3D U-Net-Assisted Automated Facies Classification From Seismic Volume of Amguri Prospect, Upper Assam Shelf, NE India

Abstract

Traditionally, facies classification from seismic data heavily relies on manual interpretation, a time-consuming and subjective task. To circumvent this laborious traditional workflow, an automated 3D U-Net-based framework of facies classification from 3D seismic data is presented. Initially, the seismic data were cleaned using a dip-steered median filter (DSMF), and facies of different geological ages were labelled to generate facies masks. Afterwards, a 3D-UNet facies-prediction model was trained by feeding the DSMF-filtered seismic volume as input and the corresponding facies mask as the target. After successful training, the feasibility of the facies classification model was tested over the entire seismic volume. In the training phase, ∼94% accuracy was achieved, and test-phase accuracy was evaluated using accuracy metrics and structural similarity index (SSIM), signal to noise ratio (SNR), R², root mean squared error (RMSE) and mean squared error (MSE) parameters, indicating that the U-Net-derived facies classes are well corroborated with the manually interpreted facies over the entire seismic volume. Industrial-grade seismic data from the Amguri prospect of Upper Assam Shelf, India, were analysed in this study. The accuracy-based ranking of the facies classes is: Overburden > Basement > Lakwa > Tura > Sylhet > Kopili > Geleki > Barail Coal Shale (BCS) > Barail Main Sand (BMS). The model shows higher accuracy, improved delineation of facies boundaries and better spatial continuity, consistent with known geological formations. The demonstrated 3D U-Net-based paradigm is an effective and scalable approach for automated facies interpretation from seismic data, reducing interpreter bias and improving reservoir characterisation in complex geological environments.