This paper addresses the question of temporal learning in spatially distributed wireless sensor networks (WSN). We propose to fuse WSNs with the echo states network learning concepts to infer the spatio-temporal dynamics of the data collaboratively measured by sensors. We prove that a WSN topology described by a bidirected graph is strongly connected, which is a sufficient and necessary condition for implementing in-network distributed learning. For strongly connected networks we develop a systematic method to satisfy the conditions resulting in echo states in sensor networks. The effectiveness of the learning approach is demonstrated with several controlled model experiments.