We're sorry but this page doesn't work properly without JavaScript enabled. Please enable it to continue.
Feedback

Mapping Soil Organic Carbon In Soil Profiles using Imaging Spectroscopy

00:00

Formale Metadaten

Titel
Mapping Soil Organic Carbon In Soil Profiles using Imaging Spectroscopy
Serientitel
Anzahl der Teile
4
Autor
Lizenz
CC-Namensnennung - keine kommerzielle Nutzung 3.0 Deutschland:
Sie dürfen das Werk bzw. den Inhalt zu jedem legalen und nicht-kommerziellen Zweck nutzen, verändern und in unveränderter oder veränderter Form vervielfältigen, verbreiten und öffentlich zugänglich machen, sofern Sie den Namen des Autors/Rechteinhabers in der von ihm festgelegten Weise nennen.
Identifikatoren
Herausgeber
Erscheinungsjahr
Sprache
Produzent
Produktionsjahr2022
ProduktionsortWageningen

Inhaltliche Metadaten

Fachgebiet
Genre
Abstract
Imaging spectroscopy has the potential to enable measurement of soil properties in intact soil profiles at spatial scales previously not possible. There are unique challenges associated with imaging spectroscopy compared to point spectroscopy. Particularly, signal noise and the influence of non-compositional effects on the spectra. One way to manage these effects is by using wavelet transforms, which are a signal processing technique. Combining wavelet transform processed spectra with machine learning techniques can be used to improve predictions of soil organic carbon throughout the soil profile with imaging spectroscopy. In one study, intact soil cores were analyzed using a SisuROCK automated hyperspectral imaging system in a laboratory setting collecting shortwave infrared reflectance data. Predictive models were then built for soil organic carbon using a combination of wavelet analysis and Bayesian Regularized Neural Nets. The combination of wavelets, machine learning and imaging spectroscopy enabled mapping of soil organic carbon throughout the profile, and identification of the magnitude and depths that rotational treatments were having an effect. This webinar is based on the following publication: Sorenson, P. T., Quideau, S. A., Rivard, B., & Dyck, M. (2020). Distribution mapping of soil profile carbon and nitrogen with laboratory imaging spectroscopy. Geoderma, 359, 113982 https://doi.org/10.1016/j.geoderma.2019.113982.
Schlagwörter
Computeranimation
Computeranimation
Computeranimation
Technische Zeichnung
Diagramm
ComputeranimationDiagramm
Computeranimation
Technische ZeichnungDiagramm
Computeranimation
Transkript: Englisch(automatisch erzeugt)
So thank you very much for the introduction. Tom and I guess with that, just for time, I'll dive into things so. This presentation will be talking about mapping organic carbon and so profiles using imaging spectroscopy, so this is this work
that I completed as part of my PhD and published in geoderma which I'll present the citation at the end. So, with this presentation I'm going to just go over kind of the objectives of the project, the methods we used I'm going to spend some time talking about signal processing technique that I used as part of this project
Wavelength Transforms and then spending a little time talking about that to just give a little more background about them. I think that, particularly for kind of noisier data that we get with imaging spectroscopy, they have a lot of potential value.
And I think they're worth kind of exploring and considering as part of the soil spectroscopy toolkit I'll present some results, as well as the conclusions. So the overall objective of this study was to map soil organic carbon throughout the soil profile and then to investigate rotational effects on soil organic carbon distributions.
One of the things I was putting this project together, I had been thinking about was that there's a lot of great papers on method development with soil spectroscopy and I think there's still a need for that. My personal I guess hypothesis was that I think we're getting to the point with spectroscopy that it can help us answer research questions in a way that we couldn't with other more conventional tools, particularly with the resolution of data that we can get.
So, just to give you a little bit idea about the site before I kind of jump into the methods for the research project. So this site is located in the province of Alberta in Canada, which is in Western Canada.
