Mapping Soil Texture Using Machine Learning With Terrain Features

High resolution soil maps are essential for soil management. It is necessary for applications such as precision farming, environmental remediation, and hydrological modeling. Proper management of soil as a natural resource can potentially boost agricultural productivity, which in turn is beneficial to the economy. This study aimed at predicting soil texture, represented by relative percentages of clay, sand, and silt, by modeling the relationships between known soil samples, and their topographical environment. The environmental features that were explored included a range of hydrological and morphometric features, representing the terrain characteristics in the Berg river primary catchment. These features included channel network distance, terrain ruggedness index, and topographic wetness index. A random forest machine learning algorithm was developed to determine which terrain features were the most beneficial in predicting the individual soil texture classes and the most important feature in the prediction was found to be elevation.

  

The soil texture samples, together with the most important terrain features were used to build a model to predict the individual soil classes. Sand and silt were shown to have the most robust models, but the clay predictions did not achieve as good results.  Three different elevation models were tested of which one was the 5m Stellenbosch University DEM (SUDEM). Future work will involve evaluating higher resolution DEMs such as the DEMSA2 provided by GeoSmart.

  

By using the methods as described in this study, soil texture can be mapped with higher accuracy, in less time, and will eliminate a degree of the subjective capturing process inherent in traditional soil mapping techniques.

 

 

 

 

 

 

The maximum percentage value for clay was 32% and the highest concentrations were in areas of elevation. Lowland areas and the regions around rivers had the lowest predicted concentration of clay. Sand, in contrast, had its highest predictions (maximum of 90%) in flatter areas and regions nearer to the coast. Areas with high elevation and steeper slopes can be seen to have low predicted concentrations of sand. The lowest concentration of sand predicted was 44%, which is higher than the maximum of clay or silt. The distribution of clay and silt appears very similar. Both have higher predicted percentages in areas with higher elevation and lower concentrations in flatter regions. The distribution of silt, however, is more uniform than that of clay. Areas of water accumulation, in general, can be seen more distinctly as having low silt predictions, and higher predictions can be seen in regions where there is a low probability of being near flowing water.

 

This technology is based on research carried out by Hanu Mostert, under the supervision of Dr Eric Mashimbye, as part of a BSc Honors Geoinformatics degree within the Department of Geography and Environmental Studies at Stellenbosch University.

image
2022 © UX Design & Development by NetMechanic