AoR 10: Matt Reeves, Rangeland Forage Prediction Tools

Range forage production can vary widely from year to year. What if you could get a production prediction 3 months ahead of the growing season and make management decisions from it? Tip's guest today, Matt Reeves (USFS) is part of a group of researchers who have developed the Rangeland Production Monitoring Service, a backward-looking prediction tool that is not meteorological forecasting, but a machine-learning approach to a seasonal outlook based on historical climate data. This episode includes an interview with Matt followed by Matt's talk at the Society for Range Management conference a couple weeks ago.

The slides referred to in Matt’s talk are available at the show notes link:



>> As we have discussed before here on The Art of Range, forage production on rangelands can vary quite a bit from year to year. It can vary from more than double to less than half of a long term average for much of the west. While this variability, and not just aridity, it is part of what defines rangelands. Satellite derived production data showed greater variability after about the year 2000. This year to year fluctuation has significant implications for ranchers. If you are going into a year with 50% of average production, you need to respond with some management adjustment. That might be calling cows, shortening the grazing period, early weeding, delayed turnout, a reduced stocking rate, et cetera. If you are able to anticipate an increase in forage production, it might be a good gamble to exercise a different kind of flexibility. It might be staying in a range pasture a little longer. It might be keeping the same grazing period, but increasing how many animals go in, or letting a traditional stocking rate serve as a temporarily light stocking rate in order to effectively provide a year with a little rest, where not all plants get grazed or they're just grazed more lightly than usual. Researchers have been working on ways to one, measure forage production using satellite data, and two, predict whether a given year will yield more or less than the average based on several factors, not just precipitation. My guest today, Matt Reeves, has spent his career swimming in the deep end of the decision support tool pool. He will reference a number of data products that are currently available to the public, which we will link to in the show notes.

[ Music ]

Welcome to The Art of Range, a podcast focused on rangelands and the people who manage them. I'm your host, Tip Hudson, range and livestock specialist with Washington State University Extension. The goal of this podcast is education and conservation through conversation. Find us online at

[ Music ]

My guest today on The Art of Range is Matt Reeves. Matt Reeves works for the Forest Service out of Missoula, Montana and does work on modeling. Matt, welcome to the show.

>> Thank you, Tip. And I do work in Missoula. I work with Rocky Mountain Research Station, which is the research arm of the Forest Service.

>> Matt, how did you get into that kind of work?

>> Well, I got into that kind of work. I started, I graduated from Washington State University as a  in rangeland management and went on to get a Master's and then my Ph.D., ended up in Missoula where I presently am. And the reason I gravitated towards that work is I found it very interesting that we could use satellites and weather information to tell us something about the Earth and the vegetation that you might not be able to see from the ground level.

>> So your piece of work was on vegetation model.

>> Yep, my Ph.D. work was primarily Eastern Montana and [inaudible] national grasslands. And what I did there was develop productivity models that would enable us to understand the impact of weather on forage basis in those areas. But those models can easily be, and have been, transformed for the rest of the United States as well.

>> Yeah, that's interesting. Over the last, I guess I'd say 25 years it seems like, there's been an increasing recognition, the extent to which variation from year to year in precipitation amount and timing really drives both forage production and I guess more punctuated changes in plant communities around the west. And seminar, like Nathan Sayre did in his book, Politics of Scale: The History of Rangeland Science, that that interannual variability and unpredictability has probably always characterized western rangelands, and is one of the reasons why they're a little bit more difficult to apply fixed management recommendations to. Is that born out in what you have found with looking at historical information, trying to model it for the future?

>> Yeah, absolutely. One of the main things, though, that I'd caution against, overgeneralization, in that yes, variability has always been in the system. But I think we've seen more variability entering the system in the last two decades, and we have, I think, quite a bit of evidence to suggest that in some of the tools we're going to be looking at here pretty quickly on this podcast, we've discovered that, in fact, I think variability is increasing in some of our major rangeland areas. And I don't know what the causes of that might be. I have some guesses. But variability has always been there, but it may be increasing.

>> Right, right. So, what you guys have been trying to do is to tie, I guess put together models of both vegetation productivity that are measuring in real-time, and combine that with climate prediction models so that we can try to give a forecast within a management scale timeframe that people could respond to who have to make a living making decisions on the ground, is that right?

