 few administrative tasks. For those of you who are joining us for the first time, we organize our webinars in a webinar format, so you are not able to show your camera or unmute yourself. However, we do encourage you to communicate with each other in the chat. If you have any questions, please use the Q&A function. Some of those will be answered during the webinar, and we'll hopefully have enough time to answer a few of those live after. If you have any technical difficulties, please reach out to myself or my colleague, Isabelle Verbeck, and we'd be happy to help you with that. Before digging into our spectroscopy material today, I'd like to provide just a little bit of background on the Global Soil Partnership and go so on. So the Global Soil Partnership, or the GSP, was founded in 2012 to work towards healthy soils and soil sustainable management for everyone. So this is done through a mechanism to bring all of the stakeholders together to collaborate on soil issues. It's done through a unique framework of exchanges of experience and discussion among seven regional soil partnerships, over 500 partner institutions, seven technical networks, and 161 national focal points from FAO members. So as some of you who've worked with us before know, we collaborate or work in conjunction with the ITPS, which is the Intergovernmental Technical Panel on Soils, a group of soil experts that provide scientific and technical advice to the GSP. And then every year we have our annual GSP Plenary Assembly, which is our decision-making body, where all of the representatives from our FAO members come together and our GSP partners work together to review and prioritize soil actions. So Glossalon stands for the Global Soil Laboratory Network, and it was established in 2017 in answer to a growing need for harmonized soil analytical data with the mission of building and strengthening laboratory capacity and soil analysis. So they began with wet chemistry, and then in 2020, in answer to demand from many of our partners for growing interest in soil spectroscopy or dry chemistry, the Glossalon Initiative on Soil Spectroscopy. In 2021 and 22, we've begun work on consolidating the capacity development and activities on soil spectroscopy, such as the series of webinars of which this is the 12th online video courses and additional training material, which is available on our website. And so today, we're very excited to have our speaker with us, Dr. Leonardo Ramirez Lopez. He's the head of data science department at Buki Laboratory in Switzerland, and over the last 15 years, he's conducted research on infrared spectroscopy all over the world, such as in Colombia, Brazil, Germany, Belgium, and Switzerland. In particular, he's devoted to share his developments through open source software, and some of you might actually be familiar with his work as he's developed two very powerful R packages that are widely used in soil spectroscopy prospector and re-symbol. So his main research motivation is to bring soil spectroscopy to a fully operational level, and we're thrilled to have him here today speaking on working towards operational large-scale soil libraries. And his colleague, Estefina Fernandez Perez, who is an NIR expert also at Buki will be joining us today as well. So without further ado, I would like to give the floor to Dr. Ramirez Lopez. Yeah, thank you very much. So I'm going to share my screen. I'm very excited to participate on this series of webinars. I have to share my screen first. Yeah, everything okay? Can you? Okay, excellent. We can see okay. Yes, good. So yeah, today I'm going to talk about, yeah, my point of view on how to make, how to benefit from soil spectror libraries, how to make, how to use them to make operational spectroscopy in the lab, okay? So yeah, this presentation is based on the work that we have done with many people, and especially with the co-authors of this presentation, okay? So this is the slide, the introductory slide. This is the original introductory slide. You see I put some flies there, but it's not my present because my presentation stinks. You will see that, I mean, why I put some flies there later on in the presentation. I really love insects. I was really passionate about entomology when I was doing my bachelors. And yeah, they are very interesting insects from the evolutionary point of view. And yeah, today I'm going to talk a little bit about evolutionary methods, and that's why I put them there. Okay, I will jump directly into this slide that basically introduces how spectroscopy, like spectroscopic models are usually developed and then deployed. I already made a presentation like two years ago or one year ago where I presented this, but I thought, yeah, I will present this again because it is how it's basically, it's the basis of what I'm going to talk later on how to make this operational, library operation. So the first thing that we usually do is to collect data. We spend a lot of time collecting data, okay, in most of the projects that involve the creation of a spectral library's data collection is a very long journey, but organizing the data is another very, very long journey. So I would say that in my experience, more than 50% of the time that you have to invest into developing a spectral library goes into organizing data, cleaning the data, harmonizing the data, and so on. So yeah, sometimes it's even 80% of the time that you have to invest into developing a spectral library. And then what we usually, then this step is followed by the calibration of spectral models, and we have, I think that the scientific community has become very good at it. So we have been able to gather a lot of information on how to calibrate and how to validate spectral models. So we are really, really good at it. However, and basically this is an evidence of this is that the amount of scientific publications that have been released over the past 20 years or 53 years. So, but