 Good morning, good afternoon or good evening, depending on where you are. It's very early for some of you, very late for other and we thank you for that. It's my pleasure to welcome you to the third session of the Glossel and Soil Spectroscopy webinar. My name is Isabelle Verbeck from the FAOS Global Soil Partnership Secretariat. Our today webinar will present the definition and role of soil spectroscopy for laboratory and field measurement, and we reflect on possible novel approaches. Before starting, I would like to remind you that the session is organized in a webinar format in which participants cannot activate their audio and camera. This meeting is recorded and recordings and presentation will be uploaded on the Glossel and web page. We invite you to post your question in the Q&A box, which will be moderated by the colleagues. In addition, there is also a chat box that is already very active. It can be available and can be used for interaction between participants. Please use the chat responsibly. For any technical issues, write to me on the chat and I'll be happy to help. I would like also to invite you to our newly created Facebook group, the Glossel and Soil Spectroscopy forum. Before exploring the future for soil spectral interference with our renowned speaker of today, I would like to give first the floor to my colleague Yipeng, who will provide you with a bit of background on the GSB, the Global Soil Partnership and Glossel. Yipeng, you have the floor. Thank you, Isabelle. Hello, everyone. My name is Yipeng. I'm from Global Soil Partnership FAO. Probably some of you already know me before, because this is already the third webinar. The next few minutes, I'm going to briefly introduce who are we and why we organize this webinar. All the webinar are organized under the framework of the Global Soil Partnership. Global Soil Partnership normally we call GSB. GSB is established in 2012 to position soils in a global agenda through collective actions. Our main objectives is to promote the sustainable soil management and improve soil governance to guarantee health and productive soils. All our activities are downscaled through seven regional soil partnerships. Since we are partnerships, so our activities are so supported by our partners and under the guidance of the intergovernmental technical panel on soils. As you can see, we are working with, we work with a wide range of topics. For more information, you are very welcome to visit our website. On the top of all these work subjects. We also have a different networks such as Global Soil Laboratory Network, Glossalon. Glossalon established in 2017 to build and strengthen the capacity of laboratory in soil analysis and to respond the need for harmonizing soil and ethical data. Glossalon started to work on the chemistry, wet chemistry analysis, focusing on training, harmonization of standard operating procedures, and the execution of inter-laboratory comparison. Last year, we launched the Glossalon initiative on soil spectroscopy, also launched the international network on fertilizer analysis. For more information of these two initiatives, you are very welcome to visit Glossalon website. You can also write email to me or to my colleague Lucrezia. Since last year we launched this Glossalon Soil Spectroscopy initiative, and the main objective of this initiative on soil spectroscopy is to develop national capacity in soil analysis. We decided to start organizing a series of a webinar due to we have a many different background of the colleagues. Some of them in a very advanced level, some of them don't know that don't have that much knowledge on a soil spectroscopy. In the first two webinars, if you attend, you knew the first two webinars are mainly introduced what is this technology and what we can do with this technology in a real application. And the third one today, we invite our guest speaker to give some future perspective of this technology. In this webinar, the fourth and the fifth will mainly introduce the experience of the soil spectral library from France and Brazil, because one of the most often asked the question from the countries how to build a soil spectral library and how to use a spectral library. The last speaker of the last webinar of this several of them webinar will be given by the eye open door and he will mainly talking a little bit about the measurement. But this is just the beginning of our webinar. And next year we are currently preparing other webinars for next year. So, so please regularly check our website. We will have a more information in a soil spectroscopy and the coming webinars. And the last two webinars presentation and the video recording already online so please also feel free to visit our website. And now I would, now I would like to introduce today our speaker of the today, Professor Alex McBrentley has, he has been a practicing soil scientist for more than 40 years. I think in my generation of the young soil scientist that almost all of us grew up with his knowledge, I have to say, now, if you, if you have haven't read his publication, I don't think you learned enough of the soil knowledge. He works on the pedometrics, digital soil mapping, digital agriculture and a soil security. He always tries to develop new ideas to better understand and manage soil. Alex and his colleague at the University of Sydney have been working on soil spectroscopy as a way of efficiently generating new soil data and information for more than 20 years. And I also would like to introduce another panelist today, Professor Boudiman Minasini, speaker from our last webinar. Also Dr. Alexandra Wodox, Dr. Jose Padarin, and Dr. Edward Young. All of them from University of Sydney and I have intensive experience in soil spectroscopy and spatial modeling. They will be here to help us today and answer some questions in the QA session. So please feel free to post your questions in the QA box. Without the feather ado, I would like to give the floor to Alex. So Alex, floor is yours please. Isabel and Yi, thank you very much. Good morning. Good afternoon, good evening. Good night depending on whether you're in the Americas, Caribbean, Europe, Asia or Oceania. So welcome to you all and I say a special welcome to all the colleagues whom I know and who've worked with over the years. Just waiting for my screen to pop up. I'm trying to share my screen but it's coming soon. Don't worry. Hopefully it will appear and we can get going. There we go. Can somebody give me a wave and tell me that's okay. Everybody can see that. Thank you very much. So this is the third wet webinar in this Global Soil Laboratory Network series and I thank the Global Soil Laboratory Network and Global Soil Partnership for inviting me to give this webinar today. Before I do that, I need to do an acknowledgement of country. We would like to acknowledge the Gadigal people of the Eora Nation, the traditional custodians of the land from which we are webcasting this presentation. We recognize their continuing connection to soil, land, waters and culture. We pay respects to the elders past, present and emerging. So I'm giving you this seminar from Sydney, Australia, where we've been in lockdown for the last 14 weeks, which is why my hair is a bit long and I can't go to the hairdresser to get a haircut. So I apologize for that. Hopefully in a few weeks we'll be able to go and get a haircut. So today I'm going to talk about a future for soil spectral inference. And we'll try and explain what some of these words mean. And I'll also acknowledge that myself and my colleagues are from the University of Sydney here in Sydney, New South Wales, Australia. And we have just acknowledged several contributors to this. So the work, not so much the work, but the idea is that we're talking about today, we've had lots of contributors. Mario and Ed, Brendan, Bootyman, Wartini, Hozi, Ruta and Alex, a couple of them have joined CSIRO, but we continue to work on this topic and we plan to work on this topic for a lot longer. So if we have a look at our research farm up in northern New South Wales, beautiful vertisols. This is a real picture, a real pedagliff. So the word soil layer is real. It was created by digital agriculture if you like, just to, we did that for World Soil Day a couple of years ago. And really what spectroscopy is about, how can we go out into that landscape and efficiently map out, gather information about the soil and its condition. That's really what the hope of digital agriculture, hopefully very efficiently, perhaps without even sending any samples back to our laboratory. The goals of today's webinar are partially didactic. In other words, I'm going to teach you a few things which you might know or you might not know. And it's partially prognostic. In other words, we're going to make some guesses about the future. And they probably are just guesses. Before I get into it too much. I've just got one recollection and I'm getting to that age where I start telling stories about the past. So about 20 years and two weeks ago, exactly 20 years and two weeks ago. So that's roughly September the 11 2001. So one of those dates where people remember where they were that day. I happened to be in the Midwest of the United States at one of the leading universities they are giving a seminar. And I'll recall very vividly that part of that seminar I talked about mid infrared reflectance spectroscopy. And that would be the future of routine laboratory analysis. So that was 20 years and two weeks ago. So what do we learn from that. Well, we'll learn a couple of things. One is, and you can be too, you can be too far ahead in predictions, and he doesn't. Another one is