EVALUATION OF RANDOM FOREST–BASED ANALYSIS FOR THE GYPSUM DISTRIBUTION IN THE ATACAMA DESERT
Keywords: GIS, remote sensing, random forest, satellite imagery, indices, digital soil mapping
Abstract. Gypsum-rich material covers the hillslopes above ∼ 1000 m of the Atacama and forms the particular landscape. In this contribution, we evaluate random forest-based analysis in order to predict the gypsum distribution in a specific area of ∼ 3000 km2, located in the hyperarid core of the Atacama. Therefore, three different sets of input variables were chosen. These variables reflect the different factors forming soil properties, according to digital soil mapping. The variables are derived from indices based on imagery of the ASTER and Landsat-8 satellite, geomorphometric parameters based on the Tandem-X World DEM™, as well as selected climate variables and geologic units. These three different models were used to evaluate the Ca-content derived from soil surface samples, reflecting gypsum content. All three different models derived high values of explained variation (r2 > 0.886), the RMSE is ∼ 4500 mg∙kg−1 and the NRMSE is ∼ 6%. Overall, this approach shows promising results in order to derive a gypsum content prediction for the whole Atacama. However, further investigation on the independent variables need to be conducted. In this case, the ferric oxides index (representing magnetite content), slope and a temperature gradient are the most important factors for predicting gypsum content.