then, yeah, the work shouldn't stop there, of course, we have to make those models operational, we have to deploy those models, we have to bring those models into what so-called production, right? And to do that, usually, ideally we should pack those models. For example, if we develop a model for total carbon, total nitrogen and cutting exchange capacity, we have to pack them into kind of an application or into a container that can be validated, I mean validated in terms of the accuracy of the application, and see what is the, maybe we define some acceptance criteria for the models that we, whether we accept the models for production or whether we accept the models for operational use or not. And then once we say, okay, let's accept the models, then we release the application so that it can be used for routine analysis. So, what I said before is that we have become really good at calculating models, but I think that steps for five and six, we are not yet into, let's say, into a knowledge level that allows us to quickly bring spectroscopy to an operational level. So, that's the reason what I'm gonna talk about, how to make them, how to make operational, how to make a spectroscopy operational at the laboratory. I think that I have to adjust my mic. If you give me a second, I'm gonna go to the settings, here, then I think here, sorry for that, see how can I, okay, maybe if I speak louder, then that would work. But if at some point you cannot hear me, then please let me know and then I will try to go back to the settings. And okay, one second. Okay, so again, so to recap. So, spectroscopy, we have become really good at building spectral models, but we have not been really, I mean, we are not yet there when it comes to deploying such models and to make, as well as spectroscopy operational. Another thing that I want to stress in this presentation is that, yeah, most of the papers that, in most of the papers about the spectroscopy, you read that it is a very cost efficient method. So, and yes, that is right, but it also depends. So, once the method is operational, yes, it is very cost efficient. But now when it comes to do the process, to build models, to build the spectral libraries, I mean, this process is very, very expensive and tedious. So, let's say the operational side of spectroscopy is very cost efficient, but the development side of soil spectroscopy is very expensive. So, I think that that is one of the reasons why labs find it hard to implement spectroscopy. So, they usually focus on routine analysis and they sometimes they don't have really time to invest into developing, say, like operational soil spectral models. So, because sometimes also for these users, for lab users, this involves a lot of many time consuming iterations to get these applications to an estate in which they can be used. And this creates, of course, frustration. So, they sometimes they don't want to invest towards that. I mean, because of the, maybe because of budget reasons. Another thing that this whole process involves is, involves a lot of, or requires a lot of knowledge about chemometrics. Our chemometrics is a complex subject which requires time and effort to be learned and taught as well. So, those things are really critical when it comes to implementation of soil spectroscopy and we have to take them into account. However, I think that with all the open source libraries that are already out there, that are available, that we can access and then we can use. I think that with all this information that we already have collected in the past, I think that we can build operational models for laboratories in the field. So, how we can benefit from spectral libraries? How we can make these soil spectroscopy, how can we make soil spectroscopy operational? That is the question that I mean, that we should try to answer. And we shouldn't really focus too much on the, say, calibration, model calibration side. Because we are, again, we are really good at it. We should focus on bringing spectroscopy to the laboratories for routine analysis. So, how we can benefit from it? So, imagine in a laboratory, the success of the implementation of a spectroscopy depends on many factors. It depends on the quality of the hardware that you are using. Usually, the quality of the hardware for all the instrumentation is more or less really good. Software data, of course, to a less extent in chemometrics, but it also depends a lot on automation. And also depends a lot on people. But today, I'm going to focus on the automation part. And I'm going to focus on the topic of library search methods. So, because I think that library search methods are critical to make spectroscopy, again, operation. So, and I'm going to talk about, more specifically, about evolutionary genetic search. So, this is a method that we have developed over the past couple of years. And it has proved to be quite useful when it comes to profiting from, from, sorry, the spectral libraries. So, just to give you an introduction of what I'm talking about, I mean, when it comes to genetic search, imagine that we have individuals, many individuals that we can call them as well, subsets, subsets of samples. But let's call them individuals. And then we have genes. The genes are basically, can be, can be the samples that we have in our spectral libraries. So, the goal in genetic search is to breed one or more individuals that are feed enough to be integrated or to serve a target population. So, the goal is that we have a lot of genes. Imagine, the genes are the samples, again, in our big spectral libraries. And the goal is to identify the best genes that are suitable to make, say, feed, the feedest individuals. So, okay, in the, we have, for breeding, we need to breed many offspring. So, the first one works more or less like this. So, to every individual, basically subset, we have to find a kind of a weakness score. So, we find how, we try to assess how every individual