that you never really get your predictions right. But I think we saw from last week's seminar that mid infrared reflectance spectroscopy probably is the way to do much of routine laboratory analysis and I could stop the seminar there. And that would that would be enough I think. So a little bit of background in soil spectroscopy. I always like a little bit of history. So in the 60s, the work probably began in the 60s by the 80s we had the idea of digital spectra and not in soil science but in chemistry itself. So the beginning of what we might call chemometrics and all the techniques that you can use to manipulate spectra. And then, probably in the 90s and more like 2000s, we start looking at near infrared mid infrared spectra of soils and the first work on spectral libraries. So there in the 2010s 2020s, we're really interested in the potential of field spectroscopy and inference. What can you learn from spectra. So if we go back a little bit in time. No, sorry, let's just talk about the word spectrum. I only, I only put this up because I hear people talking about spectra a lot. But they don't seem to mention that the what the singular word is spectrum. You have, when you take, you take one spectrum or multiple spectra. So it's important that we understand that. And spectral is the adjective that comes from spectrum spectrum. So one spectrum, a thousand spectra, a spectral library. What is, what is a spectrum? Well, it's a response, and it could be the reflectance, it could be the absorbance, but it could be something else. Conductivity, permittivity as a function of some systematic proportion of a continuum. And, and that could be described by its wavelength or its frequency, particularly when we're talking about the electromagnetic spectrum. The spectrum is the spectrum sample that fix wavelengths or frequencies and then captured as a series of numbers. And it's a digital spectrum that we work on, and it's a digital spectrum that chemometrics works on, and that we can store in our databases and we can manipulate. So that's an important concept. I don't know if this is the first spectrum, this is the first spectrum that I know about. And this is a visible near infrared reflectance spectrum of a soil and Newtonian silk loam. And it's a paper from 1965 in soil science. The purpose of this paper was to understand the energy balance of soils, wasn't to try and do chemistry, such as you try and do physics. But what you learn from me is that there are distinct. This is a reflectance spectrum. There are distinct bands of absorption, which are largely due to particular moisture content on OH bonds, but also the systematic response to soil moisture. So, so that was the first this NIR reflectance spectrum that I know about, and that's that's actually the beginning of that work. We go around along to about 2006. This is paper here by my colleague, former PhD student Raphael Viscara Roselle and Dennis Walvort from the Netherlands and from Altera. Here we we looked at the visible the near infrared in the mid infrared spectrum. And what you see in there isn't is interesting. The visible doesn't have much many features in it. It tells you the soil color but it doesn't tell you a lot because it doesn't have many features. The near infrared has more features. More more bonds in there. But the mid infrared has a lot more features. It's it gives you the fundamental frequencies. And that's why, in one sense, the mid infrared is the one that's got the most information for us because it has more features any. But the whole spectrum, you can put the whole spectrum together from the visible to the mid infrared or even from the ultraviolet to the mid infrared. That whole spectrum has got lots of information which you can infer soil properties and soil soil composition from. And if we go back to last week from Boody's talk Booneyman's talk and we learned there that lots of things that we're interested in for describing soil condition and capacity and can be got at from the mid infrared spectrum. The ones in gold we can do very well. The ones in blue we can do a bit better ones in black are more difficult. And it's those things that are those properties that are really well described by the solid composition of soil are well estimated those that depend on the arrangement of those solid materials and solely the gas and liquid materials. Like the aggregate stability and so on are not so well predicted. And those things which are dependent on the soil solution are really not that well predicted so I think we can understand all of that. But there are other kinds of spectrum. The Raman spectra the gamma radiometric spectrum the portable or the X-ray fluorescence spectrum, they all look at different parts of the of the electromagnetic spectrum, the, as you can see, the X-ray fluorescence spectrum which tells you about particular elements, but lots of features they need, it's easy to calibrate, hardly needs calibration, a gamma radiometric spectrum more noisy, but can be calibrated for particular elements or soil properties. The Raman spectrum, once again, tells you about particular particular elements, particularly iron, I think. So there are lots and lots and these these three spectra are for the same for the same soil sample. So, so there are lots of kinds of spectrum and we can even look at the the electric spectrum of a soil. And we can look at the electric spectrum either through its conductivity or through its permittivity or capacitance. As you can see there are not so many features in this, but we know that systematic changes. If we look at GPR or TDR, these are two methods for measuring soil water, the changes with frequency in GPR and TDR can help us to be able to predict soil moisture distribution. So there are, there are, there is information in this part of the spectrum. And in fact, if we go back to our 2003 paper that introduced digital soil mapping, we had this diagram, which went through the whole electromagnetic spectrum and talked about all kinds of parts of the spectrum and how they could predict different soil properties and allow us to do different relationships. And really the future is about investigating in detail the whole spectrum here. And it's not just, it's not just the electromagnetic spectrum, we can even have an acoustic spectrum which is not part of the electromagnetic spectrum of a soil. So the Russian work from a year or two ago, this is the sound spectrum at different frequencies of a soil. So we can characterize soil, particularly soil structure I think, using the acoustic spectrum. The point is that there are many new kinds of soil spectra that remain to be explored. So that's one of the messages from today. There's much to be done with all kinds of soil spectra. And NIR and MIR are very good and are very useful. There are many others. I want to move on to this soil idea about soil spectral inference. So what are we talking about? They're not common words perhaps. So soil inference is the prediction of a property or properties from other soil property, from another soil property or properties. The way we've done that over the last 25, 30 years is through what we call pedotransfer functions. Things that are difficult to measure, we tend to use pedotransfer functions to do that. So we're doing that I call soil inference. And soil spectral inference then is the prediction of properties or other soil properties from using the spectrum as the input or from properties predicted from the spectrum. In one step it's direct, in two steps it's indirect. And a soil inference system is a software engine, but a software for the prediction of properties from other soil properties. And a spectral inference system is a software engine for the prediction of properties where the input is a spectrum which drives the prediction of other properties. So we can simplify the definition of soil inference system driven solely or mainly from soil spectra, which I think is the promise for the future. And back in 2006 actually, we wrote a little paper about the possibility of using spectra as the main driver for soil inference system but it's a combination of spectra and spectral transfer functions and pedotransfer functions that offers the future. So at this point, at this point I'm going to hand over to my colleagues Ed Jones and Hosey Fadarian. We're both research colleagues here at the University of Sydney, who are going to give you an explanation demonstration of an inference system and a soil spectral inference system. And I'm going to stop sharing my screen and hand over to Ed I think. Okay, thanks Alex. So, while my slides are loading, I'll just explain that today I'm going to give an overview for specific soil inference system that we call SINFAS or spec SINFAS when we incorporate spectral data. So today I'm going to show you what SINFAS looks like and how it works. And then following that Hosey will give a live demo of our latest SINFAS implementation. So, first, following on from Alex's definitions. So SINFAS is a software engine that facilitates systematic prediction of soil properties from other soil properties. And systematic in this sense means methodical in procedure or plan and that it's marked by thoroughness and regularity. And I'll demonstrate that soon, but just taking a step back as Alex says, pedo transfer functions, we can't understand SINFAS without understanding these pedo transfer functions or PTFs. For today's talk, a simple definition of a PTF is any function or equation that is able to predict the soil property from other soil properties. PTFs can be simple regression equations, such as the example of the wilting coefficient shown here at the bottom of the screen. And this is the earliest known example of a PTF. It describes the residual water remaining in the soil when a plant starts to wilt. And as we can see, it's a function of the source texture, and we can estimate this for all known texture combinations as shown on the right. So this is just a simple example to help you understand better. The new PTFs we develop are mainly more complex machine learning models or spectral models. But the main thing to understand is that PTFs are the building blocks of SINFAS. And SINFAS uses these PTFs to build relationships, as Alex said, to predict properties that are difficult to obtain using properties that are easier to obtain. And this is what it looks like when you run SINFAS. So on the left, we have our small input database. And this can be derived from lab measurements or estimates from spectral data. And then we have a pool of PTFs all jumbled up on there. And then on the right is our output with a large amount of output properties. This could be hundreds of columns wide. But importantly, it has quantified uncertainty. When we run SINFAS, it will first identify the input properties that we have. Then it will search through our PTF pool to determine which PTFs can be run using these initial properties. We then execute these PTFs, which gives a new set of properties with quantified uncertainty in this, the first generation of SINFAS. As this is a dynamic implementation, it will automatically update depending on the PTFs that we have available in the pool. And now after this first generation, we have a large number of available properties, which means new PTFs can be run in the second generation. And here, we can start to see we're forming a SINFAS network, which is a network of properties connected by PTFs. And when we start to make predictions on predictions on predictions, you'll start to understand why it's important to quantify the uncertainty of the prediction and propagate this correctly through the network. So we keep iterating through generations like this until there are no more PTFs that can be executed and our output is returned. You could also overlay additional analyses on top of this output as part of an expert system to make further inferences on the data. But that's the subject for another talk. So we can see that SINFAS is all about knowledge organization. It can be used for the description, indexing and classification of soil data. Oftentimes, the prediction component can seem trivial. So if we take this example of this simple PTF in the box, I think anyone could calculate this in Excel. But what happens when we add uncertainty to the equation? The calculation becomes non-trivial. And as demonstrated before, when we start linking PTFs in the SINFAS network and make predictions on predictions, the uncertainty propagation is extremely important as it will determine the quality of your final prediction. And the quality of your prediction will determine its usefulness for different purposes. And what are some of the uses of SINFAS? So the most obvious use case for SINFAS is that it provides a very low cost method to turn a small amount of soil data into a large amount of soil data. You can also backfill missing data points. And I'm sure we've already all experienced this when we receive a data set riddled with NAs. Don't you hate that? But you can use SINFAS to provide estimates of these missing values with quantified uncertainty and get all of the additional properties as a bonus. You can also estimate properties that are difficult, expensive or impossible to obtain. And by impossible, I mean if a soil sample has been lost or used up and there isn't enough remaining for a particular analysis. You can also estimate complex and obscure properties such as the real and imaginary components of the dielectric constants of free water or the integral energy of the soil moisture characteristic, which is an estimate of the total energy required to extract water from field capacity down to wilting points. So try asking a student or a commercial lab to do that for you. Now I don't think they'll be too impressed, but with SINFAS it's easy. As part of this work, we've been digging through literature to resurrect old and overlooked PTFs. Are all of these properties going to be useful? Maybe not. But with SINFAS, we're making it easy and accessible to put this information into people's hands so that they can quickly find out for themselves. And who knows what hidden gems there might be out there that might inspire further research streams. So I know this seminar series is all about spectral analysis and SINFAS is perfectly structured to be combined with spectral soil analysis in the form of a spectral soil inference system or spec SINFAS. And why is that? As Booty demonstrated last week, you can predict a wide range of properties using spectral data, which means that you can use this data to initiate SINFAS. This allows you to leverage the vast amount of information stored in both PTFs and soil spectra. Further, more importantly, as Booty demonstrated last also demonstrated last week, spectral data can be obtained in the field, meaning results from spec SINFAS can also be obtained in the field. And you can use these results to make decisions in the field. With a little bit of development, this can occur quickly, cheaply and efficiently. So we've actually been working on this behind the scenes for quite some years and initial implementation was constructed in a large Excel file, which was functional. But as you can imagine, the structure of an Excel file is quite rigid and it's difficult to maintain as new PTFs become available. It's also difficult to incorporate machine learning models. Grant Trantner and Jason Morris, two PhD students here at Sydney, produced dynamic implementations. And I have produced a spec SINFAS system in R as part of my PhD research. But they never really left our computers to be released onto the world. Our latest implementation provides APIs for connectivity with Python and R, or can also be run directly on a web interface to make it more accessible to everyone. And now I'll hand over to Jose to expand on our current implementation and give a live demonstration. Jose. Thanks, Ed. Hope you can hear me. Let's see. So I'm going to give you a short overview of our current implementation of SINFAS and some examples, interactive examples. This is not working for this one. Can you see my screen? So this is like, I already mentioned that a little bit of how SINFAS works. So we have some input data and we match all that information with our pool of properties and transfer functions to make all the predictions and uncertainty propagation. Of course this input data could be soil properties or could be directly soil spectrum. So I'm going to show you a little bit some of the requirements that we have for different soil properties. So starting first with the soil properties we have. Here we have a little example that you can see some of the information that we require for each of the soil properties so you can see things like description units. So we have some of the information with their specific names on sometimes physical constraints. Basically to organize all the information that we need to run SINFAS. Probably the most important piece of information that we have here. It's each property is linked to a laboratory code. In this particular example we have a code from the soil chemical methods from Australia. We also have some other laboratory methods there in case two different laboratory codes pointing to the same soil property. Of course this is very important in the context of a global laboratory network because of course in many countries we have different methods and we would like to use the correct transfer functions to predict using the correct properties obviously. The next thing that I want to show you is like the structure of a better transfer function. As you can see here on the right screenshot, we have a lot of information as well, including statistics that we can obtain when we're developing the transfer functions. We have the target properties and the predictors that we can use for this specific better transfer function and as you can see here they are linked directly to a specific laboratory code to make sure that we use the right property. But for the prediction we have different requirements. In the case of traditional better transfer functions we have things like an expression, arithmetic expression with the corresponding codes. So we have in the case of more complicated models like a spectral models we have machine learning methods or deep learning models that will depend that will be directly linked to certain instrument model. The second thing that we have like mentioned before is that they have different, we have support for different uncertainty propagation methods that will depend and the better transfer function that we're using. So here I'm going to show you a small demo of different clients that we can connect to SINFERS. So SINFERS is like the computing engine. We've created an API to to connect the engine with many different types of clients, including our Python, or like a web page interface. So this is an example now of of a Python example. Just to make sure you can you can you see the screen and the go. Hello, anyone. Yes, we can. Oh, sorry. Yeah, I thought the connection. Okay, thanks. So I'm just going to load some this is a Python interface. And just to have like a really relatively small example here we're just going to have a simple table with two soil properties in this case sand and clay and just two samples. So of course we can inspect all the different, all the