 Hello, my name is Ginger Cole. I'm a GIS programmer analyst with the Natural Capital Project at Stanford University and I'm the developer of the rangeland production model. In this video, I'll be describing the basics of how the model works, what kind of information it produces and how to run it. And I would encourage you to refer to the resources links that are in the agenda for the hackathon for durable resources like the user's guide, a quick start guide, and all of the inputs that you can use to run the model for our study area in Mongolia. But let's talk a little bit about what the model is and what kind of information it provides. First context for a model to study rangeland production. I believe everyone listening to this video will know that rangelands are critical globally and they support livelihoods and other ecosystem services such as carbon sequestration, forage for wildlife, and biodiversity. Our goal in developing this model was to be able to put rangelands and the support that they provide for biodiversity, wildlife, water, climate regulation, and human nutrition and livelihoods on the map. We know that rangeland production is controlled by a complex interacting system of climate and human management, mostly driven by livestock grazing. But the role of stewardship or livestock management in determining these key benefits that we get from rangelands is not typically represented in most ecosystem services models. Rangelands are often treated like one homogeneous class of land use. When we know in reality there's a huge amount of heterogeneity across landscapes in the ecosystem services that they provide driven by highly variable climate, soil, and management. The ecological value of rangelands is intuitive and very easy to understand, but the drivers that control those provision services are very complex. This model can help to explore questions like these, like all models developed by the natural capital project and all models in general, we break down those complex interacting drivers into some simplified relationships. The model can be used in decision contexts to study questions like how productive is the rangeland on this landscape? What is the impact of domestic livestock grazing on rangeland condition? And what is the impact of the rangeland condition and productivity on those livestock? And the model can be used for scenario analysis to answer questions like what will future rangeland condition be and future animal condition be with changing climate and changing livestock management. A very big picture overview of how the model works is shown in this diagram. On the left, climate and grazing livestock are the main drivers of the model, controlling forage production and rangeland condition. And then on the right, the main outputs of the model are rangeland condition and animal condition, describing how the rangeland system is influenced by climate and grazing pressure. Note for now that rangeland condition and animal condition are very vague terms, but we will get into the details of how the model represents them later. So the model is basically designed to be used to answer the question. Given a set of possible changes in climate and changes in livestock management, what are the likely consequences for rangeland condition and for animal livestock condition? There are three main sub-models, the forage production sub-model, animal spatial distribution sub-model, and the animal off-take and diet sufficiency sub-models. And we will talk about each of these briefly during this video. The model extends over both space, as you see in this map, and time. So each of these processes are simulated by the model at each monthly time step and for each pixel of the study area. Again, I'd like to reiterate that the model has a monthly time step. So it's up to you how many months you would like to run the model for, but as you're running the model, each of these sub-models is simulated per pixel per month. First, the forage production sub-model. This piece of the rangeland production model is an adaptation of the sentry model, which is a very well-established and old model originally developed for grasslands. Sentry takes climate and soil drivers and predicts production of new above and below ground biomass. It also predicts senescence of live biomass and decomposition rates above and below ground. These are mostly driven by precipitation, the main driver of productivity, and also soil nutrient availability, temperature, and a set of parameters that describe the plant response to all of these. From this sub-model, we get standing biomass and forage nutrient content. Standing biomass is the metric that we use in the rangeland production model for rangeland condition. Next, the animal offtake or grazing sub-model. This is also adapted from an existing model. It's adapted from the graze plan model, and it essentially takes forage biomass and the density of animals on each pixel, and it predicts how much forage will they consume. This is driven by things like the animal characteristics, their need for forage, their demand, their ability to access the forage as predicted by the model, and also the nutrient content of the forage. So if forage is predicted to be of high nutrient content, animals will actually be able to eat more. From this, we get a prediction of how much forage is consumed by animals on each pixel for that monthly time step. Next, after the model predicts what animals will eat, we compare the nutritive content of that diet to the animal's nutritional needs. We do this with another simplified metric called diet sufficiency. Diet sufficiency is simply the ratio of the intake of metabolizable energy to the animal's requirement for metabolizable energy in that time step. It essentially gives us an indicator. Given the predicted intake of forage by animals on each pixel, to what degree does that diet meet their nutritional needs? Because it's a ratio, it varies around one. And if diet sufficiency is lower than one, that indicates animals are not meeting their nutritional needs