SOILAI

Soils support plant life and form an important aspect of ecological biodiversity. Soil data, however, is often absent from digital terrains but nevertheless remain important in areas of computational ecology. Digital elevation maps are a common artefact used to represent real-world terrains, however they only capture the elevation of a terrain which for certain applications might prove insufficient. In order to better represent real world terrains, we set out to devise a technique capable of producing soil property maps of the same dimensions and resolution as a source digital elevation map. Furthermore, the produced soil property maps are required to have plausibly come from the same terrain as the source digital elevation map.

We detail a method adopted from Digital Soil Mapping that employs Random Forests in order to generate soil property data given an existing digital elevation map. Furthermore, we supplement these soil property maps with associated relative prediction interval maps. These relative prediction interval maps are produced using quantile random forests and are used to convey the uncertainty associated with the outputs of the random forest model. We demonstrate the system’s ability on a series of digital elevation maps, and reason about the quality of Random Forrest outputs.