Ensemble kriging for environmental spatial processes

Remote-sensed and reanalysis databases are valuable sources of environmental data that support a wide range of engineering applications. However, the sizes of such databases are often measured in terabytes (TB). Whereas these datasets with high spatial resolution are usually stored on the servers of national laboratories, the large data volume can be inconvenient for individuals who wish to work with the data. To that end, it is important to investigate how much redundant information the dataset contains, e.g., are the time series from two adjacent pixels statistically different? We use kriging, a spatial interpolation technique, to quantify such redundancy. More specifically, if the kriged environmental processes are sufficiently accurate, one can circumvent the need to work with the original high-spatial-resolution data, and use only a dimension-reduced version of the data. The empirical part of the paper considers the National Solar Radiation Data Base (NSRDB), which provides half-hourly, gridded, satellite-derived solar irradiance data, with a spatial resolution of 4 km by 4 km, spanning 1998-2017, with a total size over 40 TB. NSRDB is a valuable dataset for solar resource assessment applications. The beam normal irradiance (BNI) process is reconstructed using data on various dimension-reduced lattices. The trade-off between spatial resolution and data accuracy is studied.