OpenTEAM In-Depth - Remote Sensing Methods for Measuring Soil Organic Carbon at the Field Scale

OpenTEAM In-Depth - Remote Sensing Methods for Measuring Soil Organic Carbon at the Field Scale

Show Video

all right well we'll get started uh welcome everyone to an open team in-depth webinar uh we have a good team from the regen network here today um and just to open it up uh this is really uh meant to help build the knowledge base of the open team community in a way that fosters coherence and collaboration so we can better serve our growing and diverse membership um so through this today we'll illustrate and document the dynamic nature of our community inspire and share new concepts and technologies and i think a few weeks ago there was demand for kind of a an overview of remote sensing remote sensing 101 which will also be kind of presented today uh so we have three presenters from regen network sam bennetts giselle boomin and sophia liker all from the science team uh so i'll turn it over to the three of you we'll have uh up to uh 30 to 45 minutes of a presentation and the rest for questions and answers so we'll have some time for interaction as well i see your screen perfect you should be ready to go thanks again for for being here and presenting today great wonderful um yeah so as laura said thank you all so much for joining um my name's sophia i'm from the region network science team we're really excited to really dive into some of the great work we've been doing over the past year uh specifically looking at kind of the benefits of remote sensing why we're using remote sensing and then taking a deep dive into our remote sensing methods for measuring soil organic carbon so to start us off um i'll do a quick overview on what we're going to cover today as laura mentioned i know there's a wide range of people who are joining us and who may have more background on remote sensing some of you have may never heard of remote sensing so first i'm going to do a little bit of a background on what is sensing why are we using it some of the benefits of remote sensing and then i'll be passing out passing it off to sam who will be presenting our actual grassland methodology and then a case study where we actually use this methodology to quantify soil organic carbon in a farm in australia over in new south wales and like laura mentioned we'll open the floor at the end for questions and these may be about remote sensing if you have questions for us specifically on you know different satellites or remote sensing background or if you have questions on our methodology itself so we should have plenty of time for questions at the end so make sure you save any questions you have for us and we can kind of have a more open discussion so to dive right in on what is remote sensing i figured i would start it off with just a quote from the usgs and they define remote sensing as the process of detecting and monitoring the physical characteristics of an area by measuring its reflected and admitted radiation at a distance typically from satellite or aircraft and special cameras collect remotely sensed images which help researchers sense things about the earth and this may all sound really complicated um but ultimately the goal with remote sensing is looking at the physical characteristics of earth from afar and you can collect remote sensing data from a wide variety of tools so you can use gather remote sensing data from drones you can gather remote sensing data from airplanes or helicopters but really today what we will be focusing on and our kind of approach to remote sensing is using data collected by low earth orbit satellites such as sentinel 2 which sam will touch on a little bit later so there's a lot of i won't get into all the details on the advantages and disadvantages to each of these but just know that there's a lot of remote sensing beyond just the satellites we'll be talking about today so kind of to dive right into some of the technical background around remote sensing specifically when we talk about remote sensing using satellite imagery is there's two main metrics that we'll often talk about and that sam will talk about when we get into the actual meat of the methodology and the first is spatial resolution and when we're talking about spatial resolution really what we're talking about is the size of a pixel and you can kind of think about this as different levels of photographs so you can have a image from a high resolution camera you know really expensive like multi-thousand dollar canon camera for example and you know that those pictures that you get from a really high resolution camera are going to be very detailed um when you zoom in you'll be able to see a lot of the detail within the image versus maybe a lower resolution camera like your iphone or something which might have a little bit more of a course of a resolution and when you zoom in it might get a little blurry so the same thing applies when we're talking about spatial resolution of satellite imagery so satellite imagery can range in spatial resolution we often talking about it in the units of meters but it can range from like 250 meters all the way down to one meter or even sub one meter and you can see down here in this little bottom infographic how that reflects the output the pixel output display so you can see that as you decrease the pixel size you're actually able to more accurately reflect what's happening on the landscape so at the one meter scale you really can see in detail what's going on in the landscape but when you move up you know more to the 30 meters um it's a little bit more course of resolution so again we'll talk about when when talking about remote sensing you'll often hear the term spatial resolution and really what we're talking about there is just the size of the pixel so that's the first metric that's often used and the second metric is temporal resolution and essentially what temporal resolution is is just the time between images so again thinking about as we are gathering images from satellites which circle the earth it's going to take a certain amount of time for the satellite to take a snapshot of a certain location such as say san francisco and then circle the earth and revisit to that same place where it can take a snapshot of the same place and so again this is just when we talk about tempura resolution it's about this revisit time between when you're gathering images and the revisit time can range from say five days for a certain location um on the globe to