A method and device for identifying dust pollution based on deep learning

CN122797818APending Publication Date: 2026-09-22CHINA NAT ENVIRONMENTAL MONITORING CENT
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611300058.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-26
Publication Date
2026-09-22

AI Technical Summary

Technical Problem

[0003]现有沙尘污染事件识别方法主要包括以下几类:浓度阈值法依赖单一固定阈值,难以区分沙尘过程与本地扬尘等非沙尘过程;遥感识别法受云层遮挡和时间分辨率限制,难以实现全天候业务化监测;经验规则法主观性强、跨区域推广困难;现有技术存在的技术问题为:多源环境数据未能得到充分利用和有效融合、深度学习模型的决策过程缺乏可解释性、模型输出形式与业务需求脱节

Benefits of technology

(1)通过对空气质量监测数据、气象数据和卫星遥感数据进行时空配准处理,解决了多源异构数据在时间和空间维度上的对齐问题;通过对不同类型数据分别进行针对性特征编码,实现了对各类数据物理含义的充分挖掘;通过特征融合得到统一环境特征表示,使模型能够综合利用地面监测的高精度污染物信息、气象扩散条件信息和卫星遥感的区域空间覆盖信息,实现多维度信息的互补增强,提升了沙尘事件识别的准确性。

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122797818A_ABST
    Figure CN122797818A_ABST
Patent Text Reader

Abstract

This application relates to a method and apparatus for identifying dust pollution based on deep learning, belonging to the field of environmental monitoring technology. The method includes: acquiring multi-source environmental data of the area to be identified, including air quality monitoring data, meteorological data, and satellite remote sensing data; performing spatiotemporal registration processing on the multi-source environmental data, and encoding the processed multi-source environmental data to obtain air quality features, meteorological features, and satellite remote sensing features; fusing the air quality features, meteorological features, and satellite remote sensing features to obtain a unified environmental feature representation; standardizing the unified environmental feature representation and inputting it into a pre-trained deep learning recognition model to output dust pollution identification results; analyzing the output of the deep learning recognition model using interpretability analysis methods to obtain the contribution of each environmental feature to the dust pollution identification results, and generating structured interpretation information based on the contribution.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application belongs to the field of environmental monitoring technology, and specifically relates to a method and device for identifying sand and dust pollution based on deep learning. Background Technology

[0002] Dust pollution events are characterized by their suddenness, wide impact, and high degree of harm. How to quickly and accurately identify dust pollution events based on daily monitoring data has become an important requirement for the refined development of environmental monitoring services.

[0003] Existing methods for identifying dust pollution events mainly fall into the following categories: concentration threshold methods rely on a single fixed threshold, making it difficult to distinguish between dust events and non-dust events such as local dust storms; remote sensing identification methods are limited by cloud cover and temporal resolution, making it difficult to achieve all-weather operational monitoring; empirical rule methods are highly subjective and difficult to promote across regions; existing technologies have the following technical problems: multi-source environmental data are not fully utilized and effectively integrated, the decision-making process of deep learning models lacks interpretability, and the model output format is out of touch with business needs. Summary of the Invention

[0004] Based on the above analysis, the embodiments of the present invention aim to provide a method and apparatus for identifying sand and dust pollution based on deep learning.

[0005] The objective of this invention is achieved as follows: A first aspect of the present invention provides a method for identifying sandstorm pollution based on deep learning, comprising: Acquire multi-source environmental data for the area to be identified, including air quality monitoring data, meteorological data, and satellite remote sensing data; The multi-source environmental data is subjected to spatiotemporal registration processing, and the processed multi-source environmental data is feature-encoded to obtain air quality features, meteorological features, and satellite remote sensing features; the air quality features, meteorological features, and satellite remote sensing features are then fused to obtain a unified environmental feature representation. After standardizing the unified environmental feature representation, it is input into a pre-trained deep learning recognition model to output the dust pollution recognition result. The output of the deep learning recognition model is analyzed using interpretability analysis methods to obtain the contribution of each environmental feature to the dust pollution recognition result, and structured interpretation information is generated based on the contribution.

