Multi-Attribute Classification Based On Sparse Autoencoder - A Gas Chimney Detection Example

Abstract

Multi-attribute classification technologies, including supervised and unsupervised methods, play an important role in seismic interpretation. Providing that enough labeled samples are available, supervised methods usually obtain some credible results. However, for seismic data, the unlabeled data is huge but the labeled data is usually limited. Therefore, the combination of supervised and unsupervised methods is a feasible idea. As a common unsupervised feature learning algorithm used in deep learning, sparse autoencoder can realize automatic feature extraction using the unlabeled data. In this abstract, we introduce the sparse autoencoder and design a semi-supervised learning framework for multi-attribute classification applications by combining unsupervised feature learning and supervised classification. In the proposed framework, the original data of both labeled and unlabeled samples are used to train a sparse autoencoder at first, and then encoded to get some features for better representation. After that, the features of these labeled samples are used to train a classifier. At last, we apply this classifier on these unlabeled samples to obtain the final classification results. To demonstrate the validity of the designed technology, we give a gas chimney detection example. Results show that the multi-attribute classification technology based on sparse autoencoder outperforms the traditional multilayer perceptron method.