So this is actually on long term research plots they were established in the 1920s to look at developing farming practices for these types of soils. In Western Canada, this particular plot with them was looking at an incorporating forages into the mix and. And that was done in the 1970s, so these plots have had this treatment on them for about 40 years, the types of
soils are in the Canadian system ortho grey Luva cells I don't expect necessarily many of you to be familiar with the Canadian system. They translate as bore all in the USDA which my apologies i'm not as familiar with the USDA
system and an albic Luva cells in the world reference base so they're kind of key features of a clay translocation out of the horizon, so the rotations on the study, I wanted to investigate were. agro ecological rotation, they call that which is really a forage legume grain mix a continuous forage rotation continuous grain and then.
sort of the traditional which is not really done much anymore, but we follow next in the prairie, so it gave us. same soil but different rotational effects, I was hoping, would give us a nice picture of different carbon dynamics throughout the profile so in terms of collecting.
The data, so I have a picture of example of of just some soil cores. So these would be the cores the cylinders that we were imaging so the spectra collected with a sisu rock hyperspectral imaging system, so this is produced by a company specimen in Finland.
It actually has two cameras of visible light and a shortwave infrared but for this study I focused on the shortwave infrared data. So it collects data from 1000 to 2500 nanometers and 256 spectral bands. And I guess the what the really exciting part for me was was that with this setup we were able to get a
point two millimeter spatial resolution for the pixels so really dive into to the the variations and find spatial scales in the sample. Just for our reference data for calibration was all done by dry combustion using a cost tech elemental analyzer.
You know, while this is kind of specific just to mention how we build the calibration model, so I tagged. Take where we're going to get our samples for each of the image so once we collected the image, then we took like a slice for lining up to this point for the entire slice. homogenized it and then for building calibration models took the average reflected spectra for the region where the sample came from.
Because that is one of the challenges is how do you calibrate your data to your lab samples, so one of the challenges I guess I would say with. Some of the imaging spectroscopy which you know if you're used to working with potentially it's the ASD equipment.
In the vis near infrared or Mr data the spectra can be quite a bit noisier than what you might be used to seeing so this, for example. Is the is a single pixel spectra that I retrieved, for an example, and this is the spectra after averaging so a lot of this.
A lot of this fine scaled noise from a single spectra doesn't really translate out when you average it out. So thinking about how we deal with some of the noise in the spectra is a really kind of important issue.
With some I think, at least with this system in particular, but the nature, I think, of imaging spectroscopy is you're just going to more likely have noisier spectra. So you can see a lot of this fine scale variants which does create challenges for things that just taking derivatives because those are going to amplify or sorry amplify some of the noise.
So one of the techniques that i'm using this study that i'm going to now kind of maybe spend a little bit time talking about. Because, hopefully, this is one of the takeaways is that I use a process called wavelet transforms on the signal, so this is our signal processing tool that have been around for quite a long time in the digital signal processing world.
And what it does is to let you transform a single spectrum into a number of coefficients, known as wavelets and each then of these wavelet coefficients. will let you capture variants in the spectra different scales, one of the advantages of it is that the shapes and magnitudes are preserved of features using wavelet analysis, you can use the results for modeling unlike.
say continuum renewable removal, which is really valuable for visual spectral analysis, but you won't necessarily get consistency, depending on the baseline variants so. What will let you do is this is just sort of an example figure i'll show
you some from spectra as it's going to let you break down variants than at different scales. So there's a range of different types, but in terms of the two kind of starting places to explore would be what's called continuous wavelet transforms and discrete wavelet transforms so continuous wavelet transforms preserve.
You know, for every one of your like wavelengths you're going to retain a value for those wavelengths whereas discrete will reduce the number of data dimensions, because there is redundancy in the data. The advantage of the continuous wavelet transforms are they're more easily compared to your original spectra and you can.
More easily some different scales together to use multiple wavelet scales. The discrete wavelet transforms are a little harder to to translate to the original spectra to make comparisons and to sort of explain Okay, well, we have this feature that was important so that's probably this.
This covalent bond that was driving driving that signal. And there's a range of what are called different what are mother wavelets or what they're called, but there are different wavelet functions that are fit to the data so.
If you want to explore this there's a whole range of them, the second order Gaussian wavelet is the one that I. stick to using with spectral data, and the reason for that is that a lot of individual spectral features can be represented by Gaussian or quasi Gaussian functions. And I want to stop and point out that a lot of these ideas around wavelets are really
drawing off of the work from Benoit Rivard's group at the University of Alberta, who was on my. On my committee for my PhD and this paper from 2008 if you're interested, is a good kind of theoretical or more a better kind of bigger picture discussion of wavelets and how to apply them to spectroscopy and some of the concepts.
