Using machine learning to update soil survey

Loading...
Thumbnail Image

Date

2007-04-02T13:35:18Z

Journal Title

Journal ISSN

Volume Title

Publisher

Research Projects

Organizational Units

Journal Issue

Abstract

Soil survey in the recent past seems to be taking a paradigm shift with the advent of various geospatial and pedometric techniques. The impetus for this includes both the usability and the limitations associated with traditional soil survey products. This research explores the use of machine learning and GIS tools for updating an existing soil survey of Monroe County in Southeastern Ohio. A soil-landscape modeling framework was adopted to predict soil series based on a number of high-resolution geospatial environmental correlates. Base data layers included the existent soil survey, climate attribute surfaces (precipitation and evapotranspiration), historic vegetation, terrain attributes derived from digital elevation models, and bedrock geology. The old soil survey was randomly sampled to generate a pre-classified training set containing the target soil series and their environmental correlates. Two machine learning algorithms (J48 and Random Forests) were used to build classification models. The built models were then applied to the entire county to generate digital soil maps. The models predicted the correct dominant soil series about 60 percent of the time. When compared with the existent soil survey map, the digital soil map was able to predict even the components with in a soil mapunit. Machine learning can efficiently be used to get new insights into the traditional soil maps and can be used as a guide for further field investigations.

Description

Keywords

machine learning, soil survey

Citation