The Neural Network Prediction Based On The Thermal Anomaly Information

Earthquake prediction is one of the main fields of earthquake science research.Due to the impenetrability of earth’s crust、the infrequent of earthquake、the uncertainty nonlinear and complexity of earthquake physical process,the earthquake prediction is an extremely difficult task.The earthquake prediction has been recognized as one of the ten big problems in the world.This article attempted to consider thermal anomaly information as a source of earthquake prediction,by building a neural network,made a predictive testing of the earthquake’s three elements,had achieved a good results.This paper based on the MODIS data which has synthesis of eight days with 1km resolution, used Robust Satellite Techniques method of time-domain and spatial-domain to extracting thermal anomaly information before the earthquake separately,and combined the fault information divided by the China’s three blocks,considered the time-space relationship between thermal anomaly information and the fault zone,thought carefully about the information of the neural network input neurons,using in and around China’s 100 earthquake cases whose magnitude are more than 5 and 70 random samples of aseismic for training and simulation by BP neural network and wavelet neural network.On this basis,simulated 20 earthquake cases and 15 random samples of aseismic to study the possibility of earthquake three elements’ s prediction.At last, compared the predicting outcomes for different neural networks of the same kind of anomaly extraction algorithm.at the same time,compared the predicting outcomes for different anomaly extraction algorithm of the same kind of neural networks.To study which thermal anomalies algorithm and neural network is better for the earthquake prediction.compared some index of the prediction results,which including the accuracy rate、the missed detection rate、the false alarm rate、the magnitude error(belowed Level 3 and belowed Level 1)、the seismogenic time error(less than 30 days and less than 10 days)and the epicenter location error(less than3°and less than1°).We found that the predicting results of the RST algorithm in time domain is better than the RST algorithm in spatial domain,the predicting results of the wavelet neural network is better than the BP neural network.In the end,also explained from the level of probability.