Implementation and evaluation of existing knowledge for digital soil mapping in Senegal

Abstract

Digital soil mapping approaches that require quantitative data for prediction are difficult to implement in countries with limited data on soil and auxiliary variables. However, in many such cases there is a wealth of qualitative information available, such as profile descriptions, catenas or general purpose soil surveys. This type of information opens possibilities for more qualitative approaches to digital soil mapping when quantitative mapping is unfeasible. In this study we used a classification tree approach combined with literature and a small dataset of 40 point SOC observations to map the topsoil organic carbon (SOC) content for a data-poor environment in the Senegalese Peanut Basin. A literature review provided an overview of the driving factors of soil variability in the Peanut Basin. Geomorphology, topography, vegetation, and land use were identified as the main factors explaining the spatial variation of SOC in the Peanut Basin. These factors were represented in a classification tree by variables that were derived from a digital elevation model and a satellite image. Threshold values and actual predictions for the classification tree were based on literature and the small soil dataset. Next the classification tree was used to create a map of SOC for the study area. Using cluster random sampling, 155 locations were sampled for validation. Validation of the model results showed a poor model performance with large prediction errors. Error analysis showed that although the variables that were used to predict SOC were important sources of variability, a larger soil dataset is needed to better calibrate the classification tree model. Calibration of the classification tree on the basis of the validation dataset produced much improvement and acceptable results after cross-validation. It is concluded that digital soil mapping on the basis of existing knowledge and general auxiliary information is feasible, provided that a sufficiently large and appropriately collected soil dataset is available for calibration.