[0006] Furthermore, the air quality monitoring data includes PM2.5 concentration, PM10 concentration, coarse particulate matter concentration, PM2.5 / PM10 ratio, and air quality sub-index; the meteorological data includes temperature, humidity, wind speed, wind direction, and air pressure; and the satellite remote sensing data includes aerosol optical thickness (AOD) data.

[0007] Furthermore, the spatiotemporal registration processing of multi-source environmental data includes: performing time synchronization processing on air quality monitoring data, meteorological data, and satellite remote sensing data according to a preset time granularity; and establishing spatial correspondence between air quality monitoring data, meteorological data, and satellite remote sensing data according to a preset spatial index.

[0008] Furthermore, the feature encoding of the processed multi-source environmental data includes: encoding the air quality monitoring data for pollution features to obtain air quality features characterizing the composition and concentration relationship of pollutants; encoding the meteorological data for diffusion conditions to obtain meteorological features characterizing atmospheric diffusion conditions; and encoding the satellite remote sensing data for spatial coverage to obtain satellite remote sensing features characterizing the spatial distribution information of dust.

[0009] Furthermore, the feature fusion of the air quality features, meteorological features, and satellite remote sensing features to obtain a unified environmental feature representation includes: inputting air quality coding features, meteorological coding features, and remote sensing coding features into a feature fusion module; the feature fusion module determining fusion weights based on the importance of each type of coding feature; concatenating the various coding features and performing weighted fusion based on the fusion weights to obtain the unified environmental feature representation.

[0010] Furthermore, the standardization process for the unified environment feature representation includes: standardizing the unified environment feature representation using the Z-score standardization method; the Z-score standardization method uses the mean and standard deviation of the training set of the deep learning recognition model as standardization parameters, and performs centering and scaling processing on each dimension of the unified environment feature representation; the standardized unified environment feature representation is then expanded into a three-dimensional tensor form to adapt to the input format of the deep learning recognition model.

[0011] Furthermore, the deep learning recognition model is a one-dimensional convolutional neural network model, including a first convolutional layer, a first max-pooling layer, a second convolutional layer, a second max-pooling layer, a flattening layer, a fully connected layer, a random deactivation layer, and an output layer; the pre-training process of the deep learning recognition model includes: acquiring historical multi-source environmental data including historical air quality monitoring data, historical meteorological data, and historical satellite remote sensing data, as well as dust pollution labels corresponding to the historical multi-source environmental data; performing spatiotemporal registration processing on the historical multi-source environmental data, and performing feature encoding on each type of registered historical environmental data to obtain historical air quality features, historical meteorological features, and historical satellite remote sensing features; and encoding the historical air quality features, Historical meteorological features and historical satellite remote sensing features are fused to obtain a unified historical environmental feature representation. A year-specific dataset partitioning strategy is adopted to divide the unified historical environmental feature representation and its corresponding dust pollution labels into training, validation, and test sets. The training set is used to iteratively train the one-dimensional convolutional neural network model. During training, different values ​​for the number of convolutional kernels, the number of convolutional layers, the dropout rate, the batch size, the learning rate, and whether or not a batch normalization layer is introduced are combined to obtain various hyperparameter combinations. Comparative experiments are conducted on these hyperparameter combinations, and the model corresponding to the hyperparameter combination with the best performance is selected as the pre-trained deep learning recognition model based on preset performance evaluation indicators.

[0012] Furthermore, the output of the dust pollution identification result includes: inputting the standardized unified environmental feature representation into the pre-trained deep learning identification model; the deep learning identification model sequentially extracts and reduces features through a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a flattening layer, a fully connected layer, and a random deactivation layer, and then outputs the probability value of the dust event through the output layer; if the probability value is greater than a preset classification threshold, it is determined to be a dust pollution event; otherwise, it is determined to be a non-dust pollution event.

