Exploring soil pollution patterns in Ghana's northeastern mining zone using machine learning models

Abstract

• Positive skewness and high kurtosis values suggest uneven distribution and localized contamination hotspots. • Pollution indices indicated varying degrees of pollution across different parts of the area. • MARS models effectively predict pollution indices based on heavy metal concentrations. • Low RMSE and MAE values suggest high accuracy and generalizability of the models. This study assessed the pollution status and effectiveness of machine learning models in predicting pollution indices in soils from a mining area in Northeastern Ghana. 552 soil samples were analysed with an Energy Dispersive X-ray Fluorescence (ED-XRF) spectrometer for their elemental concentrations. Four pollution indices; Nemerow Integrated Pollution Index (NIPI), degree of contamination (Cdeg), modified degree of contamination (mCdeg) and Pollution Load Index (PLI). Additionally, the Multivariate Adaptive Regression Splines (MARS) machine learning approach were used. The high CV%, skewness, and kurtosis values show a high degree of variability and uneven distribution patterns which denotes dispersed hotspots that can be interpreted as an influence of gold anomalies and illegal mining activities in the area. V (120.86 mg/L), Cr (242.42 mg/L), Co (30.92 mg/L) Ba (337.62 mg/L), and Zn (35.42 mg/L) recorded values higher than the global and regional contaminant thresholds. The NIPI shows that 46.74% and 26.81% of samples are slightly and moderately polluted respectively. The Cdeg analysis supports these findings, with 36.96% and 41.49% of samples classified as having “moderate” to “considerable” contamination, respectively. The PLI indicates progressive soil quality deterioration (43.84%) of samples reflecting substantial environmental disturbance. The pollution indices show the effect of illegal mining on Shaega, Buin and other areas in the eastern boundary of the study. The MARS models developed for the study demonstrated high predictive capabilities with an R 2 value of 0.9665 for model 1 (NIPI), and RMSE and MAE values of 0.8227 and 0.4287 respectively. For model 2 (Cdeg), R 2 value of 0.9863, RMSE and MAE of 1.0416 and 0.6181, respectively. Model 3 (mCdeg) produced an R 2 value of 0.9844, RMSE and MAE of 0.1225 and 0.0670. These findings suggest MARS models can be an integral tool for soil quality analysis in cooperation with pollution indices. The study suggests that remedial and legislative measures be implemented to address the issue of illegal mining in the area.