Upscaling soil and vegetation measurements in an alpine catchment using remote sensing and statistical modelling

Abstract

Soil texture is an important soil physical property that determines water availability, nutrient availability and growth of vegetation. The growth of plant species and development of high-alpine soils are shifting in alpine catchments due to changes in conditions as a result of e.g. warming at higher elevation, shorter periods of snow cover and less precipitation falling as snow. Hence, understanding hydrological behaviour in high alpine catchments is crucial. Therefore, insight in the spatial distribution scale of soil texture and vegetation in an alpine catchment is required. In this research, field data obtained during 3 weeks of fieldwork in the Meretschibach catchment was used in combination with remotely sensed data for statistical modelling. The Random Forest model was used to predict soil textures and vegetation classes on a catchment scale including accuracies and variable importance for all models. The soil textures in the study area both determined and predicted were dominated by high sand fractions: sandy loam, loamy sand and sand. The most important variables for prediction of soil properties were slope and elevation. In contrary, the most important variables for predicting vegetation and rock cover percentage on surface were spectral bands and NDVI. The RF classification model for predicting vegetation showed the best performance and poorer performance for soil textural classification with a misclassification rate of 19.5% and 61.1% respectively. The performance of RF regression model was most accurate for prediction of rock cover percentages with an R2 of 0.57 and a NRMSE of 0.17. The results demonstrate that field data in combination with RF models can be used to determine the spatial distribution of surface characteristics. However, it suggests that discovering statistical trends for in-soil parameters is challenging. The findings did suggest that there is potential in narrowing the training data to the most important prediction variables. Future research could also use different models to see which model is superior for upscaling soil texture for this specific site.