Webinar: Convergence of Wetland Science and Technology: Predicting & Field-Mapping Wetlands

all right looks like attendees are starting to file in good morning or excuse me good afternoon everyone uh thanks for joining us today my name is mike connors i’m the director of customer success at ecobot this is the fourth in our webinar series on the convergence of technology and wetland science and we’re especially excited about this one as ecobot is an esri emerging partner and we have esri co-hosting with us today we’ve got a very interesting topic predicting and field mapping wetlands and a really great panel of speakers and industry experts with us uh who will introduce in just a few minutes uh so before we get started just wanted to remind everyone who’s joined us before we’d like to make this as interactive as we can be over zoom so uh please feel free to use the tools that we have in the chat and the q a to ask any questions that you have and uh if at all possible we’ll roll those those questions into the topic that we’re speaking about and uh we can always you know have some follow-ups and some breakout sessions later if we can’t get to everyone’s questions um i also want to make a note that we’re recording the webinar so if anyone needs to come late leave early just want to catch something again we will be sending out a link to all registrants uh with some instructions on finding uh the video over the next day or two so um i will kick it over to uh my co-host now my colleague jeremy chevy who is a professional wetland scientist with nearly two decades in the industry as well as the co-founder and chief scientific officer at ecobot jeremy it’s all you great thanks i’m gonna share my screen here so we can kick this off properly and uh again super excited to take this time to be joining with you all here again this is our this is our fourth webinar and we’re pretty excited with the level of interest in this so i’m jeremy chief scientific officer with ecobot and i am your other co-host daniel martin i’m consultant and project manager at esri with a background in ecology and landscape architecture where i focused on mitigation great and today we’re going to be looking at prediction models and field mapping of wetlands and so i thought i thought we’d start off this way this is why we’re all here water water is life our human body is made of up to 60 percent or more of water 71 of the earth is covered in water but only three percent of it is fresh and only half a percent of that water is actually utilizable accessible fresh water that we can use and survive on that’s a crazy ratio to think about sixty percent of our body and half percent of the earth and so i’m betting that’s why everyone is here on this call today because we’re all looking at ways that we can better manage and plan for that half percent so today we’re going to we’re going to do a brief introduction to our panel and our presenters uh very brief touch into why we’re doing this again a very brief intro into esri’s arc hydro tool set and also into ecobot but the majority of our time is going to be focused on our two case studies uh and at the end of our case studies we’ll have a very brief touch in to our next webinar coming up in september and possibly into october and then we’ll have some time at the end for discussion and to field questions and as mike mentioned at the beginning if you have questions feel free to throw them into the chat box or the q a uh bubble down at the bottom of your screen there that will uh and some some people might be able to handle those questions on the fly uh as we go we have a great panel lined up today including sandra fox senior project manager with st john’s river water management district and megan lane the chief scientist at u.s fish and wildlife service joining us panelists and presenters is uh nevin durash and gina o’neal nevin is a senior biologist for ese partners with over 18 years of field biology experience and 15 years of environmental consulting experience he has a diverse environmental consulting background including work and natural resources gis and due diligence projects with recent work focused on utility scale renewable energy projects this work has included sampling and assessing a variety of biota including wetlands and waters as well as wildlife such as birds bats small mammals amphibians and invertebrates devin’s work on these large-scale projects has been supported by integration of gis into each step of field data collection in order to efficiently plan conduct and report field survey data and jean o’neil she joined esri in 2019 following the completion of her phd in civil and environmental engineering from the university of virginia her dissertation focused on the development of a wetland delineation model involving work in lidar dem analysis python geoprocessing workforce hydrologic and geomorphologic modeling and machine learning at esri gina continues to leverage these skills by supporting spatial analysis projects related to remote sensing and hydrology among others in the natural resources sector in her free time she enjoys hiking in her new home in colorado and learning to bake bread so thank you to all of our panelists for being here today and and gina may pivot and start shipping out bread to everyone i’ll be ready for it um why are we here the pivotal role of this webinar series has been to really help bring our attention to what we are really doing and what we’re really doing i’m an ecologist i look for patterns but it’s not just looking for patterns in nature looking at social patterns looking at digital looking at data patterns and in the ecosystem of wetland science and technology we are converting biological information and natural communities into digital data and the purpose of this really feeds into being able to make better policy make more intelligent planning work with conservation preservation making sure that there is room for generations to come we’re giving a voice to aspects of our world that don’t necessarily typically have a voice and by creating that data we are strengthening we’re bolstering that voice we’re harnessing the power of this new ecosystem for natural resources planning monitoring and conversations so as as jeremy was just articulating the picture emerging in wetland science is one of interconnectedness and shared efficiencies this new digital ecosystem is where a network of hardware software and science work together in the planning and conservation of natural resources before launching in today today’s presentations we just want to take a quick moment and give you a little introduction to arc hydro and ecobot so at a high level you can see where these tools are located within that digital ecosystem so what is archive well first off it’s free which is always great um archivo operates on top of the esri platform for those of you familiar with the esri platform it’s important to point out that this is not an extension it’s not a solution but it’s an add-in that primarily operates on arcmap and pro you just need to download the archive setup and that installs the toolbox and the associate resources the idea behind archivo is to provide a gis framework for the development of integrated analytical systems for water resources and there are several key elements to this so it’s a framework not a solution for a specific water resources problem it’s geared towards analysis and not cartography and it actually includes multiple resources such as a data model tools such as the whim which we’ll be hearing about today workflows on how to best use those tools for particular use cases and the community of users with common interests and patterns of use the arc hydro geonet page is a gateway to all these resources in the platform to share information and experiences here are a few more links and resources but no need to worry about writing them down now as we will have the links from the presentation summarized afterwards online for you guys to visit yeah so the same place that you’ll be able to go to pick up the recordings there will also now be a section uh starting next week where you’ll be able to find these links for the resources that we are utilizing throughout our webinars i just want to very briefly touch base into uh the you know give a brief intro into ecobot so ecobot is in a field application that works on mobile tools that is allowing wetland scientists to complete their work at about half the time offline and what we’ve done is we’ve streamlined the level of effort required to complete u.s ace wetland delineations the data collection reporting and field