A method for extracting pure microwave radiation anomalies on the surface before earthquakes based on deep learning

Abstract

The present invention discloses a method for extracting pure surface microwave radiation anomalies before an earthquake based on deep learning. The mapping relationship between multivariate parameters in data such as surface temperature, surface vegetation index, surface soil moisture, rainfall, and total cloud volume in the historical earthquake-free period and surface microwave radiation is established through deep learning, and then the actual value of microwave radiation under the influence of non-seismic parameters is calculated using the above mapping relationship on the target day before the earthquake, and finally the pure seismic microwave radiation without non-seismic influence is obtained by difference. The present invention fully considers the potential impact of seismic activity on surface temperature, establishes a mapping relationship between multivariate input remote sensing parameters that are not easily affected by earthquakes and surface temperature, obtains the non-seismic surface temperature of the target day, and then obtains the surface non-seismic microwave radiation and the pure anomaly value caused by the earthquake based on this; the multivariate non-seismic parameter information of the target day is fully considered, rather than only considering the surface microwave radiation of the target day; the rationality of the prediction is guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of satellite microwave remote sensing technology, and in particular to a method for extracting pure microwave radiation anomalies on the ground before an earthquake based on deep learning. Background Art

[0002] Satellite microwave radiation imagers have the characteristics of multi-band multi-polarization, high sensitivity, high time resolution, and little influence by clouds and rain, and can detect changes in radiation information on the earth's surface in the microwave frequency band. The microwave radiation information received by the satellite contains not only the dynamic change information of the temperature and microwave emissivity of the surface material, but also can reflect the medium change information from deep on the surface to a certain extent due to its penetrating properties, so it is often used in the study of geological disasters and surface processes.

[0003] The surface microwave radiation received by satellite sensors is affected by both the target radiation source and the transmission path. On the one hand, the surface temperature and emissivity are affected by factors such as vegetation cover, soil moisture, and ambient temperature; on the other hand, the surface microwave radiation is also affected by factors such as thick clouds and precipitable water during transmission to satellite sensors. In the special context of the enhanced surface microwave radiation before an earthquake due to seismic activity, the above factors will affect the accuracy of the identification and extraction of earthquake additional microwave radiation. Therefore, how to suppress or quantitatively remove the influence of the above non-seismic factors has always been the top priority of the study of seismic microwave radiation anomalies.

[0004] The currently commonly used method for extracting earthquake microwave radiation anomalies is to use a large amount of historical data to establish an ideal background field on a target day before an earthquake. This background field is assumed to contain no earthquake information, and then it is differentiated from the actual microwave radiation observation value on the target day before the earthquake to obtain the so-called earthquake additional radiation anomaly. For example, the invention patent with patent number ZL202110403474.0 provides a method for constructing a pure microwave radiation background field on the surface by setting up a method for removing outliers in the background field. The invention patent with patent number ZL202111327328.0 provides a method for constructing a monthly average background field using historical microwave radiation data of the same period to extract earthquake microwave radiation anomalies. The above patents all attempt to use historical data to construct a reliable background field to obtain additional earthquake impacts. However, many current solutions, including the above-mentioned patent, ignore the fact that in the process of obtaining anomalies through differential algorithms, non-seismic interference information exists not only in the background field (subtrahend), but also in the actual observation value of the target day before the earthquake (minuend). This makes the obtained seismic anomaly (result) still disturbed by non-seismic information and therefore cannot be identified as a pure surface microwave radiation anomaly value. Summary of the invention

[0005] The present invention aims to overcome the shortcomings of the prior art and provides a method for extracting pure surface microwave radiation anomalies before an earthquake based on deep learning. Through deep learning, a mapping relationship between multivariate parameters in data such as surface temperature, surface vegetation index, surface soil moisture, rainfall, total cloud volume and surface microwave radiation in historical earthquake-free periods is established. Then, the above mapping relationship is used to calculate the actual value of microwave radiation under the influence of non-seismic parameters on the target day before the earthquake. Finally, pure seismic microwave radiation that does not include non-seismic influences is obtained by difference.