different information that we haven't seen first. And this is just like a quick view of the API. This is just made to like for developers to inspect it but I can show you here that we have like a long list of better transfer functions. So if we get any of these better transfer function obviously we can get the details including the predictors and targets. So we're going to use this specific better transfer function that has a unique identifier within our database to make certain predictions. So we have the data with those two soil properties clay and sand. And if we run this using the specific better transfer function. In this case we're predicting field capacity. We can also see the predictions plus the uncertainty. We can also do the same with this is another better transfer function to predict the permanent wilting point. You can also see the predictions and the uncertainty. And we've built a same first. And easily propagate this uncertainty. So let's say we want to calculate the, the available water content of us all that will be subtracting from the field capacity the permanent wilting point. And if we do that we get a prediction. And most importantly we have like an automatic propagation of uncertainty. So here I'm just showing you like how to use single better transfer functions but I really want to use this in the, like the whole inference system. So using just to remind you had just two columns sand and clay, those are the initial properties in this case, and we can query the database that we have to see what kind of predictions we can make with same first. And you can see here that starting from from the top we have two properties and clay, and we can predict in the first generation that it already explained how this works, we can predict a series of nine different salt properties. And then in the following generations we can use like those predictions and the initial data to make more predictions for instance here we have available water content that depends on field capacity and the permanent wilting point. Of course, here we just have the plot of this but we can just upload all this data to the same for engine and get some response. So if we actually use the initial data that we have and we send a request to same first we get back all the information that was shown here in the previous graph. So we have the two initial properties and then we have a series of salt properties that are being predicted in different generations with their estimates of uncertainty. Of course this is starting just from two salt properties that probably we can obtain in the laboratory but like here we're talking about of course the spectral inference and why using a spectrum could be an ideal case in this kind of salt inference system is because we can predict like many salt properties from a single spectrum. In this case we have linked a model like a machine learning model specifically a deep learning a convolution neural network to predict multiple properties simultaneously. So starting for from a single instrument they read the reading of this instrument I respect. We can obtain six different properties including total nitrogen clay content pH and so on with with different levels of accuracy. And from those property then we can use that as a like in last the first generation of sin first and predict multiple salt properties. So you may see like there's a difference with the previous graph that I show you. We have some properties that are in a different color here in orange, and that means that sin for has multiple better transfer functions to predict that specific property. So that's something very important that we wanted we wanted to include in in sin first because that allows us to mix different models and potentially improve the predictions of let's say NIR or MIR predictions. And talk about a series of classifications like play class ABCD depending on how accurate where the predictions, and we can potentially use this system to take some of the properties that are poorly predicted to a higher category. So of course this is just a graph and I can show you how it works. If we submit this data to to sin first. This is just reading some data from a collection of spectra that I have. This is just two samples and you can see here they have the in the column names we have the name of the instrument. And if we plot this, you can see that the two spectra. We can answer this information to sin first, like in the same way that we did before just with two properties. And we get in the background, this is processing all the machine learning models that we have the back in this case is this is NIR data. And from all that's just a single spectrum that we have we can get 23 different soil properties. And of course with their estimates of uncertainty. We have a very small example of how we will use sin first from in this case by Python. We have similar things for our if you, if you like programming. But of course we're trying to do this and like to be easily accessible to more people, and we're also planning to create like some sort of website like this one, where you can select like a CSV file with the data in this case just for soil properties, and by by clicking just here infer, we'll send that information and we get back all the prediction, but you can see, well these are just a few samples so we get a response quite quickly. And if we open this, you can see that we have like the initial force sold properties, and we get all the predictions that sin first return. So ideally here we want to have a really good engine of course that does all this prediction uncertainty propagation and an API for other people that is addressing communicating software between the engine and their specific software. And this could be obviously applicable to things like collecting data in the field so you could have like, like an Android or an iPhone application that gets information from a field of spectrometer and you can connect to the engine of sin first and get predictions in real time. So that's the small demonstration of what we've implemented so far. And now the floor is for Alex again that's going to talk a little bit about the future of all this. Thanks, Ed and Jose. Can everybody hear me again. Yes. Yes. Good. Thank you. Thanks Ed and Jose. I think that gives a little bit of flavor about where we're headed by by bringing together the idea of spectra and predictions from spectra and perotransfer functions bringing those all together into a spectral inference system so that's where we're that's probably the cutting edge of where we're at at the moment, and we're working hard to to make that a reality. So thanks Ed and Jose for that little glimpse of the future. I don't seem to move on the next slide yet, but anyway hopefully it'll, it'll let me do that eventually. It's stuck. My slide seem to be stuck. What can I do. Alex, maybe you can try to stop sharing and sharing again. Okay, I'll try that. I'll try that and I'll share again. Yes, thank you very much. See, you know it's much more about this than I do. That's fantastic. And just, and Jose's point at the end was, we can get, we can get various levels of prediction from from the spectra, but we might be able to improve some of the predictions by bringing the perotransfer functions. So it's a combination of the two together that will might give us a lot more properties and better predictions and obviously as somebody asked in the in the Q&A, we do need to test the quality of these predictions by actually actually making real measurements to test whether it's working. It is stuck. It has stuck again. Why is it sticking? Hmm. Don't know why it's stuck again. Maybe you just don't use full screen for now. Don't do full screen right. Okay, I'll try again. First time we've had any problems on the and try one more time. I think it's stuck in full screen. I think that's part of the problem. Let me see. Let's see if we can. Okay. No, it keeps sticking. I don't know why. If you want, I can share the front here. Why did that happen? Yeah, maybe you can share the slides, the final slides. Of course I've changed them since I sent the ones to you. But anyway, that's life. Can I? Okay. All right. Can you see my screen now? Okay, can we can we go back one then please? This one? Yeah, so I'm going to talk about the future. I'm sorry that happened. I don't understand that. And all I was going to say about the future is that I don't know anything more about the future than you guys do. Everybody out there. So this is just, this is just a few things that I think, but you probably think just as well as I do about the future. Let's go to the next slide. So we a couple of years ago gave a talk at the soil science society of America. I had a PhD lecture and I talked about these digital soil science and beyond. And basically that paper has just been published. So we look at the diagram on the left. We talked about that we probably think the future will have digitally enabled regenerative soil science. Generating soil digitally digitally enabled extraterrestrial soil science. We'll be doing soil science on Mars or other planets. And we might also have digital soils. In other words, soils embedded with digital devices and digital capability inside the soil itself. So that's what we see as a future of digital soil science. But as part of that, if we look on the bottom left here, we have things like Internet of Things sensors machine learning robotic measurements. We think that spectroscopy is the best tool to drive the generation of information through Internet of Things sensors and robotic measurements in order to drive this new digitally enabled soil science. So spectroscopy