Problem with LiDAR Point Cloud Segmentation

[luisz]

Hi everyone,

I am currently working on an environmental simulation model in QGIS for a university study project.

To achieve this, I need to generate spatial input data using CloudCompare to process airborne LiDAR point cloud data. Specifically, I require the Digital Elevation Model (DEM), Digital Surface Model (DSM), and Canopy Digital Surface Model (CDSM) for my study area. This involves segmenting individual objects such as trees, buildings, and the ground surface from a LiDAR 3D model provided by the federal state of Nordrhein-Westfalen.

I followed this tutorial https://youtu.be/3QigbJGuHY8?si=aJPBRCFG0-CGEPo8 from minute 04:48 to 07:00, where a LiDAR model from Australia is successfully segmented into the desired components. While I was able to replicate these steps with a similar LiDAR scan from Australia, the process does not work as smoothly with the LiDAR scan of my study area.

When I set the Active Scalar Field to Classification under Properties > Scalar Fields in CloudCompare, I primarily see two colors under SF display parameters (as shown in the attached image), whereas there should be more (as demonstrated in the video). Selecting either of these two colors results in incorrect segmentation, preventing me from accurately isolating trees or buildings from the model.

Why is this happening? Is the LiDAR scan not accurate enough? Can anyone offer assistance or suggest a better method for segmenting the model into the desired components? If someone could suggest what I should do differently, I would greatly appreciate it. Alternatively, I can provide the .laz file of the area if someone is willing to try it for me.

Thank you in advance for your help!

Best regards,
Luis

[daniel]

The 'classification' field is part of the original data. This is generally generated by the company in charge of acquiring the data, or later by another party, before the data is published.

And sadly, the quality of this classification data varies a lot (depending on what was asked to the data vendor, the need, the quality of the classification tool that was used, etc.).

It seems your data has a minimal classification (which is still better than nothing I guess ;). So you would have to generate a better classification if you wish to follow the exact same workflow as in the video. For vegetation, yon can try your luck with either Canupo or the new 3Dmasc plugins. But it will ask for some involvement on your side to understand how these plugins work. Else, there are other tools that can do that in an easier way I believe (but mostly commercial tools).

[DroneN3rd_508]

As you continue to work with point clouds you're going to see just how important it is to have a whole lot of different methods & tools for segmenting the data, which is needed to properly classify the data. Most of my datasets are sent to me by the people who collect it in the field and then process it, resulting in some half-way decent classification that only extends to (2) class codes: "Unclassified" (class code "1"), and "Ground" (class code "2").

My approach is more "ground up": DEM first, then DSM after.

First split the entire cloud into sub-clouds, by the classifications that exist. With the scaler field set to "Classification": Edit -> Scaler Fields -> Split cloud (integer values).

I always start with "class #2" (ground), and start cleaning that up to eventually have what I need to create a DEM. This can mean manually segmenting/removing or a combination of tools & plugins, but eventually you're going to want to get to a point where you will create a subsample cloud of the ground. I usually set the subsample to "distance" and input "10.00", which is enough for a good DEM of the area you're doing.

The short of it is: watch the tutorials you find online to learn the tool, but never expect it to be as effective as they show. Remember, the person doing the video has spent a fair amount of time managing a dataset that generates great results.

Its really just: segment -> classify -> merge -> repeat....