En français, Modèle numérique de surface (MNS) et Modèle numérique de terrain (MNT).

From DSM(Digital Surface Model) to DTM(Digital Terrain Model) is a traditional topic, i.e. DEM(Digital Elevation Model) filtering. Among the traditional methods, the most famous method is using progressive TIN (Triangular Irregular Network).

With the development of deep learning, many methods and dataset are published, in this document, the methods and dataset is listed.

Because the input may be different, this will significantly influence the method, so the methods will be classified by the input. The input can be point cloud, DSM, DSM+Image.

Point cloud based usually means LiDAR based, because LiDAR point cloud is accurate and with penetrability. On the other hand, the noise distribution of dense matching is different among methods, so DSM filter on dense matching based point cloud is rare.

Using deep learning, the DEM filtering is converted into a classification problem, ground and non-ground point. Actualy, this method is not a direct method, after classification, the DTM is generated from TIN using the ground points.

• 2D CNN based

  convert point cloud into image, then use 2D CNN based classification.

• 3D CNN based

  with the great sucess of PointNet, a lot of methods are proposed, for example, PointNet++, KPConv, RandLA-Net and so on. These methods are proposed based on computer vision dataset.

  From DSM to DTM, Nearest Neighbor Network combine two works, i.e. Full 3D CNN and PointNet.

• other work

  (1) two-level fusion network first use scene detection and then use deep learning to do point based classification using the scene recognition.

  (2) geometry-attentional network

  use the attention in 3D network.

• DNN

  (1) denoise problem

  convert the problem from DSM to DTM to a denoise problem in 1D.

  (2) classifcation problme only classify the non-ground object in urban scenes.

• 2D CNN

  (1) Encoder-decoder

  use evevation, internsity, return number and height different as input, and the out put is the label(ground or non-ground).

• DNN

  (1) manul feature + FCN

  calculate several man made features and then use FCN to obtain the final classification result.

• normalized Digital Surface Model (nDSM) based

  (1) ResDepth

  combine image and residual DSM to improve the DSM result.

For the AHN2 the ground and non-ground are saved in two files, can be found here.

This dataset is from open LiDAR dataset.

In AI4GEO, the dataset summary is vey detailed, can be found here.

A summary of the LiDAR dataset can be found here.

If you think you have any problem, contact [Teng Wu][email protected]