mapping of wetlands through our intuitive mobile application and our partnership with esri tremble and many others ecobots integration with measuring mapping technologies allowing people to not only map wetlands in the field offline but also when syncing via bluetooth to sub meter accurate gns receiver gnss receivers we can and is really allowing people to increase efficiency massively also our recent merger with wet form has really helped position us as the front runner in the field application niche and this is the new in the new digital ecosystem for wetlands scientists and so with these massive levels of efficiency gain that we’re seeing and we’ll hear a little bit more about today ecobot has over 10 000 wetland delineation reports that have been created and those go from few acre wetland delineations and it scales all the way up to landscape level and potentially up to nationwide wetland assessment projects as well so we’re excited to be helping people save 50 to 60 of their level of effort so today this is why most events are here we want to hear from our presenters so we’re going to take a look at two case studies we’re going to look and see how the combination of arc hydro of gina’s whim tool and ecobot are having like how is this impacting landscape level projects we’re going to see how these wonderful tools combined to create those massive levels of efficiency that i was just speaking to and so we will have time at the end for discussion and questions and answers but without further ado i’d like to introduce our first speaker and colleague nevin dursch from ese nevin take it away good uh morning or afternoon depending on which side of the country y’all are from me uh we’re just right here in the middle in texas so uh glad to have everybody joining us so yeah i wanted to give uh kind of an idea with this presentation of you know where where the field biologists are that are doing day-to-day wetland delineations in you know particularly in the consulting world kind of where we’re at right now and kind of a tip towards where we’re going uh you know really the goal with uh our projects uh all of you who are consultants know and even you know everybody knows you got to maximize your time in the field and uh try to you know make the most of it um and so what we’re trying to do here is really just how can we multiply force and really make the best of the time we’ve got in the field so with that jeremy if you want to go ahead and advance so a lot of the projects we work on particularly here in texas we’re dealing with a large-scale wetland delineations uh particularly for uh renewable energy projects where we’ve got extremely large footprints uh two three uh even up to twenty thousand acres of land that we’re looking at and trying to identify you know wetlands and wetland complexes within uh and a lot of these require extensive delineation work uh it’s oftentimes well ahead of any design for the project or anything like that um and it’s you know the amount of time needed to actually do that in the field can start multiplying pretty rapidly and you know just a general cost to do that uh can get out of control very quickly as you as you scale up in acreage in particular in the texas coastal wetlands we’re dealing with a lot of widespread wetland complexes often where we’ve got a bunch of isolated features that really don’t connect to anything downstream uh where they’re you know their isolated depressions so with that uh jeremy if you want to go ahead and advance uh we’ve got you know fuel data we need to collect following the 1987 manual you know running transects uh luckily we have relatively uniform habitats that we’re working in a lot of times it’s even abandoned agricultural fields whether it be rice sugar cane something along those lines uh the complicating factor for us in many cases though is that we’re dealing with numerous smaller parcels uh that have been cut out over the years and everything and resulting in you know a number of different baselines that we have to run and it starts expanding the number of transects very quickly uh and again we’ve got you know kind of a interesting situation with these isolated wetlands that we’re dealing with in the area in advance thank you so in order to get out into the field you know we want to prepare as best we can uh to really maximize that time as i mentioned so you know a lot of y’all are familiar with uh should all be familiar with at this point uh you know the native data products that are available so this is you know data that it doesn’t really require a whole lot of synthesis or anything you can pretty much review the data see what’s out there you know you’ve got national wetlands inventory uh the usgs topographic maps and uh nrcs soil surveys looking for hydric soils and those the only issue some of these can be dated uh the resolutions may be you know where the scale that we’re looking at may not be a small enough scale to pick up uh the the types of wetlands that we’re looking at in a lot of cases so that’s one of the challenges that we face initially with those products uh once we move on to aerial photography which is really improved uh i know just over the course of uh of my career uh you know it used to be that the usda at nape the national aerial imagery program you know came out every couple of years uh you were lucky some years it was three bands some years it was four band that sort of thing now it’s become you know very consistent in what’s presented to us and we have the options uh for three and four band data um but it does require some interpretation and some knowledge you know along the way of uh interpretation particularly with the color infrared imagery and being able to read that uh some of the data also available uh depending on the location some of the high-res imagery that you get from local jurisdictions uh i know around here in central texas where we work some of the cities often fly down to six inch aerial resolution which can give you great definition the only issue being that those you don’t know the the exact frequency that those are flown at uh they they generally aren’t flown quite as often so they may not be as up-to-date as say some of the usda imagery then as far as the advanced data products go and this is you know been growing recently we’ve got we’ve always had the national essay always but for a long time now we’ve had the national elevation data set uh which has been a good starting place but the resolution is pretty coarse on their uh you know generally uh 10 to 30 meter resolution and it it just doesn’t provide quite the definition for some of these smaller wetlands with discrete hydrology but as of late lidar data has become much more readily available again speaking specifically to texas the stratmap program has been implemented by the texas natural resources information service tenris and they make their lidar data available freely to the public and it’s all online at this point it used to be in the past you actually had to mail them a hard drive and wait for that thing to come back around and you know you needed to know exactly which tiles and all that now it’s become extremely easy to just go pop in download the data you need and you’re good to go so a lot of the data the lidar data it’s actually been driven by uh the need for updated fema mapping uh with the newer floodplain models uh coming online and that’s that’s been really driving the the acquisition of a lot of that data and the publication of it moving forward go ahead so um for those of y’all who aren’t quite familiar with lidar it’s basically you get a few different data products out of it at least and this is a screenshot from the tenoris texas natural resource information services from their website but typically the the three data products that are available are the digital elevation model uh and that’s usually a bare earth uh digital elevation model they’ve removed uh clutter from uh you know trees buildings those sorts of things you get hypotography so the contour lines uh and as well as a lidar point cloud that actually has uh classification information for that data one of the caveats i guess with uh working with this data you’ll find very quickly that um it can occupy a huge amount of uh space i know uh storage is getting cheap on computers but it’s still uh it can start adding up pretty big for for these larger uh projects we’ll go ahead so back to the projects that we’re working on um they’re typically located a lot of the larger projects at the time right now are in the human gulf coastal prairies region uh one of the interesting issues