is very important for the future of soil science and digital soil science and digital soil management in fact. Next slide. Thank you. So let's just speculate a little bit about routine lab spectroscopy and we are that's Eddie. Next slide. So we have laboratory spectroscopy. And I guess to do laboratory spectroscopy. The idea is to do it on whole soil samples, not to have to separate extract and so on. And we have a whole set of techniques at the moment that we can use for whole soil samples like the UV, this NIR, MIR, XRF laser induced breakdown spectroscopy laser induced fluorescence spectroscopy, Raman, etc. So there's a couple of issues that for all of these that we have to be concerned about sample preparation and calibration and these are the issues that people have been working on for NIR and NIR for the last few years. Next slide please. In terms of sample preparation, it does get back to the mundane questions of, do you have to dry the sample or how much do you have to dry the sample? Does that have an impact on the spectrum? And grinding the size of the sample, the size of the material, does it matter how big the material is that's presented to the spectrometer? And we know from experience that MIR is much more sensitive than sample size or the size of the material. So we tend to grind it much finer than we do for NIR. And so that question remains for some of the others. I think, so these are two issues that need to be systematically experimented for all of these different new kinds of spectroscopy because we think that these other methods will also add information in generating soil. And of course, the advantage of spectroscopy is, once you prepare the sample, it's non-destructive except for limbs, limbs is obviously got a bit of destructiveness to it. You can put it in other spectrometers and get further information. Next slide please. Sorry, Jose, can you please mute yourself? I guess you are busy replying questions. Okay, good, thanks. And then the course is the whole issue about calibration. And that will be discussed more in the next few webinars coming up. But really the questions really come down to how local do you want the calibrations to be or how global do you want them to be? And there's always some balance between local. You would like them to be local because they predict better. You might want them to be global because she can use them anywhere. You might not be quite as good, but you can use them anywhere. But then, how many observations do you need to make local or global calibrations? And the answer is probably not 10 or 100, more like a thousand or 10,000. And that's one of the issues that we have to do more work on calibration. And we have to develop calibrations for these new kinds of spectroscopy. Next slide. And then the bigger promise for the future is, can we do routine field spectroscopy? So if we go to the next slide. Now, why would you do routine field spectroscopy as opposed to doing it in the lab? And the answer is, and doing it in the lab because you've got more control conditions might give you a more precise result. But there is a principle here that often analytical chemists don't appreciate what field soil scientists do appreciate. When you have real and significant natural soil variation, in other words, the soil itself varies across the landscape, across a field. If you have low cost, low precision soil data generated, say from a spectrometer, and we can do that at lots of places that can produce much higher value specialized soil information. In other words, I could, if I can take for the same cost 50 observations with a spectrometer and a spectral transfer function, or two observations by taking that back to the lab and measuring for the same cost, I can make a much better map with the field observations than I can with the lab observation. And that's the beauty and the promise of particularly field spectroscopy. And I say spectroscopy is one way probably the best way of generating such data. The next slide please. And so there are so many possibilities with field spectroscopy when we start thinking about it. At the moment we have things like gamma radiometrics, ground penetrating radar, NIR, XRF, MIR to some degree, it's more sensitive in the field, things like lives and so on. These can all be used in the field that we believe in the future. Soil, the soil moisture effect, the issues that we have to deal with here, soil moisture effect, how to present the sample, the calibration and so on. Next slide please. We're going to share the recording, don't worry. The recording is going to mainly be Alex saying, next slide please but in between those there'll be a few bits of information, I hope. So the soil moisture effect is soil moisture is likely to offer the most interference in most of these devices, but it depends. So with gamma radiometrics soil moisture affects the observations, affects the spectrum much less than it does, for example, NIR. So there are different effects. And also for XRF it affects it less also. So remove the moisture effect using various algorithmic techniques. So things like direct standardization or empirical parameter orthogonalization. So there are techniques for doing that. The fancy words don't matter, but there are techniques to remove the soil moisture effect. Next slide please. And then of course the sample presentation is a question. Even though we're in the field, can we actually put the spectrometer in front of the soil either in situ, exactly where it is or do we move it around a little bit to bring it closer to the spectrometer. So there's a couple of questions there. Do we need to homogenize the material in any way to make it easier to take the spectrum, or can we do it on a completely undisturbed bit of soil. How are we going to look at it? Are we going to do it laterally? In other words, look at the spectrum, pull the spectroscope laterally through the soil or perhaps push it vertically into the soil. So there's a lot of astronomers like the lateral pedologists like the vertical, but both are a possibility and need to be looked at in terms of how the mode of operation. Next slide please. Calibration, similar to the lab, but a bit more challenging because there's much more variability in the field. The conditions can't be quite as standardized. So the question, when it comes to calibration of these field devices, is global calibration feasible? Can we actually ever get enough information to get a global calibration? And is it actually useful? So there are some questions about calibration of field spectroscope, I think. Next slide please. Mobile mobilization. In what kind of mode should we use these devices? Obviously the easiest is a vertical probe that you push down into the soil, but we could amount that probe on an autonomous vehicle. It can drive around and do this thing automatically. We could put it on a very low flying aircraft. And we could do it as an endochoric soil probe. Endochoric, which is a word I made up yesterday, means inside the volume of the soil itself. In other words, it operates inside the soil. So next slide. So here's an example of a low flying aircraft flying about 30 centimetres above the soil surface with an EM induction device. This thing only has a few channels, but there's no reason why this couldn't, we couldn't have a similar device that produced full spectra in the electron, in the radio frequency part of the spectrum flying over the soil, taking spectra. This one is very useful for mapping soil moisture. Next slide, please. Can you see because the previous slide should be this one. I think there was a slight, you know, this little thing, it's just an imaginary thing of a spectrometer that crawls through the soil by itself and takes spectra. But it might also be diverted to actually make some changes to the structure of the soil or such a device. So that's kind of way into the future kind of idea. But that's what I called an endochoric device. Now, if we get on to the interface between chemometrics and pedometrics between the soil statistics and the statistics of the spectroscopy. Next slide. The whole question about spectral data sharing and calibration, which you're going to talk about in the future. Next slide. I think we'll go to the next one. So a couple of years ago, a lot last year, Hosey particularly wrote a paper about what's a good model for sharing data. And in this case, in the case that we're talking about sharing soil spectra or sharing soil samples, which soil spectra can be derived and we do understand. I think everybody understands that the sharing of spectra the sharing of data is offers difficulties to institutions these days. And we need to think about methods that allow us to share data and learn from them, but not have that data in the control necessarily a bit of a centralized body or a centralized institution. So I think we're moving away from that idea. So we move on to the next slide. And I found this paper published recently in nature, which talks about the same problem but in the context of doing diagnosis using DNA sequencing, but the same thing applies, and where people are working in all over the world. So here we could change the little DNA type icons there for spectra. So I think they're about spectra and a little squares with spikes on there. You can think about that as the spectral models, which are the diagnosis so you get spectral models and you got spectra. You can think about it in that way, we can think about different modes of learning, local learning that's all of us sitting in our labs making little pedo transfer little spectral libraries for our own, our own lab and making spectral transfer functions