Terrace walls detection from a Pleiades Digital Elevation Model
Résumé
In Mediterranean cultivated areas, terrace walls consist in historic man-made levees on agricultural plot margins that were initially settled to limit soil erosion and favor plot cultivation by decreasing the terrain slope. Some models and indicators integrate now terrace wall maps to diagnose the hydrological risks or the biodiversity at catchment or region scales. In addition, terrace walls also appear as inter-plot elements that can induce overestimations in cultivated areas that are declared by farmers. Consequently, there is a great need for mapping terrace walls at a large scale for various purposes. However, terraces walls are generally poorly mapped in national map agencies geodatabases. Consisting in a linear terrain elevation disruption, a terrace wall can be identified using geomophometry methods or image processing algorithms on a high resolution digital elevation models (DEM). But these methods have been, up to now, only tested on airborne LiDAR DEMs. Thanks to the Pleiades satelitte agility, stereo Pleiades coherent image pairs offer new opppotunities to build up DEM with a very high spatial resolution allowing to detect terrace walls within cultivated and poorly vegetated areas. In this study, a DEM built from a Pleiades stereo images, acquired in February 2013, was available all over the Peyne Catchment, in the Hérault Department. This 90 km² vineyard catchment is representative of Meditteranean Piemont areas, prone to flash floods and soil erosion. This catchment presents a dense network of terrace walls (400 km), having various heights (from 30 cm up to 8 m) that have been exhaustively mapped in 2010. Thanks to the opensource MicMac suite developed at the IGN, a 0.7 m DEM was built by photogrammetry techniques. A low regularization have been applied to ensure a dense DEM. Next, two terrace walls detection methods were compared. The first method enhances local elevation contrast in DEM using Gaussian filtering, then a line segment detector algorithm (LSD) that delineates the terrace walls from the filtered DEM. The second method is using statistical thresholds on geomorphological filters that delineate areas corresponding to terrace walls. For both methods, implementation is easily accessible and a parameter calibration step is needed. Using available ground truth all over the catchment, the detection performances are thus computed for both methods and results are compared to those obtained using a lidar DTM.
