In remote sensing and geospatial analysis, gap-filling represents a key challenge, as missing data frequently occurs due to factors such as cloud cover, sensor malfunctions, or inconsistencies in data acquisition. Remote sensing images are often large, complex, and the data represented strongly non-stationary, further enhancing the difficulties of gap-filling. An additional challenge is that with large gaps and in the absence of predictive covariates, gap-filling methods do not have enough information to achieve a high level of local accuracy. Here we present the chessQS algorithm that produces realistic spatial structures, even in cases where there is not enough information available for a locally accurate reconstruction. It is based on Multiple-Point geoStatistics (MPS) and is adapted for very large domains by using an automatic segmentation into overlapping tiles, in a chessboard-like pattern that addresses both computational and non-stationarity limitations at the same time. The method is tested on a tree-species classification based on AVIRIS-NG hyperspectral imagery over the Gruyère Pays-d’Enhaut Regional Nature Park in Switzerland. Cross-validation results show that chessQS reproduces complex data structures very well, takes care of non-stationarities, and is capable of gap-filling a very large image rather quickly. The inclusion of an overlap between the tiles, to propagate results to neighboring simulations, provides continuity across the whole image. After completion, the original gaps are not distinguishable. The method can easily be used on other regions and datasets (continuous or categorical), and requires only few parameters, making it very flexible and easily applicable. • Introduces geostatistics-based chessQS for large-scale gap-filling. • Handles non-stationarity via chessboard tiling with overlaps. • Preserves complex spatial structures across simulations, even with limited predictors. • Flexible and stochastic method for categorical and continuous datasets. • Applied to tree-species mapping based on AVIRIS-NG hyperspectral images in Switzerland.
Gerber et al. (Fri,) studied this question.