My theory is one of the reasons, maybe some of this hasn't it hasn't it must attention is that. I think if you look at the literature, the machine learning models have had a lot of focus. Which I think is probably rightfully so, is, I think they have a bigger benefit for model performance, however, I still think these tools have a role, particularly in sort of our.
field spectroscopy imaging spectroscopy things where you have a lot of variance in some of those conditions that you're scanning and. So, if we look at some examples we take this, this is just an example of a shortwave spectra from this study in the shortwave region, so these absorption features are relating to the water at 14 and 1900 nanometers.
So you're breaking down your different scales are all to the power of two so when you're looking at the first scale it's really variants across like two two bandwidths. scale two you're looking at four three looking at eight so you're
starting to get smoother and smoother features because they're corresponding to wider wider. aspects in the spectra so we can often think about these scale one is really representing a lot of the noise in the spectra so this lets us denoise the spectra by removing it.
And then, as we move again, this is the original spectra to higher scales you start to get the broader features. So very much by the time you get to these higher spectra such as our sorry higher scale such as six seven and eight we're really looking at those. Like baseline things relating to the baseline so overall illumination different scattering effects from particle size.
A lot of those aspects, so the wavelet analysis with the spectra lets us both remove the noise and also lets us removes. or at least reduce some of those non compositional effects in the spectra that we don't really we're not trying to model directly.
So that was the kind of goal of the wavelet transform was to remove those non compositional effects and also the noise. So, to give you an idea and this study the optimal looking at trying to testing through cross validation on a training data set only.
ended up being the sum of the second, third and fourth order wavelets that were optimal so we're taking this this example spectra here. unprocessed and we're then transforming it into this. This wavelet coefficient values where removing some of that fine scale noise that we can't really see on this plot,
but if we were to zoom in you'd see as well as removing some of the underlying structure in the data. Just to mention, if you want to explore these tools in our there's the wm TSA or wavelet packages either and then Python it's pie wt is the packages to be used.
So I guess to kind of close up my discussion on the wavelet transforms here, I just wanted to close up by I had mentioned discrete wavelet transforms which reduce the date dimensionality of the data, so, for example, we're looking at. The scale like the third order scale for continuous wavelet transforms here we've got we retain
a value for each one of our bands and the discrete wavelets because a lot of the. A lot of the curve fitting. leads to data points that actually aren't necessary anymore, because you know that they're going to be
resulting from a particular mathematical function, so the dimensionality can be reduced in this case because it's. 256 original bands and scale three is leads to. A value of eight you're able to reduce from 256 to 32 band or 32 values, so an eight fold reduction
in the dimensions of the data, while preserving things such as, for example, this feature here relating to water this feature here relating to water, you get some of these the nature of the the wavelet analysis does lead to these kind of.
peaks appearing, but you are retaining absorb. person, if I may interrupt you, so the other two like there's many mathematical methods to transform so like principal components and just using derivatives what's the what's the advantage of wavelets what why why wavelets why not do just principal components.
So you could use, for example, the continuous wavelet transform and do principal components, so you could use the continuous do a data dimensionality reduction. That way, the reason i'd say to use this approach over other processing tools being such as the taking a derivative, because the derivative is going to increase the noise.
In the data. Whereas I mean, then you can do other signal processing tools such as so that's key goal a smoothing which again you have window size choices there. I have found, at least with this type of data that the combination of of noise
reduction and that baseline removal at the higher order wavelets i've got better results than doing. The kind of derivatives and so that's key goal a smoothing I haven't actually tested the discrete wavelets as a dimensionality reduction tool as compared to principal component analysis or partially squared regression I think that's worth. That would be worth exploring in the literature someone examining the use of these tools for for dimensionality reduction.
So kind of moving on out of that which i'm happy to answer more questions later just for time in terms of the predictive model in this study I tested a range of different predictive model types.
What I wanted to point out is actually the Bayesian regularized neural nets, I had the best results for this data set, I think. Probably many of people on the call will probably agree that the optimal model will vary depending on the data set like sometimes
i've had support vector machines that sometimes key with models, the thing I wanted to point out about the Bayesian neural nets is that. So there are two layer neural net and what's particular about them is they're using what's called the new when would row algorithm to sign initial weights and then a ghost Newton algorithm for optimization. I found them to be a lot easier to optimize and fit to data and had a
lot more success than using some other types of neural nets, particularly with smaller data sets so. it's something if you're want to start exploring different neural nets I would actually recommend it something you kind of play around with and see. Prior to doing any predictions I then did a three by three median focal filter on.
