Predictive modeling of total Hg background concentration in soils of the Amazon Rainforest biome with support of proximal sensors and auxiliary variables

Abstract

Mercury (Hg) is a very well-recognized potential pollutant worldwide, and in the Amazon biome illegal artisanal gold mining practices tend to maximize this environmental problem. Considering the large extent of this biome and the evidences of Hg contamination, the main objective of this work was to predict the total Hg background concentration in Amazon soils using proximal sensors with aid of Random Forest (RF) and Support Vector Machine (SVM) algorithms, taking the total Hg content by direct analyzer as a reference. Soil texture, fertility properties and terrain attributes (elevation and slope) were considered as auxiliary variables in the modeling process. For that, nine contrasting sites of the Amazon rainforest biome were selected. In each site, eight locations (30 m apart) were carefully selected to collect composite soil samples at three depths: 0–20 cm, 20–40 cm, and 40–60 cm, totaling 216 soil samples. The pXRF data were separated into calibration (70%) and validation (30%) datasets for validation purposes. Total Hg in the studied soils ranged from 21.5 to 208 μg kg−1. RF models had better performance than SVM ones in Hg content prediction, showing the highest R2 and lowest RMSE values. Clay fraction content, total Al2O3, low degree of crystallinity Al forms, and stable elements (Nb, Zr, and Ti) positively correlated with total Hg. The combination of pXRF + magnetic susceptibility + soil texture data provided the best prediction models, however, pXRF data alone successfully predicted total Hg with R2 of 0.83. Saving time and cost and non-generating chemical effluents in prediction of total Hg background concentration in Amazon soils may be performed, constituting a secure basis for environmental regulation.