Random Forest with semantic tie points for classifying landforms and creating rigorous shaded relief representations

Abstract

In this study, we tested Random Forest (RF) with semantic tie points to classify discrete landforms, which would be ultimately useful for increasing the accuracy of shaded relief representations. We therefore focused our efforts on the following landforms: rock outcrops, screes, alluvial fans and low relief areas.A characteristic of RF, which makes it a good algorithm for geomorphological mapping, is its ability to generate thousands of classification trees. Each tree provides a classification and the value classified by the majority of the trees is the final output of the algorithm. Furthermore, having multiple classifications provides a measure of uncertainty. This is very important because it gives practitioners an idea of areas where the method is less accurate, which would require more effort in terms of sampling or surveying.This method was applied in two mountainous areas of Switzerland, the first where we trained and calibrated it and the second where we used it for classification. The results were validated using existing geomorphological maps, and they show that this method can obtain good training accuracy with a relatively small starting dataset. Moreover, both calibration and classification present a percentage of agreement with existing geomorphological maps of over 70%. Some geomorphological classes, alluvial fans and screes, present a classification accuracy that is lower than the calibration, which is in line with previous tests found in the literature. However, for other classes, i.e., rock outcrops and low relief areas, the accuracy increases, suggesting that this method can be employed extensively and relatively securely for these landforms.