On the image the hyperspectral image, and the reason for that was to reduce some of that noise that i've shown in the initial spectra. And then one of the I guess the big advantages of collecting the whole course is that we could then do some comparisons to look at things such as moran's I to look at a good spatial aggregation of carbon.
Separate from sort of physical aggregates and also use tools such as spatial generalized. linear models or least squares in order to look for treatment effects and really tease out where exactly are treatment effects are happening in the profile. So I just for time i'm going to spend more time kind of going through the actual imagery that we are able to generate but overall are.
Was quite happy with our fit of our carbon and nitrogen models here with an R squared of 0.94 for carbon and 0.88 for total nitrogen. And this was then independent validation results we'd split our data into a training
and test set and did all of our optimization on the training data only. We also did a clay contents as well, which had an R squared of 0.8 not quite as good. My thoughts there are that I have to explore it more, but I think that potentially some of the the carbon features kind of dampening some of the clay features might.
might have been a factor, and the other reason is just there was more error in our in our clay training lab data than our carbon data, I think clay content the calibration methods are always kind of a challenge. So we look at our results. we're able to pull out.
kind of exactly where our horizon boundaries were occurring, based on the carbon content and also the clay content in this profile just one of these example profiles it's difficult to see with this scale, but if we. kind of zoom in and if we change the scale, you actually can see that we have a lot of fine
scale variation and so carbon contents and a lot of kind of aggregation of so carbon it's by no means homogenous. And that's one of the things that we can really pull out with this method same thing with clay content as well, so these soils again are characterized by a loss of clay from the horizon into the horizon, and we can see that there is is some variability with the aggregates.
The other thing I think the really valuable part about imaging spectroscopy for. This project is that we're able to really precisely measure what depth we're starting to see the treatment effect differences, so this continuous forage rotation.
led to increase carbon all to about five centimeters and then the forage or the forage green mix we had increased carbon down to about 11 centimeters. And we're able to see kind of exactly where that was occurring because of the fine spatial resolution of our data.
This was tied to increases in carbon nitrogen ratio and also we're able to pull out things like the carbon wasn't uniformly distributed was exhibited spatial aggregation and was increasingly aggregated when we're looking at those those two forage mixed treatments. And then, one of the things is we didn't see much difference in carbon contents at depth between these treatments.
Again I wouldn't conclude that as a, this is what means for forage part of this is probably because that high clay B horizon kind of restricted carbon additions deeper in the profile, which could be very different in a ternizemic or mollisolic soil compared to these lubosolic or borol soils.
Just kind of cognizant of time here, I think, kind of overall from this study and from some other work I've done I've been really finding wavelet transforms have been a valuable tool. When you have data sets that have noise or non compositional effects that
you want to consistently remove and standardized and I think the other thing. Maybe as a wider community that I've been thinking is that I think that reflectance spectroscopy and imaging spectroscopy is really starting to mature into a tool that we can use to help us. To help answer research questions and we should be thinking about it that way, and how can
we use this higher resolution of data to better understand soil on landscapes or within the profile. And I think also these imaging spectroscopy tools are going to allow us to really understand sort of dynamics and soil at very fine spatial scales in a way that you can't if you've got a piece the samples up to get discrete measurements.
So I have made the code from this project available on github. And then i've also made code available for spectral pre processing scripts using wavelets that are annotated better and. and cleaner code than some of this other code, so if you're looking at exploring some of these tools, you can go there i'm happy to answer any questions.
And then i've also included the citation from this study if you're interested in in reading that and getting more details so with that i'm happy to answer any questions if there are any and. Thank you very much for the opportunity to present here.
Thank you keep the slides on because the questions will be. slide so just keep them on. yeah president thanks so much for a fantastic talk and i'll see i'll give the audience. Another wavelet stuff is maybe a bit of a whirlwind there's a whole world of detail to go into with that. Okay well all the audience comes up with some questions i'll ask you a quick question how transferable do you think the model is that you built.
