Normalized Digital Surface Model (nDSM is often used in the production of urban Land Cover maps. Since they provide above-ground height information for each pixel, they are very useful to distinguish between single-storey and multi-storey building, for example.

nDSM is produced by subtracting the terrain elevation (DTM) from the surface elevation (DSM). While DSM can easily be generated from stereography, the production of DTM is much more labour intensive. Its production is mostly performed through specific software such as Geomatica where a domain expert has to visually identify groups of above-ground objects (trees, buildings, etc.) to be removed from the DSM, but also sometimes deal with terrain artefacts. Then, the expert has to find the adequate combination of filters and interpolation to remove the identified objects. On large scene, this process can take up to several weeks.

Recently, some research [1] tackle this issue and proposed Machine Learning methods to derive DTM from DSM. They used SRTM data (a well-known free DSM product at a spatial resolution of 30 m) in their experiments. The conclusion is that FCN is the best performing approach to achieve the objective. They code is shared on this repository : https://github.com/mdmeadows/DSM-to-DTM.

This research will aim at developing a similar approach but adapted to use with Very-High Resolution products at a sub-meter spatial resolution. A literature review will be conducted and the most interesting approach will be implemented. Assessment report will be produced for validation. The model will be trained with a Brussels dataset including stereoscopic shots (GSD stereoscopic views (GSD=5cm) and a LiDAR survey at 40 pts/m², following a North/South transect.

contact: GRIPPA Tais tais.grippa@ulb.be, Olivier.Debeir@ulb.be

Reference

[1] Meadows, Michael, et Matthew Wilson. 2021. « A Comparison of Machine Learning Approaches to Improve Free Topography Data for Flood Modelling ». Remote Sensing 13 (2): 275. https://doi.org/10.3390/rs13020275.