Novel Approaches to Soil Mapping Using Remote Sensing and Ensemble Learning Methods.

There is currently increasing concern about the impacts of land degradation on ecosystem services and the ability of biological systems to meet human needs, as well as about impacts on biodiversity. However, the extents and severity of soil degradation, which is arguably the most important aspect of land degradation today, are not well understood at landscape to regional to continental scales. Conventional soil maps generally do not help bridge this information gap and information on soil degradation risk factors such as soil erosion prevalence across landscapes, as well as soil functional properties or soil condition, are needed. In this paper, ensemble learning is explored for mapping of soil organic carbon (SOC) and pH, along with soil erosion prevalence using remote sensing data. Ensemble learning is a paradigm where several learners, or models, are trained to solve the same problem and it can be applied to solve both regression and classification problems. In the current study these methods are applied to the development of maps based on ground data from 38 sites covering the main climate zones of Africa, except true desert areas, by relating ground measurements and soil properties to moderate resolution Landsat satellite imagery. Further, the uncertainty of the generated pH and SOC maps is assessed by calculating the coefficient of variation (CV) between each prediction model in the ensemble for each image pixel.