CN-121480877-B - Multi-mode weather pollutant prediction method and system integrating optical aerosol remote sensing image and ground weather time sequence data

CN121480877BCN 121480877 BCN121480877 BCN 121480877BCN-121480877-B

Abstract

The invention belongs to the technical field of climate data analysis, and discloses a multi-mode weather pollutant prediction method and system for fusing optical aerosol remote sensing images and ground weather time sequence data, wherein the time sequence characteristics representing the atmospheric aerosol change of the area are obtained through aerosol optical thickness time sequence images provided by remote sensing satellites, and the time sequence data of ground weather observation are obtained; the method comprises the steps of decomposing through STL, respectively obtaining deep representation of each aerosol mode and weather mode through a transducer encoder, constructing unified multi-mode representation to obtain comprehensive characteristic representation of each weather variable, optimizing an adjacent matrix through maximizing a reward function to establish a weather causal network, and predicting the weather variable by adopting a long-period memory network. The invention has obvious advantages in causal modeling and prediction tasks, and has effectiveness and practicability in the tasks of atmospheric causal discovery and target pollutant variable prediction.

Inventors

LIU JINDUO
LU YILIN
JI JUNZHONG

Assignees

北京工业大学

Dates

Publication Date: 20260508
Application Date: 20251124

Claims (6)

1. A multi-mode weather pollutant prediction method integrating an optical aerosol remote sensing image and ground weather time sequence data is characterized by comprising the following steps: acquiring an aerosol optical thickness time sequence image provided by a remote sensing satellite, obtaining a time sequence feature representing regional atmospheric aerosol change through ROI analysis, and acquiring time sequence data of ground meteorological observation; Using an STL decomposition algorithm to separate out time sequence features of atmospheric aerosol change and multi-scale features mixed in time sequence data of ground meteorological observation, and respectively obtaining respective deep layer representations of an aerosol mode and a meteorological mode through a transducer encoder; carrying out alignment fusion on the aerosol mode and the weather mode, and constructing a unified multi-mode representation to obtain an aerosol mode updating representation fused with weather information, namely, a comprehensive characteristic representation of each weather variable; Generating an adjacency matrix of the causal directed graph based on the comprehensive characteristic representation of each climate variable, and optimizing the adjacency matrix by maximizing a reward function so as to establish a climate causal network, wherein the method for constructing the adjacency matrix comprises the following steps: by calculating the probability that causal edges exist between each pair of variables Thereby giving candidate DAG structures under the current strategy for any two variables And , The calculation formula of (2) is as follows: ; Wherein, the 、 Representing variables respectively And Is used for the feature vector of (a), 、 As a matrix of weights that can be learned, As a parameter vector that can be learned, Representing a Sigmoid activation function; Then, defining the reward function of the environment as the AIC score of the candidate graph and the sparse penalty thereof: setting candidate causal graph as Is common to Variable node, the Personal node Is as the father set of With the local resolvable idea of log likelihood, AIC is written as: ; Wherein, the To estimate at maximum likelihood The local log-likelihood of the lower one, Is the first The number of parameters of the local model; to encourage sparse structure, introduce Norms penalty, then reward function The method comprises the following steps: ; Wherein the method comprises the steps of In order to be a contiguous matrix, The number of directed edges is indicated, Controlling the sparsity; Based on the climate causal network, a long-short-period memory network is adopted to predict climate variables, so that future pollutant variables are predicted.
2. The method according to claim 1, characterized in that: the method for obtaining the time series characteristic representing the regional atmospheric aerosol change through the ROI analysis comprises the following steps: at discrete moments Acquiring images within a region of interest (ROI) Wherein Is a set of pixels in the ROI, For each time, the number of pixels is equal to Calculating the area mean and the area variance of the area: ; ; from this, a time series characteristic representing the change of the atmospheric aerosol in the region can be obtained 。
3. The method according to claim 1, characterized in that: the STL decomposition algorithm disassembles the observed sequence into three parts of trend, season and residual error: ; ; ; ; ; Wherein, the In order to observe the sequence of the sequences, Indicating a long-term trend of smoothness, For a seasonal term that is smoothed across a seasonal window, As a residual term, Representation of sequences of pairs The trend portion is calculated using a moving average method, ) Is a seasonal fraction obtained by a periodic extraction method.
4. The method according to claim 1, characterized in that: the method for constructing the unified multi-modal representation comprises the following steps: feature vectors of aerosol modalities as queries With characteristic matrix of meteorological mode as key Sum value By calculation of And (3) with Matching degree pair of (2) Weighted summation to obtain an aerosol modal update representation fused with weather information The method comprises the following steps: ; Wherein the method comprises the steps of For the hidden spatial dimension of the attention mechanism for scaling, To transpose the symbols.
5. The method according to claim 1, characterized in that: the method for predicting the climate variable by adopting the long-short-period memory network comprises the following steps: For each target variable node to be predicted in the climate cause and effect network, an input sequence is constructed according to the historical time sequence data of all father node variables, namely at time At the moment, the target variable All parent node variables of (a) Is combined into an input vector: ; The LSTM network recursively updates the hidden state by using the input sequence, thereby capturing the dynamic influence relation of the father node variable to the target variable, and the state update of the LSTM is expressed as follows: ; Wherein, the Is that The hidden state of the moment of time, A state transfer function of the LSTM unit; Then, the prediction output is obtained by linear mapping of the hidden state: ; Wherein the method comprises the steps of And The weight matrix and the bias vector of the output layer respectively, To target variable At the moment of time Is a predicted value of (a).
6. A multi-mode weather pollutant prediction system for fusing optical aerosol remote sensing images with ground weather time sequence data, which is characterized by comprising an original sequence acquisition unit, a characteristic representation unit, a multi-mode fusion unit, a causal network unit and a prediction unit, wherein the method is applied to any one of claims 1-5: the original sequence acquisition unit is used for acquiring an aerosol optical thickness time sequence image provided by a remote sensing satellite, obtaining a time sequence characteristic representing the atmospheric aerosol change of a region through ROI analysis, and acquiring time sequence data of ground meteorological observation; the characteristic representation unit is used for separating out a time sequence characteristic of atmospheric aerosol change and a multi-scale characteristic mixed in time sequence data of ground meteorological observation by using an STL decomposition algorithm, and respectively obtaining respective deep representations of an aerosol mode and a meteorological mode through a transducer encoder; The multi-modal fusion unit is used for carrying out alignment fusion on the aerosol modes and the weather modes, constructing uniform multi-modal representation and obtaining aerosol mode updating representation fused with weather information, namely comprehensive characteristic representation of each weather variable; The causal network unit is used for generating an adjacent matrix of the causal directed graph based on the comprehensive characteristic representation of each climate variable, and optimizing the adjacent matrix by maximizing a reward function so as to establish a climate causal network; The prediction unit is used for predicting the climate variable by adopting a long-short-period memory network based on the climate causal network, so that the future pollutant variable is predicted.

