Machine learning for the analysis of indoor radon distribution, compared with ordinary kriging

Abstract

Having a reliable forecasting tool is necessary to correctly identify radon prone areas, especially in cases where the variable of interest is the indoor radon concentration. An appropriate characterisation of the features of the buildings becomes fundamental. In this work, the results obtained (in global and local scale) using the following approaches for estimating the concentration of indoor radon at locations that were not sampled were compared: geostatistical model, based on ordinary kriging, and machine learning (ML) technique. In the first case, algorithms designed for the specific and fine treatment (by modelling the variographic structure) of the spatial component of the phenomenon were used, whereas in the second case a model that can also exploit information linked to other variables that characterise each single dwelling in which the measure was conducted was used. For locations having large errors, the ML approach provides better results, due to the information related to 'soil contact' and 'building material'.