Land-Surface Segmentation as sampling framework for soil mapping

Current sampling methods require a large number of samples to account for spatial variation of environmental covariates, which often conflicts the available financial resources. Thus, efficient sampling strategies are desirable. The aim of this study was to evaluate the potential of land-surface segmentation in stratifying a landscape into homogeneous areas, which can be used as support in optimizing soil sampling. The experiments were carried out in a study area where soil samples were available. Land-surface variables were derived from DEMs and segmented with a multiresolution segmentation (MRS) algorithm, into objects considered as homogeneous soil-landscape divisions. Thus, one sample was randomly selected within each segment, based on which predictions of the A-horizon thickness and soil types, were made. Predictions based on the land-surface segmentation sampling schemes outperformed predictions based on simple random sampling and conditioned Latin hypercube, respectively. INTRODUCTION Spatial resolutions of soil maps for about 70 % of the Earth’s ice-free land surface are too low to help with practical land management [1]. Conventional survey methods involve too much resources to be cost-effective in high-resolution soil mapping. Digital Soil Mapping (DSM) is an appropriate framework for producing detailed soil maps based on quantitative relationships between soil properties or types and their ‘environment’ [2]. Efficient sampling designs play an important role in DSM [2], as they have a significant impact on the accuracy of the maps [3]. Classical sampling methods (e.g. simple random sampling, systematic sampling and stratified sampling) as well as the model-based sampling strategy require a large number of samples to account for the spatial variation of environmental variables [4]. As sampling is constrained by financial resources, efficient sampling strategies are desirable [5]. Increasingly available geospatial information (e.g. satellite imagery, geology maps, Digital Elevation Models (DEMs) can be exploited as environmental covariates to optimize sampling locations [5] within the framework of a soil-landscape model [6]. However, sampling with support of environmental covariates has not been fully developed in DSM [5]. A number of recent papers [e.g. 4, 7, 8] demonstrated the value of purposive mapping based on such covariates in producing more accurate predictions by using fewer, but more representative samples. Land-surface segmentation (LSS) is a relatively new technique to partition land-surface variables (LSVs) obtained from DEMs into contiguously homogeneous areas in multivariate feature space [9]. The most popular segmentation algorithm is Multiresolution Segmentation (MRS) as implemented in the eCognition® software [10]. The algorithm merges spatially contiguous pixels or cells into segments based on local homogeneity criteria [10]. The resulting land-surface objects incorporate scale, spatial autocorrelation, anisotropy and non-stationarity in their definition of homogeneity [11]. There have been only a few attempts to map soils based on LSS. The only approach of segmentation to optimize soil sampling [12] showed that a segmentation-based sampling (SBS) scheme produced better distribution of sampling locations over the area of interest, as compared to simple random sampling and regular sampling schemes. It is clear that the potential of LSS to DSM has not been fully employed and the applicability of this technique to optimize soil sampling has only been touched upon. Therefore, we aimed at evaluating the potential of LSS in stratifying a landscape into homogeneous areas, which can be used as support in optimizing soil sampling.