[0006] To achieve the above object, the technical solution of the present invention is as follows:

[0007] A method for extracting pure microwave radiation anomalies on the surface before an earthquake based on deep learning, comprising the following steps:

[0008] Step S1: acquiring long-term satellite microwave radiation data and quasi-synchronous multi-input remote sensing parameter data in the target area according to the time and location of the target earthquake, and performing data preprocessing; the multi-input remote sensing parameter data includes surface temperature, surface humidity, surface vegetation index, rainfall and total cloud volume;

[0009] Step S2: Build a deep learning model , establish the mapping relationship between satellite microwave radiation and multivariate input remote sensing parameters in historical non-seismic periods;

[0010] Step S3: Build a deep learning model , establish the mapping relationship between surface temperature and surface humidity, surface vegetation index, rainfall and total cloud volume in historical non-seismic periods;

[0011] Step S4: Obtain the surface humidity, surface vegetation index, rainfall and total cloud volume data of the target day within the earthquake cycle of the target area, and use the deep learning model in step S3 to calculate the surface humidity, surface vegetation index, rainfall and total cloud volume data of the target day within the earthquake cycle of the target area. The non-seismic surface temperature of the target day is predicted;

[0012] Step S5: Based on the non-seismic surface temperature of the target day within the earthquake cycle obtained in step S4, and the surface humidity, surface vegetation index, rainfall, and total cloud volume data of the target day within the earthquake cycle, the deep learning model in step S2 is used to calculate the surface temperature of the target day within the earthquake cycle. Predict the non-seismic surface microwave radiation on the target day;

[0013] Step S6: Obtain the actual value of the surface microwave radiation on the target day within the earthquake cycle of the target area, and eliminate the non-seismic surface microwave radiation on the target day within the earthquake cycle obtained in step S5, so as to obtain a pure microwave radiation anomaly on the target day before the earthquake without the influence of non-seismic factors.

[0014] One beneficial effect of the present invention is that, by fully considering the potential impact of seismic activity on the surface temperature, a mapping relationship between multivariate input remote sensing parameters that are not easily affected by earthquakes and the surface temperature is established, and then the non-seismic surface temperature is obtained in step S4 and applied to step S5, thereby ensuring the rationality of the prediction.

[0015] Another beneficial effect of the present invention is that since the data used to predict the non-seismic surface microwave radiation on the target day in step S5 comes from the quasi-synchronous data of the target day rather than historical data, the non-seismic surface microwave radiation on the target day also includes the influence of the multivariate input remote sensing parameters on the surface microwave radiation, and the non-seismic parameter information on the target day is fully considered, so that the microwave radiation anomaly obtained in step S6 only includes the additional influence of the earthquake.

[0016] Furthermore, the specific method of data preprocessing in step S1 is:

[0017] Step S1.1: unify the spatial resolution of all data and crop them uniformly within the target area;

[0018] Step S1.2: Considering the persistence of the influence of soil moisture and rainfall on surface microwave radiation, the soil moisture and rainfall data are calculated as the mean of all observations during the period of time T and a few hours before T; T is the time when the satellite collects microwave radiation data on the target day in the target area;

[0019] Step S1.3: According to the target earthquake occurrence time, 60 days before the target earthquake occurrence date to 30 days after the earthquake occurrence date is set as the earthquake period, and the data during this period is set as the earthquake period data;

[0020] Step S1.4: The period before 60 days before the target earthquake is divided into the historical non-seismic period; according to the historical earthquake catalog and the selected target area, the 30 days before and after the date of the historical earthquake with a magnitude greater than 5.0 in the target area are set as the historical earthquake period, and all historical data in the historical earthquake period are eliminated to obtain the historical non-seismic period data.

[0021] Furthermore, the step S2 comprises the following steps:

[0022] Step S2.1: Build LSTM-Transformer model ; The LSTM module consists of several memory units, each of which controls the flow and update of information through the input gate, forget gate and output gate; Among them:

[0023] The forget gate determines how much information in the memory cell state at the previous moment needs to be forgotten, and its output value Between 0 and 1, Sure, is the activation function, is the weight matrix of the forget gate, is the hidden state at the previous moment, is the input at the current moment, is the bias vector of the forget gate;

[0024] The input gate determines how much new information needs to be written into the memory cell state, and its output value Between 0 and 1, Sure, is the weight matrix of the input gate, is the bias vector of the input gate;