[0013] Further, the step of analyzing the output of the deep learning recognition model using interpretability analysis to obtain the contribution of each environmental feature to the dust pollution recognition result, and generating structured explanation information based on the contribution, includes: performing interpretability analysis on the output of the deep learning recognition model using the SHAP method, calculating the SHAP value corresponding to each environmental feature of each input sample; determining the contribution direction and magnitude of each environmental feature to the dust pollution recognition result based on the SHAP value, wherein environmental features with SHAP values ​​greater than a first preset threshold are determined as contribution directions that increase the dust probability, and environmental features with SHAP values ​​less than a second preset threshold are determined as contribution directions that decrease the dust probability; sorting each environmental feature according to the absolute value of its SHAP value; selecting environmental features with an absolute SHAP value greater than a preset contribution threshold as target environmental features; determining the feature value offset state of the target environmental features, and generating the structured explanation information according to a preset explanation template based on the feature value offset state and contribution direction of each target environmental feature.

[0014] A second aspect of the present invention provides an apparatus for identifying sandstorm pollution based on deep learning, comprising: The data acquisition module is used to acquire multi-source environmental data of the area to be identified, including air quality monitoring data, meteorological data and satellite remote sensing data; The feature extraction and fusion module is used to perform spatiotemporal registration processing on the multi-source environmental data, encode the processed multi-source environmental data for features to obtain air quality features, meteorological features and satellite remote sensing features; and fuse the air quality features, meteorological features and satellite remote sensing features to obtain a unified environmental feature representation. The dust pollution identification module is used to standardize the unified environmental feature representation and then input it into a pre-trained deep learning identification model to output the dust pollution identification result. The interpretability analysis module is used to analyze the output of the deep learning recognition model using interpretability analysis methods, obtain the contribution of each environmental feature to the dust pollution recognition result, and generate structured interpretation information based on the contribution.

[0015] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the method for identifying sand and dust pollution based on deep learning as described in any embodiment.

[0016] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method for identifying sand and dust pollution based on deep learning as described in any embodiment.

[0017] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: (1) By performing spatiotemporal registration processing on air quality monitoring data, meteorological data and satellite remote sensing data, the alignment problem of multi-source heterogeneous data in time and space dimensions was solved; by performing targeted feature encoding on different types of data, the physical meaning of various types of data was fully explored; by obtaining a unified environmental feature representation through feature fusion, the model can comprehensively utilize high-precision pollutant information from ground monitoring, meteorological diffusion condition information and regional spatial coverage information from satellite remote sensing, to achieve complementary enhancement of multi-dimensional information and improve the accuracy of dust storm event identification.

[0018] (2) By using interpretability analysis to analyze the output of the deep learning identification model, it is possible to quantitatively reveal the specific contribution of each environmental feature to the dust pollution identification result, thereby verifying the rationality of the model's decision-making logic and breaking through the bottleneck that deep learning models are difficult to trust and deploy in environmental monitoring operations due to a lack of interpretability.

[0019] (3) Transform the contribution of each environmental feature into structured explanatory information, so that the model output results can be transformed from black box probability values ​​that are difficult to understand into structured explanatory texts that business personnel can directly understand and use as a basis for manual judgment. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings.

[0021] Figure 1 This is a flowchart of the method for identifying sand and dust pollution based on deep learning provided in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the year-based dataset partitioning strategy provided in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the performance evaluation metrics for the training and verification processes provided in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the evaluation indicators for the dust pollution identification results provided in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of a waterfall diagram provided in Embodiment 1 of the present invention; Figure 6 This is a schematic diagram of the device for identifying sand and dust pollution based on deep learning provided in Embodiment 2 of the present invention; Figure 7This is a schematic diagram of the electronic device architecture provided in Embodiment 3 of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. It should be noted that, unless otherwise specified, the implementation methods and features in the implementation methods in this disclosure can be combined, separated, interchanged, and / or rearranged. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] Example 1 A specific embodiment of the present invention, such as Figure 1-5 As shown, a method for identifying sandstorm pollution based on deep learning is disclosed, including the following steps: S1. Obtain multi-source environmental data of the area to be identified, including air quality monitoring data, meteorological data and satellite remote sensing data.

[0024] In this embodiment, the air quality monitoring data includes PM2.5 concentration, PM10 concentration, coarse particulate matter concentration, PM2.5 / PM10 ratio, and air quality index; the meteorological data includes temperature, humidity, wind speed, wind direction, and air pressure; and the satellite remote sensing data includes aerosol optical thickness (AOD) data.

