In many applications massive amount of data are collected in space, and it may not be practical to store and process all the data on a single machine. A lot of attention has been paid recently to develop statistical inference in a distributed manner to achieve computational and statistical efficiency. Much of the existing literature on distributed statistical inference splits the data randomly and assumes a homogeneous model across all local machines. In practice, many times data exhibit spatial heteroscedasticity, and it makes sense to allow the data model to have global parameters as well as local parameters which may vary in space. In this paper we develop new computational strategies under this more general framework which has proven convergence property and can achieve statistical efficiency. Simulation studies and real data analysis of remote sensing data are used to illustrate the advantage of the new method.