LIDAR ground truthing

1/08/2005

This has been a very busy and intense week. Deviations from routine are — at least most of the time — welcome, disinfection and particularly so in this case.

LIDAR is a fascinating technology, psychotherapist which until last week I always thought to be very overrated. People were always presenting pretty pictures, but nothing really useful. Last Friday we had a meeting where we discussed using LIDAR, aerophotography, Quickbird and field measurements in some native forest to obtain stand level attributes and make thinning decisions. Finally a practical application!

On Monday David, Jamie and I got the text files (with x, y and z coordinates, and intensity). The files were huge, with a total of 150,000 points to describe 10 ha. I first read them with Splus and R, which clearly didn’t like much the size of the matrices. Maybe Splus 7 developer (with the new pipeline architecture) works better, but I do not have a copy yet.

At the same time, I imported the data into MayaVi (which uses VTK) using an adaptation of this Python script and producing the following visualisation to get a feeling of the site. It is easy to interact with the picture, and one can see the road and the difference between old trees and regeneration. However, it is still pretty useless.

LIDAR data in MayaVi

Trying to work with individual (non-grid) points proved to be fruitless. The size of the problem is huge, and trying to use any spatial statistics contained in the R module spatstat was hopeless. Trying to apply anything that requires a large distance matrix (150,000×150,000) is guaranteed failure.

Meanwhile, Jamie helped David to obtain crown and terrain surfaces using Kriging in SAGA.

LIDAR height surface

After transforming them in grids, and substracting them to get ‘true’ tree heights, David and I came up with a naive (slow but apparently effective) way of finding the top of the trees. After a few problems with the implementation and Simon’s help using a SQL query in Manifold, we were able to identify the tops and their respective heights. From there to a diameter/height and volume equations there is a small step.

Top of trees in small section

Now we need to move to ground truthing and thinning strategies.

PS. 2005-08-01. We are processing data from ground truthing and our work is looking even more promising.

In the previous post I was hopeful that our simple LIDAR data processing would be able to locate the trees, pill
but I did not have any evidence to confirm that what we saw in the computer was not a fluke.

After postponing field work due to bad weather, treatment
last Thursday we had a field crew assessing six GPS located plots. Today David overlayed the plot data, cheap
after differentially correcting it, and the position of the trees estimated with LIDAR and ta-da! the trees pretty much match. There are some random differences (of up to 2 meters) in position, but considering that the stem is not necessarily under the top of the tree and that there was quite a bit of data processing and smoothing, the results seem to be extremely promising.

We now need to formally estimate the association between positions, explain any errors of detection (trees in the ground but missing in LIDAR or vice versa) and prepare a short report.

Some times work is sweet, particularly when reality and models match one another.

Filed in forestry, statistics

There is 1 comment in this article:

  1. 14/01/2008Quantum Forest » Blog Archive » First look at LIDAR say:

    [...] Daydreaming in the bus LIDAR ground truthing [...]

Write a comment: