Mapping Invasive Grassland Plant Species with Hyperspectral Remote Sensing

0:00 Speaker 1: To map an invasive plant in grass plants. And finally, at the end of this talk, I'm going to talk about data availability considerations for mapping invasive plants using remote sensing. So remote sensing of invasive plants is not new. In fact, it is a few decades old. One of the first studies that used remote sensing technology to Map invasive plants was done in 1975, where researchers used airborne color photography, color infrared photography, to detect an invasive plant called water hyacinth.

0:41 Speaker 1: This is an aquatic invasive plant. So remote sensing on invasive plants is not new, but since then, since 1975, there has been a lot of progress on remote sensing technology, on the methods we use to map invasive plants, and now remote sensing of invasive plants is becoming more and more common. What makes remote sensing a viable approach to monitor invasive plants? Well, there are many reasons, but the primary reason is that remote sensing can help us collect data across large spatial domains over time with consistent formats.

1:32 Speaker 1: which is a great advantage compared to field-based techniques. If you want to collect data on invasive plants based on field techniques, we need to go to the field. We need to visit a large number of points and document the status of invasive plants. But unfortunately, these field-based techniques are costly, are really time-consuming, and we often do them a few number of times. We cannot repeat them. Again, because of this, high cost of field-based techniques. So remote sensing has a lot of potential to monitor invasive plants.

2:13 Speaker 1: Another important question we need to answer before conducting a study on remote sensing of invasive plants is, how can remote sensing help us monitor invasive plants? So there are a few tasks where remote sensing can help us monitor invasive plants. The first task is mapping the presence of invasive plants. Some people call this direct mapping of invasive plants. We basically use remote sensing and map where invasive plants are present. Another task remote sensing can help us in the context of monitoring invasive plants is providing useful data for predicting the distribution of invasive plants.

3:01 Speaker 1: Some people refer to this as indirect mapping. A great example of this is spatial distribution modeling where we use remotely sensed data along with other environmental data to understand and identify areas that are prone to future invasion. And another task remote testing can help us with is assessing the impact of invasive plants. For example, understanding the impact of invasive plants on ecosystem functioning. For example, primary productivity. So each of these tasks, you know, we can spend hours and discuss them.

3:43 Speaker 1: And because of that, for today's talk, I'm just going to focus on the first task. mapping the presence of invasive plants. Although remote sensing has a lot of potential to map invasive plants, it is not necessarily simple. It is not necessarily straight forward. Here is one example, a hypothetical conversation that could happen between a remote sensing scientist and and application partner. So there is this remote sensing scientist. They have a lot of remotely sensed data. They are interested in studying invasive plants, and they approach their application partner, for example, a landowner, a manager, an ecologist, and tell them, hey, I have a lot of remotely sensed data.

4:37 Speaker 1: Are you interested in studying invasive plants? And the application partner responds, yes, of course. Can you help me detect individual cheatgrass plants in my land. So cheatgrass is the name of an invasive plant in some regions in the US. And the remote sensing scientist responds, I would love to, but these plants are too small to detect remotely. And then the application partner responds, oh, so you mean we cannot use remote sensing to study invasive plants? Now, when it comes to mapping invasive plants using remote sensing, often the expectation is that we can use remote sensing to detect and map every individual species.

5:25 Speaker 1: But it is rarely the case, especially in ecosystems like grasslands where we are dealing with really small plants. Detecting and identifying individual plants is difficult, if not impossible. And this begs the question, how can we effectively use remote sensing to map invasive plants in grasslands? There are a few ways. One of the easiest ways to effectively map invasive plants in grassland ecosystems is by taking advantage of contrasting phenology of invasive plants compared to native plants. So phenology, we can simply define it as the life cycle events of a plant.

6:12 Speaker 1: For example, when they green up, when they senesce. So if you can find a time window where, for example, an invasive species is green and native plants are not green and collect our remote sense data during that time period, the odds of success in mapping invasive plants will be higher. Now the important point here is that this phenology changes a lot from year to year because of many factors like temperature or precipitation. So in order to identify this time window, we need to have access to remotely sensed data with fine temporal resolution.

6:52 Speaker 1: In other words, we need to collect remotely sensed data frequently. Here is one example. On the left side, you're looking at a native plant community. On the right side, you're looking at an invasive plant community. So both of these communities are from the same site. They are only a mile or two apart. And these pictures were taken only a couple of days apart. But as you can see, the native plant community is completely finessed. But the invasive plant community is photosynthetically active. and very green.