you know up to 10 or 20 days so again when we're talking about satellite imagery really the two metrics that are used are spatial resolution and then the second is temporal resolution and then moving over to the right again this is a really basic breakdown of remote sensing but i think it's important to touch on the um fact of what we're talking about when we talk about spectral bands so when we talk about the different satellites we use we'll often talk about the number of spectral bands they collect and essentially spectral bands just align with different ranges within the electromagnet magnetic spectrum so you can see the on the top is just you know basic breakdown of the electromagnetic spectrum and as we all know there's a certain range which encompasses uh visible light and within that visible light you can break um the wavelengths into different ranges which are collected by a satellite so satellites often collect bands in the red and the blue and the green visible light spectrum but the cool thing about these satellites is they're also collecting images or snapshots outside of the visible light spectrum so you're not just seeing you know the normal picture coming in with the red green and blue band but you're also getting information from bands outside of the visible light spectrum such as from near infrared or shortwave infrared so when you get all this information from these satellites really what you get is a stack of different pictures encompassing different bands so you have a picture of the red band you'll have a picture of the shortwave infrared or the blue band so when we talk about spectral bands really we're just talking about these different snapshots which align with different ranges within the electromagnetic spectrum so that's just the basic breakdown of the kind of technical background and details when it comes to remote sensing so moving forward in terms of what can remote sensing tell us so not only do we have these bands like i mentioned in the last slide you can actually combine these bands to create and calculate indices and these can tell us about a range of different information on the actual ecosystem health so i'm just going to explain one which is probably the most famous one which is ndvi the normalized difference vegetation index and this is just a metric or a measurement of vegetation health so on the left you can see that healthy vegetation when you're looking at different reflectance of the bands it's reflecting about 50 of the near infrared band but reflecting only about eight percent of the red band but then on the right when you're looking at more of a stressed piece of vegetation or an unhealthy plant it's reflecting about 40 of that near infrared band and 30 of the red band and essentially what you do is you can combine all this information into this ndvi equation and get a final output metric which will move you over to the right side of the screen where you can see that all right if you get an ndvi value of say 0.66 to 1 you usually can assume that the vegetation whether it be trees or crops are fairly healthy but if you get a metric that's around zero to like .33 you usually can consider it somewhat of an unhealthy plant and i think indices i bring up indices because not only we do use indices that sam will touch on within our methodology for tracking soil organic carbon but we also use it for some ecosystem health analyses so this is just ndvi i listed a couple other indicators you can detect with various remote sensing indices below such as looking at fire impacted areas or the water content held in vegetation you can also use indices to look at bare soil presence within an ecosystem or even look at the health of overall forest so again indices are just another way you can use remote sensing data to assess what's happening on the surface of the earth and then from here i wanted to touch a little bit on some examples of what can you use remote sensing for why is it used what are some of the benefits of it so first you can use remote sensing to map land cover across an area and on the right you can see just a landsat image and landsat's a nasa satellite but it's a image of a river and you can use remote sensing and classification algorithms to actually determine a classification of what's happening so using that image that you gather from the satellite you can then use algorithms to determine all right these pixels align to pixels with water versus you know the brown is indicating barren soil so you can use classification algorithms to map land cover across an area as one one of the ways you can use remote sensing but you can also use satellite remote sensing to monitor say the rate of wildfire spread or map the structure of forest to understand overall forest health i think in terms of open team and a lot of people on this call may be interested of you know how do you use satellite remote sensing for assessing the health of crops within your fields or even monitoring the change of surface water so again there's a whole slew of other examples on how you can use remote sensing but these are just a few to to kind of understand the different ways you can use remote sensing data and overall analysis and then moving on uh some of the benefits of remote sensing is you can remote sensing provides tools to monitor landscapes at a variety of scales again going back to the whole idea of the resolution or the pixel size you can use really high resolution data to monitor say an area of a couple acres or you can use satellites that have higher or more coarse resolution to even map like land cover across an entire state for example so again remote sensing can be used at a wide variety of scales and also one of the main benefits is there's a lot of free and open source databases that have remote sensing data so um landsat on the on the right that's a nasa derived remote sensing data set but also esa or the european space agency has a lot of data sets in terms of what we use but also even the usda has remote sensing databases that are high resolution for the united states so again all these open source databases have resolution to that are high enough quality that are fairly good to be able to assess um the health at the overall field skill and then finally before we really dig into our methods i wanted to touch on one of the the largest benefits at least i believe in terms of why we use remote sensing and that is using remote sensing for time series assessments so um i really love this infographic on the right