Conflation of sparsely sampled and gridded elevation data: scale mismatch and semantic differences

Abstract

Data conflation refers to the methods and processing of information fusion whereby multi-source data are integrated to derive required information that is thought to be of increased accuracy, finer resolution, better homogenised semantics, fuller coverage, enhanced representational and computational efficiency, or improved utility for certain purposes than any single data source alone. As spatial data can be conceived of as realisations of random fields, multivariate geostatistics provides a coherent framework for data conflation. As important metadata of spatial data, scale, which is considered synonymous with data support in this paper, and semantics, which concern the meanings given to 'elevation' in elevation data, for example, are complicating factors for data conflation, and need to be handled properly. This paper's novelty lies in its handling of semantic differences via corrections to biases in local means, and its trend-residual decomposition-based structural modelling of terrain elevation data, constituting a data-driven, geostatistical strategy, which significantly increases accuracy or consistency in conflated data. Using SRTM (Shuttle Radar Topography Mission) and GTOPO30 data at a site in northwest China, experiments revealed that effects of the semantic difference on data conflation can be reduced through properly estimating and incorporating biases of local means in primary versus secondary data. It was also found that structural modelling adaptive of scale and non-stationarity is effective for achieving accuracy in data conflation: adaptive scale modelling is done through proper trend-residual decomposition of elevation data and their residuals while non-stationarity in spatial variability is handled through localised mean specification and covariance modelling.

Geostatistical approaches to refinement of digital elevation data

“…Collocated cokriging (MM2) (19) was then carried out using residuals of R 1 and R 2 , and generated the predicted R 1 's residuals, which should be added to the R 1 's trends surface and then Z 2 's trends surface (interpolated at Z 1 's resolution) to derive predicted Z 1 surface, as shown in Figure 1(e) (see also Ref. (18)). Error measures are shown in Table 2, indicating that collocated cokriging generated the best of all results obtained with Z 1 and R 1 -R 2 datasets.…”

“…This allows for the standardized cokriging (14,16,17,18) and, more importantly, the univariate kriging with bias-corrected Z 2 data.…”

“…To compare geostatistical methods for elevation data refinement and to highlight complications of scale mismatch and semantic differences for kriging-based refinement, two sets of digital terrain data (which were used also for Ref. (18)) were obtained over an area of approximately 45 km by 45 km near Zhangye in Gansu Province, China. One was from the SRTM, the other from GTOPO30, with the former used as the primary data-set and the latter as the secondary data-set, as shown in Figure 1(a) and (b), respectively.…”

“…A set of sampled data for Z 1 were obtained from the Z 1 DEM by selective sampling, in particular, the VIP (very important points) algorithm in ArcGIS (see also Ref. (18)). The sampled data-set for Z 1 consists of 250 SRTM grid cells, shown as dots in Figure 1(a).…”