7:34 Speaker 1: So if we can identify this time window and collect our remote distance data during this time window, then identifying and mapping the invasive plants will be way easier. But as I said, the tricky point here is that this optimal time window can vary from year to year. Another way we can detect invasive plants remotely is when invasive plants have different characteristics, different functional traits compared to invasive plants. So I use the term functional traits. Functional traits are simply attributes expressed by an individual in their environment.

8:17 Speaker 1: For example, nitrogen content, plant height, seed mass. These are examples of functional traits. Now, something really interesting is that these functional traits can affect how we see plants from above. In other words, plant functional traits affect remote sensing signals. And if we want to follow this approach, this trait-based approach to detect invasive plants, we need to have access to remotely sensed data with fine spectral resolution. I will talk about this more in the coming slides, but this trait-based approach is another way we can map invasive plants.

9:00 Speaker 1: Here is an example. You're looking at two plant communities. The one on the left is a native plant community. The one on the right is an invasive plant community. You can clearly see that visually these two communities are very different. The plant types are very different. The leaf angle, even the leaf colors are very different, which might indicate that one of these species might have higher pigment content compared to another plant community. So all these functional traits together affect how we see these plants remotely from above.

9:44 Speaker 1: Something really important is that I want to mention here is that the success of a phenology-based or trait-based approaches depends on the prevalence of invasive plants in the study area. What does it mean? Here is one. example. If you are dealing with large and homogeneous patches of invasive plants, like the one you see on the left side, it is easier to see this from above, and it is easier to map this from above using remote sense data. Whereas if you're dealing with small and sparse patches of invasive plants like the one you see on the right side, it will be more difficult to detect it from remotely sensed data.

10:29 Speaker 1: And as a result of that, when we are doing remote sensing studies of invasive species, it is ideal to have access to remotely sensed data with fine spatial resolution or fine grain size or pixel size. it will increase the success of our mapping efforts. Now that we have enough background knowledge on remote sensing of invasive plants, I want to talk about the main focus of this presentation. which is hyperspectral remote sensing. This is a specific type of remote sensing where we can collect remotely sensed data with fine or high spectral detail.

11:14 Speaker 1: Some people call this also imaging spectroscopy. They both refer to the same thing. So we can put our hyperspectral sensors on a platform, for example, a satellite or an aircraft or a drone, and then collect remotely sensed data with fine spectral resolution and then use these data to detect or map invasive plants. But before doing a deep dive on remote sensing of invasive plants using hyperspectral data, I want to take a minute or two to briefly talk about optical remote sensing of plants. So we have sun as the source of energy or electromagnetic radiation.

11:59 Speaker 1: This electromagnetic radiation strikes the target, in this case, our plant communities. Then some interactions will happen between electromagnetic radiation and the target, and a portion of this energy is reflected back. Then our sensors on satellites or aircraft or drones can collect this information. And then at the end, we will have a spectral signature like this, like the one you see on the right side. Some people call this reflectance. Some people call this the spectral profile or spectral curve.

12:33 Speaker 1: But what you see in this graph on the y-axis is something we call reflectance. This is the proportion of energy that is reflected back from the surface. And it is often reported in percentage from 0% to 100% or as a ratio between 0 to 1. And what you see on the x-axis is different regions of the spectrum, visible, near-infrared, short-term infrared. And something really interesting about this spectral profile is that it can give us a lot of useful information about plants. For example, it can give us information about water content of a plant, or its structure, or its pigment content.

13:17 Speaker 1: If we can build a sensor that can capture this information with high level of detail, meaning many narrow bands on the x-axis, we will have a hyperspectral instrument. So this is different from multispectral remote sensing, where we have only a limited number of spectral bands. Now, I'm going to present to you our case study where we used hyperspectral remote sensing to map an invasive plant in grassland ecosystems. The name of this plant is Lespiriza cuneata. Its scientific name is Lespiriza cuneata.

13:57 Speaker 1: It also has some common names, for example, sericea or Chinese bush clover. To give you an idea how this plant looks like, I'm going to show you a very short video that we shot a few years ago. And in this video, you can see how this plant looks like in real world. So here is one example, that tall, dense stand that you see. It is Lespedeza cuneata. And here is another one right in front of you at the bottom left of this slide. Another example of Lespedeza cuneata. So this invasive plant has become a major problem in several states in the US.