that we’ve got with that is we’ve got a lot of soils that are mapped as non-hydric but with hydro components so essentially the nrcs soil surveys just aren’t performed at a resolution that really helps us out with identifying these smaller pothole type weapons that we’ve got one of the other challenges that we run into is a significant uh agricultural modifications to the landscape levees ditches irrigation canals uh those sorts of things um it’s one thing when they’re actually active and maintained but once they’ve been abandoned it can really throw a wild card into just trying to determine on a landscape what’s going on hydrologically where water is trying to flow where it can’t flow anymore and trying to determine that can be extremely difficult especially if you don’t realize that until you get out into the field uh you can spend a heck of a lot of time just trying to chase down what direction things are going because it is at least in the area we’re working uh topographical relief is almost non-existent uh the the levees and canals are generally the minimum and maximum of the topographic relief in the area go ahead so as i said um with the low topographic relief additionally the the wetlands we’ve got there i’m typically talking about freshwater emergent wetlands here um you know we say coastal prairies but we are uh outside of any uh uh saline areas uh typically uh one of the other challenges that we deal with down there is uh we’ve got you know being historically farmland in a lot of cases we’ve got widespread and abundant invasives it forested wetlands down there are almost exclusively chinese tallow additionally even a lot of the prairie areas are just vastly dominated by uh cyperus uh the woodrest flatsedge there’s about 50 different names for that one it seems like but we’ll go with usda’s for today but you know having uh fat in fact wet plants being so widespread uh and generally just the the human environment and everything um you know almost everywhere you go down in this area you’re going to have a dominance of fat species even in upland areas it really tends towards uh the hydrophytic vegetation and as i said the you know the map soils uh generally don’t uh show the definition for hydro that being said a lot of the the soils once you get on the ground uh particularly in the areas we’ve worked with historic uh uh agricultural work uh almost everything ends up looking like hydric soils that we get into so move forward so um as i was mentioned so if you rule out our well or our vegetation and our soils because those are fairly consistent inside and outside of wetlands in a lot of cases oftentimes we get down to hydrology being our real key indicator for these types of wetlands as to what’s going to call it in or out and we’re talking about extremely minor changes in topography where we end up with a geomorphic position and that’ll will allow us to make a weapon call either direction and it it’s we’re talking a matter of inches in most cases uh where that occurs uh here on the right you can see example of one of the pothole wetlands that we’ve got this one’s a little bit deeper it’s actually got some water some open water on it but we end up with a lot where uh the surface features are a little bit more subtle and generally dominated by emergent vegetation go forward so with that all in mind with these larger projects you know as i said multiplication of force try to do what we can to to make the best of our time in the field we want to look at some gis tools that are available for uh you know identifying where these potential wetland locations may be ahead of the field study you know if if nrcs isn’t showing them if nwi isn’t showing him you know what other you know and even some of the false color imagery just doesn’t have quite the definition for us uh what are other uh information that we have available and i’ve been uh involved in gis for uh all of my career and bounced back and forth i had started arcgis i was just trying to think it was seven point something that i started arcgis on if that uh dates things a little bit uh but i’ve also worked in the with qgis using some of the tools and libraries that they have uh one of the ones in particular is grass which was developed by the army corps of engineers originally both and as and saga which was uh developed by uh some folks over at the university of hamburg in germany both of those tools uh actually have standalone interfaces that can be used as well but they’re also incorporated into some of the toolboxes in qgis go ahead so understanding you know the types of wetlands that we’re really trying to identify out there ones that aren’t necessarily connected to downstream uh hydrologic features that are isolated uh you know we we’ve wanted to try to figure out how we can locate those ahead of time and you know reading through some of the literature and everything really these these can be identified as sinks throughout the landscape and uh both the the arc hydro tools and the qgis tools have some options available for identifying sinks and a lot of times they’re used for preparing data sets for further analysis but what we wanted to do is just to see how you know kind of how it shows up with with the data we’d already collected in the field we took our original lidar dems filled sinks on those and then just made a subtractive difference uh grid that essentially shows us the depth of the sinks and it’s been really helpful for us to kind of tease out you know some of these deeper depressions that we’ve got even some of the shallower ones and with that i don’t have a scale on here and that’s uh cartographically obviously a big no-no i apologize to the arcgis folks but um we’re talking even the so with the pothole wetlands that are shown the deepest one comes up maybe a foot below the surrounding landscape so really really small differences in topography that we’re talking about here let’s move forward one of the other tools and this is really where we’re trying to figure out what’s going on with some of that historic uh agricultural work that’s happened in these areas uh looked over to the the topographic wetness index it basically you take a model where you determine the attachments uh within the area based on the digital elevation model as well as the slope for for those off that digital elevation model and calculate the topographic wetness index for it and it’s it’s generally meant for steeper sloped areas but we did find that it did a very good job of highlighting where these agricultural improvements had been made historically the image that you can see down at the bottom many of those uh the straight-line dishes are almost invisible when you’re on the ground it’s you know we’re talking maybe six inches of topographic difference there and you know when it’s all covered in sedges deep rooted flat sedge uh you can’t see them at all and so it’s very helpful in determining where these areas are and where water tends to collect on the landscape there and just through a mention in there there’s a saga has their own wet wetness index that uses a kind of a modified catchment scheme to to map the uh the catchments there but it’s it’s very similar to the top graphic witness index so with that we took the the twi maps and uh our sync difference dm out into the field and you can go ahead so we also combine that with some of the natural color and color infrared data to really hone in on what what parts of the area needed additional work we targeted uh specific uh wetland areas that kind of met the criteria of having you know a aerial signature that looked like it had potential as well as where the the two indices that we used were showing increased signs of wetness and lower topography and used that across the site to go and actually target uh which wetlands we wanted to look at and had a lot of success with that uh one of the drawbacks that i see right now though is we’re using these as essentially a qualitative assessment we’re we haven’t advanced to the point where we’ve got a quantitative cut off where we can say okay if the wetness index is you know x number then that’s where we need to draw the line uh we we don’t have anything like that that we’re using at this point um you know the other drawback is that it does require some knowledge significant knowledge of the site already and just generally the local landscape i don’t think that these tools are in a place yet the tools that we are using and as we’ve implemented uh that they would be in a place where you could do just a desktop assessment for an area and uh be able to write off large-scale chunks of