Impact of Data Imbalance on Reservoir Fracture Identification

Fractures in oil and gas reservoirs are crucial for hydrocarbon storage and fluid flow, significantly affecting recovery rates. However, fracture identification is a typically imbalanced classification problem. This paper proposes a novel resampling method, Safe points-Synthetic Minority Over-sampling Technique (Safe SMOTE), which improves upon the traditional Synthetic Minority Oversampling Technique (SMOTE) by identifying “reliable point pairs” to synthesize new minority class samples. Specifically, when nearest minority class neighbors are in a sector devoid of majority class samples, they are considered reliable point pairs, ensuring the generated samples are more representative. This approach effectively reduces the risk of misclassification for synthetic samples and enhances the accuracy of classification models. Additionally, an evaluation metric is introduced that considers both the training and testing set to better assess the resampled dataset. To validate Safe SMOTE, fracture data from Iraq's T Oilfield was analyzed. Its performance was compared with six resampling methods using Transformer (TS) and Random Forest (RF) classifiers. Additionally, the optimization and combination of two loss functions (Focal Loss (FL) and Dice Loss (DL)) were explored. The results indicated that the TS method outperformed RF overall. Furthermore, in the context of RF, the combination of Safe SMOTE with the loss functions surpassed other resampling methods. These findings further confirm the efficacy of Safe SMOTE in addressing imbalanced datasets.