Mapping Soil Depth in Sri Lanka Using Statistical and Machine Learning Models

Although harmonised soil profile data are available, no attempt has been made to comprehensively map soil depth to bedrock across Sri Lanka. In this paper we describe the development of a digital soil depth map for Sri Lanka using a soil profile dataset and an extensive array of environmental covariates. We compared the ability of kriging and machine learning models to assess and reproduce the variability of soil depth across the country. The result of 10-fold cross-validation showed that both methods could map the variability to a limited degree. Kriging methods tend to reduce the variability to produce smoother patterns, while random forest methods could capture the range of all possible values for soil depth without the need for any spatial random process modelling. Random forest method has the advantage of less pre-processing requirements if good number of covariates are available. Kriging in contrast did not perform well with the presence of covariates. The resulting map can be used to provide essential depth information for environmental models that are sensitive to soils with variable depth in Sri Lanka