Abstract: To address the strong heterogeneity of carbonate fracture-cave systems and the limited characterization capability of single seismic attributes, this paper proposes a fine characterization method that fuses multi-constraint structure tensor with variable-scale dynamic optimization. First, cubic convolution interpolation is employed to correct the spatiotemporal sampling scale discrepancies of seismic data, thereby eliminating calculation biases in gradient structure tensor. Fracture and cave confidence attributes are constructed based on tensor eigenvalues, and Gaussian-smoothed instantaneous amplitude is introduced as a weighted constraint to enhance the anti-noise capability of the attributes. Second, based on optimal surface voting, a variable-scale dynamic optimization algorithm is proposed. By adaptively adjusting the 3D analysis window and integrating multi-scale features, this algorithm effectively resolves the fragmentation issue in fracture identification, balancing microscopic details with macroscopic continuity. Furthermore, drilling and logging data are utilized to calibrate structural boundaries, ensuring the reliability of the prediction results. Practical application results indicate that the proposed method significantly improves the clarity, continuity, and accuracy of fracture-cave identification. It is capable of finely characterizing subsurface geological structural features, providing robust technical support for oil and gas exploration in complex carbonate reservoirs.