because i think it really encompasses the strength of why we can use remote sensing to track change over time so you can see that it looks like they're pulling out a certain portion of a image from say washington and you have a certain image from a certain date and then from there because these satellites have been circling the earth and taking these snapshots over a period of time you actually can compile a whole series of these snapshots or create like a stack of what we call images and within these images you can even go down to a pixel level um or even a certain portion if you want to look at a certain parcel within this image and then track the change over time by looking at the spectral value aligned with that specific pixel or for that specific field scale from there you can take that and then look at the change of the spectral index over time to kind of track you know what was happening in the landscape itself so you can see that this certain image they're saying they're tracking from 1984 to 2016 but you can see it around like 2002 or 2005 something happened a change happened and you can see that by looking at the actual pixel level at the pixel level or the field level that there is a drop in the spectral index and so depending on what ecosystem type you're looking at or what was happening there you could say hey this could have been a drop due to if it's ecosystem health say a wildfire or did deforestation take place all these things you can do to it you can assess using remote sensing so for us when we use um time series assessment for change over time what we use these for is to look at the change in percent of soil organic carbon at the pixel level and also look at the change of ecosystem health over time so again i hope that provided a little bit more background on remote sensing for those who may not have known as much about the benefits but now i'll pass it on over to sam and he'll dig into our actual methodology for tracking soil organic carbon over time sweet um well yeah now that sofia has given us an overview of remote sensing and how it can be used uh giselle sophie and i wanted to share some of the work that the science team has been doing so for the last year giselle sophie and i have been working on a methodology to measure greenhouse gas reduction in co-benefits and greasing systems more specifically our methodology is a remote sensing approach to measure carbon sequestration and identify or quantify solar organic carbon stocks at the field scale and we're doing this for grasslands under the practice of prescribed raising uh so our methodology relies on multi-spectral satellite imagery with a spatial resolution of 20 meters or higher and then we couple that with machine learning algorithms to kind of estimate soil organic carbon stocks across an area and look at those change over time like sophia just showed this is an open source approach and the idea with it is to minimize the sampling efforts and costs so you can go to the next slide sophia um so this is an overview of our grasslands methodology and it can really be broken down into three main stages so the first stage is soil sampling and stratification the second stage is data preprocessing where we incorporate remote sensing data and then the third stage is modeling and soil organic carbon quantification where we use that ground truth data and satellite imagery to estimate stocks across the farm and then ultimately look at change in time so the first thing we do when we adopt a new project is we evaluate the project area to assign soil sampling locations this is a pretty important stage in the process in more traditional methods you might use exhaustive sampling where one might take hundreds of samples across the farm so that's going to represent the spatial variability of the soil because you're collecting so many soil samples but with remote sensing it's a little bit different we're requesting fewer soil samples so we want to make sure that those soil samples are pretty representative of the spatial variability of the soil so during this process we're going to perform what's called stratification where we essentially divide the project area into different parcels or strata that will share similar soil characteristics here we might use variables such as elevation and topography vegetation or above ground biomass hydrologic differences or patalogic variables such as soil type and percent silt percent clay and so taking all those variables will kind of make this classification and you can see here in the image on the left how this project area has been divided into all these parcels and so what we're assuming is that the soil type within those parcels is relatively similar so it might share you know similar topography or it might you know have a lot of vegetation in that area or a little bit or the soil type is generally the same so once we divide that we'll randomly us we'll randomly pick uh some of these strata and then assign soil sampling locations within that so then the project manager or the land manager will go out and collect these soil samples at those assigned locations and then send those to a lab and so at the lab they're going to be testing for soil organic carbon and bulk density which is used to measure soil again and carbon stocks um and then for the scope of our methodology we're also assessing other soil health indicators such as ph you know nitrogen um sulfur you know a bunch of other variables so once that's done we'll go into this data pre-processing stage so when they collect the soil samples we also have them collect the soil sample location and so the soil sample location is what ties the remote sensing data to the ground truth data so the image in the middle shows uh i think it's just a satellite image and you have the soil sample location overlaid on top and essentially what we're doing is we're taking our satellite image and our other remote sensing data whether it's like a percent clay raster or maybe an elevation like dem uh digital elevation model and we're saying at that soil sample location we're going to extract all that remote sensing data and then we're going to turn it into this nice data set where we have our percentile organic carbon and all of our predictor variables so if you were to look at this in a table you might have rows with your different soil sample locations and then a column was percent soil organic carbon and then each variable listed in a new column so now that that data set is set up we can go into the modeling