In order to answer, let’s put aside important, but broad issues:
- The fact that identifying and segmenting trees is a very complex analysis which depends on many things (things related to the type of vegetation, and quality and amount of available data, for example).
- That processing 'large point clouds' in R is a real concern (due to memory limitation), and still depends on how large the study area is, how large the point cloud is; the goal of the analysis; if the point cloud can be preliminarily thinned, filtered, spatially sampled; the hardware; etc.
- That the question is not entirely clear; if working with a catalog and using CHMs are mandatory; the context of the analysis; type of output; reproducible code, etc.
lastrees with a catalog, as explained by JRR, it is not currently possible (version 2.0.3):
the difficulty come with the edge of the processed chunks. Processing the dataset into independent chunks implies that there is no easy wall-to-wall continuity. At the edge of a chunk you will have the first half of a tree labelled 123 and the second half of the same tree from another chunk will be labelled 456. Wall-to-wall continuity is the reason why lastrees does not have a LAscatalog version yet.
Using a CHM with
lastrees is possible as long as one chooses a segmentation algorithm which takes it as an argument. For example,
silva2016 (not me),
watershed, etc. The following is a well crafted answer about this: Exporting crown boundaries from tree segmentation in R?. Also, carefully read the package's documentation. It will explain that such algorithms can be run independent of the point cloud; i.e., at most, the LiDAR data can be used to register the results from the CHM-based segmentation.
And since this is a subject which still is in the academic frontier, to deepen exploring other possibilities, the scientific literature is our friend. For example, see the articles recommended in Extracting tree crown areas from remote sensing data (visual images and LiDAR).