Spatio-temporal kriging is a powerful predictive approach for dealing with data that are spatially- and temporally- referenced. Sensor data streams are often in this form since they comprise timestamps and sensor IDs. Proximity in time and space also explains correlations and regularities in the measurements, which can be leveraged in prediction.
Modern data streams pose unprecedented challenges in computation that curse the application of kriging. State-of-the-art mostly revolves around subsetting and subsampling techniques that reduce computational costs. In some suitable setting it can be useful to resort to separable models, made up of a spatial model within time frames and a time series model within locations.
Composite estimators would estimate the spatial and temporal models separately. This estimator is easier to evaluate than the maximum likelihood estimator and more robust to model misspecification. Autoregressive models in the time domain make both estimators more tractable.
This article proposes a composite estimator and an efficient implementation of the maximum likelihood estimator that can help to process large datasets from sensor networks. An example is presented, based on real data from an indoor monitoring application. Results apply to a broad range of cases where data accumulate fast and steadily.