[0025] Specifically, air quality monitoring data are derived from hourly monitoring data from national-level air quality monitoring stations in various cities within the target area. The daily average value is calculated from the hourly data and used as the city's daily air quality monitoring data. The coarse particulate matter concentration is the difference between PM10 and PM2.5 concentrations, representing the degree of injection of large-diameter mineral particles during dust storms. The PM2.5 / PM10 ratio reflects the particle size distribution characteristics of particulate matter, and this ratio usually decreases significantly during dust storm events. The air quality sub-index is calculated according to the AQI sub-index calculation method specified in the "Technical Regulations for Ambient Air Quality Index (AQI)" (HJ 633—2026), and the difference between the sub-indices of PM10 and PM2.5 is taken. Meteorological data are derived from meteorological reanalysis data or ground meteorological observation stations. Satellite remote sensing AOD data can be derived from polar-orbiting or geostationary orbit satellite remote sensing products, and the spatial regional average is calculated by date to serve as the city's daily AOD data.

[0026] S2. Perform spatiotemporal registration processing on the multi-source environmental data, and encode the processed multi-source environmental data by feature to obtain air quality features, meteorological features and satellite remote sensing features; perform feature fusion on the air quality features, meteorological features and satellite remote sensing features to obtain a unified environmental feature representation.

[0027] In this embodiment, step S2 specifically includes: S201. Perform spatiotemporal registration processing on multi-source environmental data, including: The air quality monitoring data, meteorological data and satellite remote sensing data are synchronized in time according to the preset time granularity. Establish spatial correspondences between air quality monitoring data, meteorological data, and satellite remote sensing data based on a preset spatial index.

[0028] Specifically, using cities within the target area as spatial index units, the air quality monitoring data of all national control stations within each city are arithmetically averaged to obtain city-scale air quality monitoring data; meteorological reanalysis data are extracted by bilinear interpolation according to the latitude and longitude coordinates of the city center to obtain city-scale meteorological data; and satellite remote sensing AOD data are spatially averaged according to the boundaries of city administrative divisions to obtain city-scale AOD data.

[0029] S202. Perform feature encoding on the processed multi-source environmental data, including: Pollution feature encoding is performed on the air quality monitoring data to obtain air quality features that characterize the composition and concentration relationship of pollutants; By encoding the diffusion conditions of meteorological data, meteorological characteristics that characterize atmospheric diffusion conditions are obtained; Spatial coverage coding is performed on satellite remote sensing data to obtain satellite remote sensing features that characterize the spatial distribution information of dust.

[0030] Specifically, the values ​​of five dimensions—PM2.5 concentration, PM10 concentration, coarse particulate matter concentration (the difference between PM10 and PM2.5 concentrations), PM2.5 / PM10 ratio, and air quality sub-index—are arranged in a fixed order to constitute air quality characteristics. This characteristic vector represents the concentration level of pollutants in the atmosphere, the particle size composition characteristics of particulate matter, and the typical pollution characteristics dominated by coarse particulate matter in dust events. The values ​​of five meteorological elements—temperature, humidity, wind speed, wind direction, and air pressure—are arranged in a fixed order to constitute meteorological characteristics. Among them, wind speed and wind direction jointly determine the horizontal transport direction and diffusion rate of dust, temperature and humidity affect the vertical diffusion and sedimentation process of dust particles, and air pressure distribution reflects the restrictive effect of weather background on dust formation and transport. The aerosol optical depth (AOD) value is used as a satellite remote sensing feature. AOD represents the total extinction effect of aerosols in the entire atmospheric column. Due to their strong light scattering and absorption characteristics, dust aerosols usually cause a significant increase in AOD value during dust events. AOD represents the spatial distribution and intensity of dust.

[0031] S203. The air quality characteristics, meteorological characteristics, and satellite remote sensing characteristics are fused to obtain a unified environmental characteristic representation, including: Input air quality coding features, meteorological coding features and remote sensing coding features into the feature fusion module; The feature fusion module determines the fusion weights based on the importance of various coded features; The various coding features are concatenated and then weighted and fused according to the fusion weights to obtain the unified environment feature representation.