a project without any sort of field verification but they they’ve we found them to be significantly helpful in you know as a field uh base map data source uh particularly when you get into a pinch trying to run and delineate a wetland and just figuring out you know what’s the topography doing because as you can see from the the picture on the right that’s the actual view that you get on the ground about 90 percent of the time from these areas and you just you’ve got to have some other map sources to to really help you tease out what you’re going to be doing in the field from there you can go ahead so what does that actually mean for the time we spend on the ground uh we had a project uh that uh this case studies generally based on uh over 3 000 acres that required routine wetland delineation we recorded over 250 wetland data forms uh in ecobot mapped over 90 aquatic features that’s 14 days in the field with two person team and recorded a maximum of 31 wetland points during a single day um so it really upped our efficiency we’re able to you know travel through areas and you know we’re able to try do the transects as required by the corps of engineers uh with that additional knowledge of the hydrology in the area of where the sinks actually are located where water tends to gather a lot more than if we had gone in blind or with just the typical base map data go ahead and so you know moving forward looking to test this in the field additionally in order to further determine what the wetland limits and how that may correlate to the sync maps and the twi and saga wetness index models and just use whatever gis tools are available you know i think the next presentation is gonna definitely tip us in another direction as well on uh other tools that are available and uh with that that’s what i’ve got we’ve got one of our little friends there from that uh wetland site actually a little uh alligator that decided to join us on one of our delineations thanks wow nevin thanks so much for sharing uh that was quite the project so our our next presentation will be uh jean o’neal discussing her wetland identification model take it away gina thanks daniel so hi everyone my name is gina i am a technical consultant with esri and i’m really excited to have this chance to present to you guys uh the wetland identification model so this tool set is a product of my dissertation work while at the university of virginia and i’m really lucky to have the chance to continue developing it and implementing it as part of archive with esri um so with that i will go ahead and get started next please so just a little bit of background um next so i’m sure i don’t really need to pitch to this group why having accurate wetland inventories is beneficial um so of course just this knowledge of wetland distribution and abundance is really important for protecting remaining wetlands and not just that but it’s also uh required before any large land development projects through section four for the clean water act so it’s on a lot of people’s radars and i’m just looking quickly at that figure on the right of the drivers of wetland loss in the u.s these are things that are not really going away as far as urban development rural development um so this is a pressing need and will continue to be so although it is beneficial to have wetland inventories the process of carrying out manual field surveys are very costly and time consuming especially when you have these very large areas and furthermore while we are locking the us to have this national scale repository of wetlands i.e the national wetland inventory for the purpose of finding these smaller environmental planning skill wetlands on the nwi can be insufficient in addition to that nwi was not intended for mapping wetlands for regulatory purposes and it has been found to omit a considerable fraction of wetlands so for the purpose of protecting remaining wetlands we need to kind of look elsewhere next please and that’s where remote sensing or coupling of remote sensing and machine learning can come in so in about the last decade this somewhat recent collection of remote sensing data really offers new opportunities for being able to identify wetland and their characteristics uh quickly and at large or varying scales so among the remote sensing data that is useful um there is multi-spectral imagery radar and light detection and ranging or lidar data um and then of these in the development of the wetland identification model we really honed in on lidar data so as you can see on the right um the green on that map is showing where in the u.s lidar data is available and with this image being a few years old now i’m sure that that map is more and more green so not only is lidar data freely available it’s also really high resolution so when we use lidar returns to create this digital elevation model we can describe the ground at a resolution of about two meters um and from that from those dms we can describe flow convergence and patterns of near-surface soil moisture all things that could tip off where wetlands are likely to be and finally machine learning models can be used to learn the different rules or topographic characteristics that make up a wetland and then use those rules to unseen areas to predict new wetlands next please um and so those two principles bring together or those two principles are coupled for the wetland identification model or the whim so basically the whim aims to use both lidar derived wetland indicators and machine learning to identify likely wetland areas specifically at this environmental planning skill next please so the whim is implemented as an arc hydro tool set as you can see so there’s three main portions to this process there’s a pre-processing a predictor variable calculation and then a classification and accuracy assessment and in order to run the whim um you need three key inputs so one of those being a high resolution dem or likely a lidar drive dem second being um a raster that shows you where surface water or streamlines or blue lines are and then third is a ground truth raster which essentially is i’m going to teach your model how to identify wetlands by comparing things to these ground truth areas of known wetland and non-wetland so again um this process is automated and implemented as an archive tool set which you’re seeing on the right there so each of the different components are made into a script tool and then to execute the entire process in one step you can also trigger these tools in succession using the model builder that it’s implemented within next please so now going into um some more detailed parts of the wim workflow and starting with pre-processing so um i would say a key component that the wim offers is some different methods for applying smoothing to your input dems so when we’re working with these really high resolution dms we can run into an issue of micro typographic noise where the more and more derivatives that we take from the dem the more noise and highly detailed changes in topography that we’ll notice and might actually deter or might actually get in the way of just focusing on those topographic changes that indicate wetland formation and you’re seeing an example of that on the right so here you’re seeing the same transect along a wetland and where that wetland hits is outlined in red and along that transect the curvature derived from a wider dm and you can see how that curvature changes when you changes along the wetland transect when you apply no smoothing to the input dm and going on as you apply a mean a median a gaussian and then a pronamalic smoothing this is all essentially to show that by introducing this pretty key pre-processing step of smoothing your dm before taking the derivatives from it you can really prepare the typographic surface um in the way that best represents wetland formation and kind of scales down uh those topographic variations that aren’t as important for this application next please and now going into the second part of the workflow so deriving these different predictor variables um and so there are three predictor variables used in the whim and each of them are used as a proxy to show where water is likely to flow and where it’s likely to stay so the first one is one that nevin also talked about the topographic wetness index or twi and this input is essentially looking at the ratio of an area or the comparing an area’s tendency to accumulate water to its tendency to drain water so again where in our landscape is water likely to flow to and where is it also likely to stay there next please the next one we use is curvature