Hi Jon,

Pity it's not full waveform as it cut down many options. Now I don't think backscattering would be easy to calculate and it might be also less helpful with discrete scanner data (I'm not even sure whether it's possible/practical to do). (BTW the incident angle I meant was the angle of the beam as it hits the target surface, thus it requires modelling the target surface normals).

One simple thing you could do (unless it's already done with scanner software or any post-processing) is correct intensity by distance (sometimes it's called R-squared correction), when you calculate for each point the distance from the scanner and correct the intensity appropriately.

Now with intensity as the only additional point attribute you probably have to analyse geometry of points and their neighbourhoods.

Jon - and when you visualize the point cloud - can you clearly distinguish visually the two classes: biomass and trunks/branches?

Do you get multiple points for branches/trunks, so it's possible to detect relatively smooth/linear surfaces? How many points do you get for branches/trunks? Is it easy to visually detect such objects?

The vegetation should give very rough surface that should be distinguishable from wood/trunks.

Other property is that vegetation/leaf/biomass should be somewhat penetrable by laser giving you many returns, forming very rough surface, that should be clearly different than solid reflections you get from wood/branches. (There is also the case then you have leafes/biomass, just partly covering wood/branches behind them - that again depends on your data and whether you need to cover such cases also with your model).

So I would start explore those geometric properties more, starting just with visual and exploratory analysis of the data.

Typically when you can clearly see with your eyes what you're looking after in your data, then it's relatively easy to come up with the algorithm to do the same automatically. When you cannot see that with your human eye, then the task is really hard.

If you can see in the point cloud what is what, then you can label and extract the training samples and then look at the data using statistical/machine learning methods trying to fit a model for classification. With discrete (non full waveform) you just have much less information about each point, so you need to analyse geometrical properties of many neighbouring points, which is always more complex approach.

My experiences are mostly with full waveform data, when we have really many point attributes, so it's really easy to make even more deriven features and create a model using machine learning algos with relative ease.

With discrete scan data that methodology seems much less suited, and I'm afraid I cannot provide much more ideas without looking at your data.

Regards,

Adam

Hello Robert, 

Thanks for the information and direction to the resource. I had actually already considered the leaf off/on approach, or something similar in any case, but a layer of complexity that I wanted in my investigation is that I can separate the vegetation/biomass from data captured as a single scan. However, I could use this approach to create training data in any case. 

Ill spend some time on the CHM approach and see where that takes me. Thanks again. 

Regards

Jon. 

Hi jon

I haven't worked with intensity values explicitly, but there is a journal paper by my friend Alberto Ramirez   who has considerably worked with intensity data to estimate canopy gap fraction.

Testing the Application of Terrestrial Laser Scanning to Measure Forest Canopy Gap Fraction  

You may have a look at this paper and see it may serve your purpose or contact the Salford University team who have done considerable research on intensity data for vegetation studies. 

hope this helps

vishal 

Hello Vishal, 

Thanks for the pointer to the article, it looks like there is some very useful information in there. I shall read it with interest.

Kind regards, 

Jon. 

Hello Jon,

Are you determined to use the intensity values alone for the task, or you can also consider other attributes of scanned points?

One approach could be to calibrate the data to obtain real reflectance (or backscattering coefficient) values instead of intensity alone. For that you will also need to estimate incident angle of the beam, that also requires to estimate surface normals at the point.

But anyway using intensity (or reflectance) alone for the task makes it quite hard.

I would suggest - if that's not excluded by your methodological assuptions - to consider also geometric attributes of the point's neighbourhood and other available attributes (if the scans were full-waveform then the echo width and echo number information should be also of great help). We typically extract many features - related to geometry in the point neighbourhood or various transformations and aggregations applied on point cloud attributes - then feed that data together with some training examples to machine learning classifiers to do feature selection, modelling and then prediction (we use a custom software workflow for that).

If your task is essentially a 2-class problem then you should get really good accuracies using that approach (I would expect accuracies in the range of at least 80-90%).

Hello Adam, 

Thanks for the information and overview of your methodology. Just to clarify, I am not trying to look at intensity alone, it's values are to be combined with other metrics from the point cloud in order to gain additional information about the subject of interest. 

The scanner I'm using is a single return terrestrial laser scanner, not full waveform unfortunately, so the range of potential data I will have available is somewhat limited hence the need to combine the intensity values with the additional metrics. 

I'm interested in your comments about backscattering coefficient, so I will look further into reproducing this, however, one point I should make is that my early investigations show me that the incident angle of the beam will be varied as presumably, the laser is being reflected from different points on the scanner mirror. Do you suggest to overcome this problem to estimate a single mid-range angle and apply that to all the data? Is that a common practice as far as you are aware?

I appreciate your comments, and look forward to hearing from you again. 

Kind regards, 

Jon. 

Hi Jon,

Pity it's not full waveform as it cut down many options. Now I don't think backscattering would be easy to calculate and it might be also less helpful with discrete scanner data (I'm not even sure whether it's possible/practical to do). (BTW the incident angle I meant was the angle of the beam as it hits the target surface, thus it requires modelling the target surface normals).

One simple thing you could do (unless it's already done with scanner software or any post-processing) is correct intensity by distance (sometimes it's called R-squared correction), when you calculate for each point the distance from the scanner and correct the intensity appropriately.

Now with intensity as the only additional point attribute you probably have to analyse geometry of points and their neighbourhoods.

Jon - and when you visualize the point cloud - can you clearly distinguish visually the two classes: biomass and trunks/branches?

Do you get multiple points for branches/trunks, so it's possible to detect relatively smooth/linear surfaces? How many points do you get for branches/trunks? Is it easy to visually detect such objects?

The vegetation should give very rough surface that should be distinguishable from wood/trunks.

Other property is that vegetation/leaf/biomass should be somewhat penetrable by laser giving you many returns, forming very rough surface, that should be clearly different than solid reflections you get from wood/branches. (There is also the case then you have leafes/biomass, just partly covering wood/branches behind them - that again depends on your data and whether you need to cover such cases also with your model).

So I would start explore those geometric properties more, starting just with visual and exploratory analysis of the data.

Typically when you can clearly see with your eyes what you're looking after in your data, then it's relatively easy to come up with the algorithm to do the same automatically. When you cannot see that with your human eye, then the task is really hard.

If you can see in the point cloud what is what, then you can label and extract the training samples and then look at the data using statistical/machine learning methods trying to fit a model for classification. With discrete (non full waveform) you just have much less information about each point, so you need to analyse geometrical properties of many neighbouring points, which is always more complex approach.

My experiences are mostly with full waveform data, when we have really many point attributes, so it's really easy to make even more deriven features and create a model using machine learning algos with relative ease.

With discrete scan data that methodology seems much less suited, and I'm afraid I cannot provide much more ideas without looking at your data.

Regards,

Adam

Hello Jon,

I don't know the wavelength of the sacnner you used, but in my Master's dissertation we developed a software called K-Clouds that use K-means to classify data on point clouds. The main goal was Geology but we did some tests in vegetation with good results, been able to separate leafs and branches. My dissertation is available on http://www.hindawi.com/journals/tswj/2014/539029/ and we used an Optech Ilris 3D TLS 1535 nm wavelenght. Hope this can help.

Regards

Leonardo

Hello Leonardo, 

Thanks for the information.  I shall have a read of your thesis and get back to you with any questions if that's OK.

Kind regards, 

Jon.