Are fine resolution digital elevation models always the best choice in digital soil mapping?

Abstract

Digital Elevation Models (DEMs) are fundamental in assessing soil variability and are regularly used in Digital Soil Mapping (DSM) as a scale benchmark for all the other environmental covariates. They are conventionally used at their original grid resolution with a conventional 3×3 window size for the computation of the derived terrain attributes. The choice of scale frames the analysis and shapes the end result suggesting that better attention and quantitative knowledge of scale may improve predictive performance. Previous studies have shown the influence of pixel size but have not investigated in detail the interacting effect between window and pixel sizes. The aim of this study was to examine the scale dependency of soil classification performance at the landscape scale using two machine-learning techniques commonly applied in DSM: artificial neural networks and random forest. These were applied in three different areas in terms of their geomorphology and soil type located in Ireland. A series of DEMs representing different scales were created from the original 20m DEM by smoothing and re-sampling it with different window and pixel sizes for a total of 143 combinations: the original, 10 smoothed but not re-sampled, 12 re-sampled and not smoothed, and 120 smoothed and re-sampled. These were used to generate 11 terrain parameters from which 4 points per km2 were randomly extracted and used to predict soil classes. The overall prediction accuracy in the three study areas varied between 35% and 60%. Pixel size was found significant in all areas, the interaction between window and pixel sizes significant in morphologically rough areas and window size was significant only in flat homogeneous areas at coarser resolutions (above 140m in this study). In general, predictive performance was best at very fine and very coarse scales in morphologically varied areas, coarse scales in flat homogenous areas and relatively scale invariant in mixed areas. We conclude by examining whether this empirical approach is appropriate to compare scale combinations to obtain a better prediction accuracy of DSM techniques.