Data dissemination protocols in cyber-physical systems must consider the importance of data packets in protocol decisions. Importance of data cannot generally be accurately represented by a static priority value or deadline, but rather must stem from the dynamic state of the physical world. This paper presents a novel congestion control scheme for data collection applications that makes two key contributions. First, packet importance is measured by data contributions to the accuracy of estimating the monitored physical phenomenon. This leads to congestion control that minimizes estimation error. Second, our protocol employs a novel mechanism, i.e. spatial aggregation, in addition to temporal aggregation to control congestion. The protocol is generalized to multiple concurrent applications. Our approach employs different granularities of aggregation in transporting spatio-temporal data from nodes to a base station. The aggregation granularity is chosen locally based on the contribution of the transmitted data to the reconstruction of the phenomenon at the receiver. In an area affected by congestion, data are summarized more aggressively to reduce data transfer rate while introducing minimal error to the estimation of physical phenomena. We implement this scheme as a transport layer protocol in LiteOS running on MicaZ motes. Through experiments, we show that the proposed scheme eliminates congestion with an estimation error an order of magnitude smaller than traditional rate control approaches.