which again is a pretty common one so this one just looking at the second derivative of the topographic surface and using that to teach the model or to allow the model to learn what parts of the landscape are convergent or divergent and how does that coincide with wetland formation next and the last predictor variable is a depth to water index um so this is a newer wetness index that has been used for wetland identification and this is where that streamlined data comes into play so with the dtw the model is assuming that areas that are very flat and uh typographically close to streamline to streamlines are more likely to be saturated so again just another indicator of where is that near surface soil moisture most likely to be next please and so just kind of coming back out to the bigger picture again um each of these three predictor variables are all being derived from that lidar dem and then before it is ingested into this machine learning portion of the workflow they are combined which you’re seeing on the right there as this three-band composite image where each band of that raster is drawing information from one of those three predictor variables next please and that brings us to the core part of the workflow which is the machine learning or the random forest classification so just to give an overview of what random forest is they are based on decision trees so here you’re just seeing an animation of what a decision tree does so having a feature space on that x and y coordinate x and y coordinates um and then knowing the categories that each of those different icons belong to the decision tree is basically trying to identify splits and all those features so that at the end it can categorize all of the input data into groups that are as different as possible to each other and that within each group all the features are as similar to each other as possible next please and so how does that translate to um our application of wetlands so if we’re looking up in that left uh upper left hand corner um imagine that we have all of our topographic data and we’re going to select that scattering of pixels to train our model and we know that those red pixels are wetlands and we know that the black pixels are non-wetlands so we would um run these this collection of pixels through our decision tree and the decision tree mates might start to look at the dtw are these pixels near uh near water and are they flat if not it’s probably a non-wetland moving down to curvature are these pixels in an area that’s convex yes or no and it makes a more defined selection and then lastly it may look at that typographic wetness index and see if there’s a high flow accumulation within those pixels and so we would learn these rules for what a wetland looks like and after training itself on this kind of testing or training set it would be able to apply the same rules to a new area where we do not know wetland and non-wetland next please and so the random trees or the random force algorithm implemented in esri products as the random trees classifier it takes that decision tree building block and takes it a step further so beginning on the left-hand side with that same collection of known wetland and non-wetland pixels it will take bootstrap samples of that training set so that means that we’re going to split that one group of known wetland of ground truth pixels into several different groups where we’re going to select with replacement back into the main group and from each of those bootstrap samples each one is going to build its own decision tree and so by having each of those independent decision trees we we remove the risk of the model being very influenced by or being biased to the input data and the model gets to see many different iterations of what a wetland could or could not look like and the final prediction is a result of a majority vote of all of those different decision trees so the advantages of this random trees classifier is that it is really good protection against having a model that’s over fit to one area because it does take the different bootstrap samples um like i said it reduces biases and input data and then it also has a really handy metric at the end that tells you the variable importance so essentially it will tell you if we were to take away the twi what would our drop in accuracy be and we’ll do that for all of all of the variables next please and another part of this um workflow is that it offers a flexible training sampling scheme so uh you can split up your ground troop data into a training and testing set where only the training is seen by the model and you can choose to either sample from your entire area of interest which you’re seeing on the left there so just picking random selections of pixels there um and using those to train the model or you could use something that’s a little bit closer to reality which is on the right and limit that training data sampling from just the sub area so on the right imagine that um all of the box there needs to be surveyed at some point and we’re going to send boots on the ground to tell us where the ground truth wetlands are just within those yellow limits and use that data within the yellow to train the model and then we’re going to apply the model on the larger scale of that unknown area next please and so that brings us to the final portion of the workflow on the accuracy assessment portion and i just wanted to quickly touch on this as accuracy assessment for any machine learning model um it’s important to specify that this machine learning model is trying to detect anomalies in the landscape so typically you’re going to have a lot more non-wetland area than your wetland area so the wind was designed keeping in mind that we want to have a very transparent and representative accuracy assessment that won’t mislead the user thinking that an area that is predicted to be all non-wetland is very accurate just because most of the area is not wetland so that points us to these two key metrics that the accuracy assessment portion of the workflow uses so one is high recall on the left there so if you imagine that blue area delimits um delineates uh your predicted wetland and then that red outline is where the true wetland is this would be an example of high recall because we have captured a large portion of that true wetland and then in the middle we have um precision and so precision or having high precision precision would look like being very strict or very conservative with your predictions but you’re going to try to be correct as often as possible so what you’re seeing there where the predicted wetland is very tightly within the true wetland boundaries we’re not making a lot of overpredictions that’s a high precision scenario and then ideally a model would have high recall and high precision so again that ideal scenario where we have encompassed a lot of the ground truth wetlands but we’re also really scaling back our overprediction next please and so now just going through some example outputs from the whim um so the wim provides two main outputs one being those kind of binary predictions wetland or non-wetland and that’s what you’re seeing on the left um so on the left the gray areas are where the model has predicted non-wetland and the blue areas is where the model has predicted that there is a wetland and the red outline shows where those wetlands actually are in the landscape so we can definitely see that we are in some areas um pretty tightly adhering to those true wetland boundaries in some other areas we’ve predicted predicted a wetland where there isn’t actually one so maybe some over prediction so maybe some other input data would need to be introduced there to remove that overpredictive tendency of the model um on the right hand side you’re seeing the other type of output that the limb has which is this wetland probability raster and this output instead of categorizing each cell into a binary wetland or non-wetland it is instead showing you on a scale of zero to one what the model’s probability is that that cell belongs to the wetland class so here you can kind of see you know of the entire wetland prediction where was the model most confident that it is a wetland so this could be really important for decision makers in deciding we want a model with really high precision and we really want to limit overprediction so maybe we just focus on the cells where the model is at least 80 positive that there’s wetland there and then of course vice versa next please and just wrapping up now so um go taking a step back and looking at what the whim in its current form can