Mapping the major soil-landscape resources of the Ethiopian Highlands using random forest

Abstract

Geospatially explicit information of soil-landscape resources of Ethiopia is lacking or fragmented for much of the country. Recently, massive soil data were collected, however these are limited to properties related to soil fertility and valid for the topsoil only. Understanding the country’s soil-landscape resources, including their qualities and constraints beyond the topsoil, remains key information for systematic and reliable scaling up of evidence-based agricultural best practices including soil fertility management recommendations. The objective of this study was to produce a coherent dataset of the major soil-landscape resources of 30 highland woredas (districts), contributing to the Agricultural Growth Program of the Government of Ethiopia. The study started with an exploratory survey to identify the major (most common) soils occurring across the landscapes followed by a full survey to assess the distribution of the identified major soils. Representative soil profiles were characterised from soil pits and classified as Reference Soil Groups (RSGs), with prefix qualifiers (PQs), according to the World Reference Base for soil resources (WRB). A large number of soil profiles was classified from auger observations. Observed soil classes at both RSG and RSG + PQ level were combined with spatial explanatory variables (covariates), representing the soil forming factors in the landscapes, and their relationships were modelled and validated by random forest. A multitude of tree models was trained using each profile for calibration in approximately two third and cross-validation in approximately one third of the models. Cross-validation showed that RSGs were predicted with a reasonable overall purity of 0.58 and RSGs + PQ were predicted with a purity of 0.48. The most relevant covariate in the models was the Geomorphology and Soils map of Ethiopia at 1: 1 M scale disaggregated into soil-landscape facets. Next models were used to predict soil classes across woredas which resulted in a 250 m resolution raster map of the most probable major soils. This raster map was generalised into a polygon map of major soil-landscape resources. The purity of this final map was estimated to be 0.54 for RSGs and 0.45 for RSGs + PQ. Soil properties relevant for agricultural interpretation, such as depth, drainage, texture, pH, CEC and organic carbon and nutrient contents, were mapped according to the RSGs depicted on the soil-landscape resources map with a RMSE/mean ratio of on average 42%. We conclude that soil expert knowledge and conventional soil-landscape survey combined with random forest modelling results in an attractive hybrid approach. The approach proves cost-effective and sufficiently accurate and can be used to inform scaling up of evidence-based agricultural best practices.