14:43 Speaker 1: The phrases you see on this slide, the top nitrogen plant, greedy grass, fiendish, these are phrases landowners use to describe this invasive plant. And in fact, in some cases, this invasive plant have become the land owners' biggest land management cost. They need to spend a lot of effort and energy to control or slow down the spread of this invasive plant. And as a result of that, a lot of land owners have expressed the need to have a cost-effective approach to map the presence of this invasive plant in grasslands ecosystems.

15:35 Speaker 1: So if you are interested to learn about the distribution of this invasive plant in the US, you can go to this website at the bottom of this slide, add maps and type the name of invasive plant or any invasive species that you are interested in, and you can make a simple map. In this map, what you're looking at is basically the distribution of this invasive plant, the counties where this invasive plant has been observed. In total, about 20 states have been affected by this invasive plant, and it is a major problem in several states like Kansas, Missouri, and Oklahoma.

16:15 Speaker 1: So like any other invasive species, this invasive plant also has some interesting background stories. This plant was brought to the U.S. in the 1890s as a cheap forage to feed cattle. It was also brought here to control for soil erosions. But the question you might be asking is that, what makes Lespedeza cuneata invasive? In order to answer this question, we need to go through the characteristics of this plant. This plant is a prolific seed producer. Some research has shown that each stem of this plant can produce up to 1,000 seeds, and they can stay active in soil for up to 20 years.

17:00 Speaker 1: It can also tolerate droughts, Laspirida cuneata also reproduces under poor soil conditions. For example, if you have acidic soil, this invasive species can probably survive. It is also a nitrogen-fixing legume, and it has, on average, higher nitrogen content than many native grassland species, which means that it is more nutritious. It is also taller than many native plants, and you can see an example of this at the bottom of the slide inside that white dashed rectangle. That is Lispedeza cuneata, and you can clearly see that it is much taller than native plants.

17:45 Speaker 1: It means that it can outcompete native plants for light and replace them. And probably one of the most interesting characteristics of Lispedeza cuneata is that it has high concentrations of phenolics. specifically condensed tannins. So tannins are chemical compounds. They are defense mechanisms. And grazing animals don't like condensed tannins. It makes them sick. So when Sericea or Lepidiza cuneata is mature, many grazing animals will avoid it because of high concentrations of condensed tannins.

18:24 Speaker 1: So when you look at the list of these characteristics, it makes sense why this plant was brought to the U.S., and why it is now a major problem. So a few years ago, we were really interested in using remote sensing to study these invasive plants. We wanted to use remote sensing to map Lespiriza cunyara across large spatial domains. But it was not as easy as we thought. We had a few challenges that we needed to think about. the biggest challenge we had, we call it spatial scale mismatch. What does it mean?

19:06 Speaker 1: If you look at, let's say, the Kuniara individuals, they are relatively small compared to the pixel size of commonly used remote descent data like Landsat. Even if we have... tall, large, and homogeneous patches of Laspiriza cuneata, many of them are still very difficult to map remotely, again, because the size of these patches are smaller than the pixel size of remotely sensed data like Landsat. For example, Landsat 8 and 9, they have spatial resolution or pixel size of 30 meters, which means that every pixel covers an area of 30 meters by 30 meters.

19:46 Speaker 1: And it indicates that detecting Laspiriza cuneata patches that are smaller than this size can be very challenging. So that was our first challenge. The other challenge we had was limited temporal resolution of remotely sensed data. So initially we were interested in using a phenology-based approach to map this invasive plant because after talking with landowners, we learned that this invasive plant usually greens up a few weeks before native plants and dies, senesces, a few weeks after native plants.

20:24 Speaker 1: But the problem was that this time window, the timing of this green up or senescence will vary a lot from year to year. And in order to find that optimal time window, we needed remotely sensed data with fine temporal resolution, but we didn't have that type of data.

20:47 Speaker 1: So how did we overcome these challenges? After thinking about these challenges, we realized that the easiest approach, the best approach, the most effective approach to map Las Pires de Cuñada is through collecting our own data from an airplane. So what we did, we used an aircraft and collected fine resolution remote defense data. In this case, the spatial resolution or pixel size of remote defense data was one meter. We also collected hyperspectral data, which means that we collected data with fine spectral resolution from an airplane.

21:28 Speaker 1: So at the end, we had data with fine spectral and spatial resolution, which is a great data set for mapping invasive plants. But what was the idea behind our project? So if you remember a few slides ago, I talked about functional traits and the fact that they can affect remote sensing signals and how they look like. So the key hypothesis behind this project was that Lispedeza cuneata has specific remotely observable functional traits that can be used to distinguish it from other native plants in the remotely fenced data.