and their condition may be declining. If diet sufficiency is above one, one or greater, it indicates that the diet is sufficient to meet animals' nutritional needs, and we can expect that their condition is improving. Next, grazing, of course, impacts standing biomass in the model through animal offtake. Animals consume forage. There's therefore less standing biomass at the end of the current time step. However, the growth of grass predicted in the next step, predicted by our adaptation of the century model, is also impacted by this offtake by herbivores. So grazing impacts forage production in the following time step of the model through altering the root-shoot ratio, total production of forage, and also altering the nutrient balance through return of nutrients to the soil in urine and feces by the animals. These are all controlled by parameters that are required data inputs to the model, but example published versions of these parameters exist for the century model, which is how we adapted this routine. The last sub-model to describe is the animal spatial distribution sub-model. This is performed inside the model, but I just want to give you a brief idea of how it works. So to model the animal offtake of forage on each pixel, we first, of course, need to know how many animals are on each pixel, the density of grazing animals. The best information that we have, though, about livestock densities is often at the level of administrative units, such as zooms in Mongolia, where each zoom contains hundreds or thousands of pixels. So we need a way of disaggregating the animals across pixels within zoom, going from this top map on the top that gives a total number of animals per administrative district down to number of animals per pixel. In the model, this is done by looking at the difference between potential biomass, which is predicted by the model without grazing, and observed biomass, which is taken to be indicated by a remotely sensed vegetation index. For our application here, we're using NDBI. So this illustration with the green rasters shows that we assume that areas with a small mismatch between potential biomass and observed biomass are lightly grazed, but areas with a large mismatch, such as high potential biomass but very low observed biomass, are understood to be heavily grazed. And this is how we disaggregate animals across pixels in order to predict more accurately the amount of forage taken off by animals on each pixel. The main outputs of the model are rasters covering the study area, and for each monthly time step for which the model was run. So for each of these, you will get rasters of standing biomass, that's our indicator of rangeland condition, and diet sufficiency, our indicator of animal condition. Since each of these are rasters, it is possible to construct time series, such as you see on the right. However, that would need to be done in post-processing of model outputs. Also, I want to mention that all of the model outputs are, of course, associated with some uncertainty. Although the model outputs are given as fixed quantities, with a model like this, it's difficult to precisely quantify the uncertainty of the model outputs. However, we know that the model is impacted by uncertainty in the driver variables. For example, the rate of precipitation is uncertain, to some degree. It's also impacted by uncertainty in parameters. For example, the rate of new forage production, given a certain level of precipitation, is uncertain to some degree. And also some structural or process uncertainty. For example, the factors that control how animals choose their diet. Like all invest models, the rangeland production model must be calibrated before we can have confidence in the absolute values of its outputs. However, even if there are no empirical data available to calibrate the model, we can have more confidence in the relative values, such as relative values across time, like you see in these time series on the right, and also relative values across space, like you would see in rasters, where the model can help to identify potential hot spots of productivity or grazing intensity, and cold spots, or areas of relatively lower productivity or grazing intensity. So next, I'm going to move into how we will be using and running the rangeland production model for this hackathon. And this information is replicated in the quick start guide. You can see the link to that guide in the agenda for the hackathon. You can use that as a guide to installing and running the rangeland production model. Rangeland production model is not distributed as part of the invest model suite. So you will need to find and download the installer for the rangeland production model separately. You can follow the link in the agenda to the installer, or follow this link given in the PowerPoint presentation here. Each release contains the installer to install the model on your computer, full documentation, a set of sample inputs, and the source code for the model. But for this hackathon, we will be using tailored inputs for the model study area. So don't use the sample inputs that are distributed here with the model. Most users will want to install the rangeland production model using the setup executable that you can download following the link in the agenda. This will be installed just like any other program on your computer, and it will install an application called rangeland production. Unfortunately, we don't yet have a Mac installer. If there are any users with Python experience, we also have a Python package that you can install from GitHub, but I will not be describing this use for the model because I think most people will probably simply use the model installer or the setup executable. User interface for the rangeland production model is very similar to invest models. So by this point, you will be somewhat familiar with this user interface. Most of the inputs that you need to supply are just file paths to the location of the file on your computer. Remember for our purposes for the hackathon, all of those