Ambient temperature and relative humidity–based drift correction in frequency domain electromagnetics using machine learning

Electromagnetic instrument responses suffer from signal drift that results in a variable response at a given location over time. If left uncorrected, spatiotemporal aliasing can manifest and global trends or abrupt changes might be observed in the data, which are independent of subsurface electromagnetic variations. By performing static ground measurements, we characterized drift patterns of different electromagnetic instruments. Next, we performed static measurements at an elevated height, approximately 4 metre above ground level, to collect a data set that forms the basis of a new absolute calibration methodology. By additionally logging ambient temperature variations, battery voltage and relative humidity, a relation between signal drift and these parameters was modelled using a machine learning (ML) approach. The results show that it was possible to mitigate the effects of signal drift; however, it was not possible to completely eliminate them. The reason is three-fold: (1) the ML algorithm is not yet sufficiently adapted for accurate prediction; (2) signal instability is not explained sufficiently by ambient temperature, relative humidity and battery voltage; and (3) the black-box internal (factory) calibration impeded direct access to raw data, which prevents accurate evaluation of the proposed methodology. However, the results suggest that these challenges are not insurmountable and that ML can form a viable approach in tackling the drift problem instrument specific in the near future.