A Knowledge-Guided Deep Learning Model for Extracting Pollution-free Seismic Microwave Brightness Temperature Anomalies

There has long been research on the phenomenon of abnormal microwave radiation emitted from the Earth's surface before a major earthquake. However, the enhanced microwave radiation received by satellite sensors is affected by a combination of factors such as surface vegetation, soil moisture, land surface temperature, and atmospheric environment. So far, it has been difficult to remove non-seismic interference through quantitative physical modeling, leaving only the earthquake-related additional components for earthquake precursor analysis and short-term earthquake prediction. To tackle with this, we developed a knowledge-guided deep learning model that leverages a large amount of remote sensing observation data for training, incorporating prior knowledge of earthquake anomaly analysis. In the modelling process, a large amount of multi-source data, such as surface microwave brightness temperature (MBT), land surface temperature (LST), surface vegetation index, soil moisture index, atmospheric water vapor content, cloud cover, land cover type, digital elevation model (DEM), and geological type, were collected, and a regression model between multiple factors and surface MBT were firstly established through deep learning methods. In the same way, another regression model was developed between non-temperature parameters and LST by using historical records. During the seismic window of one month before and after the target earthquake, the LST was obtained by using non-temperature data through the second regression model, and then was substituted it into the first regression model to get the MBT value that does not include the additional effects of earthquakes. Eventually, we can obtain the additional MBT value due to seismic activity by calculating the difference with the actual observation, which represents the earthquake-related MBT anomaly. Since the deep learning-based modeling is based on long time series data and the output results of the model already include the contribution of multiple factors on the surface to the MBT, the differential results are mainly affected by the additional impacts of the earthquake, so they can be considered 'pollution-free'. In other words, there is no need to use additional auxiliary data to discriminate and separate the non-seismic disturbances. For a specific target area, such as the Tibetan Plateau, after establishing a model based on historical data using the aforementioned method, we can obtain real-time earthquake MBT variations as the input data is continuously updated. This can be used to analyze and identify potential earthquake precursors, and consequently, for short-term earthquake prediction.