files will be downloaded from the rpm input folder on box. There are some help texts that you can read by clicking on the blue icon next to each input and just click run to launch the rangeland production model. It's very important for the hackathon, you can take the input values directly from the rangeland production model exercise, and I suggest that you do this, copy and paste them directly because it's very important that you get the text of those inputs correct exactly, and unfortunately a small typo in the text of those inputs may cause the model to crash in a way that is not going to be easy to figure out what happened. So be very careful when copying the input paths for the hackathon. Also, all of the inputs that we have provided for you to use for the hackathon are documented in an appendix of the exercise. So if you'd like to see how I generated all of the inputs for this study area, you can find the details there. I'll just briefly talk about what the inputs are and how they work together. First, there's the workspace, which is where all of the outputs will be created. The time period to simulate with the model is specified by the number of months, the starting year, and the starting month. The area of interest is like all invest models. That is the bounding area for which the model will simulate. Soils, data, control, plant nutrient availability. And for our application, I downloaded them from the ISRIC soil grids website. The model requires monthly temperature, precipitation, and a vegetation index, such as NDVI. We only need one month's worth of temperature data. For example, we need temperature data for January, February, March, April, etc. For precipitation, the model requires precipitation data for each month that the model simulates. For example, if you wanted to use the model to simulate 24 months, you would need to provide precipitation inputs for each of those 24 months. Same for the vegetation index. You would need to provide a vegetation index raster for each of those 24 months. Describes in detail how the site and plant functional type parameters are mapped to space. If you're interested in understanding this, you can see the quick start guide for this explanation. But for our application, I'll just note that we are using a single set of site parameters and a single set of plant functional type parameters across the entire study area. I'm not going to go through how these site parameters and plant functional type parameters are mapped to space. Next, we have animal inputs. These inputs must describe the total number of animals inside each polygon. For our case here, we're using zoom boundaries, and we're using livestock statistics from the National Statistics Office, giving the total number of livestock in sheep forward units per zoom. And then these must also be associated with some parameters given in the animal parameter table. The model requires initial conditions be given by you, the user. These can either be supplied as rasters, one raster per state variable of the model, or as tables. And for our application here, we're using tables. You can see examples down on the bottom of the screen where the each state variable considered by the model is initialized with a single value for the entire area of interest. This final box in the model interface that says save state variable rasters is optional. If you check that box, the model will save a raster of each state variable for each month that it is run. So this will create a lot of rasters on your computer. However, it can be very useful for debugging, or if you're interested in looking at some of the intermediate outputs of the model. Usually, I would recommend that you do not save state variable rasters and you simply look at the main model output. So the output created by the model will be created in a folder called output in the directory that you supplied as the workspace. A summary results folder is inside that folder, and that creates outputs aggregated across space or time. And then there are also rasters in the output folder giving output values for each month that the simulation was run. Those outputs include rasters of animal density, disaggregated, rasters of diet sufficiency for each month, potential biomass, which is predicted biomass in the absence of grazing, and standing biomass, which is biomass remaining at the end of that time period after animals have consumed what they will intake. Summary results are what I would suggest that you look at first. And in the rangeland production model exercise, these are the results that the exercise will walk you through looking at for the exercise. So this will show you summary results such as mean potential biomass, mean standing biomass after animal offtake, and mean diet sufficiency of animals. All of these are taken as the mean across time steps, a whole time period that the model was run, and across pixels inside that feature. The last summary output that I'll highlight here is mean animal density. This is the mean across model time steps per pixel. So it will give you an idea of where the model has estimated that there more or fewer grazing animals. To sum up, the rangeland production model is useful to augment monitoring efforts across space and time, because the model gives some estimate of both rangeland condition as indicated by biomass and animal condition as indicated by diet sufficiency on a continuous basis across pixels of the study landscape and across months of the year. We hope that the model can be useful to augment field monitoring efforts. Also, the model can be useful to analyze impacts of changing climate and changing animal densities. It's set up to be relatively easy to run the model with increased or decreased precipitation and increased or decreased animal densities. This can give us an idea of how much rangeland condition or animal condition is sensitive to changes in precipitation and changes in total animal numbers. Okay, with that, please check out the rangeland production model exercise and the quick start guide. Good luck!