Description

Multi-mode weather pollutant prediction method and system integrating optical aerosol remote sensing image and ground weather time sequence data Technical Field The invention belongs to the technical fields of climate data analysis, remote sensing and artificial intelligence, and particularly relates to a multi-mode weather pollutant prediction method and system for fusing optical aerosol remote sensing images and ground weather time sequence data. Background Climate causal discovery aims at identifying causal relationships between climate elements from observed data and reconstructing a climate causal network in the form of a directed acyclic graph (DIRECTED ACYCLIC GRAPH, DAG). The process has important significance for analyzing internal mechanisms of a climate system and supporting climate trend prediction and intervention decision. For example, by constructing a climate causal network, causal influence relationships of different climate variables can be visually presented and used for simulating influence of a certain factor change on the whole climate system, so that potential influence of the climate change is evaluated, corresponding countermeasures are formulated, and basis is provided for climate variable prediction. In the application of air pollution early warning, prevention and control and the like, the accurate prediction of the pollutant concentration is also of great significance to the establishment of prevention and control measures. The existing climate cause and effect discovery method is mainly divided into two types, namely a traditional statistical machine learning method, and the method comprises a Grangel cause and effect test method, a nonlinear state space model, a cause and effect network learning algorithm, a structural equation model and the like. The methods infer causal relationship based on statistical test and a mathematical model, have the advantages of simple model and easy interpretation of results, but have a plurality of limitations. For example, they are sensitive to noise and outliers and often require large amounts of data, which perform poorly when dealing with high-dimensional complex climate data. The other is a deep learning method, which is excellent in extracting a complex pattern of large-scale data and nonlinear relation in recent years. Some studies began to use deep learning for the construction of climate causal networks and have made preliminary progress. The depth model can automatically learn layering characteristics and capture intricate and complex dependency relations in a climate system, and is a powerful tool for constructing and understanding a climate causal network gradually. However, these studies focused mainly on causal structure identification, and lack an integration scheme to apply the causal structure found to climate variable prediction. Meanwhile, the climate system often relates to multi-source heterogeneous data, namely, on one hand, satellite and foundation remote sensing provides image data of large-range and time-varying atmospheric components, and on the other hand, ground meteorological observation provides time-sequence meteorological pollution data such as local PM2.5 and NO 2、SO2. These data modalities contain complementary information, for example, aerosol remote sensing images reflect the spatial distribution of the atmospheric environment, and meteorological time sequences record the evolution of local climate elements. If relying on only a single data source, the full view of the climate phenomenon may not be adequately captured, resulting in incomplete or inaccurate causal relationship mining. Based on the above-mentioned current situation, there is an urgent need for a method that can effectively integrate multi-source climate observation data, extract cross-modal correlation features, efficiently and robustly discover causal relationships in a huge structural space, and perform climate variable prediction. Disclosure of Invention Aiming at the problems that the existing weather causal discovery method depends on single-mode data to cause information underutilization, causal identification inaccuracy, structural search difficulty, failure to apply causal structures to weather variable prediction and the like, the invention provides a multi-mode weather pollutant prediction method and system which are fused with an optical aerosol remote sensing image and ground weather time sequence data. The invention provides the following technical scheme: a multi-mode weather pollutant prediction method integrating optical aerosol remote sensing images and ground weather time sequence data comprises the following steps: acquiring an aerosol optical thickness time sequence image provided by a remote sensing satellite, obtaining a time sequence feature representing regional atmospheric aerosol change through ROI analysis, and acquiring time sequence data of ground meteorological observation; Using an STL decomposition algorithm to se