data subset and the data in the pseudo-missing data subset are respectively normalized to generate the pre-training model data set.

4. The method for acquiring missing data of remote sensing pollutants according to claim 1, characterized in that: The codec model is obtained by training the pre-trained model dataset, specifically: Dividing the pre-trained model data set into a training set, a validation set, and a test set according to a preset ratio, and training the codec model using the training set; Compare the output results of the test set obtained by the codec model with the corresponding samples in the validation set to obtain a loss function; When the loss functions obtained multiple times meet a preset termination condition, the training of the codec model is terminated, and the codec model after the training is terminated is obtained.

5. The method for acquiring missing data of remote sensing pollutants according to claim 4, characterized in that: The preset termination conditions are specifically: The loss function obtained for a specified number of consecutive times does not decrease; And / or, the loss function obtained for a specified number of consecutive times does not fluctuate; and / or, reaching a predetermined minimum loss threshold; and / or, reaching the maximum number of training rounds; Among them, the loss function includes but is not limited to: mean square error, mean absolute error, mixed linear quadratic loss function, logarithmic hyperbolic cosine loss function and quantile loss function.

6. The method for acquiring missing data of remote sensing pollutants according to any one of claims 1 to 5, characterized in that: After acquiring the remote sensing pollutant missing data corresponding to the remote sensing pollutant data according to the pseudo missing data subset and the codec model, the method further includes: retaining missing data of remote sensing pollutants corresponding to the remote sensing pollutant data, and setting the remaining data in the remote sensing pollutant data to 0; Retain the non-missing data in the pseudo-missing data subset, and set the missing data in the pseudo-missing data subset to 0; Performing an element-by-element matrix addition operation on the remote sensing pollutant missing data corresponding to the remote sensing pollutant data and the non-missing data in the pseudo missing data subset, and splicing them according to longitude and latitude; The spliced ​​data are used as the repair results of remote sensing pollutant data.

7. The method for acquiring missing data of remote sensing pollutants according to claim 2, characterized in that: The auxiliary data include but are not limited to wind speed, wind direction, temperature, air pressure, altitude, longitude and latitude meteorological and geographic information data in the same area and at multiple times.

8. The method for acquiring missing data of remote sensing pollutants according to claim 6, characterized in that: The codec model is specifically a neural network model based on an encoder-decoder architecture, wherein the encoder includes a three-layer two-dimensional convolutional neural network and a self-attention mechanism layer; the decoder includes a three-layer two-dimensional convolutional neural network.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method for acquiring missing data of remote sensing pollutants as described in any one of claims 1 to 8 are implemented.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the method for acquiring missing data of remote sensing pollutants as described in any one of claims 1 to 8 are implemented.

Images