By Vincent Spijker
Landforms are the natural features of the earth’s surface and they play an important role in many scientific fields as they have an effect on geomorphological, hydrological, ecological, biological and physical processes. This makes the process of being able to easily produce landform classification maps important. Many methods have been described to automate landform classifications based on Digital Elevation Models (DEM); however these methods are often semi-automated and have not been validated. In this study, two in landform classification methods, described in literature, have been turned into automated landform classification models. Based on a validation study in the Rocky Mountain area, in the Front Range of Colorado near Boulder, these models have been validated, tuned and the best classification method was selected.
Both methods have a DEM as input and a classified landform map as output. The models have been tested at 3 supports (10x10, 20x20 and 100x100 meter). The first model is based on multiple methods that use DEM derivatives to classify landforms. The value of each cell is determined based on the derivatives values of that cell. The second model is based on the principle of geomorphons in which the landform of each cell is determined based upon the distribution of relative altitudes of its neighbors. The validation is based on a set of 240 field observations. The initial average classification accuracy of the DEM derivatives based model was 61.1%. For the geomorphon based model this was 61.4%. To investigate the cause of these low classification accuracies multiple factors influencing the classifications were examined. Forest cover can reduce the quality of a GPS signal, decrease the DEM quality and make it more difficult to do classifications in the field; however, no relationship between forest cover and the classification accuracy was found. A low classification accuracy of one stratum can lead to a major decrease of the average accuracy of the total model. None of the strata had such low classification accuracies that it significantly influenced the total accuracy. What did have a major effect on the classification accuracy was the GPS positioning accuracy. The GPS accuracy was found to result in many mispositionings and misclassifications in the field, resulting in lower classification accuracies during validations. After correcting for the GPS positioning error by increasing the validation window to 3x3 cells, the average classification accuracy for the DEM derivatives based model and the geomorphon based model became respectively 82.9% and 81.8%.
The average classification accuracies of both models are almost equal; therefore, the decision on which method is best depends on the users purposes. Different models best describe different landscapes and landform classes. The automated character of the models give them an advantage above many methods that are not or only semi-automated.
Keywords: Landform Classification Models; Automated models; Classification accuracy; GIS; Front Range Boulder, Colorado; GPS