Semi-automatic disaggregation of a national resource inventory into a farm-scale soil depth class map

Abstract

Knowledge of soil depth spatial variability is important for land use management especially in dryland agriculture regions, which rely on climate and soils to provide adequate water and nutrients during the growing season. Soil spatial variability can be predicted from legacy soil data through machine learning techniques producing quantitative soil maps requiring minimal resources. South Africa has a country wide 1:250,000 scale resource map known as the Land Type Survey (LTS) which includes soil properties such as soil depth, soil class, root limiting layer, clay content, and texture. Each LTS polygon (land type), is comprised of unique soil – terrain patterns and is therefore, not a true soil map. This study aims to disaggregate the LTS into a farm-scale soil depth class map through a two-step disaggregation approach. First, landform elements were predicted through a pattern recognition algorithm known as geomorphons. Geomorphons, together with the original LTS were overlaid to produce polygons with unique distributions of soil. The polygons were disaggregated further to produce a raster map of soil depth classes through a soil map disaggregation algorithm known as DSMART. The first most probable class raster achieved an accuracy of 68% and for the two most probable class rasters, an accuracy of 91% was achieved. The two-step approach proved necessary for producing a farm-scale soil map. The result of this study is significant as it produced a soil depth class map from a national resource map at a scale and resolution (10 m) suitable for farm management.