and this is we're going to be finding correlations between the ground truth data and all these predictor variables to try to find correlations that we can use to present or predict percent soil organic carbon there are a variety of ways to do this originally we were using linear and power regression models so in this case we would take one variable in this example band six and then we would compare that to percent soil organic carbon to find a correlation and if there was a correlation we could use that to estimate uh soy organic carbon at unsampled locations more recently we've been using machine learning models such as random forest and what's really nice about these is you can incorporate an entire data set into these predictions so rather than looking at just a single variable now that you can incorporate all these variables and identify correlations that might not be seen visually and this is really powerful because incorporating like topographic data in a bunch of different bands has actually helped increase our accurate accuracy and you know create better predictions so sophia you can do the next slide sweet so this is kind of an example of how we go through this on the left we have our soil sample locations and then you know this first arrow we've finally found our correlation with our machine learning algorithm and then we're going to use that to predict soil organic carbon at unsampled locations so the image in the middle shows this prediction so we've masked out everything that's not grassland so all the trees roads waterways um have been masked out and now this is just grassland and it shows the spatial variability of soil given carbon across the area the darker red is higher percent soil organic carbon and the more yellow is a lower amount and what's really cool about this is you can see that for every 10 meters like every pixel we have a different prediction which is really cool because you know you're accounting for that that spatial variability over a short area then we use the bulk density and the soil depth to estimate soil organic carbon stocks and the image on the right shows like our final stock calculation so you can see through all these parcels how the amount of uh stocks change and how much carbon has been sequestered the next slide and then finally what we're going to do is we're going to compare um you know we're going to perform this time series analysis where we're comparing two years so we have 2019 on the left and then 2017 on the right and you can see some of these parcels have gotten darker and that's representing that more carbon has been sequestered in that time period which is really cool and then to calculate our final creditable carbon change we're also going to take into account the animal emissions and the model uncertainty so we have had pretty good results um with our predictions so far but there is a certain amount of uncertainty associated with estimating percentage elegant carbon in these machine learning algorithms and calculating bulk density so we want to make sure that we take all that into account in our final um kind of carbon stock quantification uh just a few more notes i think you can go to the next one sophia um so the image type that we've been or the satellite image that we've been using most frequently is sentinel 2.

uh we use this mostly because it has a high spatial resolution so it has that it says 10 to 60 meters um but it's really 10 to 20 meter spatial resolution so we can have uh predictions every you know 10 to 20 meters it also has a pretty good revisit time um every five days but we usually get um an image at least once a month um you can go to the next slide and this is just a kind of a slide to show how spatial resolution um how important it is so on the bottom you have landsat and that has a spatial resolution of 30 meters and you can see that it's more granular and so you're getting pretty good predictions and you can see the spatial variability but the image on top sentinel with uh 10 meter resolution is just a little bit more granular so you can can have slightly better predictions finally i guess another application that we've been using remote sensing for within our methodology is assessing co-benefits so one of the co-benefits that we look at is kind of the ecosystem resilience um in vigor and what we're doing is we're comparing the project area to the landscape surrounding it to see if the regenerative practices go beyond carbon sequestration the image on the left shows a land cover classification i did so the blue represents water the dark green is forest or kind of vegetate like heavily vegetated areas the light green is grassland and then the red and brown are man-made objects and roads and what we did is we chopped out the project boundary and we're looking at the ndvi or vegetation health within the project boundary and then comparing it to the surrounding landscape we're also doing that for bare soil index so we're just seeing the amount of bears you know bare soil exposed and seeing if over time you know beyond carbon sequestration we're seeing you know more solid soil and yeah so with that i guess we will open it up to questions and i would invite questions pertaining to remote sensing what it is kind of the basics but then also to our methodology um so yeah i think we can stop sharing and people can start asking away well first i just want to thank sophia and sam for for this i think there's a lot of folks who are really looking for some of that uh that basic remote sensing information um i thought maybe i'll kick off while folks are formulating questions uh around uh essentially you're versioning process because obviously these are evolving and particularly uh from the open team community you know where a lot of uh this will be you know we have uh hubs that are on the ground doing a lot of direct sampling um and so what are the kind of data training data sets that would be most useful in terms of uh improving confidence in in prediction in general and this is sort of just both sort of in general from a road sensing but also sort of more specifically in terms of your your own process uh really just like lots of soil samples the success of a lot of these machine learning algorithms or really any regression technique is just you know the amount of data inputs that you have the more data that you have the more training or the better train the model can be and the better predictions you can have so just having these larger training sets um is really nice so