Explainable Deep Reinforcement Learning (XDRL) holds significant potential for clarifying the decision-making logic of agents in complex tasks. Current XDRL research increasingly focuses on multi-granularity policy explanation methods that integrate both local and global decision-making insights. However, most of them treat local and global explanations as separate processes, overlooking their interdependencies and compromising logical consistency. This paper introduces a Multi-granularity Policy eXplanation method based on Saliency map Clustering (MPXSC), which computes both local and global policy explanations for a DRL agent in a unified, end-to-end process. MPXSC begins by employing super-pixel perturbation to generate saliency maps for all the agent’s states, representing its local explanation. These maps are then categorized by agent's actions and clustered based on local saliency features. Subsequently, the key states within each cluster are identified and serve as decision rules for their respective actions. Collectively, these results constitute the global explanation for the agent's decision-making. The superiority of MPXSC is validated through objective experiments on both local and global explanations, with a case study visually demonstrating the method’s compelling explainable evidence regarding DRL model prediction outcomes.