[0025] Candidate memory cell status It is a new information vector generated by the tanh activation function, which is used to be partially written into the memory unit under the control of the input gate , is the weight matrix of the candidate memory cell state, is the bias vector of the candidate memory cell state; update the memory cell state The memory cell state is updated through the forget gate and input gate. , is element-wise multiplication;

[0026] Output Gate Determines how much information of the current state of the memory unit needs to be output to the hidden state, , is the weight matrix of the output gate, is the bias vector of the output gate; the hidden state at the current moment Determined by the memory cell state and the output gate, ;

[0027] The Transformer module consists of several encoders Encoder and several decoders Decoder , the output value of the LSTM module is transformed and input into the Transformer module; according to the input data Extract its features ; Based on the extracted features And its time series, get the output features , and calculate the module output value , is the weight matrix of the decoder, is the bias vector of the decoder;

[0028] Step S2.2: For pixel i in the target area, the multivariate input remote sensing parameters of all time series in the historical non-seismic period are used as the model The input data of satellite microwave radiation is used as the model The output data is normalized; the input data is first converted into a three-dimensional tensor suitable for the LSTM module, and the three dimensions correspond to the number of samples in a batch, the number of time steps, and the feature dimension of each time step; the combination of the above input data and output data is divided into a training set, a test set, and a validation set, and the model established in step S2.1 is tested using the training set. Train and evaluate the model using the validation set Performance during training, using the test set to evaluate the model The ability to predict surface microwave radiation; when the accuracy reaches the required level, the mapping relationship between satellite microwave radiation and multivariate input remote sensing parameters in historical non-seismic periods is obtained :

[0029]

[0030] In the above formula and are the historical non-seismic periods in pixel i All satellite microwave radiation values, surface temperature values, surface humidity values, vegetation index values, rainfall values ​​and total cloud cover values.

[0031] Furthermore, the step S3 specifically includes the following steps:

[0032] Step S3.1: Build the LSTM-Transformer in exactly the same way as step S2.1 ;

[0033] Step S3.2: For pixel i in the target area, the surface humidity, surface vegetation index, rainfall and total cloud volume data of all time series in the historical non-seismic period are used as the input data of the overall model, and the surface temperature data are used as the output data of the overall model; the above input data and output data are combined and divided into a training set, a test set and a validation set.

[0034] Use the training set to train the model established in step S3.1 Train and evaluate the model using the validation set The performance during the training process is evaluated using the test set. The prediction capability of surface microwave radiation is based on the prediction capability of surface microwave radiation. When the accuracy reaches the requirement, the mapping relationship between surface temperature and surface humidity, surface vegetation index, rainfall and total cloud volume data in the historical non-seismic period is obtained. :

[0035] .

[0036] The beneficial effects of the above steps are: the LSTM-Transformer model combines the advantages of the LSTM model and the Transformer model. The LSTM model can effectively capture the long-term and short-term dependencies in time series data, and is more stable when training deep networks through the gate mechanism; the Transformer model can capture the complex feature relationships in time series data, and its self-attention mechanism enables it to capture dependencies at any position in the sequence. Since there is no need to process the sequence step by step, the Transformer has high parallelism during training and inference; the combination of the two is used to improve the performance of sequence data tasks, which can improve prediction accuracy and speed up data processing.

[0037] Furthermore, the step S4 specifically includes the following steps:

[0038] Step S4.1: For pixel i in the target area, the surface humidity, surface vegetation index, rainfall and total cloud volume values ​​of the target day t within the earthquake cycle are normalized;

[0039] Step S4.2: Input the normalized data in step S4.1 into the deep learning model The model prediction value is obtained from the image, and the non-seismic surface temperature in pixel i on the target day is obtained after denormalization;

[0040] Furthermore, the step S5 specifically includes the following steps:

[0041] Step S5.1: For pixel i in the target area, normalize the surface humidity, surface vegetation index, rainfall, and total cloud volume values ​​on the target day t within the earthquake cycle with the non-seismic surface temperature obtained in S4.2;

[0042] Step S5.2: Input the normalized data in step S5.1 into the deep learning model The model prediction value is obtained from the image, and the non-seismic surface microwave radiation in pixel i on the target day t is obtained after denormalization;

[0043] Furthermore, the method for eliminating non-seismic surface microwave radiation in step S6 is specifically as follows:

[0044] Then the pure surface microwave radiation anomaly ,otherwise Directly set to null value; is the actual value of surface microwave radiation on the target day t during the earthquake cycle in the target area, It is the predicted non-seismic surface microwave radiation of the target day within the earthquake cycle.

