A new method named cluster-based GSA is proposed to enhance the sensitivity analysis of models with temporal or spatial outputs. It is based on a tight coupling between Global Sensitivity Analysis (GSA) and clustering procedures. Clustering is introduced to characterize the different behaviors of the model outputs by grouping them into clusters. The cluster-based GSA produces variance-based indices that quantify how the model inputs drive the model outputs toward a given cluster or how they influence variation along a direction defined by two clusters. Aggregated indices are proposed to summarize the overall influence of model inputs on changes of clusters. The method is applied on two models having temporal outputs: a toy example and an environmental model simulating the decomposition of soil organic matter (CANTIS). In both cases, the influence of the model inputs on the different behaviors of model outputs was efficiently reported by this approach.