ABSTRACT
Missing soil parameters are common in soil investigation and published literature due to the fact that undisturbed soil samples used in a laboratory test cannot be used for another testing as the soil has been disturbed and no longer representative of the actual field condition. The issue of missing soil parameter is further exacerbated in the event of high heterogeneity of the ground, where soil parameters obtained from a specific depth in the ground may not be reflective of another depth in the same borehole. Therefore, soil parameters without direct measurements have to be inferred from other available soil parameters. This is the part where machine learning becomes useful. However, many machine learning models do not handle missing data automatically. This paper presents the attempt to modify existing machine learning algorithms to achieve the ability of filling missing values in soil datasets.
