Explorative analysis of varying spatial resolutions on a soil type classification model and it's transferability in an agricultural lowland area of Lombardy, Italy

Abstract

In Digital Soil Mapping (DSM), assessing the transferability of soil type classification models across different spatial resolutions is a pivotal step in ensuring their robustness and applicability to diverse terrains. This study investigates the impact of spatial resolutions on soil type mapping within an intensively used agricultural lowland region in Lombardy, Italy, based on a Random Forest algorithm. Employing Digital Elevation Models (DEMs) at resolutions of 5 m, 10 m, and 25 m, this study aims to identify the optimal spatial resolution for accurate soil type maps and explores the transferability of models across different resolutions. The nested Leave-One-Out Cross-Validation (nested-LOOCV) results indicate a substantial impact of resolution on model performance, with higher resolutions demonstrating superior accuracy. The model developed at 10 m resolution emerges as the most robust performer, achieving an overall accuracy of 40.3%. Model transferability analysis reveals challenges when transitioning from finer to coarser resolutions, while models at coarser resolutions adapt favourably to higher resolution data. The implications extend to DSM, emphasizing the need for careful consideration of spatial resolution in model development and transfer. The findings provide valuable insights for researchers and practitioners, urging tailored approaches based on the scale and objectives of the study area. The study encourages future research to focus on advanced techniques enhancing model transferability within DSM. Overall, this research contributes to the optimization of soil classification models, advancing our understanding of soil taxonomy in agriculturally vital lowland areas.