or every subset, how weak it is. So, and this score is assigned to every gene in the, of the individual population. So, then here I have a table that more or less tries to explain how, how this is done. So, imagine the first individual, again, is a subset. The first individual, for the first individual, you can assess how good the subset or this individual is. And then to all the genes that you have inside that individual, you assign, you label each gene with the, with the gene, sorry, with the, with the weakness score that you found for that individual. So, individual two, you do the same to your individual end. Okay. Note that every gene is not present in all the individuals. So, at the end, you can, if you, if you do this, at the end you can have a kind of a, a gene weakness score. Okay. So, and this gene weakness score you can, you can, this gene weakness score you can, you can estimate the, for example, the probability density function and then you can assign a kind of a weakness tolerance threshold. And then you can suppress all the genes that bring weakness to the individuals or to your subsets. So, yeah, that is the goal. Basically, suppress the weakest genes and breed then a new offspring with the fetus genes identified in the previous offspring. So, the idea is that you, you breed as many offspring, offspring as necessary until a gene pool is identified. And these gene pools should produce the fetus individuals. Okay. So, this is more or less how, how this type of search, genetic search is, is not. So, how can we use that in spectral libraries to understand, for example, target populations? I have to say that this, this work on genetic searches is inspired by this paper published in 2017. It is about artist local. So, what we did was we, we used the code of artist local, the code methods in artist local and then we adapt them to, to the evolutionary search theory. So, now to explain you in a more, let's say graphical or like more applied example, I'm gonna, I bring these, these, these two sets. Now, the first one, imagine that it is the reference spectra that you have in your soil spectral library on the left and then on the right, you have your say sample set of samples that come from a region or from a lab for which you need to produce a model. So, how, how, how this genetic search can be applied? So, I'm gonna go a little bit technical here, but then the goal is that you can use, for example, PLS, you can use partiality square regression to do this. So, you have, so in, in your library, you can build a PLS model on every single individual or subset that you randomly choose. Okay. You can, you can, you compute the PLS scores, you can compute the PLS loadings and then you compute here. Then you can use this PLS model to project the spectra of the target population or the target set into, into the principal component, into a principal component or like into a PLS space. So, this PLS space is represented by this S. And then what you can do, I mean, this can be seen as a kind of a compression of the spectra in the, of the target population using a PLS model that it is built from the library. And then what you do is that you, you project or project back the, or like, so you, sorry, you have a, a, a compressed spectra here and then you decompress, basically compute the, like, do a back transformation to the spectra. And then you estimate the, like, the, the, the compression error, the spectral, the decompression error. So, and then with this spectral decompression error, you can, you can estimate the spectral reconstruction error for your target population using the, the, the, the, the model that you built from the library. So this is, this Q or spectral reconstruction error can be used as measure of weakness of every individual. Again, every individual is basically a subset of your library. You can also use the spectral dissimilarity to assess the, the weakness of your, of your individuals. So you can use basically the Q and the D. When you do that, you can then apply the, the genetic search method into your library. So what we, what, what I did here, it was that, okay, I use this, this, this, this method that we developed to find, sorry, one second. I use this, I use this, this genetic search method to find in the African soil spectral library that it is represented by basically, but every black dot that you see here in this map to find in this library, the best genes or samples that can be used to produce a model that predicts total carbon aspect like infrared model that to predict total carbon in the Chuappa region in Congo. And you can see that this region is very isolated geographically isolated from the rest of the black dots of the Apsis library. So that is, that was a challenging example. So here in the, on the right, you have the spectra of the Apsis library. And at the bottom, you have the spectra, the IR spectra of the Chuappa region in Congo. So what, what we did is that we identified 300 genes, which are basically samples out of a pool of and almost 2000 samples. So meaning that we extracted 300 samples and like basically that we extracted only, we use only 15% of the samples in the library to produce a model for the Chuappa region in Congo. So we breathe, we breathe 12, sorry, 18 offspring with a total of, we evaluated a total of 2700 individuals and we did that like 100 times. So at the end, we evaluated around 200, 270,000 individuals, meaning 270,000 subsets. Just to find the 300 fittest genes or the 300 best samples to produce a model in the spectral, in the spectral library to produce models for the Chuappa region. And this is basically what we obtained. Let's look first at the right validation plots. Okay. This GS local is basically the name of the, of the method that we developed stands for genetic search for local applications. And on the x-axis, we have the predicted total carbon and on the y-axis, we have the observed total carbon. And