Abstract: The condition of surrounding rock is a critical factor in determining both the selection of construction methods and the control of project costs in tunnel engineering. Accurate classification of surrounding rock grades is essential for achieving safe, efficient, and economical construction. Currently, rock mass classification methods predominantly rely on single-source indicators and heavily depend on the experience of construction personnel, making it difficult to quantitatively characterize the structural features of surrounding rock under complex geological conditions. Based on a systematic review and in-depth analysis of current mainstream intelligent recognition algorithms for joint extraction from tunnel face images and intelligent analysis of drilling data, this study builds on the Mask Region-based Convolutional Neural Network (Mask R-CNN) framework by incorporating the Transformer attention mechanism and Deformable Convolutional Network (DCN). This led to the development of an enhanced Mask R-CNN-TD (Mask R-CNN-Transformer-DCN) intelligent classification method for tunnel face surrounding rock. Furthermore, leveraging the strong correlation between tunnel face rock quality and drilling parameters of rock drill rigs, a neural network prediction model enhanced with the Archimedes Optimization Algorithm (AOA) was proposed for forecasting the condition of rock ahead of the face. An unsupervised clustering algorithm for multi-structural data of tunnel surrounding rock was also designed, resulting in an integrated four-component methodology that combines "geological survey and design, geological prediction, drilling data, and intelligent image recognition of the tunnel face." A corresponding application system based on a "smartphone + edge server" architecture was developed and validated through engineering applications. The results demonstrate that: (1) The proposed Mask R-CNN-TD intelligent classification model achieves mean Average Precision (mAP) values of 72.6% and 56.1% for bounding box detection and mask segmentation, respectively, indicating significant improvements in both detection and segmentation performance; (2) The AOA-optimized neural network model for predicting rock conditions ahead of the face effectively utilizes multi-dimensional drilling data, achieving a prediction accuracy of 98.25%; (3) The developed four-component integrated classification method and application system exhibit high reliability and enable real-time intelligent classification of tunnel surrounding rock. This research provides an effective and intelligent approach for the efficient and accurate assessment of surrounding rock conditions in complex geological environments.