LiDAR Peak Analysis: What It Takes

[JoeGrim]

Teresa, I've come to the same conclusion as you, that using the LiDAR technique for forested summits provides the highest elevation that isn't under a tree.
As I visit summits in the future that have prominences a few feet short of 300', I'm planning on paying special attention to these under-tree HPs, and documenting them using photos and a hand level.

[JoeGrim]

If you're interested in any specific tiles in Park, Teller, Custer, or Fremont, let me know and I can send you whichever tiles you'd like.

Thanks! I'm mainly looking at Larimer and Boulder Counties for now, which are mostly covered by TNM, but will switch to other counties in the future.

[bdloftin77]

Re: Class 2 with NumberOfReturns = 2

Here is a screenshot of what is obviously a tree, but there are some Class 2, NumberOfReturns = 2, ReturnNumber = 2 points in the midst of the tree. These points had elevations that were higher than the area that clearly looked more like it was narrowing to the highest ground summit area, but I didn't choose them because in the summit area, the NumberOfReturns values were 1 (and class 2); I chose the highest one in that area.

So when I plot these class 2-in-the-middle-of-the-tree points on caltopo and look with the Google Satellite layer, it turns out they fall in a dead standing tree (looks like beetle kill). That left me wondering if LiDAR could see through a dead tree to actual higher ground points below it and show those as Class 2. Lord knows I have crawled under enough trees, dead and otherwise, to touch what looks like the highest ground on a summit to both wonder if I should have picked the highest of these points, and to earn any reputations as a -- let's just use the word fanatic -- that I might have out there. In reality, the point I ended up choosing in what looked like the actual summit area in the point cloud was only 2.284 inches lower and rounded to the same whole number anyway.

I looked at a severely treed summit in North Carolina. I was able to find the highest Class 2, Returns = 1 point (it helped to white out all the non-class 2 points in a duplicated layer instead of bordering the class 2 points in black). The highest class 2 point's elevation made sense for the peak. But the vast, vast majority of the summit area was obscured by trees and whited out. So, is there just no way of knowing if there was actually higher ground elsewhere under the trees that LiDAR couldn't see through them? If so, where do you draw the line and say doing LiDAR processing for a peak is useless?


class 2 returns 2.jpg

I am personally okay with using the final return of many for class 2 points. I trust that the people processing this data have classified it well, and that the last fraction of the lidar beam that returns (and is classified as class 2) is indeed a patch of ground in between tree branches or other vegetation. I know Eli has a more conservative approach in avoiding class 2 points that are the last of many returns, and that's okay. I think it's a matter of choice for the analyzer - the single return class 2 points could be seen as "less risky," but there could be useful class 2 final return points that would otherwise be ignored.

As for areas where there are no ground returns at all, and trees are blocking the lidar - yep, with the data we have, I don't believe there's a way to find the ground elevation. Some lidar projects use different lidar scanners/sensors/light waves and are able to better penetrate through vegetation and even snow. I don't believe the data we are seeing on TNM and OIT is this type of data, though I could be wrong.

As for where to draw the line, that's probably a personal judgment as well. If the main purpose is to determine a peak's ranked status, and a ground point has already pushed it to be a ranked peak, then that might be good enough in heavily treed areas. If the peak is still unranked, but close to being ranked with the few good ground returns, then it might not be feasible to determine its ranked status with the data we have at the moment. If the main purpose is to accurately determine the peak's elevation, then heavily treed/vegetated summits will definitely have less accurate and underestimated summit elevations via the lidar we're using.

I haven't looked at many peaks outside of Colorado, but many peaks here have some class 2 points even if there are quite a few trees. Elsewhere it could definitely be a completely different story.

[Teresa Gergen]

I haven't looked at many peaks outside of Colorado, but many peaks here have some class 2 points even if there are quite a few trees. Elsewhere it could definitely be a completely different story.

Yes, in CO I'm not having trouble so far finding Class 2 points even in our version of thick trees.

I think I missed a discussion somewhere on snow, which was another question of mine. Just point me there if I did. LiDAR doesn't see through it? And doesn't classify it as 2/ground? So if there was snow when the plane flew over and acquired the data, the highest ground could actually be under feet of snow, and the point you pick that's class 2 could be the bare ground next to it where the wind swept the snow over to the edge and built it up? But if there was a windblown ridge say 5-10 ft tall of snow along a ridge/summit edge, LiDAR wouldn't give an inaccurate summit height of 5-10 feet too high? How do you know if the LiDAR data you're looking at was acquired when there was snow in the area?