From this like one time sample collection, so do you think you could go at a different season different moisture levels and apply to models, or do you think you'll have to do some recalibration of the model because.
So I will say with this study I didn't mention I air dried all the course. Before imaging them, so I did remove that moisture effect it wouldn't be transferable to field moist samples. I the I think the wavelets help a bit with the moisture contents like that overall baseline removal from the water, but.
Thank you probably the water effects that dampen the individual features there's no fix for that and so maybe some of those moisture correction tools like. External parameter orthogonization might be something to be explored in conjunction, to see if that
helps normalize things, but the goal of this study was very much to generate the. The better mapping of the core so on, but no, I think there's still those questions to be thought about for transferability and. I think water content is the bane of for all of us right for field spectroscopy so.
Can I also ask a question john about the matching the spatial scale like you said the images are very high resolutions like a 0.2 centimeter right. And so, how do you then match that reflectance of that 0.2 centimeter pixel
with the calibrated laboratory value of solar organic carbon I mean, how do you match that. yeah so in that case I was matching based on an average spectra from the point that we took the sample from so. So there are some questions you could ask about well how much individual pixels match an average spectra because.
You know, we see more noise in this spectra my solution to that was to fall to use a median focal filter so we do really lose some of that spatial resolution by focal filtering, but it does reduce some of the noise and then also yeah but it's a it's a it's a legitimate question that I don't.
calibration plots the collaboration plots to show these actual laboratory estimates of solar organic carbon and laboratory needs like hundred gram I don't know. Right yeah so and these hundred gram I mean they will occupy at least you know, five by
five centimeter yeah right so so there's a for sure there's a mismatch between laboratory and the spatial resolution. There is yeah and so i'm i'm very much my laboratory my my spectra i'm using are an average corresponding to the same spatial scale as the laboratory sample but it's a good question I don't think there's.
Potentially, if you could do a laboratory analysis of very of micro samples like I don't know one by one centimeter you know and then correlate that exact pixel you could get a much better accuracy right. You probably could yeah I think that would have things for sure.
And I have one more question for me if nobody's asked you asking. One more question you did only nitrogen and carbon right. I did play content as well, I didn't focus on it. Okay, Mr and it's because you couldn't get all the other laboratory analysis, because you know you could have done also some.
nutrients micronutrients I don't know that will be also very interesting. yeah and it would be and there's been some different historical treatments I my guess my thoughts, I mean it's the whole debate of some of the macro nutrient patterns would largely have probably been a carbon. Because the macro nutrients would have been largely just associated with the client
carbon that's my theory is what we would have been mapping anyways but. It just give us an idea how much does it cost one scan you have to drive right you take the. yeah I air dried it and then so yeah the scans really the marginal cost is minuscule i'm not sure what they're
charging now the specimen I think it was maybe 250 100,000 ish i'm guessing on what the whole the whole setup. I know, Dr Rafar was involved in some of the early helping specimen get the early designs done so I don't know what they're charging for this system now. But to scan a core takes I don't know five minutes or it takes maybe one or two minutes once it's in.
It does it there's a couple questions in the chat now in the Q amp a so the first one, what are your thoughts on using hyper spectral remote sensing for mapping of solar organic carbon. So I mean I definitely think there's great potential I mean it's one getting the calibration setup
and the variants and moisture content will have to figure out and then making sure we have bear. Well, if we're trying to directly sense it right, we have bear surface only and then there's some questions about how do you translate the top millimeter to have a depth function, so there are those challenges.
If we're using it as a co variant multi multi spectral I think that I would probably need to defer to some of the you know vegetation. remote sensing experts about how useful is hyper spectrum what sensing for pulling out things like nitrogen dynamics and plants, can we correlate that. or not, but I think it's it's something that for getting better spatial representation I think it's
definitely worth exploring there are I mean there's some challenges but that's the fun part right so. yeah I know I think your answer really I kind of think I feel like there's almost this bifurcation in the field, right now, those that are really focused on directly sensing soil carbon from that
hyper spectral data and others that are just using it as an additional data layer in a larger. yeah my only thoughts on the direct sensing, which I think has great potential and you can be, then I think really confident about what the relationship is is that we're still going to need functions to translate the surface measurement to a depth profile right so.