contribute to those um following similar applications so one the limb is meant to be a flexible workflow and it’s well suited for the model development phase um so it’s certainly meant to be built on it’s still and it allows users to you know add in different wetland indicators if maybe you’re in a low relief area like nevin’s study area and you want to use vegetative indices instead of topographic drivers um in its current form the limb has shown that it has the ability to serve as maybe a wetland screening tool so although we have that moderate over prediction it has that high wetland coverage so maybe we use the whim in its current form as a screening tool where for a large area that needs to be surveyed we run the whim and in that way we can uh scale down the area that were target targets area where we want to do manual surveys uh next please and finally just some key capabilities that are new capabilities that the whim offers within the esri suite of tools so one is some dm smoothing methods that are specific to improving the landscape for wetland identification um second is some more robust outputs from the random force algorithm including that sliding scale wetland probability roster that we looked at and finally an accuracy assessment that is really targeted to um describing how good an anomaly detection application is next please and yeah so with that um thank you guys very much again it was great to be part of this webinar thanks great yeah thank you gina well well done super exciting i think a lot of people were really looking forward to seeing this this combo so pretty pretty powerful tool you’ve come up with there it’s kind of fun to look at it from the opposite side you know i come from boots on the ground and then like you’re coming from this macro scale i’m just like oh it’s pretty exciting to think here we are we’re meeting in the middle so uh good stuff uh just again touch base in september our webinars will be focused on drones for field mapping and how those pair with field applications and then in october we’re going to be looking at a webinar we’re going to take a look at another part of the tool belt in the arc hydro tools so more on that to come so keep your ears to the ground for that i just wanted to thank you all for taking your time to join us today daniel and i had a great time putting this together uh nevin and uh and and gina both were just awesome to work with so it looks like we still have a few uh minutes left and so i’d like to transition into a couple questions for the panelists uh the first is for gina and nevin as you both discussed wetland modeling so how might best performing wetland models vary by region and what additional data would be ideal for further model development gina could you share any thoughts on that yeah definitely um so i think that between my presentation and evans you actually got a really good look at how you might need to focus on different wetland indicators and different landscapes so the limb was originally developed in for a few regions in virginia where we have a lot we can rely more on that topographic uh characteristics of wetlands um whereas in nevin’s region or in more coastal regions maybe you know a whim model there looks more like one that um looks at vegetative indices uh rather than topography alone so i think that the models would vary um based on they would vary by the input data used and maybe even the degree that you smooth your input dem maybe in those coastal regions with really large isolated wetland depressions you can use a coarser more coarsely smooth dem rather than an area with those linear kind of riparian zone wetlands yeah that makes sense i’m i’m usually uh generally more focused on uh the the south texas region in general and that’s where we’ve focused our efforts for this uh type of work uh historically uh so that’s been the the area of my focus uh recently so i can’t really speak to other areas of the country but i think it would be somewhat applicable to more playa lake uh region somewhere like in the texas panhandle and going up through the great plains as well great well so i’ve got a question i’d like to to put out um that’s in respect to uh you know love to invite megan to speak in but you know looking at some of these tools that uh we’ve been looking at here today um how do you how do you see that these tools and arc hydro or some similar modeling as well as field data and mapping applications how are those things going to potentially help in the long term with uh data set generation for nwi or at least improving some of those sure i’m happy to address that jeremy so i would say that with nwi our we primarily rely on base imagery typically aerial photography or fine spatial resolution you know satellite data but we also pull in what we call ancillary data sets and we commonly look at digital elevation models in the past we use stereo imagery topography has always been a big part of what we do of course water flows downhill and so the types of products that um both nevin and gina you know generate would be you know ideal ancillary data sets for nwi yeah and then you know sandra i’m thinking on you know on sort of a regional basis you know and with some of your experience with arc hydro how do you how do you see some of these tools especially this wim tool potentially impacting some of ural’s work from a water management perspective i can definitely see applications for both of the tools that were demonstrated today of course i have to throw in the limitations of available digital elevation models and i see gina’s head nodding uh that’s going to be kind of the bane of our existence uh i am looking forward to an opportunity to test that tool but what i found most exciting from the set from gina’s presentation was her use of the word current uh form because i’m starting to see other applications and ways that this tool might be applied to other problems and i like that use the term uh flexible workflows so that you could actually be using the same kind of format to answer additional types of i’m happy to be in water supply so you know what kind of questions how this would apply so yeah these are wonderful developments for geospatial applications using our hydro related uh tools great thanks sandra gina anything that you want to add to that um yeah i would just kind of reiterate what sandra was saying um it is certainly meant to be flexible and it’s in no way a complete uh black and white this is the way that everyone should do it in every landscape um so hopefully if users try out the whim they’ll see um or they will uh be encouraged you know try their own input data or try to find something other than wetlands it really depends on if you have ground troop data for streams for example let’s see what kind of topographic features or other could be used to identify or predict where streams are likely to be great thanks and so one of the things you know i’m trying to keep up with all the things coming in on the chat and the q a here too there’s a lot coming in whatever we don’t handle we will nudge into future conversations as well but one of the things i do want to call out just in respect to the dems and some of the lidar information david schaefer with the army corps of engineers just chimed in a little bit ago in respect to a tool uh that might be available using the usgs 3 dep lidar image service and so he sent a link over on that you’re welcome to grab that we will also uh save that link onto the uh onto the resource page once we have posted that up here in the next few days so that you’ll be able to access that um and we’ll kind of kind of move from there so one of the things that you know just pinging off of landscape level modeling with this whim tool and other other such data sets that might become more available megan this is for you again you know where does fish and wildlife service hope to go with some of its mapping programs in the near future i know we’ve had a few conversations on that just on the phone just you and i over the past six or eight months but maybe you could share a little bit there as well sure um so you know we we are well aware that land cover changes and that nwi is not up to date you know throughout the country and that to some degree is um you know due to the fact that wetlands are hard to map and you know our process is time consuming and we don’t have the resources uh to um carry on with you know manual uh updates um you know unfortunately you know with the resources we have available and so we are looking to pull in um more automated tools i don’t envision a purely automated process anytime in the near future uh not considering the amount of classes that our data set supports but we’re hoping that some of the automated tools such that you know gina mentioned and never mentioned evan mentioned can be used to in some small but significant way augment our our workflow and so that can be done in different ways so for instance with the type of data sets that gina produced again those can be used as ancillary data sets perhaps to kind of speed up the manual decision making process or perhaps we’ll get to a point where the automated processes can be used to generate foundational data that can then be built on using manual approaches so we’re looking for that optimal combination of manual and automated processes to get us to standard compliant data in as um efficient uh a workflow as as possible efficient and cost effective or we’re kind of after the the holy grail um it’s not going to be easy um but we’re determined and we are looking forward to working with all of you to apply the tools necessary um to provide the the data that we all need to you know conserve wetlands yeah always always great having conversations with you megan i really appreciate you taking your time to join us here today i just want to let everybody know we are at the end of our hour we have a few more questions still coming in so if people are willing to stick around for a few more minutes we will just continue in this discussion if you have to drop off i understand some people have meetings they need to get to but please stick around for a few more moments someone had just asked how they can access the whim tool the wim tool is part of arc hydro so you just need to access that through geonet and so uh we will provide the link for that online as well in case you want to get around in there and play with it as far as i remember daniel that is something that is if you know once you have a as your license and you’re working with the software you can just access that arc hydro tool set as a sort of an add-on that’s correct yeah so i actually dropped the geonet link into the chat right now so if folks are interested in hopping on that uh sooner before we uh are able to get the list of resources out they can do so but yeah arc hydro is a free add-on to the esri platform uh specifically pro and arcmap and uh and the wim model is is part one of the tools that is contained within archive um i know we’ve got a couple more questions uh gina i might just bring your eye to the q and a about two-thirds of those questions are directly pointed towards you um so some of those you might be able to just answer on the fly there um yeah one of the questions that that i’m seeing that really jumps out and this is from a pretty good friend here in the asheville area so what what data set are you training the model on uh is it nwi on field collected data what are you actually utilizing to inform that machine learning part yeah um so we initially tried training on just the nwi data i’m and although when the nwi says that there is a wetland there it’s very often correct there is a wetland there but sometimes it’s just not exhaustive of all the wetlands in the area so then we were we shifted away from there so that we weren’t um giving the model wrong information on non-wetland area and so for that reason i would say it’s really important to use um manually surveyed of data that you’re really confident in both where the wetlands are and where the non-wetlands are so in developing the whim we were actually using manual surveys that the virginia dot let us borrow um so with that we had you know the exact limits that they delineated and all the wetlands that they found within those limits so we could be really confident in both the wetland and the non-wetland area okay another question that popped up that’s kind of a tangent to that but is have you tried uh utilizing whim in any developed or urban environments seems like the twi and depth of water ratio is a little more tricky there was the uh yes question yeah um we did and we found uh that we had a higher rate of the false wetland predictions along roadways so where we’re uh our topographic indicators are actually finding like a you know culvert or roadside drainage area rather than a wetland and i think in scenarios like that um really highlights how the whim could benefit from additional data sources like just land cover classification so you know to cut out all the developed area or soil moisture information um so yes it is certainly less accurate in those highly developed areas and i think in those areas it it could benefit from ancillary data sources great another question here is this one’s directed to nevin but how how do you see some of these tools you know such as combinations of whim and some of the field application tools like ecobot how do you see those uh affecting planning and budgeting for larger scale wetland delineations on the regulatory side yeah i think i think what it’s doing and this has just been my observation in general is it’s front loading the field process a little bit more of spending a little bit more time in the office ahead of time on you know really trying to dial down on where where your wetlands are where your potential areas of concern are where you’re gonna be spending a lot of time uh actually in the field ahead of uh actually getting out there and deploying folks uh to the site uh so you know it i know when we when i first started in in this work it was a lot more of like take a look at a few maps all right let’s get out there and do it and uh now you know actually spending a little bit more time doing some analysis ahead and especially as with with the large scale of some of these sites uh you you really need to be planning in advance otherwise uh you’re gonna get it out there and not know what to do yeah so a little time and energy investment invested prior going in the field may really speed up and truncate the process of what you need to do on the ground in your body yeah on the ground and then in the back end as well uh once you get back in the office um it’s you’re a lot closer to being able to you know uh do your post-processing with your gis data and you know basically finish up your report send it out and we have seen it speed up on the on the back end side of things great uh one of the questions here daniel i think this kind of goes back over to you and i think we already addressed it but somebody’s asking what type of esri licenses uh spatial or geoprocessing are needed to run the whim yeah that’s that’s a great question um and i think i might kick that uh to gina um to to correct me if i’m wrong but i believe that uh that it is within our hydro which doesn’t have any uh connection to those those other extensions gina yes um so i believe with pro 2 4 uh you did have to install um one python package scikit-learn to actually run that random forest portion of the workflow um but with pro 2 5 and later that’s no longer necessary and every all the dependencies you would need are installed automatically great thanks um somebody was asking here just comparing to some of the older methods used to develop the mwi how much more accurate this i’m going to give this to eugene how much more accurate do you think that the the whim is for coastal or flatter areas like in terms of efficiency um yeah that’s a great question we ran into a lot of issues in the coastal areas and what we found was if we apply basically if we um applied a coarser smoothing scheme to the dem and by that i mean smoothing over a lot of kind of smaller depressions and just really leaving those larger i guess significant depressions um by doing that we were able to really improve our wetland accuracy in those flatter coastal regions but it is certainly harder um to use just the topographic variables alone that’s part of this kind of classic model configuration compared to the nwi um i’m not sure specifically to coastal regions um i just know that in general the whim had a higher wetland accuracy so it identified more of the wetlands in the area but was not nearly as precise as the nwi wetlands were so it could be a trade-off depending on what is the most important end result yeah okay great thank you so one one for nevin uh the question is i just wonder if you could use this type of office delineation work to complete an off-site delineation as discussed in the course delineation manual uh with coordination with local district offices of course seems like it’s pretty accurate and significant so yeah i would say definitely preface it with uh talk to your local corps office before spending any a