A generic framework for the identification and delineation of landforms from high-for DEMs using segmentation, contextual merging, and machine learning

After successfully applying segmentation and machine learning for landform identification and delineation for concave, convex, and generic landforms (landslides, floodplains), the used approach is generalized as a framework. The approach can be implemented in any GIS software that allows scripting and is based on four steps: (i) object-based segmentation based on a specific geomorphometric variable, (ii) contextual merging if the landform is composed of multiple shapes, (iii) selection of the training data segments, (iv) statistical classification by machine learning. The framework refers to creating a set of rules for various scenarios of landform types to allow the implementation of the approach for various landforms and areas around the globe. One of the main requirements regarding the DEM is that its feature resolution be high enough to allow at least a segment to cover the target landform spatially. This requires either LiDAR or RADAR DEMs, with medium or high resolution. We tested COPDEM in areas where there is no vegetation cover and the results show that landslides, floodplains, gullies, sinkholes, and closed depressions can be depicted by the approach.