[bdloftin77]

Yes, in CO I'm not having trouble so far finding Class 2 points even in our version of thick trees.

I think I missed a discussion somewhere on snow, which was another question of mine. Just point me there if I did. LiDAR doesn't see through it? And doesn't classify it as 2/ground? So if there was snow when the plane flew over and acquired the data, the highest ground could actually be under feet of snow, and the point you pick that's class 2 could be the bare ground next to it where the wind swept the snow over to the edge and built it up? But if there was a windblown ridge say 5-10 ft tall of snow along a ridge/summit edge, LiDAR wouldn't give an inaccurate summit height of 5-10 feet too high? How do you know if the LiDAR data you're looking at was acquired when there was snow in the area?

Looks like they want lidar collection to be either snow free or with minimal snow. They probably differentiate snow from glaciers (not sure what that classification would be - class 1?)

Any snow points should be classified as class 21. You can ignore these points.

To see when a point was collected: if you have an Identify tool or are able to see point attributes in QGIS, copy and paste the "GPS Time" to an online converter, like this one.

For those who want to go down the rabbit hole of USGS 3DEP requirements here's a link with a bunch of information regarding lidar collection and distribution: https://www.usgs.gov/ngp-standards-and- ... e-contents

Collection Conditions.png

Minimum Classification Scheme.png

[Teresa Gergen]

I'm looking at a summit that has areas of open ground, shrubs, trees, a building, and rock outcrops/boulders, one of which I know is the highest natural ground. I can find the LiDAR point that has the highest class 2/ground reading, plot it on caltopo, look at the satellite imagery, and see it's not on a boulder. I would like to get a lat/long from caltopo for each boulder contender, and be able to take each lat/long and locate it in the point cloud, so I can figure out what area to be looking in for a localized class 1 highest point. Is it possible to do this in QGIS?

[Eli Boardman]

I'm looking at a summit that has areas of open ground, shrubs, trees, a building, and rock outcrops/boulders, one of which I know is the highest natural ground. I can find the LiDAR point that has the highest class 2/ground reading, plot it on caltopo, look at the satellite imagery, and see it's not on a boulder. I would like to get a lat/long from caltopo for each boulder contender, and be able to take each lat/long and locate it in the point cloud, so I can figure out what area to be looking in for a localized class 1 highest point. Is it possible to do this in QGIS?

Good question, and I was actually meaning to address something similar. The answer is an enthusiastic yes!

QGIS will easily visualize any KML. Simply create the points you want to identify in Caltopo or Google Earth, then export it as a .kml and simply drag/drop it into QGIS. The points should show up as small colored dots. You can right-click the KML layer and adjust the symbology pretty much the same way that you do for LiDAR data. Note that this should also work for KMLs that include line or polygon data--you could even delineate polygons around particular boulders and that should show up in QGIS just the same.

I've found that it's particularly helpful to export a KML of all the summit locations for soft-ranked peaks in LoJ, then drag/drop that into QGIS so I can more quickly identify which peak is which when I'm working in an area with a high concentration of nearby summits.

[bdloftin77]

I've found that it's particularly helpful to export a KML of all the summit locations for soft-ranked peaks in LoJ, then drag/drop that into QGIS so I can more quickly identify which peak is which when I'm working in an area with a high concentration of nearby summits.

Nice! I often do the same thing.

Is it possible to bring in base map/background satellite imagery into QGIS? I’ve found this to be really helpful in ArcMap.

[Eli Boardman]

I've found that it's particularly helpful to export a KML of all the summit locations for soft-ranked peaks in LoJ, then drag/drop that into QGIS so I can more quickly identify which peak is which when I'm working in an area with a high concentration of nearby summits.

Nice! I often do the same thing.

Is it possible to bring in base map/background satellite imagery into QGIS? I’ve found this to be really helpful in ArcMap.

Yep, pretty much anything that's georeferenced should be good to go. If you wanted to get really carried away, the actual 7.5-minute topo quadrangles can be downloaded as GeoTIFFs from The National Map (or Earth Explorer, or both, I can't remember). I usually just look at the peak in Google Earth though.

