Digital Elevation Data Fusion Using Multiple-Point Geostatistical Simulation

This paper proposes advanced methods based on multiple-point geostatistical simulations (MPSs) for data fusion in digital elevation models (DEMs). MPS-based methods can capture similar patterns using the spatial correlation residing in the training images and then reproduce the area of interest in terms of spatial patterning through conditional simulation. For this, the FILTERSIM algorithm, which is based on MPS, can be applied to augment the prediction results from traditional geostatistical approaches to derive fused datasets from complementary sources. However, the FILTERSIM algorithm discerns similar patterns based on a filter-based prototyping process and does not account for nonstationarity of spatial structures. Thus, an enhanced data fusion method based on a modified FILTERSIM algorithm was proposed, in which new strategies for forming prototypes were incorporated. An experiment was carried out in a study area in Northwest China to test the proposed methods. The aim is to produce fused DEM datasets at fine spatial resolution by integrating fine-resolution but sparsely sampled SRTM data and regularly gridded coarse-resolution GMTED2010 data, and then to test three data fusion methods: 1) the traditional geostatistical interpolation, which provides baseline results for comparison; 2) the conventional FILTERSIM algorithm; and 3) a modified FILTERSIM algorithm. For both the first and second methods, three different kriging approaches were used; for the third method, three new schemes for building prototypes were tested for an improvement in performance. The results show that the improved FILTERSIM method can achieve greater accuracy and retain more spatial structure than the other two methods.

I. Introduction

Terrain data are important data sources for various applications, including hydrology [1] and landslide modeling [2]. Terrain elevation data are acquired from various sources, such as topographic maps and remote measurements, which are often stored as digital elevation models (DEMs). Generally, there are three types of elevation data: point DEMs, grid DEMs, and digitized isolines. Numerous methods have been proposed for DEM interpolation to derive high-quality DEM data, including nearest neighbor [3], bilinear [4], cubic convolution [5], spline methods [6], inverse distance weighting [7], Thiessen polygons [8], and geostatistical interpolators [9], [10]. However, interpolating DEM data from a single source cannot always achieve the standard for DEM products because it may not be possible to enhance the accuracy of a DEM dataset (with a specific resolution) merely by interpolation due to the loss of subpixel (grid cell) information in the original data sampling, which cannot be recovered by interpolation. With the rapid growth of multisource geospatial information, data are usually obtained at different spatial resolutions with different accuracies. Spatial data fusion seeks to generate new datasets, which are meant to have increased accuracy, properly refined spatial resolution, or enhanced applicability and utility, from multisource data that are complementary in terms of sampling density, resolution, and accuracy. As a result, the new datasets will collectively be more informative or useful than an individual source dataset alone. In particular, there seems to be merit in the integration of multisource spatial data for improved quality in terms of accuracy and/or resolution. Here, we consider the terms integration, conflation, and fusion as being synonymous, acknowledging that fusion is the most frequently used term. Data fusion has been widely used in geographical information systems, with the aim of synergizing geospatial information from different sources. A review and comparison of all the fusion methods are beyond the scope of this paper. In this study, we only focus on the fusion methods needed to achieve data refinement. Although elevation data with low spatial resolution are easily available (e.g., GTOPO30 and GMTED2010), it is an expensive and time-consuming process to obtain elevation data with high spatial resolution over a large area (e.g., LiDAR) [11], [12]. Data refinement of DEMs can be accomplished by the fusion of two or more data sets. It has been suggested that data integration can refine the spatial resolution (i.e., downscale) [13] or increase the accuracy of DEMs [14], and it can provide a solution to the accuracy assessment [15].