[0032] In one specific implementation of this embodiment, the fusion weight is calculated through an attention mechanism, and then the air quality features (dimension 5), meteorological features (dimension 5), and satellite remote sensing features (dimension 1) are concatenated into an 11-dimensional environmental feature vector, which serves as the unified environmental feature representation.

[0033] S3. After standardizing the unified environmental feature representation, input it into the pre-trained deep learning recognition model and output the dust pollution recognition result.

[0034] In this embodiment, the deep learning recognition model is a one-dimensional convolutional neural network model, including a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a flattening layer, a fully connected layer, a random deactivation layer, and an output layer.

[0035] In this embodiment, the pre-training process of the deep learning recognition model includes: (1) Acquire historical multi-source environmental data, including historical air quality monitoring data, historical meteorological data and historical satellite remote sensing data, as well as the dust pollution labels corresponding to the historical multi-source environmental data; (2) Spatiotemporal registration processing is performed on the historical multi-source environmental data, and feature encoding is performed on the registered historical environmental data to obtain historical air quality features, historical meteorological features and historical satellite remote sensing features; feature fusion is performed on the historical air quality features, historical meteorological features and historical satellite remote sensing features to obtain a unified historical environmental feature representation. (3) Using a year-based dataset partitioning strategy, the historical unified environmental feature representation and its corresponding dust pollution labels are divided into training set, validation set, and test set, such as... Figure 2 As shown; (4) The one-dimensional convolutional neural network model is iteratively trained using the training set. During the training process, different values ​​of the number of convolutional kernels, the number of convolutional layers, the dropout rate, the batch size, the learning rate, and whether or not a batch normalization layer is introduced are combined to obtain multiple hyperparameter combinations. Comparative experiments are conducted on multiple hyperparameter combinations, and the model corresponding to the hyperparameter combination with the best performance is selected as the pre-trained deep learning recognition model according to the preset performance evaluation index. Figure 3 As shown, the performance evaluation metrics for the training and validation processes of deep learning recognition models include accuracy, loss value, recall, and F1 score.

[0036] In this embodiment, step S3 specifically includes: S301. Standardize the unified environment feature representation, including: The unified environment feature representation is standardized using the Z-score standardization method. The Z-score standardization method uses the mean and standard deviation of the training set of the deep learning recognition model as standardization parameters to center and scale each dimension of the unified environment feature representation. The standardized unified environment feature representation is then expanded into a three-dimensional tensor form to adapt to the input format of the deep learning recognition model.

[0037] Specifically, the input format of a one-dimensional convolutional neural network is (number of samples, feature dimension, number of channels), where the number of channels is set to 1. The standardized two-dimensional feature vector (each sample corresponds to a 1×11 vector) is expanded into a three-dimensional tensor (each sample corresponds to an 11×1×1 tensor) to adapt to the input format of the one-dimensional convolutional neural network.

[0038] S302. Output the dust pollution identification results, including: The standardized unified environmental feature representation is input into the pre-trained deep learning recognition model; The deep learning recognition model extracts and reduces features sequentially through a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a flattening layer, a fully connected layer, and a random deactivation layer, and then outputs the probability value of the sandstorm event through the output layer. If the probability value is greater than the preset classification threshold, it is determined to be a sandstorm pollution event; otherwise, it is determined to be a non-sandstorm pollution event.

[0039] Specifically, after receiving input in the form of a three-dimensional tensor, the first convolutional layer performs sliding convolution operations along the feature dimension using 32 convolutional kernels of size 3 to extract local correlation patterns between adjacent features, and introduces nonlinearity through the ReLU activation function; the first max pooling layer downsamples with a pooling window of 2 to reduce the feature dimension while enhancing the translation invariance of the features; the second convolutional layer further extracts higher-order feature combinations using 64 convolutional kernels of size 3; the second max pooling layer performs downsampling again; the flattening layer converts the multi-dimensional feature map into a one-dimensional feature vector; the fully connected layer contains 64 neurons and performs nonlinear combination mapping on the extracted features; the random deactivation layer randomly discards some neuron outputs with a probability of 0.5 during the training phase to prevent the model from overfitting; the output layer uses the Sigmoid activation function to map the output of the previous layer to the [0,1] interval as the probability value of the sandstorm event. Figure 4 The table shows the evaluation metrics for dust pollution identification results in each month of the region to be identified during the period of 2023-2025, including accuracy, precision, recall, and F1 score.