[0045] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the embodiments of the present invention, but do not constitute a limitation on the embodiments of the present invention. In the accompanying drawings:

[0047] Figure 1 This is a flow chart of a method for extracting pure microwave radiation anomalies on the surface before an earthquake based on deep learning according to the present invention;

[0048] Figure 2 is the surface soil moisture in one embodiment of the present invention;

[0049] Figure 3 This is a pure surface microwave radiation anomaly in one embodiment of the present invention. DETAILED DESCRIPTION

[0050] The embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0051] like Figure 1 As shown, in one embodiment of the present invention, a method for extracting pure surface microwave radiation anomalies before an earthquake based on deep learning is provided, which automatically obtains pure surface microwave radiation anomalies before an earthquake by using long-term series surface microwave radiation data observed by satellites and multi-parameter remote sensing data; the method comprises the following steps:

[0052] S1. According to the time and location of the target earthquake, long-term satellite microwave radiation data and quasi-synchronous multi-input remote sensing parameter data in the target area are obtained, and data preprocessing is performed; wherein the multi-input remote sensing parameter data include surface temperature, surface humidity, surface vegetation index, rainfall and total cloud volume.

[0053] The earthquake case selected in the specific embodiment of the present invention is a Mw7.8 earthquake that occurred in Nepal on April 25, 2015. Then:

[0054] 1) Considering that the topography of the earthquake location in this embodiment varies greatly, the target area can be set to the epicenter location (28.1 degrees north latitude, 84.7 degrees east longitude) 5 degrees east and west, 4 degrees south, and 2 degrees north, and then take the integer to form a geographical range of 10 degrees by 6 degrees, such as Figure 2 Satellite microwave radiation data and multivariate input remote sensing parameter data were clipped according to this target area, and the target date was set to April 24, 2015.

[0055] 2) The selected satellite microwave radiation remote sensing image data can be the daily microwave radiation data provided by the AMSR-2 satellite sensor, with the data starting date being January 1, 2013 and ending date being the 30th day after the target earthquake occurred, i.e., May 24, 2015; therefore, the historical non-seismic period in this embodiment is The period is from January 1, 2013 to February 25, 2015, and the earthquake period is from February 26, 2015 to May 24, 2015; the surface temperature data, soil moisture data and rainfall data can be the hourly corresponding parameter data in the ERA-5 Land dataset provided by the Copernicus Center of the European Space Agency; the surface vegetation index can be the daily surface NDVI dataset obtained by MODIS satellite observation; the total cloud cover data can be the hourly total cloud cover data obtained by Fengyun meteorological satellite observation.

[0056] 3) The spatial resolution of all data is unified to 0.25 degrees, which corresponds to the minimum resolution of the daily microwave radiation data provided by the AMSR-2 satellite sensor.

[0057] 4) Taking the time point of the daily microwave radiation data provided by the AMSR-2 satellite sensor as a reference, the soil moisture data and rainfall data in all ERA-5 Land datasets during the previous twelve hours are averaged to obtain the daily soil moisture data and rainfall data that match the daily microwave radiation data provided by the AMSR-2 satellite sensor.

[0058] 5) According to the global earthquake catalog provided by USGS, earthquakes with a magnitude greater than 5.0 are searched in the target area of ​​this embodiment, and the time of occurrence of the earthquake is extracted. The historical earthquake period is set from 30 days before the historical earthquake date to 30 days after the earthquake date, and all historical data in this period are eliminated to obtain historical non-seismic period data.

[0059] S2. Build a deep learning model , establish a mapping relationship between satellite microwave radiation and multi-input remote sensing parameters in historical non-seismic periods.

