DeepDetect: A Cascaded Region-based Densely Connected Network for Seismic Event Detection

Highlight

The DeepDetect neural network design is based on a programming approach that mimics the biological pathways in an animal brain, rather than providing the computer with step-by-step instructions to solve a problem. With a deep, dense neural network, the system has multiple layers of information to assess, progressively developing complexity as each layer is juxtaposed over another. The system develops its own understanding of how to interpret the data in phases, aided by expert human labeling at key points, and working toward more accurate interpretations as the layers build up. In this case, earthquake data is analyzed as a sequence of data points taken at successive equally spaced points in time. The goal is to achieve efficient, accurate, and a cost effective method to detect events of interest based on previously observed data.

The full paper can be found on IEEE TGRS:

DeepDetect: A Cascaded Region-Based Densely Connected Network for Seismic Event Detection

Figure: The illustration of our DeepDetect method, which involves two cascaded modules: a classification module and a regression module. The classification module is designed to separate signals of interests from others. The regression module is used to localize the events within the signals of interests.

Earthquake Detection

Event-based time series signals are quite common in various physical systems. Those events can be limited in time duration, varied by signal amplitude, and corrupted by all sorts of environmental and anthropogenic noise. The goal of this research is to detect events including picking onset times and durations of signals of interest even with amplitudes smaller than human analysts normally discern. Simultaneously, it is necessary to limit the number of “false detections” (i.e., incorrect decisions that segments of a data stream are signals of interest) to a small fraction of the true detection.

Figure : An earthquake event consists of two segments: p-wave (green) and s-wave (red). The duration and amplitude of earthquake can be varied significantly from case to case.

Multi-scale Time Series Classification

Earthquake is a unique seismic event that yields certain pattern. One of the major challenges to accurately detect earthquake is its significantly varied duration of time from case to case. To address this challenge, we design our DeepDetect as a cascaded region-based convolutional neural network. Our detection method would capture events in different sizes, while incorporating contextual information to enrich features for each individual proposal. For better generalization of performance, we utilize densely connected blocks as the backbone of the network. Because some positive events are not correctly annotated, we reformulate the detection problem as a learning-from-noise problem to improve the detection accuracy.

Figure : Three proposals (green nodes) are considered positive since they have the intersection over union with the ground-truth above the threshold. Ground-truth are indicated at the bottom with green and red dots, which denote beginnings and ends, respectively.

Results

I have successfully applied DeepDetect to different field datasets. For example, I employ my method to analyze acoustic data acquired from a bi-axial “earthquake machine” located at Rock Mechanics Laboratory, Penn State University. Four different detection methods were compared: the traditional waveform cross-correlation-based method (Template Matching), Support Vector Machine (SVM) method, one of my recent projects using supervised dictionary methods, and DeepDetect. The average precision of each method is provided in Table below. DeepDetect yields the highest average precision.

Four different detection methods are compared: the traditional waveform cross-correlation-based method (Template Matching), Support Vector Machine (SVM) method, one of my techniques using supervised dictionary methods, and DeepDetect.