Given a spatial network and a collection of activity events (e.g., emergency requests, crime reports, accident reports, etc.), spatial network activity summarization (SNAS) finds a set of k shortest paths based on the activity events. SNAS is important for critical societal applications such as disaster response and crime analysis. SNAS is computationally challenging because of the potentially exponential search space wherein there are an exponential number of k subsets of all possible shortest paths in a spatial network. Previous work on SNAS has focused on either geometry or sub-graph based (e.g., only one path), and cannot summarize multiple routes in a spatial network. We propose a novel approach, called K-Main Routes (KMR), that discovers a set of k shortest paths to summarize activities. KMR can be considered a generalization of the well known K-means technique for network space. KMR uses inactive node pruning to reduce the number of shortest paths calculated by accounting only for active nodes. Experimental evaluation of KMR using a real-world data set demonstrates that KMR with inactive node pruning leads to substantial computational savings without reducing the coverage of the resulting summary paths. A case study that compares network based with geometry based summarization on a real-world data set is also presented.