[0040] S4. Analyze the output of the deep learning recognition model using the interpretability analysis method to obtain the contribution of each environmental feature to the dust pollution recognition result, and generate structured interpretation information based on the contribution.

[0041] In this embodiment, step S4 specifically includes the following steps: (1) The SHAP method is used to perform interpretability analysis on the output of the deep learning recognition model, and the SHAP value corresponding to each environmental feature of each input sample is calculated; (2) Based on the SHAP value, determine the contribution direction and contribution magnitude of each environmental feature to the dust pollution identification result, wherein environmental features with SHAP values ​​greater than the first preset threshold are determined as the contribution direction that increases the probability of dust, and environmental features with SHAP values ​​less than the second preset threshold are determined as the contribution direction that decreases the probability of dust. (3) Sort each environmental feature according to the absolute value of its SHAP value; (4) Select environmental features whose absolute SHAP value is greater than the preset contribution threshold as target environmental features; (5) Determine the feature value offset state of the target environment feature, and generate the structured interpretation information according to the feature value offset state and contribution direction of each target environment feature in accordance with the preset interpretation template.

[0042] In some embodiments, step S4 further includes: single-sample waterfall plot analysis, selecting one correctly identified dust pollution event sample and one non-dust pollution event sample respectively, and plotting as shown below. Figure 5 The waterfall plot shown illustrates how each feature in a sample gradually shifts the predicted probability toward the final value from the base value.

[0043] Specifically, the method for judging the feature value offset state is as follows: if the standardized feature value of the target environmental feature is greater than 0.5, it is judged to be higher than the historical normal level; if the standardized value is less than -0.5, it is judged to be lower; otherwise, it is judged to be normal. The contribution direction is determined according to the sign of the SHAP value: a positive SHAP value indicates that the environmental feature will increase the probability of dust storms, marked as ↑; a negative SHAP value indicates that the environmental feature will decrease the probability of dust storms, marked as ↓.

[0044] The format of the structured explanation information is: "prob=X.XXX|feature name 1 (absolute value of SHAP in the direction of offset state contribution), feature name 2 (absolute value of SHAP in the direction of offset state contribution), ……". For example, for a sample of a dust storm event, the following structured explanation information can be generated: The graph "prob=0.923|PM10 is high (↑0.152), coarse particulate matter is high (↑0.098), PM2.5 / PM10 is low (↑0.067), AOD is high (↑0.043), relative humidity is low (↑0.031)" indicates that high PM10 concentration is the biggest driver of the increased probability of dust storms, contributing 0.152; high coarse particulate matter concentration is the second largest contributor; and a low PM2.5 / PM10 ratio also increases the probability of dust storms.

[0045] Compared with existing technologies, the deep learning-based method for identifying sand and dust pollution provided in this embodiment can achieve the following technical effects: (1) By performing spatiotemporal registration processing on air quality monitoring data, meteorological data and satellite remote sensing data, the alignment problem of multi-source heterogeneous data in time and space dimensions was solved; by performing targeted feature encoding on different types of data, the physical meaning of various types of data was fully explored; by obtaining a unified environmental feature representation through feature fusion, the model can comprehensively utilize high-precision pollutant information from ground monitoring, meteorological diffusion condition information and regional spatial coverage information from satellite remote sensing, to achieve complementary enhancement of multi-dimensional information and improve the accuracy of dust storm event identification.

[0046] (2) By using interpretability analysis to analyze the output of the deep learning identification model, it is possible to quantitatively reveal the specific contribution of each environmental feature to the dust pollution identification result, thereby verifying the rationality of the model's decision-making logic and breaking through the bottleneck that deep learning models are difficult to trust and deploy in environmental monitoring operations due to a lack of interpretability.

[0047] (3) Transform the contribution of each environmental feature into structured explanatory information, so that the model output results can be transformed from black box probability values ​​that are difficult to understand into structured explanatory texts that business personnel can directly understand and use as a basis for manual judgment.