>> Yeah, that's right. And the way I view all of these tools is that it seems to me that range management is as much of an art as a science. And I liken it to being, you know, if you're a manager or producer, it's almost like you're, you are an attorney, you're trying to make your case for some choice that's been made, and you need all the evidence to make those decisions on your side. And so I think what we are trying to provide here is some tools and ideas that help tip the scale in your favor. And we do offer the first of those tools that you mentioned is our rangeland production monitoring service, which has two components. The first would be the retrospective, in other words, the where have we been component. And the other component is the in season projections of expected forage yields. And the first of those, which is the retrospective aspect, relies on satellite information. And we've taken that satellite data and calibrated it to annual production across 600 plus million acres of rangelands in the United States. We can do that now because of the large quantity of satellite information that we get. In some cases, we have daily information, in fact, to include. So, that's the retrospective piece. And like I said, there's the idea of making projections. This is available for the northern region, so that would be, of the Forest Service, so that would be, you know, North Idaho, all of Montana, part of South Dakota, or North Dakota, and that projection system uses real-time information, including weather, and also biweekly satellite information. We take that information and make projections about expected forage yields. We don't use forecast information as some of the other systems that we're going to talk about. For example, the Grass Cast. The Grass Cast uses, you know, meteorological forecasts. Say, for example, yeah, it's going to be warmer and wetter in a month than it is today. We don't take that philosophy. Instead what we do is we look at 35 plus years of observations and we use the real-time information to ask the question, when we have been where we are right now, in the past, where did we end up? Which is a very different question than saying, what's the weather going to be like in a month or two. And that's our philosophy, is using a backward looking approach to model the future.

>> Yeah, I've often said that it would be easy to be a meteorologist because you can be wrong more than half the time, and you still get to keep your job. It looks like the objective with this is to look in the past and apply, apply  I realize that forecasters, meteorologists do that when they make forecasts. I do realize that. But it looks like the idea with this is to be right more than half the time based on what has happened before.

>> Well, that's right. And the way that you do that is you look through all of our observations over the last 30 or 40 years. For example, we know that cheatgrass dominated rangelands can respond very positively to wintertime or fall moisture in the previous year. And that can affect the growing season conditions. Well, you'd want to incorporate that in your model because if you were just using the forecast to inform that model, it would be naive to antecedent conditions, you know, in the near, in the near past. And we don't want to get too technical. I realize this really isn't a technical forum. But the way we do that is a data mining exercise. So, it's a machine learning approach where the computer can tease apart those relationships between annual production, satellite information, and meteorological observations in the past to project into the future. And importantly, the projection work that we offer, we also include a 95% prediction interval. So, we offer the user, or the consumer of this product, our measure of error. And, of course, as you might imagine, early in the year, the error is usually larger than as we get closer to the future. But I can tell you that, for example, in 2017, we had the flash drought, and much of Eastern Montana, particularly in the northeastern corner, we had called it by about May 5th, suggesting that this was going to be a pretty epically bad year. That may not be enough time to, for a lot of people to act on, but considering much of that country has a peak, usually sees a peak about fourth through the eighth of July, it does give you a little bit of time to think about how you're going to react. But the drop, each biweekly period that we did that projection, it kept going down. I had not seen that over the past 35 years. We had not seen that kind of behavior.

>> How about 10 years ago? I was visiting with some guys at the Nobel Foundation, and this was just after they had the massive drought in the south, Southern Plains there. And they said that drought was predicted by every meteorologist, every common modeler under the Sun. Everybody said, this is going to happen. There's really no question about it. And there were some people who took that advice, you know, who called heavily, who destocked. Some people, you know, took, Harlan Hughes was a North Dakota State University lifestyle economist, and one of his, one of his recommendations was, under certain circumstances, when the market is high and you've got bad conditions coming, liquidate the whole herd and buy back when it comes back. That's pretty extreme. But his point was, people, there were some people who listened to those recommendations and responded with management action. And those people were in really good shape coming back out of the drought, either because they had what little pasture they had they could sell to the highest bidder, or because they destocked enough that their range and pasture was able to ride it out.

>> Much of that Texas Panhandle, you know, Western Oklahoma area, did experience conditions we have not seen over that 35 year record. In fact, it was probably 20%, 10 to 20% worse than anything before that in terms of forage yield. And so it took a big, a big cut there. Some of those areas, though, did experience some recovery, and we can see that in the production record. For example, by 2015, a lot of those areas in the Southern Plains had recovered, in fact, exceeded pre drought state, suggesting that maybe there was some resiliency built into the system. But that isn't to say that all places came out okay. In fact, some sites, some areas we see do not appear to have recovered to pre drought conditions. And so I know it was a very hard time for a lot of people.

>> You mentioned a couple of data products that are available to people. Can you go back and describe some of the applications of this that are available for people to take a look at now?

