L'uso di tecniche di classificazione non supervisionata per la mappatura automatizzata di forme a partire da modelli digitali del terreno

Unsupervised classification techniques for automated mapping of landform elements from digital elevation models

Abstract

We report the preliminary results of a methodology for the automated mapping of some landform features. Our investigation has still to be refined and extended to a wider area to give reliable results for the study of the dissectional and tectonic history of the study region, but it is already possible to discuss some advances we have achieved in automatic landform classification for the study area. The Trigno River catchment was selected to test the method. In the study area, due to the almost complete lack of river terraces, the reconstruction of old stages of valleys development must be attempted using other characters of the valley-side slopes. Among others, the widespread, gentle erosional glacis that overhang the present thalwegs at different elevations can be used as indicators of discontinuous downcutting and as proxy of ancient base levels. Most of the methods that are currently used to characterize terrain features by means of DEMs extract the information contained in each cell of the raster and are able to quantify the local surface geometry very accurately (within the resolution limit of the DEM used). However, to make discriminations among landforms which differ in terms of spatial and contextual properties, it is essential to take into account also the properties of neighbour cells. In this sense, we propose a procedure that, instead of using a classic neighbourhood analysis, partitions the DEM by submitting slope angle and elevation value to a k-means algorithm for the identification of homogeneous landform elements. In this way, we expect to obtain patches of contiguous cells whose boundaries reflect characteristic landform elements of the study area. Mean values of (i) slope angle, (ii) tangential curvature and (iii) profile curvature of the cells belonging to each patch, are then used as «feature vectors» for an unsupervised landform classification of the area under examination. For the Trigno River test area a 10-cluster solution has been chosen by a trial and error procedure. The resulting automated mapping is judged very good as it recognizes ten different units, each of them matches very well with ones of the geomorphological units that have been surveyed in the test area. A comparison with a cell-by-cell unsupervised classification shows that the proposed method produces more accurate results. In particular, a good discrimination between summit erosional surfaces and valley bottom was achieved. Further developments, aimed to the refinement of the classification results, are also briefly discussed.