Lack of precision when using the lidr package's segment_trees function

Question

In short, I'm trying to calculate the density of trees over 7m tall, so I want to estimate the crown of each tree in my test plot. To do this, I'm trying to segment each tree over 7m to give to the crown_metrics function. I've tried different approaches with the segment_trees function, using different algorithms (dalponte2016, silva2016, etc), and I've also tried changing different parameters in the algorithms. So far, dalponte2016 has given me the best result, but I'm still not satisfied as some points are not well segmented (I tried to take a screenshot, I'm not sure it's clear enough). As you can see (red circle), some points are not segmented even though they are clearly part of the tree. The only possible explanation that I'm thinking of is that my point density isn't sufficient (4.6 points/m²). Or that I'm not using the algorithm correctly.

Here's my code (I didn't put all the attempts with different parameters for dalponte2016):

las <- lidR::readLAS('Data/placette2.las')
las.norm <- lidR::normalize_height(test.placette, algorithm = tin(),na.rm =T, res= 0.3)
las.norm <- filter_poi(las.norm, Z >= 0, Z<50)

chm <- rasterize_canopy(las.norm, res = 0.5, p2r(0.2, na.fill=tin())) 
w <- matrix(1,3,3)
chm.smoothed <- raster::focal(chm, w, fun = mean, na.rm =T)

ttops <- locate_trees(las.norm, lmf(ws = 3))
ttops <- ttops[which(ttops$Z >= 7),]

algo <- dalponte2016(chm.smoothed,ttops)
las.segmented <- segment_trees(las.norm, algo) 
plot(las, bg = "white", size = 8, color = "treeID") 

crowns <- crown_metrics(las.segmented, func = .stdtreemetrics, geom = "convex")
plot(crowns$geometry, col =height.colors(25), main ="Crown with Tree tops")
plot(sf::st_geometry(ttops), add = TRUE, pch = 3)

Result of the crown detection with tree tops:

I was expecting a better result for the segmentation and crown. Some points aren't segmented correctly and this results in a bad crown estimation for few of my trees.

I tried different algorithms for the segment_trees function with different parameters. I also try to estimate crowns for all trees and not only for the ones above 7m.

I wanted my script to be a reproducible example, but I don't really know how to do that with .LAS files. Best Regards, Lea

Answer 1

The mistake you made was to compute the local maxima from the point cloud and then use these maxima as seed on a raster that was modified with a focal function. The local maxima of the raster no longer match the maxima found in the point cloud because the smoothing moved the maxima in the raster. Then the segmentation was messed up.

Compute the maxima from your raster.

library(lidR)
library(terra)

test.placette <- lidR::readLAS('~/placette2.las')
las.norm <- lidR::normalize_height(test.placette, algorithm = tin(),na.rm =T, res= 0.3)
las.norm <- filter_poi(las.norm, Z >= 0, Z<50)

chm <- rasterize_canopy(las.norm, res = 0.5, p2r(0.2, na.fill=tin())) 
w <- matrix(1,3,3)
chm.smoothed <- raster::focal(chm, w, fun = mean, na.rm =T)

plot(chm.smoothed, col = height.colors(25))

ttops <- locate_trees(chm.smoothed, lmf(ws = 3))
ttops <- ttops[which(ttops$Z >= 4),]

plot(ttops, add = T, pch = 19)

algo <- dalponte2016(chm.smoothed,ttops, th_tree = 1)


seg = algo()
plot(seg, col = pastel.colors(50))
plot(ttops, add = T, pch = 19)

las.segmented <- segment_trees(las.norm, algo) 

plot(las.segmented, bg = "white", size = 8, color = "treeID") 

crowns <- crown_metrics(las.segmented, func = .stdtreemetrics, geom = "convex")
plot(crowns$geometry, col =height.colors(25), main ="Crown with Tree tops")
plot(sf::st_geometry(ttops), add = TRUE, pch = 3)