Masters programme | E-portfolio
Semester I & II
Remote Sensing Applications

UAV-based DSM & Orthophoto generation

Illustration of the workflow as well as the intermediate products in the creation of terrain models based on drone data. Application example based on a small-scale river bed within a forested area.

Table of Contents

UAVs are advantageous for creating low-cost elevation models with high spatial resolutions on a local scale. Subsequently, the workflow for creating a digital surface model and orthoimagery based on a set of provided drone images (DJI Mavic Pro) is presented. Metashape Agisoft was used as a software environment.

Processing pipeline & intermediate products

Prior to performing image alignment, screening for vague images was performed visually and also quantitatively by using the implemented function to estimate the sharpness level of the most focused part of each image (-> Right-click images in left pane -> Estimate Image Quality). For none of the images, the calculated quality value was below 0.83. Thus, all images were considered to be of sufficient quality for all further processing steps. Image alignment resulted in a set of 54.000 tie points, i.e. feature points matched across images, forming the sparse point cloud. Relative to the size of the area of interest (AoI), this amount of tie points seemed to be good enough in order not to reconsider the parametrisation of the tool (e.g. changing the max. number of feature or tie points per image). The estimated exterior and interior camera orientation parameters resulting from the image alignment were only preliminary results and altered again during the georeferencing of the images.
Sparse point cloud

7 ground control points (GCP) were used to enhance the existing georeferences provided by the UAV itself. Once a GPC marker was manually set on an image, the marker projections for all other images were automatically loaded using the existing tie points (guided marker placement). To increase the accuracy, each of these automatically set markers was checked again manually and moved if necessary. Overall, this resulted in a positional accuracy of less than 5cm.

As a prerequisite for digital elevation model (DEM) creation, a dense point cloud was created subsequently. Note that a DEM may also be derived from a mesh. Nevertheless, calculating a dense point cloud was preferred since…

a.) …mesh computation adds additional computational overhead. In the given case, a 3D polygonal representation of the scene does not bring any benefits. The production of a mesh would therefore not be an end in itself but only a means to produce a 2,5D DEM. But for the derivation of a high-quality mesh, a dense point cloud is needed as input data anyway.

b.) …computing a dense point cloud allows the creation of a digital terrain model (DTM). Even though the current analyses will be limited to a digital surface model (DSM), the dense point cloud leaves the possibility to create a DTM later on. This is due to fact that a classification of points from a dense cloud into ground and non-ground points can be conducted.

c.) …the inverse distance weighting technique used when creating a DEM based on a dense cloud is considered superior to the triangular irregular network technique when a mesh is used. Though, this point is only of minor relevance.

The creation of a dense point cloud requires depth maps calculated using stereo matching. Since this is a computationally intensive process, the default quality parameter of the Build Dense Cloud tool, set to medium, was left untouched. Point confidence calculation was included to enable filtering by confidence level. The initial dense cloud contained more than 6.950.000 points. Using the confidence interval [2,255] to filter out low confidence points (-> Tools -> Dense Cloud -> Filter by Confidence), a dense cloud with approx. 5.300.000 points remained as an intermediate product for further processing. As evident from the figure below, low confidence points are especially those which are located towards the edges of the AoI. Proceeding without applying the filter leads to the inclusion of the bottom-right/south-east outliers in the subsequent DEM generation.

Low confidence points (top) and final dense point cloud (bottom)

Without prior classification and corresponding filtering of the ground class, the DEM product automatically represents a DSM. Thus, DSM creation simply requires using the Build DEM tool. Again, most default settings were kept. Only the reference system was changed to UTM N33 (EPSG:32366). The effective ground resolution of the created DSM was approx. 5.7cm. When exporting the DEM, the sampling rate was set to 5cm in order to simplify any kind of subsequent calculations based on the number of pixels/pixel size.

Orthomosaic generation merges the original images projected on the scene surface. Accordingly, the DSM was used as an input surface for this tool (-> Build Orthomosaic). Default settings were applied including the mosaic blending mode and carrying out hole filling. The default pixel size referring to the ground sampling resolution was 1.4cm. Analogous to the DSM export, the pixel size for the orthomosaic export was lowered to 1cm.

Final elevation model & orthoimagery

As can be seen from the maps below, the resulting DSM and orthophoto are covering an area around the river of about 230m length and 100m width. Heights are varying between 482m and 516m. In line with the orthophoto, the highest points correspond to the peaks of trees. Note that not all trees are mapped in the DSM. This is mostly due to removing the unreliable dense cloud points in these areas. Due to the peripheral location of the trees, they are only captured by images of a single flight track, which lowers the reliability of their cloud points. In addition, the complex structure of trees limited the number of initially created tie points. Hence, it is questionable to what degree reliable statements on the tree heights can be derived from the DSM.

Digital surface model
Orthophoto with ground control points

Evaluating the results, it can be said that the provided UAV data fulfilled the following criteria that were important for their further use within the current study:

  • sufficient image overlap (to the side as well as to the front)
  • sharpness (also quantified above)
  • high resolution (allowing to create a orthophoto with less than 1cm pixel size)

Nevertheless, the following things related to the data acquisition could be improved:

  • varying image brightness and partly even saturation which made it hard to locate some of the GPCs (esp. those on the northern side)
  • shadows due to the acquisition time hampered the identification of reliable tie points
  • exterior camera orientation parameters (yaw, pitch, roll) appeared to have some serious flaws and the same was true for the altitude information, which was off by more than 60m; both are not that important for the current study as sufficient image overlap and accurate ground control points made it possible to correct these parameter