or calibration functions individually. They might work locally, but they won't work in other places, because all of us have a few data, they might not have the power of prediction. The complete opposite of that is central learning, central learning where where all the, where all the, all the spectra are put into a central location, which is probably the idea that the gloss alarm has for making spectral libraries at the moment. Not necessarily, it's a method of crescent but not necessarily the method of the future. Everybody has to hand over their spectra and their samples. Next slide. So that's the current way of making spectral libraries. Next slide. And then there are other ways of federated learning where everybody sends in their spectra sends in their, their spectral models and you make universal models from the different spectral models. In that case, you don't never hand over your data to anybody, but you make a federated spectral model. And I don't think that's been tried at all. And I think that's a good way for us to investigate in gloss alarm. To me, that would be actually a very good thing to do. And then the last one is swarm learning, where you still locally do the functions, but you do some some of the data and spectral functions do in an anonymous way get transferred between each other. And I think that that's the future for generating spectral transfer functions or calibration functions. So that's the future we're making spectral library so I think there's a big future in this kind of data sharing approach without the centralized institution, which is a kind of old fashioned model I think for doing this kind of work. Next slide please. So the other thing I'm very passionate about is using spectra for soil classification and identification. Next slide. I think the biggest challenge for the future in which soil spectroscopy can play a crucial role. It's a big challenge and spectroscopy can help a lot is to have a formal digital global quantitative system of soil classification. We don't have it. The WRB soil taxonomy is not that it's not, it's not, it's not global it's not digital and it's not quantitative. And we really need this if we're going to develop soil, soil understanding and soil management. Next slide. And of course so we've been working on some example and my update slides had a reference here but there are 23 soil variables that we can ask that we can use to do a global system. If not all of them are derivable in the field or lab from MIR or NIR soil spectra. So we can really get at most of the information that we need for soil classification and from the spectra. Next slide please. And this is an example of putting together taxa from different systems here soil taxonomy WRB Australian soil classification New Zealand. But this could be taxa from all of your systems and all of your national systems could be in the same system. We can see how they relate to the taxa of other countries and so on. And we could do all of this, put it all together and using spectroscopy as a way of generating the data. So I think that's very much the future. So I hope you all take up the mantle of a global digital soil classification system using spectroscopy. Next slide, please. There are a few dangers in spectroscopy and they're worth mentioning on the way through. So if we go to the next slide. And obviously, everybody's having a look at machine learning and deep learning, we're not really using artificial intelligence at the moment, but we're certainly using machine learning and deep learning. Next slide. And really, the question I have is really about the interpretation of machine learns soil spectral prediction models. Can we interpret the models. Can we interpret these models? Can we interpret what they're telling us? And if we can't interpret the models, what do we understand about the relationship between the spectra and the soil properties? So that's not just a philosophical question. It's a scientific question and I think we need to think deeply about that question. Next slide. The next issue is about proprietary soil prediction. So we see on this slide thing, words like agtech, soil tech, intake, there's lots of startups or small companies out there offering new ways of doing things. And one of the new ways of doing things is to more cheaply predict soil properties, which is fine. But if we're going to use those services rather than have our own laboratories or make our own measurements, we do need to understand how they're doing it and what the uncertainties are. So this is what I call proprietary soil prediction made by companies. Proprietary means that the way they do it is the intellectual property of the company. So you don't actually know how they do it, whereas if you make a pH measurement in the lab, everybody knows how you do it. So next slide. Well, do we need to know what's inside the black box? This is what might be inside the black box. Do we need to know what it is? Does it make sense? So if we're going to go down the road of proprietary soil prediction, I think as soil scientists, we do need to have some understanding of what's inside the black box. And we certainly need to be able to know how well these things are actually measuring or predicting soil properties. So we have to be very careful about proprietary soil prediction. It may be that it's not proprietary that the whole thing is open, and that's fine. No problem. Next slide. And then the last one I talked about is doing too much with too little. In other words, we read lots of papers where people have collected 50 samples, collected spectra, hundreds of wave bands, and then fit a model with 50 observations and hundreds of wave bands. And then you fit a model that's got hundreds of parameters to predict pH or something else. And I think that is, or it might be for predicting the microplastic content in soils. We'd have to be worried about overfitting or overstretching the data. Because spectral predictions calibrated with too few observations are not predictive at all. They might fit very well, but when you come to bring new material, you get the wrong answer. So it's a kind of false dawn, you know, you think you've done well, but when you go to predict it doesn't work. And spectral predictions using calibration models with too many parameters are also not predictive. So why when we do predictions, at least in our lab, we always have, we put observations to the side to see how well they predict. We don't believe just in bootstrapping and cross validations a way of telling you how well the model actually works. So I'm sounding all very negative now. I'm not negative about soil spectroscopy. I think it really is the future for digital soil science. And let me go to the next slide. I'm going to do some conclusions which hopefully are a bit more positive. There are many kinds of spectroscopy that are potentially useful for the rapid generation of soil information on whole soil and make, but at the moment many of those are largely uninvestigated. So those of you who are looking for a project for the future. There are many projects for the future in the talk that I've given you today that you can investigate. There are many kinds of spectroscopy that you can investigate and write many nice papers about. Next slide. At the moment, mid infrared, near infrared portable x-ray fluorescence and gamma radio metrics are the most deployable kinds of spectroscopy for the whole soil. And they're both useful in the lab and in the field. So these are the methods for the moment. So if you want to get started doing something and you want to use it to actually generate real data, then I would be looking at one of these methods. I should say, if you do better if you put them all together. If you don't favor one over the other, I would prefer to use all of them and get bits and pieces of information from all of them by putting it all together. Next slide please. I believe that soil spectral inference systems will be the principle mode of operation. So putting together the calibration equations, the spectral transfer functions for the pedo transfer functions will be the way we do this. It will be applicable to both laboratory and field based systems. So you'll put in the soil, you'll collect one or two different kinds of spectra, and it will really tell you about 200 soil properties with their uncertainty. And if they're very uncertain, then you might go and say, well, we're going to measure that. But that's how we feed the soil information systems of the future. Next slide please. In the laboratory calibration will be developed for many hundreds of soil properties using federated and swarm learning. I see the federated and swarm learning, the way to generate calibrated in the future, which means that you will be, you will be collaborating with labs across the world, but you won't necessarily be sending all your data there. You'll be doing it in a more anonymous way using the, the idea of the blockchain. In the field spectrometers of various kinds will be deployed on autonomous platforms to update a wide array of dynamic soil properties important for monitoring soil condition for health to help us better manage soil. Next slide. Principal applications will be for real time, agronomic and environmental decision making, but also including soil classification and diagnosis and soil monitoring. That's the end of the conclusions. Go to the next slide please. Just a little advertisement since I'm here. If you'd like to say something new, I would like to hear about it. So we have a new journal called soil security. I'm interested in publishing some of your work around soil security, soil health, soil conditions, soil