Effectiveness of a soil mapping geomatic approach to predict the spatial distribution of soil types and their properties

Highlights

• We assess landform and parent material influence on soil types and properties.

• We predict the composition of mapping units in terms of principal soil types.

• Soil spatial patterns are coherent with terrain attributes and parent materials.

• Models are more reliable in predicting absence rather than presence of soil types.

Abstract

A soil map (1:50,000 scale) was recently produced in Sardinia (Italy) using a cost-effective GIS approach. In this study we aimed to verify, in two pilot areas and by means of statistical analysis, the effectiveness of the adopted methodology in representing and predicting the spatial distribution of soil types and properties. We focused on evaluation of 1) the influence of landforms and parent materials on soil types (WRB Reference Soil Groups) and selected soil properties and 2) the suitability of the adopted methodology for calibrating a model to predict land unit composition in terms of different soil types.

Leptosols, Regosols and Cambisols were prevalent on slopes, with Leptosols being more frequent on convex slopes and Regosols and Cambisols on concave slopes. In flat areas, soil types mainly depended on the type and age of parent material, with Regosols and Cambisols prevailing on Holocene deposits and highly developed soils (mainly Luvisols) largely prevailing on Pleistocene deposits. On hard rock, Leptosols were very frequent on terrigenous metamorphic rock and frequent on granite. Besides Leptosols, Regosols occurred more frequently than Cambisols on both parent materials. Landforms strongly influenced soil depth and available water capacity. Soils on plains were deeper than those on slopes, where convex forms had shallower soils than concave forms. A similar trend applied to the available water capacity. The parent material had a significant effect on topsoil properties (thickness, texture, pH and organic carbon content) of soils belonging to the same WRB Reference Soil Group (analysis done on the most relevant WRB Reference Soil Groups, i.e. Leptosols, Regosols and Cambisols).

We calibrated and tested stepwise multiple linear regressions (MLR) and general linear models (GLM) to predict the composition of map units in terms of different WRB Reference Soil Groups. The two models gave very similar results, with distinct distribution patterns that were coherent with the relationships observed between soil groups and specific combination of terrain attributes and parent materials. Results showed that both models were more reliable in predicting the absence rather than presence of a given soil type.

Introduction

There is an increasing demand for soil information, since the knowledge and understanding of soil and how it is distributed across the landscape is considered essential for its effective use, management and conservation (Grealish et al., 2015). Consequently, soil information is also essential to help decision makers in land planning and in drafting environmental management policy (van Delden et al., 2011, Brungard et al., 2015), and may even be required by law (Vacca et al., 2014).

Soil information can be provided by soil maps, which are graphic representations for transmitting information about the spatial distribution of soil attributes (Yaalon, 1989). In general terms, soil maps can be produced in a conventional or in a digital way. Conceptually, conventional soil maps and digital soil maps (DSM) are very similar (Kempen et al., 2012), since both approaches use a soil-landscape model to predict soil at unobserved locations (Hudson, 1992). The main difference is that in conventional soil maps the soil-landscape model is a qualitative model based on soil surveyors’ expert knowledge, while in DSM the soil-landscape model is quantitative. Comprehensive overviews of DSM were provided by McBratney et al., 2003, Grunwald, 2006, Minasny and McBratney, 2016, and Arrouays et al. (2020). Among the first conceptualizations of DSM, McBratney et al. (2003) formalized the so-called scorpan model as “empirical quantitative descriptions of relationships between soil and other spatially referenced factors with a view to using these as soil spatial prediction functions”. The possibility of producing DSM strongly depends on the availability of ancillary data (Zeraatpisheh et al., 2017), including existing soil data (e.g. polygon-based soil maps and soil profile databases), which can serve as both training and validation datasets (Zhang et al., 2017). Consequently, in areas with limited existing soil data, producing an accurate DSM can be challenging (Stoorvogel et al., 2009), so this method has rarely been used for routine production mapping or addressing land management questions (Grealish et al., 2015). In these areas, pragmatic and easy-to-apply relationships for predicting soil properties under different environmental conditions, and assist in soil data collection, are needed to provide answers for the current issues that require a fast delivery of information (Gray et al., 2009, Grealish et al., 2015). Several different statistical approaches have been tested to generate quantitative predictions of categorical soil variables from limited samples with the general aim of producing soil maps for unsampled or sparsely sampled areas at different scales, from national to sub-regional (Minasny and McBratney, 2016). Grunwald (2009) provided a multi-criteria characterization of digital soil mapping and modelling approaches, classifying DSM techniques in three wide categories based upon predictor variables and modelling approaches, which can then integrate data driven statistical approaches with pedotransfer functions and dynamic mechanistic modelling of soil properties.

In Sardinia (Italy), the use of soil information and maps in land use planning is specifically required by law (RAS, 2006, RAS, 2008). Because the scale of the three available soil maps covering the island (Arangino et al., 1986, Aru et al., 1990, Madrau et al., 2006) was considered not adequate for local land planning strategies, a new project was recently initiated for the production of a new soil map, at a scale of 1:50,000. The general structure of the project and the methodology used were described in Vacca et al. (2014). The existing soil dataset (point data and maps) was considered insufficient and inappropriate to produce a DSM without resorting to the support of ancillary variables. Adopting a cost-effective approach, existing digital environmental data, along with soil data, were therefore used to delineate homogeneous spatial areas in terms of soil, geological substrate, landform, and land cover in a GIS environment.

This paper aimed to verify, in two of the pilot areas and by means of statistical analysis, the effectiveness of the adopted methodology (Vacca et al., 2014) in representing and predicting the spatial distribution of soil types and soil properties. This is considered crucial, as it affects the reproducibility of the model. There appears to be a need for clarification of the quantitative relationships between soil properties and environmental covariates in order to reduce the uncertainty of the model and allow better prediction.

The specific objectives of this paper were to (1) evaluate the influence of landforms and parent material on soil types; (2) evaluate their influence on soil properties; and (3) evaluate if the adopted methodology is suitable for calibrating a model to predict land units composition in terms of principal soil types.