>> Yeah, there are two. Again, I'll remind you. There's two avenues here. The one is the for in season forage projections that's available for the northern region, which is the Montana, Dakotas, and part of North Idaho. And that will be most useful to people looking for more information to incorporate into their annual plan. It's also useful for people developing risk management strategies. So, if you want that added piece of information, it's just one more piece of the system that helps tip the scale in your favor. Fire and fuel managers will also benefit from this because when we combine the projection with what we've seen in the past, we'll have a much better handle on the regional fuel conditions than we presently do. And I'm sure there will be other applications that are derived from that projection product in the future as we continue to provide that service.

>> And what is the name of the projection product again?

>> The projection product is just the second part. So, the Rangeland Production Monitoring Service is two components.

>> Got it.

>> And we really don't have a second name. Some people have suggested you'd call it a cowcaster or something, but I don't know that I'd go that far. But with the retrospective part of the monitoring service that tells us where we have been, it has had a lot of use. And, in fact, we are using it now to update ecological site descriptions for a lot of adjoining lands that adjoin Forest Service allotments. For example, to update those sites with more, I think more relevant and more contemporary estimates of forage production. We've also used that information to understand something of treatment effectiveness, particularly in the Kaibab National Forest. They asked if we could use that type of information to understand the abundance of forage post treatment for a lot of their PJ and a lot of the thinning that's going on right now on the Kaibab Plateau.

>> PJ is Pinyon Juniper?

>> Pinyon Juniper, that's right. Thanks for catching me on that. A lot of Pinyon Juniper thinning. Some Ponderosa pine thinning. And they wanted to know, can we see and detect a forage response from this across a lot of these areas where our limited number of range cons might not be able to visit. So, I think it's an excellent way to get a handle on what's been happening across your landscape. It may be a ranch, it may be allotments in a federal land management scenario. It may be pastures. And in just a short bit of time using the tool, you'll be able to understand what's been happening. Have you been going up, down, or sideways? And I'll tell you that a lot of times, what people see on the ground, they forgot what it was like say 20 years ago, and they don't necessarily remember how today's conditions compare to what happened 20 or 30 years ago. It's especially relevant with our Forest Service and land management range cons for example. That's the range conservationists that we'll interface with. Permittees seeking to use federal lands for grazing. Very important for those folks because a lot of times they're new to the job. There's a lot of turnover. That is a situation that can be difficult because they don't know what's happened. And so we can use this information to get them up to speed pretty quick.

>> What other tools are out there for farmers and ranchers? I'm thinking mostly of livestock producers, to be able to respond well to the precipitation and temperature anomaly we call drought.

>> Well, one of the things that comes to mind would be the National Drought Mitigation Center's Managing Drought Risk on the Ranch handbook. Folks might want to look at that too.

>> Great. And Matt, this has been useful. Matt Reeves is with the U.S. Forest Service Rocky Mountain Research Station. Matt, again, thank you for your time.

>> Yep, it's good to be with you. Thank you.

>> The following audio segment is from Matt Reeves' live presentation at the Society for Range Management Conference in Minneapolis on February 11th, 2019. This was in a symposium titled New Geospatial Technologies for Monitoring Rangelands. What kind of questions can we ask now? In his presentation before a full house, Matt references graphics that he's displaying on a screen. That slideshow is available as a PDF, and it's shown on the website. There are several ways to access the show notes. In iTunes, show notes are in the details of you in the episode window. You have to tap more to see the full text of the show notes. If you are listening in the sound cloud interface on the website,, you click on view track to the right of the episode title, and then show more to see the whole thing. And on the main website, you can click on the big button, view show notes, and you will be directed to a drop box folder where you can download a PDF of the same information.

