Abstract: The integration of artificial intelligence and geological sciences has accelerated the digital transformation of geotechnical engineering, with lithology identification emerging as a pivotal task in geological exploration and tunnel construction. While automated lithology recognition offers substantial advantages in operational safety and cost efficiency, two critical challenges persist: (1)the scarcity of domain-specific geological image datasets from real tunnel environments, and (2) the adverse impact of severe class imbalance on minority category recognition accuracy. To address these dual challenges, our study developed a tunnel construction-specific geological image dataset through systematic collection and preprocessing, followed by a comprehensive evaluation of class imbalance mitigation strategies. A residual neural network architecture enhanced by a weight-adaptive Softmax loss function was proposed, which dynamically calibrates category weights based on sample distribution to suppress majority-class dominance during training. The framework synergizes ResNet's hierarchical feature extraction with an attention-driven imbalance correction mechanism, achieving 95% overall accuracy in lithology identification. Attention heatmap visualization further confirmed the model's alignment with expert-interpretable geological features, demonstrating both favorable technical efficacy and engineering-oriented interpretability.