Study on the Parameterization Response of Probabilistic Neural Networks for Seismic Facies Classification

Abstract:

Nowadays many machine learning techniques are being applied for seismic facies classifications, as these algorithms are able to identify non-linear relationships between seismic attributes and seismic facies. These relationships are not always obvious to a human interpreter, and therefore can provide faster and more accurate results. Nevertheless, most of the success of these techniques rely on the quality and training of the input data. Previous studies have shown the accuracy of the Probabilistic Neural Network (PNN) technique in the classification of salt and conformable sediment related facies. However, in order to further test the capability of this algorithm when using more seismic facies, we chose the complex East Break and Alaminos Canyon 3D survey in the western Gulf of Mexico where we could identify large scale features, like those related to salt, mass-transport deposits, and conformable sediments; but also, subtle, smaller. and more challenging features like those related to channels and the seismic noise near the seafloor. In this approach, we additionally tested the influence of the parameterization during the preparation of the input data on the final performance of the classification. To do so, we created a series of models and cases while varying different parameters like the amount and type of facies, the combination of seismic attributes and the Kuwahara filter window size. Overall, the classifications were accurate, however, the results indicate that when classifying only large scale features, it is better to apply bigger Kuwahara window sizes, while when classifying small scale features is better to apply a bin size Kuwahara filter window size. Also, we identified the importance of including amplitude attributes and curvature attributes in the facies prediction models. Finally, we developed a workflow that can help interpreters to improve seismic facies classification models when using PNN.