>> Looking at issues that people are dealing with, one of the things that they have, a crosscutting theme, and that would be the need to maintain annual production. You know, if our lands aren't behaving as we hoped, that can cause some headaches. And so looking at annual production and support, each of these challenges in different ways. And today, I'm going to talk about how we can meet some of those challenges with this new data and tools that the Rangeland Production Monitoring Service. And I'm going to share some of the results of some analyses we've done around the country for different reasons, and talk about how just this one simple tool can reveal a lot about your landscape. So, with a good long time series, you can tell a lot about what's been happening on the landscape and see what that means for a forage base. But here we are, the Production Monitoring Service. Now, there's two components here that I want to briefly mention. The first is retrospective, which is where have we been. And that's what we use to monitor and look at trends. Are we going up, down or sideways? And then as in season [inaudible] forage projection starting March 1st, in the growing season. But today, we're not going to talk as much about the forage projection. That's a subject as something different. But it's in the northern region. And the focus here is to ask the question, where are we headed? And most of the talks we've seen today involve machine learning. This is another one of those where with we look at antecedent conditions and suggest what that may mean for linear term forage situation. So, answering the question, where do we think we're headed based on the information we've got, and that includes weather information and then both remote sensing, put together in machine learning environment. What I want to focus the rest of the time on is talk about the monitoring side. So, the retrospective version of this, which is built on the TM, and motives, and I'll explain how this is coming about. But the TM and motive satellites meets ranking 1984 to present and beyond. It's for all U.S. [inaudible] range. And so it's about 660 million acres of coverage, plus or minus. And it's built on the NDBI. That's been a common theme for many of these talks. And this one is really similar in that way. And it's available right now for anyone that has an interest. Let's get into the map a little bit. I told you it was a Thematic Map version which was a 30 meter platform. Scott did a good job talking about that as well. I worked with him on that project. So, I was real pleased with how that turned out. But it's a 1984 to present. And in 2012, Thematic Mapper had some problems, you know, nationally. So, we filled that with motives NDBI cross calibrated to the DM signal. And we are using maximum NDBIs to avoid this need of using growing season long NDBI and then having to make synthetic data. And so filling the holes, and modeling what should have been. So, we're using a maximum. And again, it's in these ranges. So, how do we calibrate this into production for something like the United States? And we focused on the ecological site description production value for high, medium and low. Get into that a little bit. And this is conducted for 110 different vegetation types here. So, here's what we did. We took, this is the SSURGO dataset with the production values associated with them. We've got the corresponding NDBI values here for the maximum, for the average, and then for the minimum values. So, this is a calibration way of converting RR, remote intensive information, into production values. And it works pretty good. So, for example, here's what we've got. You can see this is our 110 vegetation types here, one down. And we have three values, the above average, average, and the below average, and a corresponding NDBI from the same time period. So, it gives you a lot of points to begin to model with. Let's look at that calibration a little bit more. So, here's where we're at with calibration. And on the X is the NDBI, and the Y is the annual production. And this is just from the ESTs. And if you separate out a few of these outliers where there's a couple of points that I'll probably make later on, in either case, we've got R squared of about .8 here . 721 on the low range. And then down here, we estimated the error rate of no vegetation to be about .1 for those that have an interest in that sort of thing. So, what about the validation where it's really critical? This is validation data from the ARS. And you're not involved in [inaudible]. Pay attention to these points for a second on the lower end. These points here have very nice fit on the low end of production. And on the X is our observations. On the Y is the predictions. Then also take a look at the temporal. So, what's going on through time in terms of validation? A secondary X here is time in years. Secondary Y is a time series of annual production to observe. And that's what's involved with right here. The solid lines are predicted. The dotted lines are the observations. So, very nice couple in the back. And as we look at this cloud as a whole, we've got an R squared that's very high, except what's going on up here? Well, that's common. So, we looked at the common data [inaudible] situation. Any of you who used NDBI, you know it's going to saturate. We didn't find any benefit using EDI for those that have an interest in that. So, we couldn't get around the saturation. But I want to talk a bit more about that. This is where that saturation starts to happen, right? Right about in here. You can see it curving upwards. You can see it here. The good news is that's about 3,300 kilograms per hectare. Most of us working in the drier ranges of the United States don't have that problem. So, that's, I guess that's one good thing about that. We're just not that productive. So, tall grass, this doesn't  our method is not, as others have found, not as good as we'd like. You need some height information too based on what other research suggests in those very productive regions. So, let's take a look at some results here. And I want to focus on national [inaudible] levels, and then some ecological site [inaudible] units, which is a Forest Service kind of equivalent, the ecological sites. Let's start with 2011. Some of you remember seeing this before. And the reason I want to point to this is that after it dropped 2011, 2012, and what we see here is the percent change in ANPP compared with a 35 year mean. Boy, that was a tough, tough, year. And, in fact, we had never seen values that low. Conversely, this is when the Missouri flooded up here. So, when you take a look at the national perspective, you get some really interesting results. Let's take a look at the [inaudible] Rita Blanca National Grasslands right here. So, here's the temporal profile of that production system. And on the X is time. On the Y is the annual production. And I want to point out a couple of things. First, our average variability there is about 30% in terms of interannual variability. And, of course, down in here was when we saw that red. And that corresponds to this really low cap drop from 2012, lowest since 1949. Look at this difference, though. Between the bottom and the top, that's a 2 1/2 fold difference between a minimum and a maximum. So, that's one cool thing about having a kind series like this is that you begin to derive some pretty interesting metrics about variability of landscapes. Okay, a bit more regional in flavor. What you see here is the BLM allotment situation. And you can derive metrics, kind of like Scott pointed out, interannual variability. Trend through time. Our average production. And it's never average. It's either above or below the average. But the average production. And then the driest year. Walk through this a little bit with me. Cool tones represent areas with lower variability in the annual production. If I'm warm, it's higher variability. Over here, trend through time, I don't want to be red. That means I'm going down through time. 1984 to present. Cool tones, exactly the opposite. I'm going up. We're writing a paper about what's going on here and here. And if you have questions, I'll share with you some of those results. Average production, of course, red means lower. Over here, very interesting. Droughtiest year. If I'm greenish or coolish, that means my droughtiest year was further in the past. And these warmer tones mean nearer, in the future. This was the 1988 drought that encompassed the region when we had the [inaudible]. Okay, zooming in a little bit to the allotment level, this is quite [inaudible] they're having a bad set of years here. And I don't know why, what the cause is. We don't get into that. But using this kind of information, you can very quickly get a sorting of who's winning and who's losing this time period. So, [inaudible] canyon here. What we see, again, the color represents correlation through time. So, the red warmer tones means we're going down versus cool tones, we're going up. Now, zooming in a little bit more to ecological sites. For example, ecological units, down here in the Kaibab National Forest. This is within pastures. And we find something very interesting. Again, a correlation. It's from 1984. Cool tones, we're going up. Warm tones, we're going down. What's happening here on this soil? I don't know. But here's the production profile. Down and then it levels off. So, you can find some really interesting things. You ask the question, well, what's going on? And this provides a long term dataset for us to begin to ask those questions and look at priorities.

