Seismic Reservoir Prediction via Enhanced Semi-Supervised Regression with Pseudo-Labels

Abstract Seismic reservoir prediction constitutes a crucial component of petroleum geological exploration. However, the high cost associated with acquiring labeled seismic datasets has positioned semi-supervised learning (SSL) methods as an active research area, given their capacity to enhance the accuracy of seismic reservoir prediction. In this paper, we propose a semi-supervised regression algorithm with pseudo-labeling (PL-SSR), which integrates well logging data with a limited number of labeled samples to improve the accuracy of seismic reservoir prediction. The algorithm generates pseudo-labels for unlabeled samples via predictions from pre-trained models, and selects only high-confidence predictions to iteratively refine regression coefficients using the Partial Information estimator. Compared with Ridge Regression, Random Forest Regression, and Lasso Regression, the proposed method achieves the highest accuracy, with consistent results retained following variable selection. Blind well testing further validates the method’s effectiveness, demonstrating a 60% reduction in prediction errors relative to traditional SSR approaches. Notably, PL-SSR yields the best overall performance, rendering it a promising solution for cost-sensitive seismic exploration scenarios.