I see the next question time of tropical soils, the overall baseline spectra has a strong correlation with soil texture not only absorption features. Moving moving the higher order wavelets may have impacted the prediction capacity for clay yeah I don't have much experience with with tropical soils.
So I'll have to defer right to the hoses expertise there, it could be something I didn't think of, and I think it might be worth or circling back to. I think some of our clays and the dynamics and the soils here are very different from tropical soils and a lot of so.
But I think a lot of what I mean in our soil as i'm talking about here it's the carbon content carbon content and water content tend to be the two things that really drive overall baseline reflectance. But there are circumstances where I think certain clay types are darker and I think you're probably right so that's a good point to consider.
And maybe I can mention also the European space agency starting next year with the chime it's a high perspective satellite it will be public data, the same as sentinel.
And so we'll have basically a high perspective imaging from sky and we couldn't eventually have both. high perspective from the sky and on the ground right and do like a fusion like really match the exactly the same wavelengths and then and then try to correlate the same wavelengths from the satellite that we also detect in situ let's say.
This is not the question i'm just just introducing it but yeah things are changing rapidly. yeah and i'll say for bear soil sensing whether that's multi or hyperspectral. You know i'd be interested what other people's perspective, some other regions, the
challenge we have in the northern great plains in North America is that we. Have very little bear soil now because conservation tillage is pretty extensive so we almost always are going to have crop residues in the mix so maybe there's spectral and mixing techniques i'm not sure that'll be valuable but.
i'm not sure from other parts of the world right that's a different story in terms of bear soil surface frequency. And in this instrument, have you tried it with the organic source like with the 30 40% organic. I haven't it would be interesting to know if we're going to have issues with saturation i'm not sure if anybody here wants to chime in as experience with other you know with other let's say with asd
equipment or point spectroscopy organic soils I wonder what the signal the noise is going to look like but. that's a hypothesis i've never done it. yeah that's a good point and I mean I definitely I mean my experience in the nir range and you start getting very it's yeah you're you're.
This soil starts absorbing almost all the light, and so you start getting a lot of noise when there's really dark PD soils, but it would be interesting and I think that depends a lot on your. optical configuration and with kind of the light source being. And the detectors being pretty far removed from the core you might actually do all right, it will be
cool to test out, especially given the extent of peatlands in Canada, the yeah good country to be pushing that work yeah and i'd say as much as I think the wavelets are useful tool for the noisings. If if there's something in your spectra that's dampening the absorption features
it doesn't really help there right because you still have a weaker signal. So i'm just thinking if it's really wet or if it's really high carbon right that's masking the other absorption features. yeah I want to say, like the imaging the imaging of the whole like scan and it reminds me of all the problems in the salt taxonomy.
Where they they want to come back, you know, like once you make a team is you can save it forever right yeah so sometimes they want to come back to the sole profile descriptions and. And then they would like to really see like you know the transitions and reclassify yeah I think if we could scan basically every
every soil if we could scan it like this and then keep that you know basically forever that people can come back to that. It will be really major major progress, I think, for so science, you know wasn't the focus in this study, I think some work with higher horizon boundaries would be really interesting. To do, because I mean just you can see, I mean we have a lot more regular horizon boundaries, particularly.
And this was a sample that was had forges for the last 40 years, so you know if you have your plow layer right you tend to have a lot of those stark boundaries, but I think some of those irregular boundaries and less disturbed soils would be able to examine a lot of interesting dynamics.
And now this images, so they may be available to public. Like I can't make them available I will maybe have to chat Tom about how we can make stuff more you know I was, I was new to figuring out some of the stuff about availability of making data. A great great contribution and we will dedicate an article on our source spectroscopy.org website.
And we would open open it for users to test different type of analysis modeling so how many skills, do you have, by the way. I had to double check how many course, it was, I think it was like 20 course about.
Okay, and there's one question just wanting from Andrew grant about lab method, so it was done by dry combustion with a cost tech elemental analyzer. Preston I had a question about kind of scaling this up to thinking about kind of the carbon market monitoring applications, I mean what
where do you do you see this as being a really viable tool for monitoring carbon change at whether feel like project level monitoring. I don't know if imaging spectroscopy would necessarily be the right approach on that like.
