Abstract:
|
We introduce a spatial data compression method, Spatial Adaptive Dispersion Clustering (SADC), specifically designed to reduce the size of a spatial dataset in order to facilitate subsequent statistical inference. Unlike with traditional data and image compression methods, the goal of SADC is to create a new dataset that will be used as an input into spatial estimation or prediction methods, such as traditional kriging or fixed rank kriging, where the full data may be large and make spatial inference computationally infeasible. SADC can be classified as a lossy compression method, and is based on spectral clustering. It aims to produce contiguous spatial clusters and to preserve the spatial correlation structure of the data so that the loss of predictive information is minimal. We demonstrate the predictive performance of adaptively compressed simulated data for several scenarios and compare it to two other data reduction schemes: using local neighborhoods and a simple aggregation. We also present an application to remotely-sensed sea surface temperature data.
|