22:09 Speaker 1: So if you notice, I use the term remotely observable functional traits. And the reason is that not all functional traits are remotely observable. We cannot estimate all traits remotely. We can, for example, estimate pigment content remotely, but estimating seed mass directly using remote defense data is not straightforward, if not impossible. So that's why I used remotely observable functional traits in this slide. So the approach we developed was a relatively straightforward approach. It had two steps.

22:48 Speaker 1: In the first step, we identified remotely observable functional traits that can distinguish Lespedeza cuneata from native plants. And in the second steps, we used these traits to develop a remote sensing approach to map Lespedeza cuneata. So two steps, relatively straightforward. So to address and achieve The first objective, the first step of our analysis, we went to the field, we went to a grassland and collected a lot of these samples. We harvested a lot of these samples and we sent them to the lab to quantify a long list of vegetation functional traits, the ones that you can see on the table.

23:41 Speaker 1: And then we did some statistical analysis to identify which of these traits help us distinguish Lespedeza cuneata from native species? And at the end, we identified six traits, total nitrogen, chlorophyll content, carotenoids, potassium, magnesium, and canopy hide. These are the traits that we thought are useful for distinguishing Lespedeza cuneata from native plants. Now, our second step. The second step of this analysis was a remote sensing aspect of the project. So something really interesting about remote sensing is that we can use remotely sensed data, remote sensing spectral information, to estimate some important functional traits.

24:36 Speaker 1: So for example, we used a data-driven model called PLSR, partial least squares, and estimated the six functional traits that we identified in step one. And the process is simple. In these data-driven models, the input variables are our spectral information, and our output is the plant functional trait that we are interested in. So in this slide, you're looking at nitrogen content. So we ran our model. We calibrated the model. We ran the model. And when we were happy with the performance of the model, we applied this model to every single pixel in our remotely sensed data.

25:18 Speaker 1: And the outcome of this process is the contiguous layer raster of planned functional traits. So you're looking at nitrogen content here. And here's another example. This is a raster layer of nitrogen content in part of our study area. Red pixels are pixels with high nitrogen content, and blue pixels are pixels with low nitrogen content. So we did the same thing for our other functional traits as well. And after doing all of this, we used the layers of functional traits to map, to detect Lespedeza cuneata presence.

26:06 Speaker 1: In order to do this, we use something called classification. So classification is a very common approach to use in remote defense data. And what happens in classification, we basically give the input data to the classifier, we train the classifier, and then the classifier labels the pixels for us. In this case, the labels were lespidiza cuneata present, lespidiza cuneata absent. and then repeated this classification hundreds of times, and we came up with a probability map, which shows the probability of Lespedeza cuneata presence.

26:43 Speaker 1: And you can see an example of this map on this slide. The red pixels are regions with high probability of Lespedeza cuneata presence, and blue pixels are areas with low probability of Lespedeza cuneata presence. Now, something really important in remote sensing, when we make a remote sensing product, we need to assess the accuracy of our product in order to tell the end user how reliable our product is. And we did the same thing in our case. We did accuracy assessment. In other words, we assessed the accuracy of our classification product.

27:28 Speaker 1: And for this accuracy assessment, we considered two scenarios. The first scenario was when Lespedeza cuneata formed large and homogeneous patches, like the one you see on the slide. In this case, our approach performed really well. The accuracy was 94%. But something really important here is that invasive plants do not necessarily happen at large patches. They can also form sparse and small patches mixed with native plants, especially during the early stages of establishment. So, in order to do this, we went to the field, we surveyed a large number of points, we documented the abundance of Lesbiriza cundinata, and then compared our results with remote sensing results.

28:23 Speaker 1: So, our analysis showed that there was a strong agreement between remote sensing results and field-based results. However, what we found was that our approach overestimated lespidiza cuneata presence in some cases, which is presumably because of the mismatch between the size of the plants and the spatial resolution of our remotely sensed data. So these results were really promising, were really exciting. But the important question is, what can we do with these remote sensing products? One of the most exciting applications of these remote sensing products is improving and enhancing decision-making activities.

