Accurate identification of fracture–cavity systems is crucial for carbonate reservoir characterization and drilling-risk reduction. However, conventional seismic attribute methods are often limited by strong multi-solution ambiguity and low sensitivity to weak, heterogeneous, and non-stationary fracture–cavity responses. To overcome these limitations, this study proposes a fracture–cavity identification method based on a unified sparse time–frequency–phase (TFP) atom decomposition framework and applies it to the Maokou Formation in the Yi-202 well area of the Sichuan Basin. In this framework, seismic traces are represented as a sparse superposition of TFP atoms characterized by amplitude, arrival time, dominant frequency, and initial phase. An L₁-regularized sparse inversion, solved by a fast iterative shrinkage-thresholding algorithm, is first used to obtain a spectrally enhanced, high-resolution trace. Analytic matching pursuit is then employed to extract TFP atoms and jointly estimate time, frequency, and phase, followed by coordinate-descent refinement to improve decomposition accuracy and stability. Based on the refined atom dictionary, a joint frequency–phase energy volume is constructed, from which band-selective reconstructions and a Kullback–Leibler divergence anomaly indicator are derived. Compared with conventional methods that primarily rely on amplitude or single-frequency attributes, the proposed method jointly exploits frequency and phase information, thereby improving the discrimination of subtle fracture–cavity anomalies and reducing interpretation uncertainty. Field-data application demonstrates that the method can more clearly delineate small-scale fracture–cavity anomalies that are difficult to identify using conventional approaches, highlighting its practical value for carbonate reservoir prediction and drilling-risk mitigation.