[0060] 1) For pixel i in the target area, build an LSTM-Transformer model ; Since this embodiment uses a variety of data and the time series is long, the model performance requirements are high. In this embodiment, the LSTM module consists of 256 memory cells, each of which controls the flow and update of information through an input gate, a forget gate, and an output gate. The output value of the forget gate Between 0 and 1, Sure, is the activation function, is the weight matrix of the forget gate, is the hidden state at the previous moment, is the input at the current moment, is the bias vector of the forget gate. The output value of the input gate Between 0 and 1, Sure, is the weight matrix of the input gate, is the bias vector of the input gate. Candidate memory cell state It is a new information vector generated by the tanh activation function. This vector will be partially written into the memory unit under the control of the input gate. , is the weight matrix of the candidate memory cell state, is the bias vector of the candidate memory cell state; update the memory cell state The memory cell state is updated through the forget gate and input gate. . Output Gate Determines how much information of the current state of the memory unit needs to be output to the hidden state, , is the weight matrix of the output gate, is the bias vector of the output gate. The hidden state at the current moment Determined by the memory cell state and the output gate, .in, Multiply the elements.

[0061] 2) In this embodiment, the Transformer module sets a 6-layer encoder Encoder and several decoders Decoder , the output value of the LSTM module is transformed and input into the Transformer module; according to the input data Extract (input data )feature ; Based on the extracted features And its time series, get the output features , and calculate the module output value , is the weight matrix of the decoder, is the bias vector of the decoder.

[0062] 3) For the target pixel i in the target area, all multivariate input remote sensing parameter numbers excluding historical earthquake information from January 1, 2013 to February 25, 2015 are used as the input data of the overall model, and the corresponding AMSR-2 satellite microwave radiation time series data are used as the output data of the overall model, and are normalized; the input data is first converted into a three-dimensional tensor suitable for the LSTM module, and the three dimensions in this embodiment are 128, 128, and 128 respectively; the above data combination is divided into a training set, a test set, and a validation set according to a ratio of 0.7:0.2:0.1, and the established model is validated using the training set. Train and evaluate the model using the validation set Performance during training, using the test set to evaluate the model For the prediction ability of surface microwave radiation, the test set here does not participate in model training; the mean square error is used as the model The loss function is to evaluate the model using the mean absolute percentage error accuracy.

[0063] 4) After the accuracy reaches the requirement, the historical non-seismic period is obtained Mapping relationship between satellite microwave radiation and multivariate input remote sensing parameters :

[0064]

[0065] In the above formula and are the historical non-seismic periods in pixel i All AMSR-2 satellite microwave radiation values, ERA-5 surface temperature values, ERA-5 surface humidity values, MODIS vegetation index values, ERA-5 rainfall values ​​and Fengyun satellite total cloud cover values.

[0066] S3. Build a deep learning model , establish a mapping relationship between surface temperature and surface humidity, surface vegetation index, rainfall, and total cloud volume during historical non-seismic periods.

[0067] 1) For pixel i in the target area, build the LSTM-Transformer model in the same way as in step S2 .

[0068] 2) For pixel i in the target area, the surface humidity, surface vegetation index, rainfall and total cloud data of all time series from January 1, 2013 to February 25, 2015 are used as the input data of the overall model, and the corresponding surface temperature data are used as the output data of the overall model; the above data combination is divided into training set, test set and validation set in the ratio of 0.7:0.2:0.1 respectively, and the model established in S3.1 is tested using the training set. Train and evaluate the model using the validation set Performance during training, using the test set to evaluate the model For the prediction ability of surface microwave radiation, the test set here does not participate in model training; the mean square error is used as the model The loss function is to evaluate the model using the mean absolute percentage error When the accuracy reaches the required level, the historical non-seismic period is obtained. Mapping relationship between surface temperature and surface humidity, surface vegetation index, rainfall and total cloud data :

[0069] .

[0070] S4. Obtain the ERA-5 surface humidity value, MODIS surface vegetation index value, ERA-5 rainfall value, and Fengyun satellite cloud volume value of pixel i in the target area on April 24, 2015, normalize the above data, and input them into the established deep learning model. The predicted output is denormalized to obtain the non-seismic surface temperature on April 24, 2015, which is consistent with the scale of the corresponding real data.

[0071] S5: The non-seismic surface temperature of pixel i on April 24, 2015 predicted in S4, as well as the ERA-5 surface humidity value, MODIS surface vegetation index value, ERA-5 rainfall value, and Fengyun satellite cloud volume value on the same day are used as input values, and after normalization, they are input into the established deep learning model. The predicted output is denormalized to obtain the non-seismic surface microwave radiation of pixel i on April 24, 2015, which is consistent with the corresponding real data scale. In this embodiment, the input data is April 24, 2015, so the predicted value also represents the surface radiation state of pixel i on April 24, 2015, and includes the influence of surface temperature, surface humidity, surface vegetation index, rainfall, and total cloud volume on surface microwave radiation; there was a lot of rainfall in the target area in April 2015, which interfered with the study of earthquake anomalies. The surface soil moisture distribution is as follows: Figure 2 shown.