you know if a farm were to provide a bunch of soil samples or i think the long-term idea would be to regionally collect enough soil samples across you know a state or something like that that you could have a well-trained model that requires less input because you've you know had so so much training input so yeah that answers your question yeah i mean i and i guess the second part is sort of your your versioning of this and sort of how you're uh incorporating new data sets or uh specific projects that your target or regions that you're targeting to to sort of create those uh training data sets yes so so far we've only run this on just a few projects in australia but the idea is that we want to open this methodology up for anyone to use in any area so i noticed a question in there if any of this was automated which a lot of it is and our code can be found on github so um i mean whether you share that data with us or if you want to you know use our methodology to kind of train a regional model or a field model specific to your area um those would both be great options and yeah in terms of versioning i think just over time collecting more data and just having that in the archive to better train these models that we're going to use great i see a question from alden alden do you want to uh comment here sure um great work and and and don't don't take me wrong here but of course in nori we're trying to reconcile some of the big differences we're seeing between um some of the estimates that come out of these new ways of doing things and the old-fashioned um reference site sampling and process modeling way which we know has has large deficiencies so we're trying to find a path not be on one side versus the other but when i i i've been for other reasons because i had to give a presentation to australians yesterday looking at the uh the um zero soils map for australia and the soil map you just showed suggests 2017 background soil organic carbon stocks that are much much much larger than suggested by the australian government soils maps so i'm curious what you found are you saying that the the australian soils maps are just completely way off based and because if they could be that much off space then so could the american soils maps or what what what are the other possible explanations for the really really big differences between your map and their their map that that i'm not guessing yeah i can answer that um so first i'm not sure where you got the 250 tons per hectare in the project i will check as i don't have that number in my mind right now but what we saw is that we were like getting stocks within the normal ranges of of the stocks that are supposed to be there for uh managed grazing what we did find is a a very high sequestration rate in those areas which is related to this the particular management that they do there which is a and b raising it implies you know high frequency of movements a lot of animals um for very controlled periods of time and they pause the time that they they revisit those parcels and those management practices tend to increase the solar organic carbon sequestration by 10 times so when you compare a regional map to to this particular parcel you might get like very high differences um i guess those maps are more original regional scale um that what we are creating which is within the field scale uh again i i didn't ask the question with the intent of putting you on the defense please understand that i i'll i'll follow up uh gisele and um i'll just send you the link to the uh uh it's the zero um csiro australian government website which is the map that that i i'm referencing there okay yes for sure yeah just i wanted to clarify that we checked the the stocks and they were like within the normal ranges so that's why maybe we are talking about different units or comparing different different stuff great well we had a number of questions in the chat uh i'm thinking maybe uh dan kane uh uh if we had a response from gregory uh to that i wonder if uh sophie and sam if you wanted to pick up dan's question or if you'd like to reframe it for them yeah dan could you frame that a little bit more because i actually couldn't access that paper oh shoot okay i'll see if i can find a better link um yeah so this in the quick carbon program we ended up kind of doing some similar stuff too and realized that as we gathered um some more field data and sort of expanded our ability to do um sort of ground truthing of remote sensing only models and sort of not dissimilar to what you're doing combining both digital soil maps and data from from satellites um uh the sort of the accuracy of those models um began to look worse as we started to get more and more data to test it against um and this is sort of like a critique i've seen elsewhere in remote sensing and ecological modeling literature that um if you're not kind of considerate about how you you choose and place some of those ground truth samples um you may get an inflated sense of accuracy so not to say that um that you that remote sensing for estimation of a property like slow carbon isn't still useful but that kind of the how you generate that test data set matters a lot um to interpreting those rmse values that you produce so um and yeah so i've been trying to take a deep dive myself on some of this literature um and figure out okay how might this apply to the world of soil carbon so it was just eager to get to get your thoughts in that regard um and hear how you're thinking about it um i think that you might take that one because you have changed it tested a lot of models and we just did a lot of stuff um trying to avoid that any bias there yeah i don't know if i necessarily have like a clear definitive answer to that i think that we're still you know the methodology is pretty nascent in the early stages of development um and so even using machine learning models was relatively new um i have played around with like a lot of just you know how the the test train split happened and how that might have affected the model um and i did notice that sometimes you know depending on how you split the data it would have way higher accuracy than it would otherwise so i played around a bit with that but i haven't actually dug into any of the literature that um you're referring to so it would honestly be great to just like get some of that from you and have us dig in because i know that you're doing very parallel work and yeah i'd love to like you you have a better idea on that than we do at this point so