A coupled geostatistical and machine learning approach to address spatial prediction of trace metals and pollution indices in sediments of the abandoned gold mining site of Bekao, Adamawa, Cameroon.

Trace metals present in high amounts in aquatic systems are a perpetual concern. This study applied geostatistical and machine learning models namely Ordinary Kriging (OK), Ordinary Cokriging (OCK) and Artificial Neural Network (ANN) to assess the spatial variability of trace metals and pollution indices in surface sediments along the Lom River in an abandoned gold mining site at Bekao (Adamawa Cameroon). For this purpose, thirty-one (31) surface sediment samples are collected in order to determine the total concentrations of As, Cr, Cu, Fe, Mn, Ni, Pb, Sn and Zn. These trace metals are used to compute pollution indices as the sediment pollution index (SPI), the Nemerow index (NI), the modified contamination degree (mCD), and the potential ecological risk assessment (RI). OK, OCK and ANN models are compared to determine the best model performance. The best models are selected based on the values of the root mean square error (RMSE), the coefficient of determination (R), the scatter index (SI) and the BIAS. Results showed that the sequence of trace metal mean concentrations in the sediments is Fe > Mn > Cu > Ni > Sn > Cr > Zn > Pb > As. The mean concentrations of Ni, Cu, Zn and Sn are above the average shale values (ASV) and the pollution status is globally moderate to significant with a low potential ecological risk. The spatial dependency obtained with semivariogram models are moderate to weak for Mn, Fe, Ni, Pb, SPI, NI, mCD, RI As, Cr, and Sn and strong for Cu and Zn. According to cross-validation parameters, ANN model is the best method for the prediction on trace metal concentrations and pollution indices in surface sediments along the Lom River in the abandoned gold mining site of Bekao.