Search

This poster applies a dropout-regularised artifical neural network, constructed in the MXNet framework, to map arsenic enrichment in south west England. The network models the relationships between arsenic (as a centred log-ratio from XRF analyses of 3395 stream sediment samples) and high resolution geophysical data. The resultant model, trained to optimal accuracy using early stopping, achieves an R2 of 0.7 on held-out test data - a promising level of accuracy for predictions in a complex hydrothermal mineralisation system, and warrants further investigation into the development of more sophisticated network architectures for geohazard and prospectivity mapping purposes.

This poster applies a dropout-regularised artifical neural network, constructed in the MXNet framework, to map arsenic enrichment in south west England. The network models the relationships between arsenic (as a centred log-ratio from XRF analyses of 3395 stream sediment samples) and high resolution geophysical data. The resultant model, trained to optimal accuracy using early stopping, achieves an R2 of 0.7 on held-out test data - a promising level of accuracy for predictions in a complex hydrothermal mineralisation system, and warrants further investigation into the development of more sophisticated network architectures for geohazard and prospectivity mapping purposes.