i'd love i'd love to work together on it and maybe that's a good thing too with an open team to start thinking about because it's just sort of something i've seen in parallel literature from other fields like forestry and stuff like that so yeah which is really interesting too because i mean this is being applied to so many other things beyond you know soil carbon measurement whether it's you know forestry projects and there's tons of monitoring services out there that are you know claiming really high accuracies on these like really massive data sets um which is could be totally valid but it is like just a very new field so yeah yeah and i think that the chief concern i've heard from folks on on this is mostly how how accurate that is and how much you uh believe that that accuracy number um has a lot to say about the utility of these methods for change detection um and and that which is is kind of a key a key point but not just like certainly really useful at looking at patterns in space um but uh we really need to have um confidence in that and that accuracy number to then translate it to the ability to detect change over time so but yeah let's let's connect yeah great well and and a great prompt in terms of uh one of the roles with our uh active working groups around uh discussing uh uh our links with hubs and creating training data sets for remote sensing so and along those lines i want to go to rob trice and then kristoff perhaps but rob you were asking particularly about some of the management uh you know sensitivity to manage management on the landscape and that was certainly also part of this training data set and what our hubs are interested is can we detect some of these that the different types of management practices and how those start to show up in uh change over time but rob i'll let you uh frame it yeah exactly just kind of if we start to think about a b testing in different pastures how can we look at applying different practices and the co-benefits of different practices applied to the same pastors and do you have the ability to look at something like composting or if i put a particular seed into a pasture uh and how i have so it's fundamentally for me it's the ecosystem health it's looking at the animals it's looking at the practices and then fourthly are the business concerns those are the four big data sets that we need but we need to look at the the the correlation of all of those particularly the first three figure out to optimize technical technical implementation of climate smart practices so uh are you looking at do you have the ability to look at things or capture that information like compost was applied in pasture 42 but not pasteur 41 and that variability um yeah i think that i mean that's not that's not a project that we've necessarily worked on yet but i don't think that that would fall out of the limits of remote sensing you know as a field i think that the value of remote sensing is that you can use these data sets to look at spatial variability and you know you very well might be able to look at you know different indicators and say oh the you know health of this you know parcel is like way higher than this parcel and then you know that you composted one versus the other like that could be an indicator of how your regenerative practice is being applied and how it's you know changing the health of the landscape i i won't speak on whether or not you know that that correlation actually exists because we haven't done any work that work but um for things like that i think that you could definitely use remote sensing data things like animal welfare and other co-benefits that you talked about we aren't necessarily using remote sensing uh those might be like in situ whether it's you know like having someone go out and do like a field survey or something like that so great thanks so kristoff i mean i guess i continually want to sing your praises dorne for creating an open team because this is the perfect venue to have these sorts of conversations and great job science team at regen network humiliated including the open source community things collectively about like calibrating remote data uh building on rob's question like linking it back is that our sort of in-service of the outcome that i think everyone on this call wants to see achieved how you see what you learned from your pilot kind of growing and maybe that's both a question for gregory and the science team as well you were sort of breaking up a little kristoff so i'm not sure i totally understood the fullness of your what your question was to give a hot take awesome what you proved in your first pilot there's a big need to sort of calibrate input data that can leverage remote sensing technologies that will serve the entire industry we're all about how do you see this pilot kind of informing and growing in the spirit of the open team network uh and helping us all sort of go further together faster yeah awesome so i mean i'll i'll try to answer that at a high level because i because i think that's sort of the intention of it you just sort of like how does this all fit together and then maybe see if the science team wants to fill out fill anything in that i miss or or even just like share sort of different perspectives um i think you know to to sort of bring up what dan is talking about in others the opportunity here is to create sort of um i think sort of a system that gets more and more accurate and more and more precise over time because we're all um training it together so that we're able to provide ground truth data and correlate that with this sort of global coverage of remote sensing and use that as a very low-cost anchor for estimations and claims about ecological health outcomes right so i think what we've done is sort of prove that this is sort of there that there's lots of still questions but it's it's good enough for some people in the market to put confidence in the the approach which doesn't mean that it's perfect but it serves as sort of this first foundation that i think we can extrapolate and i think we sort of see a future in which the hub the regional hubs are doing a significant amount of sampling and research and making it possible to extend much lower coverage for high accuracy estimations in the regions that have similar climate and soil type so that we can actually create so that so that that can serve as the foundation for whatever sort of marketplace registry or certification process is taking place it may not even end up being a you