you can see that the predictions are quite, quite good. The uncertainty, what we can say about the uncertainty of these predictions is that the uncertainty is quite, let's say reasonable, quite low compared, for example, when it comes, when, when we evaluate the uncertainty of models, calibrated or total carbon predicted with memory-based learning methods. So although I would say that the results are very comparable. Okay. So here on the left side, I would say that this is the, the validation plot for memory-based learning. We wanted, we wanted to compare our results to something else that is, let's say, successfully used for predicting soil properties from soil spectral libraries. We also evaluated the confidence intervals. So in NBL, there are a lot of models that are produced to, to make up, to make the predictions of the, of your target set. But here in GS local, we only have one single model. And this is a PLS model, very basic model. So, and you can see that the confidence interval is quite, say, quite narrow, which means that the predictions are quite, the uncertainty of the predictions is quite small. So there are some, yeah, there are some maybe high uncertainties, the low levels of total carbon that might be associated with the detection limits in total carbon. So the advantage of using genetic search, for example, over methods of like, like NBL is that, again, here, you find these 300 samples and these 300 samples can be, can be used to produce one single simple model. In this case, these predictions, again, these predictions come from a PLS model that uses only 300 samples out of the 2000 samples of the library. So why I say that this is very, this is important because one single model is something that you can easily deploy, for example, in a lab. It's something that you can easily say is portable. It's easy to, it's easy to interpret as well. So you can gain knowledge about the ability of your target site or target population by looking or by exploring this PLS model. Okay. And so to come back to this slide about the success, I think that these type of methods, like genetic search methods are really important when it comes to automation. So why? Because they allow us to automatically detect the best, let's say the useful information that you have in your libraries for a given, let's say, purpose. So I believe that, I believe that the best way to use a spectral libraries is to use them, let's say, extract the relevant information for every particular case and not to try to, for example, feed a model that serves or that can predict every sample that you throw in it. Some of the contributions that I think that are relevant to mention here is that we focus, I mean, most of the research that we do focuses on the usability of spectral libraries rather than in the modeling of spectral libraries. Again, I think that we have reached a good understanding on how to calibrate models. So we use also another contribution is that we use search methods to better understand the target populations. Again, we extract the best information from the libraries to understand some target population, some local variation. We move away again from this concept of global model to a more localized model. So this is a kind of a pattern line shift. We make a soil spectroscopy simple at operational and we want to make it fast to implement and to service. And again, there is some, there is some, there is this saying that it says think globally and fit locally. I like it very much and it's very inspirational and it is basically based on this paper written in 2003 by Saoudan Royce. So I definitely recommend that one. And yeah, just to finalize, I want to, I want to say that yeah, we are planning to bring more of these methods into the packages that we are developing into open source software and like, like, like resemble and prospect are. And yeah, please stay tuned because next year or soon there are more, there will be more tools available in these packages and even additional packages that we are working. So yeah, that's it for today. And I think that we can move to the questions now from the audience. And again, apologies for the, for the, for the problems that I had with the sound. Thank you so much for that presentation. And just as we mentioned before, please feel free to ask questions using the Q&A function where you can even put them in the chat. So I think the first question is, do we need standardization of settings for the library to be usable? Yes, of course. I mean, when you work with standardized spectral libraries is way easier to find relevant information. When you work without standardized spectral libraries, then it is extracting relevant information from it becomes more challenging. However, I mean, I think that there are methods can can handle those disturbances that are introduced by standardized data. So like, we believe that things like genetic search can partially handle those challenges. And then from Jean Robertson, how big and how variable do you think the spectral libraries need to be for this to work? That is very relative. I mean, I think that it depends, depends whether the important question is that we need to ask ourselves is that, do you think that this library that you plan to use represents somehow or covers the variability of the target side that you plan to evaluate or for which you plan to build a model? But that is the question that I think that we need to answer that we need to ask ourselves every time we are going to consult a library. Thanks. And then we have, could you talk briefly about the computation efficiency about GS local method? Essentially, follow up question how time consuming it is it when you're dealing with a large library? Yeah, well, the whole time consuming is it that it depends on the computer you have. But, but we have, I mean, the way in which the method was implemented, it was basically to we build, I mean, we implemented the method