In this paper, we propose a new transport protocol for data collection in sensor networks that monitor physical phenomena. In a network with variable channel condition, this protocol adapts transmission reliability based on the importance of transmitted spatio-temporal data to the reconstruction of the phenomenon. Data whose omission generates a higher estimation error are transmitted more reliably. The protocol aggregates data from nodes to the base station and provides a constant expected estimation error while significantly reducing the energy consumption and bandwidth usage compared with other approaches to reliable communication. Moreover, our protocol is easy to implement on current motes. To the best of our knowledge, this is the first transport layer protocol, that incorporates predictive models, to be implemented on motes. We perform extensive experiments on MicaZ motes using LiteOS to compare the performance of our protocol against previous transport protocols and data suppression schemes. Our experimental results show that our protocol introduces a very small estimation error while saving orders of magnitude in consumed energy compared to reliable transport. In other cases, estimation error and energy consumption are simultaneously reduced by 50-85%.