Novel Feature Augmentation and Input Reshaping Strategy for Accurate Seismic-Based Elastic Property Prediction Using Deep Learning

Abstract

Data limitation and sparsity are considered the main source of non-uniqueness and ill-posedness in elastic property prediction on seismic data using Deep Learning (DL). The ill-posed regression problem can be solved by conducting adequate pre-processing steps through data augmentation, feature engineering and feature selection. In this paper, we develop a novel technique of reshaping the input data into various multi-dimensional shapes before using the data as an input for the DL model. This strategy can increase the spatial context and temporal data coverage which is important for regularizing the DL model during training process. We also investigate the model performance between a Convolutional Neural Network (CNN) and a Recurrent Neural Network (RNN), specifically the Long-Short Term Memory (LSTM) architecture on 2D synthetic seismic data. Our findings show that using the newly developed feature augmentation and input reshaping strategy with a simple LSTM model can help improve the elastic property prediction on sparse well data. We demonstrate the robustness and effectiveness in terms of visual inspection and relative accuracy measure of the proposed method by using the SEG SEAM Interpretation Challenge dataset Phase I and comparing our results with Mustafa et al.'s (2021) work.