Spatio-temporal earthquake prediction with structural recurrent neural networks

The earthquake prediction problem can be defined as a given a minimum Richter magnitude scale and a specified geographic region, predicting the possibility of an earthquake in that region within a time interval. This is a long time studied research problem, but not much progress is achieved until the last decade. With the advancement of computational systems and deep learning models, significant results are achieved. In this thesis, we introduce novel models using the structural recurrent neural network (SRNN) that capture the spatial proximity and structural properties such as the existence of faults in regions. Experimental results are carried out in two distinct regions such as Turkey and China, where the scale and earthquake zones differ greatly. SRNN models achieve better performance results compared with the baseline and the state of the art models. Especially SRNNClass_near model, which captures first-order spatial neighborhood and structural classification based on fault lines, results in the highest F_1 score.