Modeling regolith thickness in iron formations using machine learning techniques

Highlights

• We apply the Random Forest algorithm to predict regolith thickness.

• The most important covariates were the drainage density and east–west direction.

• The average regolith thickness for the Quadrilátero Ferrífero was 125 m.

• The model explained only about 40% of the regolith thickness variation.

Abstract

The Quadrilátero Ferrífero (QF) in Brazil is a region of great economic, social and environmental importance, involving conflicts of interest due to land use and heavily pressured by iron mining and urban sprawl. This is an area of environmental importance due to supporting rupestrian vegetation areas on ferruginous substrates and springs from important watersheds. Thus, studies that can bring more knowledge about this region becomes important to support future decisions based on technical information. We used Random Forests algorithm and several databases to model the regolith thickness of the entire QF region and we also created an individual model to predict regolith thickness in the lithostratigraphic unit Minas Supergroup that contains most of the drillhole samples. The regolith thickness modeled for the QF region presented an average of 125.32 m and R² of 0.38, and the specific model for the Minas Supergroup also predicted the average regolith thickness of 125.39 m and R² of 0.39. These values are in accordance with the average regolith thickness (124 m) data from the drillhole database obtained from exploratory programs for iron ore in the QF region. The most important predictive covariates included drainage density, east–west direction, terrain texture and vertical distance from drainage. This study is the first attempt to model the regolith thickness in this important region and the analysis of model uncertainty can orient future studies.

Introduction

The regolith is considered the weathering mantle that extends from the fresh rock to the surface, involving saprolite and pedolite (solum), where rocks undergo a transformation of their composition and features by removal or addition of physical and chemical materials (Taylor and Eggleton, 2001). Factors controlling the depth and distribution of regolith are complex and reflect the interplay through of different processes, including slope, land use, curvature, parent material, weathering rate, climate, vegetation cover, upslope contributing area, and lithology (Dixon et al., 2009, Tian et al., 2019). For instance, Patton et al. (2018) found a strong linear relationship between soil thickness and hillslope curvature.

Knowledge of the regolith thickness is particularly important because it can influence the way water is stored and moves through the landscape (Kew and Gilkes, 2006), the nutrient availability (Graham et al., 2010), and it may provide information about the rate and intensity of natural disasters such as landslides and erosion (Lebedeva and Brantley, 2013). Although global maps of regolith thickness are available (Pelletier et al., 2016) at 1 km resolution, high resolution maps are still scarce. Information about regolith thickness can be derived from direct measurements in the exposed erosional landscapes and hammering a rod technique (Basharat et al., 2018) and also from drillholes. There is a dearth of specific depth information in most regolith mapping (Craig, 2001, Farooq and Govil, 2014, Tripathi and Govil, 2020), mainly due to the difficulties to create a dense network of measurements of regolith thickness (Kuriakose et al., 2009, Shafique et al., 2011).

Numerical modeling techniques have been widely used in the Earth Sciences (Anees et al., 2016, Ouyang et al., 2017, Sashikkumar et al., 2017). The most recent advances involving the use of artificial intelligence, especially machine learning techniques (Caté et al., 2017, Lary et al., 2016), combined with the opportunistic direct field observations, enhance the processing of a large amount of data (big data) to extract information that helps in understanding the processes of the earth dynamics. For instance, Wilford et al. (2016) estimated the thickness of the Australian weathering mantle from drillhole data using the Cubist algorithm. Shafique et al. (2011) predicted regolith thickness using a multivariate linear model based on landform, elevation and distance to stream in northern Pakistan.

In this paper we apply the Random Forests algorithm (Breiman, 2001) to develop a nonparametric regression model (Lary et al., 2016) of the regolith thickness in the Quadrilátero Ferrífero (QF), Minas Gerais, Brazil. The QF has one of the largest concentrations of lateritic iron-ore deposits in the world (Salgado and Carmo, 2015), and there are conflicts of interest due to land use that are heavily pressured by urban sprawl and mining activities (Sonter et al., 2014). This region also has an environmental relevance for protecting rupestrian vegetation areas on ferruginous substrates (Jacobi et al., 2007) and springs from important watersheds such as Rio das Velhas and Rio Doce (Costa et al., 2003).

Within this perspective, many studies on the most superficial portion of this terrain have been intensified, including those on the characterization of soil genesis (de Carvalho Filho et al., 2015, de Carvalho Filho et al., 2010, Schaefer et al., 2016). However, there is limited information about the general aspects of the regolith developed on QF in the literature (Spier et al., 2006). Thus, this study can improve our knowledge about this interesting landscape and also provide opportunities for future decision-making based on technical information.