GLCM and k-means clustering for texture-based recognition of salt bodies in seismic images

Abstract

Subsalt imaging and interpretation remain challenging tasks for oil and gas exploration, as well as for identifying suitable sites for storing large amounts of atmospheric carbon dioxide. The complex geometry of salt bodies and the poor visibility of subsalt structures introduce significant uncertainty into manual seismic interpretation, especially when defining the base of salt formations. To address this, we present an interpretable and lightweight semiautomatic methodology for salt body delineation in 2D seismic sections. This approach combines texture features derived from the Gray Level Co-occurrence Matrix (GLCM) with k-means clustering. We performed a comprehensive parameter sweep, exploring different window sizes and cluster counts, evaluated using strict consensus labels obtained from two expert interpretations. A 5-fold cross-validation procedure was applied to identify configurations with consistently high F1 scores, achieving a maximum F1 of 0.78. To enhance robustness, we introduced a consensus-based strategy that aggregates results from multiple high-performing configurations. The resulting consensus map provides a continuous confidence measure for salt likelihood, highlighting well-defined boundaries and subtle features often overlooked in manual interpretations. This method reinforces the hypothesis that salt bodies exhibit a distinctive textural signature in the GLCM feature space and demonstrates the potential of interpretable unsupervised learning to support expert workflows in seismic interpretation involving salt tectonics.