We present a decentralized, communication efficient method for joint sensing and learning (regression) of spatiotemporal phenomena using a team of autonomous networked sensing agents. Gaussian Process (GP) priors are utilized in a Bayesian regression framework over the unknown function, with our main contribution being the introduction of a communication efficient decentralized GP inference algorithm and associated agent path planning strategy. Our method relies on reducing communication between agents by exchanging GP model parameters, instead of actual measured data. The global GP model is leveraged to maximize exploration of the sensing space using a policy iteration planning framework. The performance of the presented method is compared against state-of-the-art consensus based and GP-DDF method, and it is shown that the presented method provides significant improvements in regression accuracy while reducing communication required.