A machine learning approach to the geomorphometric detection of ribbed moraines in Norway

Abstract

Abstract. Machine learning is a powerful yet underutilised tool in geomorphology, commonly used for image-based pattern recognition. Analysing new high-resolution (1–10 m) elevation datasets, we investigate its usefulness for detecting discrete geomorphological features. This study develops a machine-learning-based method for identifying ribbed moraines in digital elevation data and progresses to test its performance versus time-consuming, manual methods. Ribbed moraines share geomorphometric characteristics with other glacial landforms, hence representing a valuable test of our new methodology in terms of differentiating between similar features, and for detecting landforms with similar characteristics. Furthermore, mapping ribbed moraines may provide valuable indications of their origin, a topic of debate within glacial geomorphology. To automatically detect ribbed moraines, we extract simple morphometrics from high-resolution digital elevation model data and mask regions where ribbed moraines are unlikely to form. We then test several machine learning algorithms before examining the best performer (K-means clustering) for three study areas of 15 km2 in Norway. Our results demonstrate a balanced accuracy of 65 %–75 % when validating versus ground-truthing. The performance depends on the availability of high-resolution elevation data in Norway that are needed to resolve the spatial scale of the target (10–100 m). We find the method effective at detecting both fields of ribbed moraines, as well as individual ribbed moraines. We propose pathways for the future implementation of this method on a large scale and for increasing the detail of information gained about detected landforms. In conclusion, we demonstrate K-means clustering as a promising method for detecting ribbed moraines, with great potential to reduce the time needed to produce landform maps.