[0072] S6. Repeat all steps from S2 to S5 to obtain the non-seismic surface microwave radiation of all pixels in the target area on April 24, 2015 ; Obtain the actual value of surface microwave radiation in the target area on April 24, 2015 ,according to Obtain the pure seismic microwave radiation anomaly values ​​of all pixels in the target area on April 24, 2015, such as Figure 3 As shown in the figure, the results no longer include the interference of soil moisture, a major non-seismic factor.

[0073] In a specific embodiment of the present invention, the target pixel i used can also be replaced by a pixel combination i, and the number of pixels in the combination and the basis for their grouping can be varied; the selected satellite microwave radiation data and the multivariate input remote sensing parameters are only a preferred combination, and the selected number of hidden units, encoding layers, and decoding layers of the deep learning model are only a preferred combination provided based on the above data combination. Any data set combination and parameter combination that meets the requirements can use this technical solution and should also be within the protection scope of the present invention.

[0074] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for extracting pure microwave radiation anomalies on the surface before an earthquake based on deep learning, characterized in that: The following steps are involved: Step S1: acquiring long-term satellite microwave radiation data and quasi-synchronous multi-input remote sensing parameter data in the target area according to the time and location of the target earthquake, and performing data preprocessing; the multi-input remote sensing parameter data includes surface temperature, surface humidity, surface vegetation index, rainfall and total cloud volume; Step S2: Build a deep learning model , establish the mapping relationship between satellite microwave radiation and multivariate input remote sensing parameters in historical non-seismic periods; Step S3: Build a deep learning model , establish the mapping relationship between surface temperature and surface humidity, surface vegetation index, rainfall and total cloud volume in historical non-seismic periods; Step S4: Obtain the surface humidity, surface vegetation index, rainfall and total cloud volume data of the target day within the earthquake cycle of the target area, and use the deep learning model in step S3 to calculate the surface humidity, surface vegetation index, rainfall and total cloud volume data of the target day within the earthquake cycle of the target area. The non-seismic surface temperature of the target day is predicted; Step S5: Based on the non-seismic surface temperature of the target day within the earthquake cycle obtained in step S4, as well as the surface humidity, surface vegetation index, rainfall and total cloud volume data of the target day within the earthquake cycle, the deep learning model in step S2 Predict the non-seismic surface microwave radiation on the target day; Step S6: Obtain the actual value of the surface microwave radiation on the target day within the earthquake cycle of the target area, and eliminate the non-seismic surface microwave radiation on the target day within the earthquake cycle obtained in step S5, so as to obtain a pure microwave radiation anomaly on the target day before the earthquake without the influence of non-seismic factors.

2. The method according to claim 1, characterized in that The specific method of data preprocessing in step S1 is: Step S1.1: unify the spatial resolution of all data and crop them uniformly to the target area; Step S1.2: Considering the persistence of the influence of soil moisture and rainfall on surface microwave radiation, calculate the mean of all observed values ​​of soil moisture data and rainfall data at time T and a period of time before T; where T is the time when the satellite collects microwave radiation data on the target day in the target area; Step S1.3: According to the target earthquake occurrence time, 60 days before the target earthquake occurrence date to 30 days after the earthquake occurrence date is set as the earthquake period, and the data during this period is set as the earthquake period data; Step S1.4: Divide the period before 60 days before the target earthquake into a historical non-seismic period; according to the historical earthquake catalog and the selected target area, set the period from 30 days before to 30 days after the date of a historical earthquake with a magnitude greater than 5.0 in the target area as a historical earthquake period, and remove all historical data in the historical earthquake period to obtain historical non-seismic period data.