capability, soil connectivity, the capital of soil, soil awareness, and so on. This is a journal for you. All the sort of things that global soil partners works on we publish in soil security. It's multidisciplinary, not just science. It's also got economics and sociology in there. But we are currently developing a special issue on soil spectroscopy for estimating soil condition and capability. And Alex Waddu is looking for paper. So if you'd like to write to him, please do that. Next slide. I mean, we're doing advertisements now. We've got a training course coming soon. Which is based on this book, Soil Spectral Inference with R, that we published earlier in the year. Alex Waddu is creating the training course. And it'll be hosted. We're creating it here at Sydney, and it'll be hosted by a gloss along coming soon, hopefully this year, if not early in the next year. We have had the problem that we've been locked down and haven't been able to get to some of our equipment and devices and laboratories in rooms so that slowed us down a little bit. Next slide. Thank you. Just reminding everybody all you soil scientists out there, the peak body for the world's soil scientist is the International Union of Soil Sciences. I'm very passionate about the International Union of Soil Sciences. And just to remind everybody, next slide. But next year will be the World Congress of Soil Science in Glasgow. And I hope you've got some nice papers about soil spectroscopy that you want to give there. I'm particularly passionate about that because some of you might be able to tell from my accent that I'm actually originally from Scotland myself. So, after clock 26 and climate change, the IUSS Congress in Glasgow. So I'll see you there. Next slide. And that's it. Thank you very much. Thank you. I really enjoyed it. I'm sorry the slides wouldn't feed through but hopefully you still got the gist of everything I said. Thank you very much, Alex. I think it's a great presentation. I enjoyed a lot. I have to say every time I always enjoyed and get inspired every time when I listen your talk, especially your, your vision. It's really inspired our young generation. We have quite a couple of questions. I would like you to answer live but before I answer the questions from the participants, I think that I would like to take some advantage of for myself as a moderator, ask one question from myself from my perspective. Like what I said just now I get inspired a lot. Actually I have a lot of questions but since I guess I should only ask one question so I would like to ask a very generic question. Just now you mentioned one thing about the user spectral information for the soil classification purposes which is very interesting for me. And as you know, we are most of us trained as a podologist that we often went to the field, spend hours for the classifying the soil profile and which is kind of fun and those discussions actually are main input for the soil classification work. And now, but that's also reason problem because a different country has a different system for the classification system for example, Australian has your system US has US system FAO has a kind of the international system. That's become to a major barrier of the harmonizing the soil types. So spectroscopy spectral information involving is definitely one of the solution for the future if we can. We can have a harmonized the classification system. That's definitely the way to to to go ahead. And the thing is, one thing is my question is in the future if we use a lot of spectral information for this purpose, which means we give a lot of job to the computer to do it. And that time I would like to listen your opinion in the future if that they really happens that because so far we use a lot of expertise from our soil scientist for the soil classification, but in the future when the day comes a lot will give to the computer, the role of the soil scientist and these expertise, what kind of role will be in this domain because I think it's quite important and nowadays many young generation playing with the computer. Very good with the programming but they go to the fear they can classify the profiles so I would like to listen your opinion about that. Yeah, it's a it's an interesting question and and but at the moment I think we have a failure. We have a failure of system over system. Because basically we are using, we're using classifications, the principles and the way we do classification is from the 1950s, and we haven't changed that. What I'm saying is we need to change that we need we need to make it quantitative. Now. And of course, when you go to the field to do classification, even with something like soil taxonomy or WRB. Many times you say, Oh, I can't do this class, I can't do this classification, because I need to send this sample to the laboratory to tell me how much Claire is here or because the criteria use some some measurement. And so I think in the future, all of that measurement stuff should be able to be done in the field. But the other thing is, and we, we only look at some properties and we don't use other properties. And I think when we can use spectroscopy, we can use a lot more properties. And we can put them together. And we've already demonstrated that you can bring different systems together by understanding the tax are bringing them together into a single system. And it can apply to the tax are from every from every classification system. So I think that's the great thing for the future. You go to the field. You, you, you stick the spectroscope, you get the data. But you, and then it will automatically tell you which kind of soil this is, but, but you must be able to to read the data and say, Oh, this is not telling me the right thing. You're giving me the right answer. You need to be able to understand that. But I think it also then allows you to think a little bit less about what is the soil but more about how did the soil get here. Have we going to manage the soil, we can change the focus of what we do. Thank you. We don't get we don't stop at the first step where we have 10 people having an argument about what we call this soil. Thank you. I've seen it many times. Anyway, I know every, I know not everyone will agree with me about what I said, but it's kind of heading in there in that direction. And I think it's big. And do, do that, and it will take us for a good new day. A good new day. Probably we can, we can even write a paper together for this. Sure. Okay, I think there are a couple of questions from from a participant. And I think you can see the question from the QA session. I'll read it out because the recording later will not show the text or later on the other colleagues can know the question. The first question is for just a curiosity. Will soil spectroscopy technology be as simple as using smartphone, meaning at the rural farmers level in future, so that we take a picture or scan and get the information directly. For example fertilizer recommendation those the lime in quantity, etc. It's a good question and I guess it's the it's the dream of the agronomist to be able to do that. I think in terms of technology. Yes. But we do still have to solve the problem about the fact that how do we get information about the soil solution. We haven't solved that problem yet, you know, we can, we can tell something about the total amount of soil of elements and so on, but we can't really tell about its activity. How active it is. And that's what's important for fertilizer recommendation so we still, we still will need some tricks. We will invent some new technology to allow us to do that, but I believe we will invent that technology. And yes, I believe it will be coming from a smart, a smartphone type device. Yes. And it's quite easy to, to turn a smartphone into at least a visible spectrometer now. If you take this, this device and you, and you, and you make some solution and shine a light through it, you can measure, you can measure the absorbance of a solution. You can use a smartphone now by just putting a cradle on the back of this smartphone. So you could do that already. So it's not so, it's not so far in the future, I think. And, and probably spectra spectrometers will be built in. They won't just be visible. They'll be also be in the near infrared as well. Thank you. I'll buy the way actually this question was asked by the, I think this question was asked by the brother of the cabindo. Okay, good. The second question is, could we use the sin for for prediction of level of a biodiversity. If no, could you tell about the investigation in this area relationship between biodiversity level and a spectral. Yeah, I can't answer this one very well. I wish Mario's here because we have done some, we have done some work on that. So there is some relationship between spectra and biodiversity that which we have observed. We have observed some relationships. But I don't know, I would put them in the in the range of reasonably to poorly predicted not very well predicted biodiversity. But I guess what the relationship is to do with the organic material, the different kinds of organic material that you see in soil and its relationship to bacteria fungi and so on. So there is some relationship but not not so far a very strong one and then what we have we have looked, we have tried to look at that because we're we're interested in in, in what drives soil biodiversity, what, what controls soil biodiversity. It's something that you can control easily by management, for example. And at the moment, the measurement of soil biodiversity requires a fair bit of work in terms of DNA extraction and then gene sequencing so