A method for gap-filling very large spatial datasets: application to AVIRIS-based airborne data

The analysis of spatial processes in environmental and hydrological sciences is often informed by remote sensing observations, provided by satellite or airborne sensors. However, the raw data obtained by such means can present gaps, for instance due to the orbital characteristics of a satellite or to specificities of the aircraft flight path. Many applications and modeling workflows require complete, gapless data. Geostatistical approaches are often used to fill these data gaps, however the sheer size of modern remote sensing datasets make the application of traditional geostatistical approaches challenging due to computational constraints (high-resolution, broad spatial coverage) and to data characteristics (complexity of features, non-stationarity).

In this work, we develop a new approach based on multiple-point geostatistics to fill gaps in very large and non-stationary data sets. It is based on a strategy of partition of the domain in overlapping tiles. This makes the problem computationally more affordable, while additionally enabling parallelization. It also alleviates issues related to non-stationarity, since the assumption of stationarity is more likely to be valid on a small tile than on a large domain.

The approach is illustrated on a dataset that is based on acquisitions by the AVIRIS-NG hyperspectral airborne sensor in Switzerland. The data present significant gaps, and at the same time the domain is extremely large, comprising over 300 million pixels. The simulation results are visually realistic and corroborate independent validation data.