29:16 Speaker 1: A great example is using remote sensing to identify and improve management practices to control the spread of invasive plants. Landowners use different management practices to control the spread of invasive plants. Some use fire, some landowners use grazing, some use herbicides, some use mowing. It is not clear how effective these approaches are, primarily because the success of these approaches have been assessed at the small scale. So what we can do, we can use remote sensing and map the spread of invasive species across the large spatial domains and then identify which of these management practices are more effective at controlling the spread of invasive plants.

30:06 Speaker 1: I want to take a minute or two to talk about data availability considerations. I think it is an important topic if we want to do remote sensing of invasive plants. So what we used in this project was hyperspectral data collected from an aircraft. But the problem is that airborne campaigns are very costly, and they also have limited spatial coverage, meaning that they do not cover large areas. And they also have coarse temporal resolution, meaning that we cannot repeat airborne campaigns several times during the growing season, primarily because of the cost.

30:50 Speaker 1: So the limited availability of airborne hyperspectral data is an issue. But the good news is that there are space-borne hyperspectral data. At the moment, we have a few operational space-borne hyperspectral imagers in orbit. For example, NASA's EMIT is one of these sensors which collect data with near global coverage, which is great. There is, however, one limitation, and that limitation is the coarse spatial resolution of these space-borne hyperspectral data. Usually the pixel size of these images is about 30 to 60 meters, which might be considered coarse for ecosystems like grasslands.

31:37 Speaker 1: And here's one example. On the top right, you can see our study area image using space-borne hyperspectral data. sensor, the resolution here is 30 meters. And on the bottom left, you can see the same area, but this time we collected data using an aircraft. In this case, the pixel size was one meter. And you can see how much detail we lose as we coarsen the spatial resolution. But the good news is that in remote sensing, there are some techniques we can use to resolve these limitations. A great technique is image fusion.

32:23 Speaker 1: What does it mean? So image fusion in remote sensing is the type of technique that can be used to combine data from different sensors. For example, in our case, we can combine multispectral data with fine spatial resolution, with hyperspectral data with coarse spatial resolution, and generate a new data set which has a fine spectral and spatial resolution. This is a topic we are working on and we have gotten some really interesting and promising results. And now I'm going to do a quick summary of what we learned and what we went through in this presentation.

33:11 Speaker 1: So in this talk, we discussed the unique role of remote sensing to map invasive plants. And we learned that unlike field-based techniques, which are costly, time-consuming, and limited in their spatial coverage, remote sensing can help us collect data across large spatial domains over time with consistent format. But of course, there are some limitations. One of the biggest limitations is the scale mismatch, meaning that the spatial resolution of our remotely sensed data often do not match the size of invasive plants in grassland ecosystems.

33:55 Speaker 1: Another limitation is data availability. For example, we use airborne hyperspectral data, but airborne hyperspectral data is not always available to research teams and research groups.

34:14 Speaker 1: But overall, the results were really promising. And what we can do, we can use all these exciting products from remote sensing of invasive plants and use them to enhance decision-making activities. One example that we discussed was using remote sensing to identify effective management practices to control the spread of invasive plants. So, I'm going to stop here. I wanted to thank you for listening, and now I handed this to Sativa.

34:56 Speaker 2: Thank you so much, Hamed, for your presentation. We are grateful for your time and expertise. and thank you to everyone who has joined us for all three parts of this series. To summarize, in this series, we covered a broad overview of remote sensing of invasive species. Part one was our introduction to remote sensing of invasives, and we were able to highlight NASA DEVELOP examples, as well as other examples in the literature. We also shared about available data and tools. Part two focused on the remote sensing of invasive aquatic plant species.

35:31 Speaker 2: And as Hamed just covered, we dived into a little bit more about invasive grassland species. And with that, I will provide the summary of our training. So overall, we introduced the extent and impacts of invasive species on biodiversity and a changing climate. We established that remote sensing can be used to map and monitor invasive species. We also identified different types of remote sensing data and products. We described the key considerations, benefits, and limitations of remote sensing of invasive species.

36:06 Speaker 2: And as I just mentioned, reviewed in greater detail remote sensing and field message used to monitor aquatic and grassland invasive plant species. As a final reminder, to receive credit for participation, please complete the homework assignment by September 11th. It will be posted to the training webpage today as a Google form.

36:28 Speaker 1: You can expect to.

36:29 Speaker 2: Receive your certificate about two months after this course. If you have any additional questions following our time together, please feel free to reach out. Our contact information as well as links to our website, social media, YouTube channel, and sister programs can be found here. We will now move into our Q & A portion. So thank you so much again for being here, and we'll go ahead and start going through the questions. So let's get started. The first question we had was, how does hyperspectral remote sensing differentiate between native and invasive species in grassland ecosystems?

