Prediction of soil classes with data collected in buffers delimited pixels in georeferenced soil profiles
DOI:
https://doi.org/10.5039/agraria.v14i2a5653Keywords:
digital mapping of soils, Random Forest, pedometer techniquesAbstract
Digital soil mapping studies (MDS) have used legacy maps as the main source of information for calibration of predictor models. However new approaches are needed with techniques that allow the use of information contained in georeferenced soil profiles, allowing the application of MDS in sampled areas that do not provide conventional soil maps. The objective of this study was to evaluate the performance in the prediction of soil occurrence of samples collected in pixels of profiles of georeferenced soils and in pixels collected in buffers with radius of 50, 100, 150, 200 and 250 m of soil profiles in the Lajeado Grande and Santo Cristo Rivers Watersheds. Two areas with availability of soil survey data published at scale 1: 50,000 were used for the study. For prediction of the occurrence of soil classes, ten predictive variables were generated from a digital elevation model with spatial resolution of 30 m. For the prediction, Random Forest algorithm was used. The predicted maps were evaluated for accuracy in relation to soil profiles and the reproducibility of conventional maps. The use of the sampling pixels collected in the buffers did not significantly alter the overall accuracy of the predicted maps in the Lajeado Grande river watershed, but allowed a 15.6% the gain at overall accuracy the Santo Cristo river watershed. The use of georeferenced profiles resulted in predicted maps with overall accuracy greater than 75% and agreement of reproducibility equal to or high than 67% in relation to the conventional map.
Downloads
Downloads
Published
How to Cite
Issue
Section
License
This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 Unported License.