there's a lot, a lot of algorithmic work in interpreting the gene sequences. Okay. Thank you. One of the question is also from previous during your talk, one colleague asked what is your opinion on using current available remote sensing data or method for assessing soil properties because I think this is a question from many colleagues and Right. Okay, then so this is, this is, this is one of the easier questions. And you guess what I'm going to say. So, so lots of remote sensing methods, particularly those that work in the visible near infrared and visible near infrared and remember, when he talks remote sensing the words that they use. The words that they use for the different parts of the spectrum are different. So when they talk about the thermal infrared that corresponds, that's different from what we talk about the mid infrared and so on so the language of remote sensing and the language of spectroscopy are different so you need to remember that. So, and most of this, most of the methods that look at, let's say the visible on the infrared or basically the measure the top millimeter of the soil, because they're the reflectance measurements. So if you have a bare field like the one behind his head there, and you can say something about the soil properties on the surface, but what does it tell you about the soil at 5 centimeters or 10 centimeters, or 50 centimeters. What's the model for telling you that you're not measuring that. So that's the first part. The model only tells you about the soil surface. Now there are some other techniques. And in the radio frequencies that penetrate into the soil. And, but they, at the moment they mainly tell you about the moisture content. That's why we get soil moisture, but they, they only penetrate a few centimeters into the soil. So at the moment remote sensing is good for making inferences about the soil surface, but not about deeper in the soil. That's remote sensing at the moment. And that's the major, that's a major barrier because the reflectance can reach the surface information, even has some problem with the vegetation cover this noise. And that's not even talking about the vegetation. Correct. So, you know, I don't, I think the advantage of spectroscopy is you're going to look directly at the soil at different depths. That's what you're trying to do. Another question is, we can get a cheap and quick prediction by spectral technology. How is the cost efficiency compared to the conventional analysis in the lab. Another question is, can we use spectral for soil classification I think in the second one we already discussed the first one probably something generic because the cost is the most one of the most concerned for many lives. If you're only interested in soil pH then I would buy pH meter and measure it and measure it. But if you're interested in a range of soil properties, then spectroscopy will give you a whole range of soil properties on the same sample. The second thing is it's non destructive. So I think that, you know, and I think that's the advantage when you, when you want a whole range of soil properties. And I think the, the other thing I didn't talk about is, imagine you, you store all the soil. And some new, new thing, new property becomes important. So a new contaminant becomes important. And you want to know, it's the soil of the world contaminated with this material. Or was it contaminated recently or in the past. You take the soil from the soil lab, the soil store, you take the spectrum and you make calibration, then you can map the whole world soil for contaminant contaminants. So it's really about the fact that it gives you much more flexibility about what properties you might measure. And at the moment, you might want to look at that for plastics. Because once you've got a calibration for the spectrum, and you can look at the old spectrum and say, what's the plastic there? It's the plastic in this spectrum, for example. You don't need to really go back and do all the samples again, you just need to look at the spectrum. That's the kind of advantage. Yeah, I think, yeah, just now I think that information is quite important because we kind of, we can map the past baseline map, which means there's opportunity. Yeah, because if we have an archive from 100 years ago, we can just take it out, and then we get the information from 100 years ago and then we have a comparison what is the status now, what was 100 years ago. But you can also have the spectra. And then you can, even with spectra that you have now, you can say, well, I can estimate new properties as the, as the calibrations become available. As the calibrations become available, you can redo, relook at the spectrum, generate more properties. How much glyphosate is in my soil? How much glyphosate residues in my soil, for example. That's a very important question. Everybody should measure that. It's called AMPA, look for AMPA. One more question is, say, how can the results obtained from SINFA be validated and can the Peter transfer function be extrapolated from the local to the wider region when applying this system. Yeah, well, the only way to tell whether the thing is working is to go and make some measurements. So I always believe that you should go and take some samples and see whether whether it's predicting properly. Go and take some new samples and predict. So remember, just to talk to everybody, I'm a soil scientist, I'm not a statistician. I actually believe in that we should go out in the field and collect soil samples and make observations to the testing. What was the second part of the question, Yi, please? The, how the Peter transfer function be extrapolated from a local to a wider region. So what we have in the system is a kind of region of confidence of the Peter transfer function. So we know something about the space in which it's calibrating. And if you go outside the space, the calibration space, then the Peter transfer function either doesn't give you a prediction or gives you a huge uncertainty. So we already try to recognize that you're trying to use it in a situation where it's not calibrated. That's part of the inference system. At least that's what should be in there. I hope it's in there, Jose. It is. It is. Don't worry. That's the main thing. Okay, I think this will be the last question. Our webinar today was quite successful almost two hours. People are very interested. The last question is, do you think numeric soil classification is the future and the first spectroscopy. We can infer soil properties based on spectral and then classify soils in direct way. We can also make a classification directly from spectral, which way is better in your opinion. Okay, good question. And we've tried to make classifications directly from the spectra. I think, I think if you do that, and this goes back to what he said earlier, then you deny all the pedagogical knowledge that we have generated for 100 years. So we know, we know which soil properties are important. So I think it's better to use the spectra to generate the soil properties that are important and discriminate different kinds of soil. We need to use those properties to give us the classes. And so I don't necessarily believe that you should go directly from the spectra, make spectra, make classes from spectra without soil properties. I think soil properties is a way of interpreting the soil, the data. So, and that's why a lot of the work has to be about calibration about pedotransfer functions about inference systems. That's why I talk about inference systems, because it's really about getting at the soil properties that we understand the soil properties and how they operate. And the processes is their soil knowledge. We shouldn't throw away the soil knowledge that we have already gained over 100 years. Thank you, Alex. I think I fully agree with you. And also, that's one of my major concern because I'm now I'm in this position and many of the country and the many of the partners asking us for the, for the training, different training, digital soil mapping, spectra modeling, and they always asking the what kind of algorithm machine learning and you can make the good prediction and really mentioned about using our existing soil knowledge to explain the result, explain the map. Otherwise, the map is just a beautiful picture and it can be very dangerous for the policymakers actually. Well, when we do any of this work, how do I judge, how do I judge whether it's useful? Well, I can only judge whether it's useful based on my knowledge of soil. That's how I can judge. So we do need that soil knowledge to be able to to know whether we're doing something useful, I think. Thank you very much, Alex. It's been a really, it's been really enjoyable and I inspired a lot from your talk. And Michael, I think you are the participants, thanks very much for your for your join our webinar today. And we will send the email to inform you the, we will upload the presentation and the recordings to our website and we will send you the email to inform you shortly. The week is now posting the link to the other three webinars scheduled on soil spectroscopy, and then you are all invited for the, the rest of the sessions. And remember to check this page regularly as well because there will be more webinars coming, not only for the spectroscopy also for the what chemistry and equipment to purging, etc. In the coming months. So we are really in the end of this webinar. Thank you again for the speaker and the panelists and all the participants. I wish you a pleasant end of the day or the evening. Thank you very much. Thank you everyone. Thanks for listening and thanks for the questions. See you next time.