Development of CNN Pruning Method for Earthquake Signal Imagery Classification

Abstract

The real-time detection of earthquake occurrences in a seismic wave, drawn by seismograph, is crucial for disaster mitigation. The earlier the earthquake warning, the more lives can be saved. One approach that can monitor and detect earthquake occurrence is binary classification between earthquake and noise signals. The use of deep learning models such as CNN (Convolutional Neural Network), is considered quite accurate to perform an image classification of seismograph signals. Nevertheless, the tendency to use the large CNN model is rated to have better accuracy than smaller models. In fact, the disadvantage of utilizing large model is the inference time and the deployment of a large model to obtain real-time inference is more costly than the smaller model. This paper aims to reduce the size of a CNN model (Resnet50) by pruning the unnecessary filters and neuron on the model architecture without sacrificing the accuracy. The task of the model was to classify two classes (earthquake and noise) of spectrogram images, the dataset is STEAD (Stanford Earthquake Dataset). To prioritize which filter or neuron to be eliminated, L2-norm was calculated on each filter or neuron weights. We assumed that a filter or neuron with the lowest L2-norm had the least significant role in the model. By pruning 90% of the filter and neuron of the model and retraining the pruned model, the inference time was improved from 22. 45ms to 3. 6ms (on NVIDIA GTX 1050) per image with the accuracy of 99.405%.