significant amount of time on that i think you know i’m a field biologist uh i really like to see some ground truthing i understand the limitations though at times uh whether it be land access for a number of reasons uh and we can certainly deal with that uh in our neck of the woods um it i would be i would want to have a good working knowledge of other wetlands that are nearby if i were to try to uh make a case for an off-site delineation with that data very close by reference sites where you actually do have ground treating uh available um so just to piggyback off of that nevin someone’s asking if if i’m not sure if this is for you but in general but any luck utilizing the antecedent precipitation tool from ace to help justify jurisdictional calls i have been using it recently one of the places i’ve actually found it helpful is with and one of the things i didn’t mention uh during during my presentation um with the aerial photography of knowing the exact date that those aerials are flown is hugely important and actually going back and running antecedent precipitation tool for the date that the aerial was flown so you can see if you’re seeing a lot of water on that you know the cir data is that was that a normal event or was it flown you know during a heavy flood event or something like that and we’ve actually got some sites where we’ve had some issues where you look at this aerial and it’s like oh my gosh there’s water everywhere here went back ran the tool and it’s like well it was significantly wetter than normal uh you know during those times and so it actually has helped provide some justification uh in some cases that no what we’re seeing is not a normal event on those aerials that were being referenced regularly throughout a document great so i’m going to jump back over to gina and again if oh go ahead sandra i’ve been taking notes what was the precipitation tool that you just mentioned i’m not familiar with that thank you so much for of engineers watching precipitation tool i can probably dig up a link here real quick and post it to the chat great yeah thank you um so so again we’re we’re pushing time here but we’ve got so many great questions so i’ve got one more i want to throw over towards gina um so one of the questions coming in from kevin stark asking about how do ephemeral and intermittent riverine wetlands play into the wetland probability surface so flow accumulation thresholds precipitation groundwater dominated streams yeah um great question i think that so right now the whim in its final prediction it doesn’t distinguish between different types of wetlands um although that is certainly a direction i would like to see it go i think to be able to yeah predict different wetland classes one of course the input ground truth data would also need to have those labels inherent to it but also additional input data additional predictor variables so again soil moisture land cover vegetation those would be the kind of extra characteristics i think the model would need to be able to distinguish between wetland classes great um i think jeremy do we have just one additional yeah megan thank you i could follow up on a question regarding kind of the what you would get out of nwi versus wim especially in coastal environments it’s going to depend on you know numerous factors so for one it’s going to depend on how up-to-date that nwi data set is you know if you’ve got nwi data from the 1980s it’s going to probably contain a higher degree of uncertainty than say recently produced nwi data also with with the wim products you’re looking at potential accumulation of water based on topography so a lot’s going to depend on what’s driving wetland occurrence what’s driving hydrology in that area there are going to be some areas where that um small variations in topography at the land surface aren’t going to drive hydrology we’re only going to drive hydrology you know during certain um within certain kind of climatic regions also you’re going to have uh the the structure um kind of influenced by things like ditches and so you might see kind of those depressions but if the ditches are there they’re going to be draining those those low points and so again the topography based um kind of wetland potential is going to be less relevant there so a lot is going to depend on the quality of nwi data and then also the landscape that you’re dealing with and so i know that’s not a simple e a simple answer perhaps not the one that that folks want to hear but um my suggestion would be to look at both data sets and to decide which one works better for your area that’s great yeah i seem to remember there was a great presentation at the sws conference in baltimore last year that was kind of looking like side by side like nwi and lidar and you know they took like three different sets i i’m gonna have to dig up and see who that presenter was because that was really well done um and i think it’s important to weigh in like the the data can have uh be useful from different sets in different places where you have varying degrees of accuracy um so thank you megan um so one of the things that i want to kind of cap off with here is you know again there’s a lot of questions that we still have not answered i was trying to answer them as fast as i could but they were coming in as fast as we were answering them um which is great means people are really engaged and it means that uh we’re we’re talking about a subject that’s very exciting for both of us uh for all of us um uh gina a few people have uh been asking about your published uh works and so you know if we will in the resources section again for the for the landing section page for this webinar we will provide links out to some of gina’s work um uh i know that you know some of the tools that sandra has been working with is another part of the archive toolbox is the is the hydro period uh tool and that’s something that we’re pretty excited about maybe taking a look at here in the near future as well seeing how that ties in because one of the questions that was asked here was you know how will this tool be you know potentially helpful in helping to identify where sites are for potentially restorable wetlands i guess we’re thinking about changes in hydrology how you know where are sites that whether it’s a mitigation bank or a large public entity like where can we do some restoration restoration projects so i don’t know maybe we had a gina how do you feel like that’s something that you can handle touching a little base or two um i mean yeah i think that the way that you use the predicted wetlands could vary um i don’t think it’s necessarily right to say that the predicted or the most probable wetland areas and use those directly as these are where the wetlands definitely are and where they are not but i could see them because the limb does do a fair amount of over prediction of wetlands that it could instead indicate yeah a good candidate for wetland restoration um so maybe there isn’t a wetland there right now but all of these different uh surface hydrolog surface hydrology drivers mean that maybe it would be a good location for a wetland great well so so so you know i think a lot of you know that recently ecobot merged with with wet forms they’re pretty excited about that and pat’s pat murphy has had a couple great things in here and he wants to challenge eugene and he said here’s a real challenge for the model run in alaska where almost everything is wetland so maybe we can get you guys get you guys in communication on that thanks for that pat oh that was a good one um so while we are we are pushing our time here daniel any last uh thoughts or questions that you want to feel before we wrap things up for the day no i don’t have anything um other than to just uh thank uh the panelists uh so very much for taking the time and joining us and all you folks who uh who joined us on the audience side uh really great um conversation that we’ve had at the end some fantastic questions so thank you all yeah agreed so yeah megan and sandra thank you both for coming in devon now then and gina great presentations and daniel i’ve had a blast with you on this looking forward to the next one as we continue to find what is most alive and cutting edge in the convergence of wetland science and technology so thank you everybody we’ll hope to see you again sometime soon thanks everyone everyone appreciate it thank you for inviting me you’re back all right thanks guys take care

chief scientist

As found on YouTube

Book Now For Environmental Consultingl In Newcastle

Leave a Reply

Your email address will not be published.