Depth-Wise Salt Deposit Classification in Seismic Images Using Adaptive Feature Engineering and Hybrid Ensemble Learning

Abstract

 Identifying subsurface depth intervals with high salt concentration is crucial in seismic interpretation. It directly impacts risk reduction in drilling and reservoir characterization. While deep learning methods have performed well in detecting salt bodies, they mainly focus on spatial segmentation and do not specifically address depth-wise salt classification. This paper introduces a depth-wise salt classification framework that uses adaptive feature engineering and hybrid ensemble learning. The framework aims to estimate the probability of salt presence at different subsurface depths based on seismic images. The approach works with salt-enhanced seismic images and extracts important handcrafted features. These include statistical intensity measures and texture descriptors derived from Gray-Level Co-Occurrence Matrices (GLCM). The learning pipeline includes subsurface depth metadata to support depth-aware classification. To improve robustness and generalization, two ensemble classifiers—Random Forest (RF) and Extreme Gradient Boosting (XGBoost)—are trained separately. Their probabilistic predictions are then combined using an ensemble averaging technique. The framework is tested on the TGS Salt Identification Challenge dataset. Experimental results show that the hybrid ensemble outperforms individual classifiers while keeping error rates balanced. Additionally, the method allows for depth-wise salt probability analysis, facilitating the identification of depth ranges with the highest salt presence. This method offers a computationally efficient and clear solution for depth-focused seismic salt analysis...