Using CNN To Classify Spectrograms of Seismic Events From Llaima Volcano (Chile)

Tensor-Based Learning Framework for Automatic Multichannel Volcano-Seismic Classification

“…For comparison purposes, the LR-SPM was also tested, as well as the standard PCA and SVM techniques using the vectorization of the tensor samples, which yielded pattern vectors of dimension 336. The proposed approach is also compared with a CNN inspired by the work [8] and with the method presented in [7], where the EMD is performed before attribute extraction, but using the same attributes of the present work. Multiclass SVM and STM were implemented using the one vs. all approach.…”

“…This method is denoted by EMD + SVM. Table V also compares the proposed method with deep learning approach, using a CNN with design inspired by [8]. These two techniques are implemented using the same attributes of the present work, but vectorized, in the following cases: no PCA, PCA-DR and MPCA-DR.…”

“…It can be observed from Table V that the proposed approach TABLE V: Correct Classification Rate (CCR) -comparison between the proposed approach and the ones of [7], [8].…”

“…PCA Type Classifier CCR EMD + SVM [7] 81.7% None CNN [8] 51.5% STuM 89.6% EMD + SVM [7] 85.2% PCA-DR CNN [8] 56.8% STuM 89.8% EMD + SVM [7] 85.5% MPCA-DR CNN [8] 62.35% STuM 91.9% using the STuM provided the best CCRs for the three cases (no PCA, PCA-DR and MPCA-DR) and, once again, the MPCA-DR combined with the STuM reached the best CCR. This is due to the fact that the techniques of [7] and [8] break the tensor structure of the data. The MPCA-DR has also improved the performance of the EMD + SVM and CNN methods, the worst CCR being achieved by the CNN.…”

“…In [7], an automatic classification system for volcano events is presented using the empirical mode decomposition (EMD). More recently, the work [8] explored deep learning by using Convolutional Neural Networks (CNNs) to classify spectrograms of seismic events from a South American volcano.…”