[0048] Example 2 This embodiment provides a device for identifying sand and dust pollution based on deep learning, such as... Figure 6 As shown, it includes: The data acquisition module is used to acquire multi-source environmental data of the area to be identified, including air quality monitoring data, meteorological data and satellite remote sensing data; The feature extraction and fusion module is used to perform spatiotemporal registration processing on the multi-source environmental data, encode the processed multi-source environmental data for features to obtain air quality features, meteorological features and satellite remote sensing features; and fuse the air quality features, meteorological features and satellite remote sensing features to obtain a unified environmental feature representation. The dust pollution identification module is used to standardize the unified environmental feature representation and then input it into a pre-trained deep learning identification model to output the dust pollution identification result. The interpretability analysis module is used to analyze the output of the deep learning recognition model using interpretability analysis methods, obtain the contribution of each environmental feature to the dust pollution recognition result, and generate structured interpretation information based on the contribution.

[0049] Example 3 This embodiment provides an electronic device, such as... Figure 7 As shown, it includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the deep learning-based method for identifying sand and dust pollution as described in any of the above embodiments.

[0050] Example 4 This embodiment provides a computer-readable storage medium storing a computer program thereon. When the program is executed by a processor, it implements the method for identifying sand and dust pollution based on deep learning as described in any of the above embodiments.

[0051] Computer-readable storage media include both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0052] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0053] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0054] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying sandstorm pollution based on deep learning, characterized in that, include: Acquire multi-source environmental data for the area to be identified, including air quality monitoring data, meteorological data, and satellite remote sensing data; The multi-source environmental data is subjected to spatiotemporal registration processing, and the processed multi-source environmental data is feature-encoded to obtain air quality features, meteorological features and satellite remote sensing features. The air quality characteristics, meteorological characteristics, and satellite remote sensing characteristics are fused to obtain a unified environmental characteristic representation. After standardizing the unified environmental feature representation, it is input into a pre-trained deep learning recognition model to output the dust pollution recognition result. The output of the deep learning recognition model is analyzed using interpretability analysis methods to obtain the contribution of each environmental feature to the dust pollution recognition result, and structured interpretation information is generated based on the contribution.

2. The method for identifying sandstorm pollution based on deep learning according to claim 1, characterized in that, The air quality monitoring data includes PM2.5 concentration, PM10 concentration, coarse particulate matter concentration, PM2.5 / PM10 ratio, and air quality index; the meteorological data includes temperature, humidity, wind speed, wind direction, and air pressure; and the satellite remote sensing data includes aerosol optical depth (AOD) data.

3. The method for identifying sandstorm pollution based on deep learning according to claim 1, characterized in that, The spatiotemporal registration processing of multi-source environmental data includes: The air quality monitoring data, meteorological data and satellite remote sensing data are synchronized in time according to the preset time granularity. Establish spatial correspondences between air quality monitoring data, meteorological data, and satellite remote sensing data based on a preset spatial index.

4. The method for identifying sand and dust pollution based on deep learning according to claim 3, characterized in that, The step of performing feature encoding on the processed multi-source environmental data includes: Pollution feature encoding is performed on the air quality monitoring data to obtain air quality features that characterize the composition and concentration relationship of pollutants; By encoding the diffusion conditions of meteorological data, meteorological characteristics that characterize atmospheric diffusion conditions are obtained; Spatial coverage coding is performed on satellite remote sensing data to obtain satellite remote sensing features that characterize the spatial distribution information of dust.

5. The method for identifying sand and dust pollution based on deep learning according to claim 4, characterized in that, The feature fusion of the air quality characteristics, meteorological characteristics, and satellite remote sensing characteristics to obtain a unified environmental feature representation includes: Input air quality coding features, meteorological coding features and remote sensing coding features into the feature fusion module; The feature fusion module determines the fusion weights based on the importance of various coded features; The various coding features are concatenated and then weighted and fused according to the fusion weights to obtain the unified environment feature representation.