in such a way that it you efficiently uses the computing resources. So for example, for this example that I was presenting before, like this hundred hundred iterations to find the best 300 samples for the Congo data, I think that it took me like about 20 minutes in a computer with 15 cores. So it was really not, I mean, it was quite computationally efficient, I would say for the amount of computations that were done like, like in, yeah, to find these 300 samples. And from Melissa Luis Gutierrez, how can I estimate the percentage of samples that I should use to calibrate the models? I think this is going back to, oh, what's your answer? No, that is quite, that is a very interesting question because this is very, at the moment, we are trying to find a method to estimate the amount of samples that you should use. And this is basically something that we, that it is currently under research. But we usually, I mean, in our experience, we usually extract from libraries from 150 to 300 samples. That's more or less, this is a very arbitrary, let's say, range of samples. But again, this is based on our experience, 100 to 300 is enough to build a really reliable model for, again, for a very specific population or subpopulation of samples. So to Rick Mitron ask, what is the level of accuracy we could achieve for soil property prediction using spectral libraries? And I know you showed the uncertainty compared to machine-based learning. Yeah, I think that, yeah, again, it depends. Remember, there was this, there is this quote that I like very much that it is about think globally and fit locally. And from my point of view, it depends on what the target population is. Sometimes if you have a really good coverage of your target population in the spectral library, you will see that your predictions will be quite accurate. But sometimes, for example, if you have a very, very, very rare set of samples, then the prediction might be not as good as you might find in the other case that I mentioned before. So it's very relative. One thing that I also wanted to talk about is the errors that we always tend to look at the RMSE. But we also have to look at the distribution of the RMSE. The RMSE is just one single value, and it is not a deterministic value. It has a distribution, and it is, and we also have to take into account how wide the distribution of this RMSE is. We should try to beat this culture of RMSE. It should also look more deeply into the models. I think this goes to your quote that you mentioned earlier, the thinking globally and acting, but I got that slightly wrong, the global and local. So Melissa asks, again, what would your recommendations be that the soil spectral libraries will be comparable worldwide? I think that is, I mean, from my point of view, again, it's not about comparing soil spectral libraries. Of course, it's more about integrating soil spectral libraries. I think that we have produced tons of data, and what we need to do is maybe, perhaps, integrate and pull all this data together to make it useful. I think that we should maybe try to focus more, again, more on the operational side, rather than on the, say, operational side of spectroscopy, rather than on the calibrating models or maybe even standardizing spectral libraries. Of course, that is important, but we also have to look at what we already have available and how we can integrate that. Because everything that we have is already quite valuable, and if we don't benefit from it, from what we have right now, we are basically, I mean, the opportunity cost that we are incurring at the moment is quite high. I mean, we are losing opportunities because we are not making, I mean, we are not profiting at the moment from libraries. Rarely you see operational models for soils in the laboratory. Leo, if you'd like to take a look at the questions, you're welcome to as well. You're good. I'm just... I don't see the questions, actually. No worries. I have zero questions in my chat box. If you... Hi, there is Estefania here. If you want, I can reduce some that I have spotted. That would be great, Estefania. For example, can you please explain why the prediction standard errors are low in the genetic search algorithm as opposed to the NBL? Yeah, that was not a fair comparison, I would say. The uncertainty estimation that I made for NBL and the uncertainty estimation that I made for GS local use different methods. So that is really not a fair comparison. I just wanted to bring the attention to the importance of also evaluating the uncertainty. And also I wanted to stress that the uncertainty of Paris local is quite low. And not to say that NBL has a wider uncertainty range. One question here by Conrad. What is exactly the point of the multiple generations, the offspring, you described? Okay. Yeah, the point is that the first offspring, for example, it contains, imagine, thousands of genes. So, and then if you go, if you breed a new offspring, the gene pool, meaning the sample, the amount of samples that you have probably becomes smaller. So the idea is that then you suppress all the genes or suppress all the samples that do not bring any relevant information for your target population. So that's the reason why we breed different offspring. So in the example that I show, we bred around 18 offspring. So we went from 1,700 samples to 300. How could we quantify the spectroscopy similarity between training and testing samples? So the similarity, there are a bunch of methods that you can use to estimate the similarity or the similarity between samples. Usually you can use, for example, PLS components to do that. Basically you estimate for each sample in your target population, you estimate the Euclidean or Mahalanoi distance to the, to every sample in your, or to the center of the population of your subsets that you bring or that you draw from the big library. So yeah, again, it's basically Mahalanoi distance standard one. Anonymous