know like a carbon offset it could be like a certified outcome in a supply chain right but but having this sort of um sort of public infrastructure for um ecological claims and estimations i think is what we're really excited to see happening and it really is only possible within a sort of an environment like open team where we could for instance be creating with dan and other people you know the sort of progressive versioning of how we incorporate more and more of the information about how to um estimate uncertainty and attach that to different claims in a completely transparent way so that the marketplace can see that while at the same time you know different hubs are able to provide data and train and then create coverage for farmers and farms in their regions again with similar soil types and climate so that sort of over time we essentially create a very um rigorous accurate and precise estimation tool to underpin our both management and and claim systems so hopefully that's kind of what you're going for kristoff um yeah i mean it kind of begs a few more questions like i hear machine learning algorithms and the hairs on the back of my neck go up because i'm like i don't really know what that means or how that's continuously improving and so is it us are the machines governing how we're dictating how these protocols evolve um yeah so a machine learning machine learning is basically about machine learning still needs us to so like train you're essentially training you're training something and you're saying you know uh you're you're you're you're teaching a correlation you're saying you know this is uh like the work the small amount of work i've done in it you know is like land classification so if i've spent a bunch of time in a particular place like in ecuador doing agroforestry classification and i have existing maps and other data sets i could go and say oh that farm that's bananas and that farm that is you know that that's actually primary forest and that is yucca and i know all of those things or like i could walk around my neighborhood here and i could say you know that's a garden and that's a forest right and those get correlated with what you can see that what what the satellite can see and so the same thing is happening where we're saying you know the the soil the we went and measured the soil health in this particular place and it has this particular multi-spectral frequency and we're like registering that and the more and more that you register those things the more it's the more that you can start to calibrate understanding it's the same kind of pattern recognition that humans go through like i go and look at a house plant and i see that it's wilting and i know through my experience that because it's wilting it needs water right it's the same thing or it's it's getting yellow and i see it with my eyes it's getting yellow and i'm like oh that means something is taking place so i mean it's not machine learning isn't i don't think anything scary as far as like we're sort of like vesting any sort of decision making or control over a process to machines it's simply over time creating a like a graph of correlations that gets more and more accurate so that we can sort of like make sense or sense make about what's happening where i i'd like to add something like very brief also machine learning algorithms the main value of those is that they are loud they allow for um creation creating like very um difficult relationships that are non-linear between all the variables and you can add a lot of variables so the model could be as complex as you want and they would detect the patterns if the patterns exist um otherwise a human cannot deal with such amount of inter correlations and non-linear relationships i would say um and with time we can we can rely maybe on some machine learning uh algorithms uh for prediction without needing something but i would say you need like at least 10 years of climate and data all correlated with high accuracy in order to create like a very robust model so we are far from that we need the calibration well and and i think there's i wanted to sort of highlight that giselle in terms of the open team community and the hubs in terms of helping to provide some of that training data set right all both management as well as the soil samples and i i guess i would prompt you with the time we have uh left uh to maybe maybe talk a little bit about again uh how your projects fit in as we're building this sort of common this sort of ag data commons for calibration how this project may fit in with some of the other initiatives that are working towards some of these same objectives again sort of i feel like this is an opportunity for some of that pre-competitive sort of work that we talk a lot about is those calibration sets in terms of land cover and use yes agree i think one of the big barriers for all this technology not being you know properly been spread i would say is the lack of geo-located samples and information from you know samples but we need like very high precision in the geo location so i would say if we can at least advocate for um that being like uh you know a request from now on that every farm that is extracting data from their fields is also collecting the geolocation that coupled with sharing data would allow to not only provide foreign permission for assessing the soil for the different projects but it will it could also feed all this training and all this knowledge uh that needs to be developed in order to to be using this technology more broadly um i've been als always saying that it would be great if there was like a big uh global experiment with a lot of observatories the the you know the ideal scenario would be to have observatories in different ecoregions with different climates and have you know for the different management practices uh farms that could work as observatories and so we could train images with the same soil types the same characteristics and management uh and provide for that calibration for accessing this organic carbon or other parameters in the farms that are within that region that would be great um so um yeah like i think that like the soil samples are really important you know the hub farms can be really helpful in providing that data but then also like the science and in you know kind of development from these methodologies like i mean working with dan and others interested in this field would be great to help you know that kind of like academic work but then also just like the