I don't know you'd really have to do a cost benefit of what the extra spatial resolution gets you versus point spectroscopy point spectroscopy is going to be more convenient in the field, I think, to deploy. I personally in a camp that I think you know as a community, we absolutely need to get
which i'm not saying we're there yet but but spectroscopy is one of the tools for carbon verification. Because you know I think quantitative validation of of carbon data is going to be essential to make sure the real and you know. The dry combustion and all the Labor involved with that is going to be quite expensive, I think there's
definitely potential, but I think some of those dynamics with water content variability is going to be a key thing. There's a question in the Q&A from Andrew still asking about like it looked like there's a lot of noise at the low end of the carbon and what the model performance was probably driven by you know the across that 7% gradient in carbon.
yeah I think that's a fair comment I think you're right, I think the performance was was started to break down at the lower lower content i'm not confident. That we can really reliably predict like below point two or point three car percent with this with this particular data set and study.
I don't know if anyone's working on really trying to nail down like low carbon measurement error like what's the detection limits took the question i've often had but yeah that's a that's a point.
yeah I think in general, I mean if you look generally across spectroscopy results when you have a very small range in your analyte invariably you don't get good predictions yeah and I don't I mean that that's that's that's a big problem if. you're really trying to intensively study a region that doesn't have a lot of variation in
that property and you're pretty much just predicting the mean value and that's not very useful. yeah, and so I think you know my view is how to fairly interpret some of these plots is that we are definitely pulling out. Where we have higher carbon in the in the top soil and that we're able to pull up pockets of higher carbon in the in
the subsoil but some of this variability could be just that error i'm not sure but definitely I think some of the sorry i'll say. You know, to say oh yeah this is point 2% carbon here, and this is point 3% carbon yeah I think it's fair to say we probably can't be confident that we're able to detect those kind of differences, but
we could probably pick up nodules of higher carbon in the subsoil which I think we could be confident do exist. The practical question best one is the the course you need a mechanical instrument to take the course right.
yeah I mean yeah you really do that's expensive that's the cover so really right because you cannot do it manually, could you take this like the one meter courses. yeah The thing, though, I think I think you'd have to circle back and look at the economics, though, because you're right it takes the instrument but it's much faster.
Like I mean if you were only doing you know, a surface, you know 10 centimeter sample with a shovel that's fast, but with the right kind of setup coring can be much, much faster so in the right soils, you could get really good cores in less than 10 minutes, you can be on to your next core but.
Canadian prairie is great for coring. Oh yeah I should say i'm biased we're talking about glacial and this is all glacial till sediment that I have, I think it was the question draped over till so there's not even the the stones to worry about. So yeah we're kind of ideal for pushing these cores in and just getting nice
you know we occasionally get a sandy or deposit and we get poor recovery but yeah. But yeah no that's not the same everywhere so. But I mean definitely I mean I would love to see this level of data like it basically every long term. agronomic trial, I would love to see this resolution of data coming out of them to really
understand this like the impact of different routing systems and stuff on carbon carbon and other properties. um I see there's one comment here about some machine learning project problems do show problems we're mustering with extrapolation.
When you calibrate the average over multiple pixels the information hotspots might be diluted so of problems finding them in the prediction map. it's important to reduce the area of reference sampling and therefore the number of average pixels for the model as much as possible. or use different yeah I mean I think that's a really, really good point and that's
something I think coming out of this talk or since i've done this study that I think. If you're building calibration models for this kind of imaging spectroscopy, how can you try to have like your calibration models be smaller I think that's all super valid. improvement that can be done on any kind of follow ups.
yeah maybe definitely say I mean, depending on your elemental analyzer I mean some of the micro element analyzers you only need a few milligrams. sample, so you can do a pretty good you could I mean you can't you can't sample at point two millimeters but yeah you could get like a one centimeter square out and easily have a good sample for analysis.
And yeah I think that would be a better way to do it as far as a future any kind of future applications. Great well. I think we've gotten through some really good questions and it was good discussion and so just
thank everyone for tuning in and it's really great thanks to Preston for giving really interesting talk. And yeah if anyone i'm happy to share code and if anybody wants wants to explore this stuff and wants help with anything i'm happy happy. willing to share the images.
Please get the DUI and happy to post them on our website and write the blog. Okay i'll follow up with you, Tom about that so. That would be fantastic. Thank you so much Preston. yeah Thank you, thank you again for the opportunity to present here so.