37:20 Speaker 2: So effective mapping of invasive plants depends on several factors. For example, contrasting seasonal phenology of invasive plants compared to native plants, or distinct biochemical, physiological, and structural characteristics or functional traits of invasive plants versus native plants. The strength of hyperspectral data is their fine spectral resolution. So that means that these data can be distinguished by minute spectral differences in native and invasive species. So if we have invasive species with different functional traits compared to those of native species, we can detect them using suitable hyperspectral data.

38:04 Speaker 2: Eventually, successful application of remote sensing data regardless of data type depends on the spatial resolution of remotely sensed data compared to the size of the phenomenon we are observing. This issue was also discussed in part two of the webinar. please also see the answer to question seven, which we will get to in just a bit. All right, so moving on to question two. Someone asked, can hyperspectral remote sensing detect early stage invasions before they become widespread?

38:41 Speaker 1: If so, how?

38:44 Speaker 2: Well, this primarily depends on whether the spatial resolution, so also the grain size or the pixel size, matches the phenomenon we're observing. If we're dealing with large invasive plants compared to the scale of remote sensing observation, then the likelihood of successful mapping is gonna be higher. If we're dealing with small individuals, similar to those in grasslands, then we will need fine resolution hyperspectral data for early detection. To do early stage detection, there are many options.

39:17 Speaker 2: One is classification, which was also covered in part two, or you can use site-based approaches similar to what was discussed in this webinar. Okay, question three. I have a question referring to the previous part, part two, specifically about the spectral angle mapper technique. I wonder how to assess that the angle between the curve in a pixel and the reference curve defines that the mapped material or substance is present in a given pixel. Is there an optimal way to determine the cutoff threshold for the angle value?

40:01 Speaker 2: So the threshold is user-defined. There is a paper going to be linked in this document by Dennison, Halligan, and Roberts. It's an early paper. And it will provide a nice overview of the error metrics and constraints for the spectral angle mapper and spectral mixture analysis techniques that were introduced in part two. Optimizing these error metrics is how most people identify the optimal threshold. So refer back to that paper.

40:35 Speaker 2: Question four, can we detect invasive plants using Sentinel-2 images? Yes, potentially. So this depends on the invasive in question, primarily its prevalence, its size, and phenology. Refer back to the questions one and two that we had just covered. Question five. How does one create a spectral graph per pixel from a multispectral-grown image? Well, a drone-based multispectral data is similar to those we get from Landsat, Sentinel-2, etc. After processing the data, so that is the geometric, radiometric correction, and atmospheric correction, we can extract spectral graphs per pixel from drone data using packages such as Python, R, MATLAB, or software packages like ENVI or ERDAS.

41:40 Speaker 2: Question six, is there a global equivalent for EMAS? Yes, so there are. One example is the Global Biodiversity Information Facility, and that link will also be made available in this document. Question seven. Fulgongrass Imperata cylindrical is an invasive that grows along highways, radiating away from port facilities. It grows in right-of-ways, then adjacent to fields. How big of a patch is necessary to appear in a remote sensing image? By the time it is large enough to be found on an image, it is difficult to eradicate.

42:36 Speaker 2: Can remote sensing be used here? So the minimum mapping unit, or MMU, should ideally dictate the appropriate spatial resolution for a given application. The MMU represents the smallest feature size that can be detected on a map. Features smaller than the MMU will not be represented. The specific MMU will vary depending on the application of management objectives. Once it's established, the spatial resolution of remote sensing imagery should ideally be finer than the MMU, although explicit guidelines are often lacking in the literature.

43:17 Speaker 2: In practice, however, the choice of spatial resolution is frequently influenced by the availability and cost of data, as well as trade-offs with temporal and spectral resolution. With coarser resolution data, the mixed pixel problem is an issue. Techniques like soft or fuzzy classification methods are commonly employed to handle this issue, where each pixel is assigned varying degrees of membership across multiple classes. A prominent approach is something called spectral mixture analysis, which models the spectra of each pixel as a linear combination of pure spectra from different landscape components.

44:03 Speaker 2: So end members like their soil, water, vegetation. This allows for the mapping of each pixel as a fraction of different end members. This topic was explored in detail in part two. Next question, how would one create a spectral signature for a specific plant from drone multispectral imagery? So after pre-processing the data, we can extract signatures for specific plants from drone data using packages such as Python R or MATLAB, as well as software packages like Envy or ERDAS. This was also addressed in question five.

