Given a set of regions with activity counts at each time instant (e.g., a listing of countries with number of mass protests or disease cases over time) and a spatial neighbor relation, geo-referenced time-series summarization (GTS) finds κ-full trees that maximize activity coverage. GTS has important potential societal applications such as understanding the spread of political unrest, disease, crimes, fires, pollutants, etc. However, GTS is computationally challenging because (1) there are a large number of subsets of κ-full trees due to the potential overlap of trees and (2) a region with no activity may be a part of a larger region with maximum activity coverage, making apriori-based pruning inapplicable. Previous approaches for spatio-temporal data mining detect anomalous or unusual areas and do not summarize activities. We propose a κ-full tree (κFT) approach for GTS which features an algorithmic refinement for partitioning regions that leads to computational savings without affecting result quality. Experimental results show that our algorithmic refinement substantially reduces the computational cost. We also present a case study that shows the output of our approach on Arab Spring data.