software behind it like regen network is all about and open team is all about open source collaboration and i'm starting to upload all the the code that has been used to automate these processes which includes the you know machine learning algorithms we're using and it would be great if people are interested to get more collaboration on that front um i know like dan kaine's you know done things with regression krieging which is a really valuable tool and that's something that we would love to dig into but also you know the bandwidth is a little bit limited so um yeah like i would invite people just to like check out our github and and start forking it and working on that front as well so and playing around with it and reaching out and asking questions thanks thanks sam and giselle we also have a somebody on the on an iphone i don't i can't tell who that is but somebody calling in who's got their hand up oh that's me hi my name is jeanette um i'm from sibo technologies um and great conversation i just wanted to speak up because my entire phd background is in digital soil mapping and so there's an entire robust field called digital soil mapping that has spent the last 20 years pretty much intensely devoted towards using you know machine learning artificial intelligence and geostatistics like regression creating to predict soil carbon at scale and so there's an incredible wealth of literature and it's still a really robust and active field and so i highly recommend kind of digging into that stuff and then from the remote sensing perspective um you know digital elevation models are one of the best covariates you can ever use to predict soil carbon so just focusing on satellite imagery will never get you there but if you start to incorporate different derivatives of elevation models you can dramatically kind of increase your accuracy largely because those elevation models relate directly to where water is moving so great great work it's exciting to see all the progress great and jeannette's so pleased you're able to join us today um uh so i want to at this point i would like to hand it back to laura to talk a little bit about sort of what's coming up next and uh some of the activities uh within open team to help continue this conversation before before you do that can i just share one sort of parting thought on our end so i think our sort of larger theory at region and i think this is shared probably broadly in open team but i don't want it to go unspoken is that we should be um [Music] as transparent as possible in sharing exactly you know how we make you know how we're making what conclusion about where and then be rigorous about versioning so that we can just transparently just like we're doing right now say okay this was our first attempt this is our first version it has the following shortcomings that we know of and it has the following short comings that the community has identified therefore the next version that we're doing in this open source iterative way will include the following steps right and then we can just keep moving forward in this sort of iterative way and build tools that really are not sort of paralyzed by trying to get it perfect first but instead can progressively evolve so that we all have access to the kind of sort of information tools that we need to to make this all come true and to know when we're actually making statements about real ecological change which is what it's all about because without that sort of foundation of uh sort of what intersubjective verification together it's going to be really hard to know who's saying what you know how accurate who's saying something about where is it just it's very hard when it's siloed and we're sort of defending reputation instead of sort of like building uh a knowledge commons together so really grateful for this conversation and excited to be just like iterating with everybody through this and um yeah grateful for our awesome science team for for putting all the work in and really quick going along with that i just shared in the the chat a feedback form um because we want to open our methodology and just our presentation for any comments that you guys have um about the methods or you know the presentation or whatever else so um i would invite you to you know if you have any feedback on specific sections or just overall your thoughts about this that would be great as we could work it into our versioning so yeah awesome thank you so much gregory and sam and sophia and giselle and everybody here and uh and gregory for framing that is well articulated in terms of uh you know why we're here together uh and uh how open teams you know functioning as a community so really appreciate that um yes and and i think to tag off of what gregory said too about building that knowledge commons and i think the need for ground truthing uh some next steps that are coming up across our community include really engaging our hubs at a higher level and um we're going to be having a series of kind of in focus conversations about things like remote sensing and how the hubs can support the the modeling that you want to do and provide the data on the ground that you need and and we know that it's going to be a process but i think this this conversation will really support what hubs can do and how they collect the data that that you need to train models um so uh yeah i look forward to continuing this conversation if anyone would like to be involved in the the science side of this conversation uh please reach out to me i'll drop my email well everyone should have my email and i can include you in that conversation and we'll continue to also update hubs on on what is possible and what can support them to have the remote sensing tools that maybe aren't available right now so easily but could be in a year from now or two years from now so um yeah a constant uh learning curve for all of us anything else to add dorne i see there's a few comments uh again i'd just like to express my gratitude for everybody's uh you know active participation in this conversation and uh and uh so we can all again learn together uh and so uh i appreciate you tuning in both for this in depth and we've got some really exciting ones uh uh lined up that i think each one sort of builds on uh on some of the last so uh um again thank you all and uh we'll we'll see you again sue thank you thank you all thanks sam gisele sophia gregory yeah thanks everyone that's awesome that's what we want to see

2021-03-17 07:07

Show Video

Other news