6. The method for identifying sandstorm pollution based on deep learning according to claim 1, characterized in that, The standardization process for the unified environmental feature representation includes: The unified environment feature representation is standardized using the Z-score standardization method. The Z-score standardization method uses the mean and standard deviation of the training set of the deep learning recognition model as standardization parameters to center and scale each dimension of the unified environment feature representation. The standardized unified environment feature representation is then expanded into a three-dimensional tensor form to adapt to the input format of the deep learning recognition model.

7. The method for identifying sandstorm pollution based on deep learning according to claim 1, characterized in that, The deep learning recognition model is a one-dimensional convolutional neural network model, including a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a flattening layer, a fully connected layer, a random deactivation layer, and an output layer; The pre-training process of the deep learning recognition model includes: Acquire historical multi-source environmental data, including historical air quality monitoring data, historical meteorological data, and historical satellite remote sensing data, as well as the dust pollution labels corresponding to the historical multi-source environmental data; The historical multi-source environmental data is spatiotemporally registered, and the registered historical environmental data of each type are feature-encoded to obtain historical air quality features, historical meteorological features, and historical satellite remote sensing features; the historical air quality features, historical meteorological features, and historical satellite remote sensing features are then fused to obtain a unified historical environmental feature representation. A year-based dataset partitioning strategy was adopted to divide the historical unified environmental feature representation and its corresponding dust pollution label into training set, validation set and test set; The one-dimensional convolutional neural network model is iteratively trained using the training set. During the training process, different values ​​of the number of convolutional kernels, the number of convolutional layers, the dropout rate, the batch size, the learning rate, and whether or not a batch normalization layer is introduced are combined to obtain multiple hyperparameter combinations. Comparative experiments are conducted on multiple hyperparameter combinations, and the model corresponding to the hyperparameter combination with the best performance is selected as the pre-trained deep learning recognition model according to the preset performance evaluation index.

8. The method for identifying sand and dust pollution based on deep learning according to claim 7, characterized in that, The output dust pollution identification results include: The standardized unified environmental feature representation is input into the pre-trained deep learning recognition model; The deep learning recognition model extracts and reduces features sequentially through a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a flattening layer, a fully connected layer, and a random deactivation layer, and then outputs the probability value of the sandstorm event through the output layer. If the probability value is greater than the preset classification threshold, it is determined to be a sandstorm pollution event; otherwise, it is determined to be a non-sandstorm pollution event.

9. The method for identifying sandstorm pollution based on deep learning according to claim 7, characterized in that, The process involves analyzing the output of the deep learning recognition model using interpretability analysis methods to obtain the contribution of each environmental feature to the dust pollution recognition result, and generating structured explanatory information based on the contribution, including: The SHAP method is used to perform interpretability analysis on the output of the deep learning recognition model, and the SHAP value corresponding to each environmental feature of each input sample is calculated. Based on the SHAP value, the contribution direction and magnitude of each environmental feature to the dust pollution identification result are determined. Among them, environmental features with SHAP values ​​greater than a first preset threshold are determined as contributing directions that increase the probability of dust, and environmental features with SHAP values ​​less than a second preset threshold are determined as contributing directions that decrease the probability of dust. The environmental features are sorted according to the absolute value of their SHAP values; Environmental features whose absolute SHAP value is greater than a preset contribution threshold are selected as target environmental features. The feature value offset state of the target environment feature is determined, and the structured interpretation information is generated according to the feature value offset state and contribution direction of each target environment feature in accordance with the preset interpretation template.

10. A device for identifying sandstorm pollution based on deep learning, characterized in that, The device includes: The data acquisition module is used to acquire multi-source environmental data of the area to be identified, including air quality monitoring data, meteorological data and satellite remote sensing data; The feature extraction and fusion module is used to perform spatiotemporal registration processing on the multi-source environmental data, encode the processed multi-source environmental data for features to obtain air quality features, meteorological features and satellite remote sensing features; and fuse the air quality features, meteorological features and satellite remote sensing features to obtain a unified environmental feature representation. The dust pollution identification module is used to standardize the unified environmental feature representation and then input it into a pre-trained deep learning identification model to output the dust pollution identification result. The interpretability analysis module is used to analyze the output of the deep learning recognition model using interpretability analysis methods, obtain the contribution of each environmental feature to the dust pollution recognition result, and generate structured interpretation information based on the contribution.