BestSegVal: Test for best performing parameters for TreeSeg...
In SchoenbergA/CENITH: Segmentation tool

Description Usage Arguments Details Value Note Author(s) Examples

Iterates over a,b,h,MIN and supports iterating over filtered chms. Uses supervised computed Treepositions to validate best fitting values for a,b, height, MIN and CHM filters to detect trees.

BestSegVal(
  chm,
  a,
  b,
  h,
  vp,
  MIN = 0,
  MAX = 1000,
  skipCheck = FALSE,
  filter = 1
)

`chm`	raster - Canopy Height Model RasterLayer derived form LiDAR data
`a`	numeric - single value, combination of values or sequence for MovingWindow
`b`	numeric - single value, combination of values or sequence for MovingWindow
`h`	numeric - single value, combination of values or a sequence for the maximum height of trees (in meter) to detect trees.
`vp`	shp - PointLayer with estimated Positions of trees (see details).
`MIN`	numeric - single value, combination of values or a sequence of minimum area for crowns. Smaller polygons are cropped Default= 0
`MAX`	numeric - the maximum area for crowns. Larger polygons are cropped. Default=1000
`skipCheck`	for development - bolean - if TRUE skips checking input data and estimation of ETA.
`filter`	numeric - single value, combination of values or a sequence for filtersize, uses a sum filter on the chm with a MovingWindow of (x*x), which must be odd., default= 1 (no filter.)

'start and go to sleep' The function has implemented error catching. If an iteration would cause a critical stop error, the loops continue. Returns NA for corrupted iterations.
ETA: before starting the iterations, up to 3 random combinations are tested and the processing time is recorded. The processing time is multiplied with the total amount of iterations to calculate the ETA.
Input for a, b, h, MIN and filter supports - numeric, single and combination: c(), or sequence: seq() to iterate over: (e.g a=0.5 ; a=c(0.3,0.5) ; a=seq(0.1,0.9,0.05).
filter - it could be helpful to filter the raw chm to 'smooth' small peaks and 'holes' in crowns. Helps to suppress the segmentation of many tiny objects (like stones) and leads to polygons with less 'holes'.
vp - validation point: supervised computed pointlayer. Use a GIS like Qgis to set points where Trees are estimated (see Tutorial).
if 'skipCheck' = TRUE the input checking is skipped and the iterations start directly.
result table
- a, b, height, MIN, chm - the used parameters
- total_seg - total computed segments
- hit/vp - amount of segments computed for trees marked with validation points
- under - undersegmnetation in absolut values, amount of segments which contain more than one validation point.
- over - oversegmentation in absolut values, amount of segments without validation point
- area - the total area of segments
- hitrate - percentage of segments which contain exactly one validation point in relation to the total validation points.
- underrate - undersegmentation in relation to the total segments.
- overrate - oversegmentation in relation to the total segments.
- Seg_qualy - quality of segmentation: hitrate @ combined over- and undersegmentations. A segment with more than one validation points is estimated to be better than segments without validation points. Therefore the 'miss' value is (over+(2*under))/2.

returns a dataframe with several validation scores (see details).

The 'brute force' approach to iterate over many parameters may result in very long time to finish. Preselected smaller samples may be more efficient (see Tutorial).

Andreas Schönberg

# load packages
require(CENITH)
require(mapview)
require(raster)
require(rgdal)
### NOTE: this example is used to show the usage of 'BestSegVal'. It is NOT used to show a workflow for best Results (see 'Tutorial' vignette for workflow strategies)
### further NOTE: to reduce time usage for the example only small iteartions are used to get an overlook for the functionallities.
# load data
chm <- raster::raster(system.file("extdata","lau_chm.tif",package = "CENITH"))
vp <- rgdal::readOGR(system.file("extdata","lau_vp.shp",package = "CENITH"))
# handle CRS string
crs(vp)<-crs(chm)
# take a look at the data (chm with estimated Treeposition)
mapview(chm)+vp
# start BestSegVal with parameters for Moving window and minimum height
x <- BestSegVal(chm,a=c(0.3,0.5),b=c(0.5,0.7),h=c(0.1,1),vp = vp,filter=c(1,3))
# check for best hitrate
maxrow <- x[which.max(x$hitrate),] # search max vale but rturn only 1 value
maxhit <- maxrow$hitrate
x[which(x$hitrate==maxhit),]
# visualise best combination (due to Segment Quality)
y <- TreeSeg(chm,0.3,0.7,0.1,CHMfilter = 3)
# show result segments with validation points and chm
mapview(chm)+vp+y
# BestSegVal based on this results keep a,b and filter 3 with MIN values to reduce oversegmentation and different heights.
z <- BestSegVal(chm,a=0.3,b=0.7,h=c(0.1,0.2,0.5,1),MIN=c(10,100),vp = vp,filter=3)
### Note that row 1 and 2 lead to much lesser oversegmnetation just differ in area.
# visualize
v <- TreeSeg(chm,0.3,0.7,0.1,MIN=10,CHMfilter = 3)
mapview(chm)+vp+v