A machine learning approach for slow slip event detection using GNSS time-series

Extracting tectonic transient displacements on the Earth's surface from Global Navigation Satellite System (GNSS) time series remains a challenge, because GNSS station displacements depend on multiple processes occurring simultaneously, along with noise that obscures low-magnitude transient signals. We present a novel method for automatic detection of slow slip events (SSEs) in time series of a GNSS network by training a supervised machine learning (ML) model for classification. The proposed methodology detects both temporally and spatially the signatures of SSEs or regional transients within a GNSS network. The time series of a GNSS network were transformed into grayscale images, from which descriptors, including Bag of Visual Words (BoW) and Extended Local Binary Patterns (ELBP), were extracted. These descriptors served as input features for two distinct ML models: Support Vector Machines (SVM) and Artificial Neural Networks (NN). To train and test the ML classification model, two 3-year synthetic datasets were generated, one with GNSS networks featuring slow slip events (SSEs) of varying location, duration, onset time, and magnitude, and the other without SSEs, resulting in positive and negative sets, respectively. For each GNSS network, an image was created by combining the east and north components of the time series, which have been previously detrended and common mode error filtered. Each image is further divided into sub-images corresponding to 60 days time windows, in order to temporarily detect the existence of a transient. For training and testing, the datasets were separated into 75% for training and 25% for testing, each with 50% positive and 50% negative cases. In the final step, we analyze the positively classified images, representing the time windows in which the classifier detected transients. Within each of these windows, we identify the network's time series with the highest velocity, indicating the stations and geographic area where the detected transients occurred. The test results demonstrate that both ML models achieved high performance using both ELBP and BoW descriptors as features. Finally, our ML models were validated on a real dataset with a transient signal recorded before the 2014 Iquique earthquake in Chile, and they effectively detected this anomalous signal. The proposed method can effectively detect transient signals related to SSEs with high accuracy, sensitivity, and specificity in both the test and instrumentally recorded datasets.