Abstract: Artificial intelligence algorithms have been developed to automatically identify the spatial structures of formation lithologies from multivariate log data. They represent a promising approach to reducing lithology logging costs and mitigating the subjectivity inherent in lithology identification. Considering the imbalanced distribution of lithology sample data and the spatialtemporal variability in the relationships between log attributes and lithologies, this study constructed a synthetic minority oversampling technique (SMOTE)-long short-term memory (LSTM) hybrid model. The SMOTE algorithm effectively balances the sample distributions of different lithologies, while the LSTM algorithm, using its deep learning architecture, extracts lithological characteristics from the log sequence data. With the borehole log data and lithology records from a sandstone uranium deposit as training data, the SMOTE-LSTM hybrid model achieved a prediction accuracy exceeding 85% in lithology classification. Compared to several other machine learning methods, the SMOTE-LSTM hybrid model demonstrated significantly improved accuracy and reliability in lithology identification.