Assessing the Effect of DEM’s resolution on Model Transferability of soil types: A case study of Lombardy region, Italy

Abstract

Mapping soil classes to retrieve information for specific soil management strategies according to capabilities and limitations of soil types is an important and very useful application of Digital Soil Mapping (DSM). DSM harnesses auxiliary data, including Digital Elevation Models (DEMs) and remote sensing information, to establish crucial relationship between soil characteristics and landscape attributes, facilitating the creation of soil maps. DEMs, offering detailed morphometric information about the Earth's surface, provide quantitative measurements of terrain features for GIS-based soil-mapping applications. Derived from DEMs, terrain attributes, including elevation, slope, aspect, and curvature profiles, along with secondary attributes like solar radiation and moisture index, play a pivotal role in characterizing spatial-specific landscape processes essential to soil formation. These morphometric attributes, integral to DSM due to their role in the paedogenetic process, have become indispensable auxiliary variables. The success of DSM depends heavily on the quality of input environmental covariates, with spatial resolution serving as a critical indicator. The spatial resolution of environmental covariates in DSM, often determined by the DEM's subjective spatial distribution, influences the modelling outcomes and processing efficiency. This study investigates the impact of DEM resolution on soil type classification and model transferability, focusing on the Lombardy region, Italy. Utilizing three different DEMs sources with resolutions of 5 m, 10 m, and 25 m, the research employs the Random Forest algorithm, and nested Leave-One-Out Cross-Validation (nested-LOOCV) techniques to assess model performance. The findings reveal a pivotal role for spatial resolution in determining model transferability, with distinct challenges observed during upscaling and downscaling. The study emphasizes the need for a nuanced approach to variable selection based on DEM resolution and provides valuable insights for optimizing soil classification models across diverse landscapes. The research contributes to advancing Digital Soil Mapping methodologies and underscores the significance of careful consideration of spatial resolution in enhancing the applicability of soil classification models.