With the widespread use of satellite imaging, a wealth of information is available to help understanding and modeling earth system processes. In particular, these data play a key role in the analysis of climate variability. However, satellite images can present data gaps due to partial coverage of the domain by the different orbital characteristics of the satellites. We present a method to fill these gaps and reconstruct the missing data with realistic property values using Direct Sampling multiple-point geostatistical simulation. The training images used are based on the known parts of similar images taken at previous dates.
Important auxiliary information can be available at a gap location when certain satellites cover the missing part but inform upon a different, but related, variable. For example, there may be locations where soil moisture is not informed, but where another satellite provides cloud coverage, temperature, or both. By using multivariate training images, our method allows maximizing the use of such auxiliary information, even if the relationships between the different reconstructed variables are highly non-linear.
The method is applied on synthetic imagery derived from a regional climate model of south-eastern Australia. The variables considered are soil moisture, temperature, latent heat flux (evaporation) and shortwave downward radiation (cloud coverage). Artificial gaps corresponding to satellites passages are made in the images (representing up to 40% of the domain). The gaps locations are then reconstructed with our method and compared with the original data. The reconstructions show that our method is able to accurately reconstruct the missing part of the satellite images, and to accurately reproduce the complex dependencies occurring between the variables considered.
From a practical perspective, the reconstruction method is straightforward and does not need user intervention for parameters adjustment. Therefore it can be automated to systematically process real-time remote sensing measurements and can be of interest for a wide range of applications in climate studies and for environmental modeling.