3. The method according to claim 1, characterized in that The step S2 specifically includes the following steps: Step S2.1: Build LSTM-Transformer model ; The LSTM module consists of several memory units, each of which controls the flow and update of information through the input gate, forget gate and output gate; Among them: The forget gate determines how much information in the memory cell state at the previous moment needs to be forgotten, and its output value Between 0 and 1, Sure, is the activation function, is the weight matrix of the forget gate, is the hidden state at the previous moment, is the input at the current moment, is the bias vector of the forget gate; The input gate determines how much new information needs to be written into the memory cell state, and its output value Between 0 and 1, Sure, is the weight matrix of the input gate, is the bias vector of the input gate; Candidate memory cell status is through The activation function generates a new information vector that is used to be partially written into the memory cell under the control of the input gate. , is the weight matrix of the candidate memory cell state, is the bias vector of the candidate memory cell state; update the memory cell state The memory cell state is updated through the forget gate and input gate. , is element-wise multiplication; Output Gate Determines how much information of the current state of the memory unit needs to be output to the hidden state, , is the weight matrix of the output gate, is the bias vector of the output gate; the hidden state at the current moment Determined by the memory cell state and the output gate, ; The Transformer module consists of several encoders Encoder and several decoders Decoder , the output value of the LSTM module is transformed and input into the Transformer module; according to the input data Extract its features ; Based on the extracted features And its time series, get the output features , and calculate the module output value , is the weight matrix of the decoder, is the bias vector of the decoder; Step S2.2: For pixel i in the target area, the multivariate input remote sensing parameters of all time series in the historical non-seismic period are used as the model The input data of satellite microwave radiation is used as the model The output data is normalized; the input data is first converted into a three-dimensional tensor suitable for the LSTM module, and the three dimensions correspond to the number of samples in a batch, the number of time steps, and the feature dimension of each time step; the combination of the above input data and output data is divided into a training set, a test set, and a validation set, and the model established in step S2.1 is tested using the training set. Train and evaluate the model using the validation set Performance during training, using the test set to evaluate the model The ability to predict surface microwave radiation; when the accuracy reaches the required level, the mapping relationship between satellite microwave radiation and multivariate input remote sensing parameters in historical non-seismic periods is obtained : In the above formula and are the historical non-seismic periods in pixel i All satellite microwave radiation values, surface temperature values, surface humidity values, vegetation index values, rainfall values ​​and total cloud cover values.

4. The method according to claim 3, characterized in that The step S3 specifically comprises the following steps: Step S3.1: Build the LSTM-Transformer model in the same way as step S2.1 ; Step S3.2: For pixel i in the target area, the surface humidity, surface vegetation index, rainfall and total cloud volume data of all time series in the historical non-seismic period are used as the input data of the overall model, and the surface temperature data are used as the output data of the overall model; the above input data and output data are combined into a training set, a test set and a validation set, and the model established in step S3.1 is validated using the training set. Train and evaluate the model using the validation set The performance during the training process is evaluated using the test set. The prediction capability of surface microwave radiation is based on the prediction capability of surface microwave radiation. When the accuracy reaches the requirement, the mapping relationship between surface temperature and surface humidity, surface vegetation index, rainfall and total cloud volume data in the historical non-seismic period is obtained. : 。 5. The method according to claim 1, characterized in that The step S4 specifically comprises the following steps: Step S4.1: For pixel i in the target area, the surface humidity, surface vegetation index, rainfall and total cloud volume values ​​of the target day t within the earthquake cycle are normalized; Step S4.2: Input the normalized data in step S4.1 into the deep learning model The model prediction value is obtained from the θ, and the non-seismic surface temperature in pixel i on the target day is obtained after denormalization.

6. The method according to claim 5, characterized in that The step S5 specifically comprises the following steps: Step S5.1: For pixel i in the target area, normalize the surface humidity, surface vegetation index, rainfall, and total cloud volume values ​​on the target day t within the earthquake cycle with the non-seismic surface temperature obtained in step S4.2; Step S5.2: Input the normalized data in step S5.1 into the deep learning model The model prediction value is obtained from the θ, and the non-seismic surface microwave radiation in pixel i on the target day is obtained after denormalization.

7. The method according to claim 1, characterized in that The method for eliminating non-seismic surface microwave radiation in step S6 is: If, in the preprocessed data in step S1, for pixel i, the values ​​of the multivariate input remote sensing parameters are all valid values, then the pure surface microwave radiation anomaly value ,otherwise Directly set to null value; is the actual value of surface microwave radiation on the target day t during the earthquake cycle in the target area, It is the predicted non-seismic surface microwave radiation of the target day within the earthquake cycle.

Images