Abstract: The gradient structure tensor (GST) serves as a common attribute to characterize the boundaries of strongly heterogeneous reservoirs, such as fracture-cavity types, playing an important role in research on carbonate reservoirs. However, due to a lack of dimensions and high difficulties in threshold determination, the accuracy of GSI directly affects the precision of reservoir delineation and reserve estimation. Given this, focusing on a fracture-cavity carbonate reservoir model, this study proposed a three-step strategy for threshold selection based on analysis of the GST attribute distribution corresponding to noisy reservoir and non-reservoir regions. First, the distribution of GST attribute values was statistically analyzed. Then, the value distribution of attribute values in the non-reservoir regions was estimated to identify their mean point, left critical point, and right critical point. Finally, the right critical point was selected as the threshold for the GST attribute of noisy reservoirs, thus delineating the boundaries of fracture-cavity bodies. Model tests and practical applications demonstrate the feasibility and effectiveness of this method. The findings provide a foundation for the application and promotion of the GST attribute in geophysics.