Caracterização de rejeitos de mineração de ferro pós-rompimento da barragem de Fundão utilizando sensores proximais para predição de atributos de interesse agrícola e ambiental

Abstract

After the Fundão dam collapsed in 2015, approximately 60 million m3

of iron-rich tailings were deposited in floodplains and river banks, impacting an extensive area. The rehabilitation of the impacted area is mandatory and, for this, one of the crucial factors is the detailed characterization of sediments and soils. Conventional laboratory-based methods are time consuming, require large amounts of chemical reagents and generate polluting residues. These drawbacks have been overcome by using different sensors to obtain some signal (e.g., fluorescence, reflectance) from the analyzed material that reflects the chemical composition. Among these sensors, portable X-ray fluorescence (pXRF) has been widely used to obtain the total elemental composition of soils and to predict attributes of interest. In this work, it was hypothesized that pXRF data may be important for predicting attributes of agricultural and environmental interest in a representative scenario of the impacted area. In addition, two other sensors that reflect the chemical composition were tested together with pXRF. They are: a portable visible color sensor (NixTM Pro color sensor) and a magnetic susceptibility meter. The objectives were: i) to characterize several samples of iron-rich tailings collected after the failure of the Mariana Fundão dam in Brazil using three proximal sensors (pXRF, Nix ProTM and magnetic susceptibilimeter); ii) assess the contribution of each sensor to differentiate impacted and non-impacted areas; iii) predict the semi-total concentration of potentially polluting elements and agronomic properties of the soil (pH, cation exchange capacity, organic carbon, macro- and micronutrients and texture). For that, 148 surface samples of tailings and/or soil collected along a 47-km section of the Gualaxo do Norte River were used. The samples reflect different vegetation conditions, land use and degree of impact. In the laboratory, characterization of soil fertility properties was carried out by conventional

methods, determination of the elemental composition by pXRF, determination of the semi- total concentration of elements regulated by environmental legislation by the USPEA 3051a

method, obtaining the RGB color parameters and magnetic susceptibility. The data obtained were submitted to descriptive statistical analysis (minimum, maximum, mean, median, standard deviation e coefficient of variation), principal components analysis (PCA) and prediction via machine learning (random forest). PCA analysis using sensor data in isolation

or combined to Nix Pro and MS data did not allow a clear differentiation of impacted and non- impacted areas. Important soil fertility properties (e.g., pH, CEC, texture, macro- and

micronutrients) were accurately predicted using random forest model. For environmental purposes, semi-total concentrations of potentially polluting elements were also well predicted using the sensors data. It is concluded, therefore, that the use of proximal sensors and prediction models can greatly contribute for in situ and rapid characterization of an extensive impacted area. Also, the approach tested here can be extrapolated and used in other similar situations.