Teresa had a question about using multiple sources (pointcloud and KML) with different coordinate systems. This <should> usually work fine without any extra steps, since QGIS tries to do "on the fly" reprojection to the current project's default CRS. However, the specific case she showed seems a little weird and I'll take a look tomorrow. It's something to keep in mind if different datasets have different projections--you can always use the "Reproject Layer" tool to double-check.

[Eli Boardman]

FYI we think we figured out the case of the pointcloud not lining up with KMLs imported from Google Earth or Caltopo. The problem seems to be that the geolocation of the imagery can be a bit spatially distorted (especially in places with steep topography) since the quality of the orthorectification performed on consumer-grade RGB imagery is much lower than the raytracing performed to geolocate the pointcloud. If that's all Greek, basically the imagery isn't always in quite the right place (it was offset by about 8 feet in this case.)
_____

Also, for the GIS/remote sensing nerds out there: a note on LiDAR classification. Spectral remote sensing data (Landsat, hyperspectral data, etc.) can be classified based on various forms of "clustering" algorithms which look for perceived patterns in the n-dimensional spectral space (the classic supervised/unsupervised classification scheme using K-Means or something similar falls into this category, as do more sophisticated techniques like Spectral Angle Mapping or Matched Filtering). However, LiDAR data is (by comparison) extremely simple: there is no information available other than the XYZ coordinates, the intensity, and various irrelevant data like the time of flight, look direction, etc. Even the intensity is mostly useless except in special research applications, since it's only a discrete wavelength.

Thus, LiDAR data is "classified" by applying things that look at spatial information ONLY, and this means that the classic approach of selecting some "ground," "tree," "water," etc. classes and performing a supervised classification won't work: there is nothing unique about these manually selected points other than their (arbitrary) XYZ coordiantes (and not all trees are located in the same place, obviously). So, LiDAR data is usually classified with some assumptions of what a smooth surface would look like versus spikes away from this surface. Imagine that you are visualizing the pointcloud in 3D: there might be a whole bunch of points covering the XY plane at almost the same Z height, then a few points in the same area that are much higher. These higher points don't fit the "ground" surface, so they get classified as vegetation (or are left unclassified). The nitty gritty of how this works is often proprietary, but suffice it to say that it usually involves some form of surface fitting. HERE is one good example from the default RIEGEL LiDAR processing package, which uses a Triangulated Irregular Network (TIN) to find a probable ground surface and classifies points near to this surface as ground.

As one of my physics teachers used to say, "what does this mean, actually?" It means that the "ground" classified points are likely miss sharp features. I could point out hundreds of examples in the Tetons where sharp summits are classified explicitly as "low vegetation" despite being thousands of feet above treeline. The ground-classified points assume a relatively smooth surface. If a peak has a particularly rough summit or saddle (lots of talus or a big sharp block), using only ground-classified points is almost guaranteed to underestimate the peak's prominence. A careful inspection of the terrain at different elevation ramps is helpful to figure out what's actually going on, and as always, some uncertainty will remain.

[bdloftin77]

FYI we think we figured out the case of the pointcloud not lining up with KMLs imported from Google Earth or Caltopo. The problem seems to be that the geolocation of the imagery can be a bit spatially distorted (especially in places with steep topography) since the quality of the orthorectification performed on consumer-grade RGB imagery is much lower than the raytracing performed to geolocate the pointcloud. If that's all Greek, basically the imagery isn't always in quite the right place (it was offset by about 8 feet in this case.)
_____

Also, for the GIS/remote sensing nerds out there: a note on LiDAR classification. Spectral remote sensing data (Landsat, hyperspectral data, etc.) can be classified based on various forms of "clustering" algorithms which look for perceived patterns in the n-dimensional spectral space (the classic supervised/unsupervised classification scheme using K-Means or something similar falls into this category, as do more sophisticated techniques like Spectral Angle Mapping or Matched Filtering). However, LiDAR data is (by comparison) extremely simple: there is no information available other than the XYZ coordinates, the intensity, and various irrelevant data like the time of flight, look direction, etc. Even the intensity is mostly useless except in special research applications, since it's only a discrete wavelength.

