Feature selection for seismic facies classification of a fluvial reservoir: pushing the limits of spectral decomposition beyond the routine red-green-blue color blend Ismailalwali Babikir* 1 , Mohamed Elsaadany 1 , Maman Hermana 1 , Abdul Halim Abdul Latiff 1 , Muhammad Sajid 2 , and Carrie Laudon 3 1 Universiti Teknologi PETRONAS, Seri Iskandar, Malaysia 2 PETRONAS, Kuala Lumpur, Malaysia 3 Geophysical Insights, Houston, Texas, USA Summary Spectral decomposition is a powerful interpretation tool that provides superior subsurface images of channels and other thinly bedded depositional systems. The analysis of the band-limited components is commonly facilitated through the red-green-blue (RGB) blend, which is limited to three volumes at a time, primarily selected based on the interpreter preference. Fortunately, machine learning technology provides the opportunity to quantitatively use many frequency volumes. We analyze twelve spectral magnitude components using multivariate feature selection techniques. The chosen subsets of features are used to classify seismic facies of a fluvial reservoir in the Malay Basin, offshore Malaysia. We find that the subset of spectral components gives a better classification result than the whole set. The sequential forward selector and the embedded selector of random forest algorithms provide the best subset of features that differentiate the desired classes. Introduction Volumetric attributes are potent tools in identifying geologic features from seismic data. Because of their ability to express geologic patterns better than the original seismic, attributes are commonly used to input machine learning models for facies classification (Marfurt, 2018). The use of machine learning to classify multiple seismic attributes increases the interpretability of the data and provides more reservoir characterization information. The application of machine learning algorithms with an appropriately selected suite of attributes offers better discrimination of thin beds, improves identification of direct hydrocarbon indicators (DHIs) and enhances the resolution to characterize reservoir and stratigraphy (Roden et al., 2015). Spectral decomposition has shown great success in imaging stratigraphic patterns. It aims to break down the seismic trace into narrowband frequencies to highlight subtle subsurface features. The interaction of the band-limited signal with the spatial distribution of seismic impedances within a group of layers makes geologic features (e.g., fluvial depositional elements) respond at different ranges of seismic frequencies. Spectral decomposition in this regard is expected to separate the amplitude response of each geologic feature from the background amplitude (McArdle and Ackers, 2012). The interpretation of spectral decomposition is commonly facilitated via the red-green-blue (RGB) color blend, where three spectral magnitude components of high, medium, and low frequency are co-visualized. Selecting band-limited frequency components for the RGB combination is a trial process that heavily depends on the interpreter’s judgment. It is challenging to analyze several frequency volumes since the RGB display is limited to three volumes. Machine learning technology has the potential to overcome several limitations in the routine subsurface interpretation workflows. It helps the interpreter efficiently handle a massive volume of data and quantitively finds connections between different data types and across datasets of different characteristics. The prediction performance of machine learning models is heavily dependent on the quality of input features. Therefore, feature selection approaches are commonly applied to eliminate redundant and irrelevant attributes. Feature selection aims to pick an optimal subset that reduces the computation time, avoids overfitting, and improves the performance of the model (Kim et al., 2019). To overcome the limitation of the RGB blend in analyzing spectral decomposition volumes, we apply machine learning classification to identify and characterize seismic (lithogeomorphologic) facies of a fluvial reservoir from the Malay Basin. Before classification, we use feature selection to choose the important attributes among several band- limited magnitude components. Dataset The study uses a 3D seismic survey from an oil and gas producing field in the Malay Basin. The survey is acquired at a shallow water region (~ 60 m), covering some 450 square kilometers of area. The bin size of the data is 12.5 by 18.75 meters, with a 2-millisecond sampling interval. The data have an SEG normal polarity. A fluvial meander system occupies the interval of interest; accordingly, we crop a sub- volume of 200 milliseconds to be used for attribute computation. We use wireline logs, reports, and some core descriptions from 10 wells for calibration. 10.1190/image2022-3752042.1 Page 1704 Second International Meeting for Applied Geoscience & Energy © 2022 Society of Exploration Geophysicists and the American Association of Petroleum Geologists Downloaded 08/20/22 to 54.80.3.144. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/page/policies/terms DOI:10.1190/image2022-3752042.1