44:51 Speaker 2: Question nine. How effectively can the maximum entropy algorithm be applied to this use case. So that max entropy is a species distribution model. Our focus of this talk was on direct mapping of invasive plants, but we can use maximum entropy and other species distribution models to model the spatial distribution of invasives to develop an SDM we will need additional environmental variables. Question 10, what hyperspectral imager did you use? So in this study, they collected airborne hyperspectral data, AISI, Senex, one kilometer, Spesim, Olut in Finland.

45:59 Speaker 2: Yeah, that information's there. The data covered the 400 to 2,450 nanometer range in 323 spectral bands. And it had a road solution of approximately 4.5 nanometers in the 400 to 970 nanometer range and 14 nanometers in the 900 170 to 2,450 nanometer range. The next question is, how do you assess accuracy? So, we can use measures called classification accuracy assessment as discussed in slide 35. The most reliable approach to conduct accuracy assessment is using field-based data and then comparing them to remote sensing classification results.

47:12 Speaker 2: Question 12, which model is used to determine nitrogen content in El Aniata? What were the modeled inputs and function for determining the nitrogen content in the field? An example of such a case would be helpful. So we use a data, researchers used a data-driven model called partial least squares. Independent variables are spectral reflectance data and dependent variables are foliar nitrogen content. And we quantified from foliage samples that they collected in the field. Next question, is IMET open source?

48:02 Speaker 2: How do we access the data? Does it need Python code, et cetera? So yes, IMET is open source. And you can check out the following links to get more information about IMET. So please refer back to this Q & A page once it is posted. Question 14, I guess the functional traits vary temporally over the growing season. Did you take that into account when mapping these traits? How does this affect the classification? So yes, this is a great point. They do vary in time. To minimize the impact of temporal variability of traits, The team collected remotely sensed data and field data as close as possible and within a few days.

49:02 Speaker 2: But the temporal mismatch between remotely sensed data and field data is a source of uncertainty. Next question, could you explain a little bit about the technique of image fusion to improve resolution? Is it pan-sharpening in another language? So there are many methods for image infusion. In this case, they use a method called hyperspectral image super-resolution, or HI-SHER. HiSure is a subcase method that extracts end member information from the coarse spatial resolution hyperspectral data and estimates the abundance fractions of the end members from the fine spatial resolution multispectral data.

49:55 Speaker 2: This was conducted in MATLAB. Next question. What are the main challenges in applying hyperspectral remote sensing for invasive species monitoring in grasslands. Yeah, so some of the challenges are data availability, pre-processing that's needed, particularly atmospheric correction, and the large data volume. It's also preferred to collect failed observations to develop models and validate them. So another challenge is collecting field data in remote locations.

50:46 Speaker 1: All right.

50:47 Speaker 2: We're gonna go to the next question. Yeah, thank you all. We have quite a few here coming in, so we appreciate your patience as we work as a team to get these sort of drafted. Okay, so question 17. I would like to ask what the process of resampling data up to three meters look like. Was it traditional, like bilinear, or perhaps another algorithm or machine learning method that was used?

51:28 Speaker 2: So this is a great question. They fused data from DESIS with those from PLANET using image fusion. Refer back to question 15 for more information. Question 18. Is there a repository where we can be able to access the indices for common invasive species or one has to craft them individually or refer to available journals? So there are open source repositories for invasive species such as EDMAP and I believe this is GBIS, which was referenced earlier in the Q & A document.

52:21 Speaker 1: Wonderful.

52:22 Speaker 2: Okay, so we have a few more questions here and we'll see if we can maybe get some response from our experts on the line if they are available to answer with their voice. If not, we will be sure to spend some time filling out the rest of the responses for these questions.

52:52 Speaker 1: So I will. Can you hear me?

52:56 Speaker 2: I can hear you just fine. Hamed, how are you?

52:58 Speaker 1: I'm doing great. Hello, everyone. This is Hamed. So, I'm going to try to answer the questions through this chat. So, how does the team see the future of invasive species mapping with the launch of Landsat NEXT? That's a great question. Landsat NEXT will be an improved version of Landsat 8 and 9. It will have spatial resolution of 10 to 20 meters and also finer spectral resolution. So we expect that at the launch of Lancet next, we will have more capability to map invasive species. Of course, as was discussed earlier, it will depend on the invasive species in question.