Thus, LiDAR data is "classified" by applying things that look at spatial information ONLY, and this means that the classic approach of selecting some "ground," "tree," "water," etc. classes and performing a supervised classification won't work: there is nothing unique about these manually selected points other than their (arbitrary) XYZ coordiantes (and not all trees are located in the same place, obviously). So, LiDAR data is usually classified with some assumptions of what a smooth surface would look like versus spikes away from this surface. Imagine that you are visualizing the pointcloud in 3D: there might be a whole bunch of points covering the XY plane at almost the same Z height, then a few points in the same area that are much higher. These higher points don't fit the "ground" surface, so they get classified as vegetation (or are left unclassified). The nitty gritty of how this works is often proprietary, but suffice it to say that it usually involves some form of surface fitting. HERE is one good example from the default RIEGEL LiDAR processing package, which uses a Triangulated Irregular Network (TIN) to find a probable ground surface and classifies points near to this surface as ground.

As one of my physics teachers used to say, "what does this mean, actually?" It means that the "ground" classified points are likely miss sharp features. I could point out hundreds of examples in the Tetons where sharp summits are classified explicitly as "low vegetation" despite being thousands of feet above treeline. The ground-classified points assume a relatively smooth surface. If a peak has a particularly rough summit or saddle (lots of talus or a big sharp block), using only ground-classified points is almost guaranteed to underestimate the peak's prominence. A careful inspection of the terrain at different elevation ramps is helpful to figure out what's actually going on, and as always, some uncertainty will remain.

Hey Eli, thanks much for clarifying this! I appreciate it. Sounds like the methods for classifying LiDAR ground points are much different than classifying typical ground imagery. This makes sense why sharp features are not classified as ground because they break ground slope rules.

Unless I’m mistaken, they still often categorize final of multiple return points as ground (if they don’t break slope rules) and also use these for their DEM bare earth models though, right? Based on your explanation, these just might be less accurate than otherwise because their algorithms are having to take a relatively large guess (depending) if not many other ‘definite’ ground points exist? Eg if a summit is covered with dense (bumpy) vegetation, and several points have multiple returns, if the final returns line up fairly well with a ground slope model they’ve generated, then they’ll classify these as ground. But as you pointed out earlier, some of these might or might not be the actual ground because they’re only based on a gradient/slope model - some of these might be lower vegetation, exposed roots, etc? So we can’t be very confident that there are ground if there aren’t many areas without vegetation to be compared to?

Also, you mentioned intensity isn’t used in point classification? Or just with ground point classification? I had thought that point intensity was a major factor in the LiDAR point classification process, but it sounds like that is incorrect.

[Eli Boardman]

Hey Eli, thanks much for clarifying this! I appreciate it. Sounds like the methods for classifying LiDAR ground points are much different than classifying typical ground imagery. This makes sense why sharp features are not classified as ground because they break ground slope rules.

Unless I’m mistaken, they still often categorize final of multiple return points as ground (if they don’t break slope rules) and also use these for their DEM bare earth models though, right? Based on your explanation, these just might be less accurate than otherwise because their algorithms are having to take a relatively large guess (depending) if not many other ‘definite’ ground points exist? Eg if a summit is covered with dense (bumpy) vegetation, and several points have multiple returns, if the final returns line up fairly well with a ground slope model they’ve generated, then they’ll classify these as ground. But as you pointed out earlier, some of these might or might not be the actual ground because they’re only based on a gradient/slope model - some of these might be lower vegetation, exposed roots, etc? So we can’t be very confident that there are ground if there aren’t many areas without vegetation to be compared to?

Also, you mentioned intensity isn’t used in point classification? Or just with ground point classification? I had thought that point intensity was a major factor in the LiDAR point classification process, but it sounds like that is incorrect.

Ah, yeah I really should have clarified that I was only really talking about the ground and not-ground dichotomy. I don't really know how most of the rest of the classes are decided (and it probably varies from survey to survey, since some have more classes than others), but I suspect you're right that intensity does play a large part in some of the classifications. In particular, I bet that the water and snow categories are based on intensity.

As far as the root/dense shrubbery aspect, I don't think there's much we can do. But then again, at what point does a rotten log cease being a tree and start being soil? I analyzed Kawaikini (the Kauai highpoint) just for fun and that summit is completely buried under a wetland bog--but I'd argue that the layer of dense moss, peat, and decomposing plant material is itself "ground."

These questions run in a similar vein to the classic coastline paradox. The notion of "the ground" is ill-defined in a physical, mathematical sense. C.F. Benoit Mandelbrot and the "theory of roughness."