attendee is asking if you think those algorithms are, well, still need to be tested by larger scale data before making them operational or if you think they are ready for application now? Yeah, this is a good question. And actually, every time I present this, I mean, I have presented this internally, for example, in the company and also to some, some colleagues here in Switzerland, and this question always comes. We have tested these intensively to build models and to make models operational in other, say, for other products, not soil, but for example, for other agricultural products. So we know, I mean, we have a good indication that it works in the field in reality, in real life. So, but yes, I guess that ideally we should do more total tests when it comes to soil spectroscopy. In soil spectroscopy, I have used the genetic search methods, not as often as I use it for, yeah, for our products, but I have found another one that I found interesting here. Sorry. If we need to select among several global soil spectral libraries, how can we select best library? Or how we know that? Or how we know which spectral library is best in our case? I think that sometimes if you don't know it, then you put everything together. But if you know, if you have a good indication that, for example, in this case, in the case that I, in the example that I presented, of course, I had, I could choose between a big spectral library and comprehensive spectral library developed for the US. But turns out that my target population was in Congo. So, yeah, it was obvious that I would choose the African one, because it's, I mean, from my ex, let's say like basic aspect knowledge, I would expect that the soils from Africa resemble the most the soils from Congo. So basically, yeah, I can limit, say, the extent of the library to samples, I mean, to samples that I know that are somehow they share some, that might share some similarities with the target population. Thank you very much. And anonymous attendee again, or perhaps it's a different one, but how would these methods work when extending models to spectra collected by another instrument with different spectral range and resolution compared to the host instrument? Yeah, of course, this is another, this is a different, let's say, this is a different topic that it would be basically how to integrate data from different instruments. So it's not much, it's not much about, let's say, search. But I think that, I mean, if you already have, for example, something like a subset of samples that you extracted from your big library, imagine you have 300 samples that you are currently using for, for routine analysis in your laboratory. If you want to extend this library with additional samples coming from different sensors, yes, of course, you can still do it and you can easily do it. I mean, the advantage of the method is that you have your samples available. You have your three samples, sorry, you have your 300 samples that you can put together with additional samples that are coming from other sources or that they are collected specifically to enhance the, for example, the accuracy of that initial model that you built from your library. So I don't see that there is any like big drawback on enhancing or augmenting the samples that you extract from libraries. Jean Robertson is asking, how much better are the GS selected local calibrations compared to other ways of selecting spectra for a local calibration? Yeah, okay, that's a good question as well. And I think that, I think that the advantage of GS, of genetic search is that our assumption is that it can handle very well disturbances in the spectra of the library that are caused by, for example, multiple data sources, for example, multiple devices that are used to collect a specific library, or it can also handle very well in some using an additional method that I didn't talk about today. It can handle also very well the disturbances in the in the in the response variables that are introduced by different laboratories. So when you have a library, maybe the response variables, the values of the response variables originate from multiple laboratories with multiple errors. I think that there are some options that you can use in genetic search to handle those disturbances. So that is the advantage basically over conventional search methods that you can filter out a lot of, let's say, circumvent the problems of disturbances, spectral and response disturbances in libraries. Okay, going up a bit on the list of questions. Do you use different soil types for the same prediction model? How local your feeds are? Too local? I didn't understand the question. Do you use different soil types like forest, meadow, arable soil for the same prediction model? Yeah, in this case, of course, again, in this case, I didn't differentiate between the soil, I mean the origin of the soil samples. So of course, you can again, you can use your expert knowledge to filter out a lot of information that you have initially in your respective library. You can limit, for example, if you are working with forest soils, then you might want to exclude some from your library soils that do not belong or that were not collected in forest or things like that. Yeah, I think that basically you can use, again, your expert knowledge to filter out first the samples that you are going to use to feed the GS local method. So with that, I think we're reaching the end of our time. So thank you guys so much for your presentation and for everyone who paid attention. If you have any other follow-up questions, feel free to reach out to me. I will be emailing all of you the link to the recording and the presentation. If you have any follow-up questions and thank you all for all your questions and your attention and for the ideas for training going forward, that's much appreciated. Bye, Dr. Stefania. Bye-bye, thank you guys so much. Thank you guys so much. Bye. Thanks, everyone.