**4. Data Processing**

#### *4.1. Close-Range Photogrammetry (CRP) Point Cloud Processing*

The next steps in the photogrammetric processing were devoted to yield as accurately as possible the 3D model of the hillside. 43.58 million points were obtained for the site area, reaching a mean density of 413 points/m2. The mean distance among points in the 3D point cloud is 7.6 cm. All the steps presented up to this point were processed with Agisoft PhotoScan Professional v. 1.4.4 build 6848. However, for filtering the point cloud, 3DReshaper was used. Noisy points were removed with the option delete points (over 20 cm). Afterward, from the segmented point cloud, tiny sets of points (below 11) were automatically deleted. Next, extensive manual filtering was undertaken to remove points identified as vegetation (trees and shrubs) or isolated points outside of the ground such as power poles or road signs. The result of this phase is visible in Figure 7, which presents the di fference in a sector without and after filtering. The final number of points was 42.75 million points.

(**b**) 

**Figure 7.** Cortes de Pallás sector: (**a**) point cloud without filtering; (**b**) cleaned and filtered point cloud.

#### *4.2. Mobile Mapping System (MMS) Point Cloud Processing*

The open-source software CloudCompare has been used to process the 3D point cloud by MMS, with the purpose of analysing the raw data acquired (Figure 8). As a default parameter, the number of points composing the sharpened MMS point cloud is 94.30 million.

In order to obtain ground and non-ground points from MLS point cloud, the data filtering algorithm Cloth Simulation Filter (CSF) was used [33]. This method allows users to obtain the "steep slope" model setting some advanced parameters such as cloth resolution, maximum iterations and classification threshold (Table 5). The cloth resolution refers to the grid size of cloth composed by particles interconnected through virtual springs, which is used to cover the terrain. The positions of the particles in three-dimensional space determine the position and shape of the cloth. The grid size has the same unit of the point cloud. The lower the value of cloth resolution, the softer the resulting mesh from the filtered point cloud is. The number of iterations is linked to the maximum iteration times of terrain simulation. Classification threshold is used to classify ground and non-ground points, based on the distances between points and the simulated terrain.

**Figure 8.** Cortes de Pallás MMS (Mobile Mapping System) point cloud: 3D point cloud (intensity scale) and the closed-loop trajectory (red line) carried out with the MMS.

**Table 5.** Advanced parameters applied to cloth simulation filter (CSF) operation to perform the segmentation between ground and non-ground points.


The final output is a segmentation of the MMS cloud, where all the vertical features, like trees, pylons and road signs, were automatically removed (Figure 9), reducing the number of points to 47.86 million.

**Figure 9.** Cortes de Pallás MMS point cloud: the result of filtering algorithm CSF (Cloth Simulation Filter), applying the shader filter EDL (Eye Dome Lighting) which help to show better the topography of the study area.

## *4.3. MMS Point Cloud Georeferencing*

Since a GNSS system was not integrated with Kaarta Stencil 2, the georeferencing of the point cloud could be realized in two ways: using the control points captured by the EDM or aligning the point cloud to the CRP model, already georeferenced. As only EDM points fell within the path trajectory, the second option has been chosen and represents a completely new experiment tested by the Kaarta manufacturing experts. The filtered CRP point cloud presented in Figure 10 was thus used next as the control system for the MMS point cloud.

After the filtering operation of the MMS point cloud, the next step consisted of roughly aligning the trajectory of Kaarta Stencil 2 with the CRP point cloud. First, the yaw angle of the starting point of the MMS close-loop trajectory in the CRP point cloud was identified and calculated. Then, the parameters of the MMS point cloud could be set up iteratively to obtain a better result in the replay operation of the trajectory line, evaluating every time the correspondence with the photogrammetry point cloud. Defined as the right alignment of the trajectory, the MMS point cloud was run following this adjusted close-loop line and automatically adapted with the same roto-translation transformation. At the end of this procedure, the two point clouds were overlapped and so they could be analyzed by executing a comparison operation between the point clouds themselves (Figure 11).

**Figure 10.** Top view of the CRP (Close-Range Photogrammetry) point cloud. In this image are visible the holes due to occlusions and vegetation.

**Figure 11.** MMS point cloud (false color scale) aligned to the CRP point cloud (red, green and blue (RGB) scale).