[ Inaudible ]

Yeah, same thing down at Kaibab. Let's look at some treatment effectiveness results in the dataset. So, on the X is time, and on the Y is production. These forests here, this is for an allotment. So, big, pretty good sized area. And these error cards here represent spatial variability over the allotment. So, our trend, we're going up, pretty good size error bars. What happened in, starting in 2005? We had some treatment going on and to get some competitive release. And the average post treatment, about 1,100 pounds per acre per treatment. 750. Kind of an interesting case. So, to wrap up some of this, we've had some good use cases so far. It's very interesting to study these profiles, because there's things in there that you don't remember. A lot of this hasn't been managing that long. So, ranchland in southwest, a lot of Forest Service regions now beginning to work with, and we're looking at especially drought and wildfire recovery and what that means for maybe some restoration efforts. [Inaudible] and bureau folks, especially in region five, which is California. Looking at some of the wild forests and bureau impact. I think it's a bit [inaudible] for people that haven't been on the ground very long. You need to get them ramped up pretty quick. You can work with them and say, "hey, this is where you're going to be working. This is what we've seen over the last 35 years." Show them the squiggly lines, talk to them about it, and they can get up to speed a little quicker. Because we have a lot of turnover in the agency. Many of you probably knew that. And finally, as a rangeland health indicator, there's no reason you can't use something like this to update your production values across the DSD. More contemporary because we're including what has been happening, you know, pretty recently. We had a workshop scheduled January 28th. It was virtual. And, of course, that didn't happen because of the shutdown. But in the next two months, we'll be rolling out another one. We're going to be interacting, looking from the different areas, talking with folks. And we also have a workshop, oh, I don't know, maybe tomorrow or the next day. You can find it in the bulletin here. We're going to unpack this a lot more.

>> Thank you for listening to The Art of Range podcast. You can subscribe to and review the show through iTunes or your favorite podcasting app so you never miss an episode. Just search for Art of Range. If you have questions or comments for us to address in a future episode, send an e mail to For articles and links to resources mentioned in the podcast, please see the show notes at Listener feedback is important to the success of our mission, empowering rangeland managers. Please take a moment to fill out a brief survey at This podcast is produced by Conners Communication in the College of Agricultural, Human and Natural Resource Sciences at Washington State University. The project is supported by the University of Arizona and funded by the Western Center for Risk Management Education through the USDA National Institute of Food and Agriculture.

[ Music ]

Mentioned Resources

Three other climate/ag tools developed by our friends at the Northwest and Southwest Climate Hubs:

We want your input

Future funding for the podcast will depend on listener feedback. Please take 1-2 minutes to respond to a 6-question survey after each episode.

Give Feedback

Taking suggestions

Have a question for us to answer on air, or a topic suggestion for a future episode? Email