53:43 Speaker 1: Primarily, the difference between the pixel size or spatial resolution of our remote distance data and phenomena we are observing on the ground. Question 21, can you please explain more on how to find traits that can distinguish my particular weed from the rest of the plant? Is it possible to have a singular trait with which can distinguish multiple weeds that are more EPS? How can I find those traits? It's another great question. So when we started this project a few years ago, We started with reading about this invasive plant, so let's see if we can get a learning about it and finding out what makes this invasive plant successful.

54:29 Speaker 1: Specifically, what are the specific characteristics that help this invasive plant spread in grassland ecosystems? And we found that some of the traits that help this invasive plant spread is high nitrogen content compared to native plants. being taller than a native plant, also being allotatic, which can release specific chemicals onto soil. They also have high tannin content compared to native plants. Now, the thing is that the list of these traits are really long. Some of these traits are not remotely observable.

55:07 Speaker 1: We can only estimate some of these traits from remote defense data. So we focused on those traits that we were able to remotely estimate based on previous literature. And then we collected 40 samples in the field and sent them to the lab and developed our model based on these traits. So to answer the question, it depends on the invasive plant that you have in hand and what traits distinguish this invasive plant from native plants. And if you can identify those traits, then there is a chance that we can estimate some of those traits using remote consent data.

55:49 Speaker 1: Can we combine multiple traits for distinguishing multiple weeds? Yes, it's another great question. So what we did in our project, for example, we used several traits. We selected a long list of traits, 10 or 12 of them, and eventually we selected six of them. So we collectively used these six traits distinguish our invasive plants from native plants. So you can use multiple traits to distinguish invasive plants from native plants. And if you're lucky, if you're dealing with invasive plants or invasive weeds that have similar characteristics that distinguish them from native plants, then yes, you can use the combination of those traits to distinguish those invasive weeds from native plants.

56:51 Speaker 1: All right. Question 23. Would machine learning with accuracy tuning using proxy data sets and those from indices such as NDVI, WTOI, be possible to enhance the effectiveness of invasive plant detection? Yes, this is another great question. As a follow-up to our project where we use hyper-spec or remotely sensed data, We developed other projects, and specifically one goal we had was how can we develop more affordable, cost-effective approaches to detect invasive plants. So we use hyperspecial data in our project, in the project that we just heard about.

57:33 Speaker 1: But hyperspecial data, especially if you're collecting them from an aircraft or a drone, are very expensive, and they are expensive. So our goal was, let's develop something easier, cheaper, and we decided to use multispectral data. Specifically, we used PlanetScope, which provides data with spatial resolution of three meters, but these are multispectral data. They have a limited number of bands. So what we did instead of using traits, We use vegetation indices that are good proxies of these traits that we identified based on our field data.

58:17 Speaker 1: So we know, for example, this trait, this vegetation index is a good proxy of nitrogen. We know that the other vegetation index is a good proxy of, let's say, chlorophyll. So we selected those vegetation indices and then used them to develop our models to map invasive plants. So the answer to your question is yes, there is potential there, especially if you do not have access to hyperspectral data. We can use multispectral data and use vegetation indices to map invasive plants. However, the choice, there should be some reasoning, some justification behind using those vegetation indices.

59:02 Speaker 1: Can you explain more about the data fusion process, planet-basis image fusion, or resources where we can learn more? Yes. We answered this question specifically in question 15. So if you can, take a look at our answer to question number 15. Hopefully, it will provide some answers for you. And there are many image fusion techniques. There are several of them. And what we used was something called hyperspectral image super-resolution. And the input for these type of fusion techniques is often a high-resolution or fine-resolution data.

59:42 Speaker 1: In other words, remotely sensed data with fine spatial resolution. And another remotely sensed data which has coarser spatial resolution. In this case, space-borne hyperspectral data. And then you need to do some tuning. These approaches require some tuning. And then after that, you will get your FUSE data. And then finally, you need to do quality assessment to see if your FUSE data is reliable and satisfactory. But we can, I will make sure to add a few papers and references for you.

1:00:26 Speaker 1: Question 25. If using ground imagery flown at 30 meters, is it still necessary to do pre-processing for the atmospheric correction, et cetera? Yes. It also depends on what type of sensor you're putting on your drone. If you are putting a hyperspectral sensor on your drone, then you will definitely need to do radiometric and X-ray correction, and of course, geometric correction. If you're using hyperspectral data, multi-spectral sensor on your drone, also yes. Often how we do this, we put a target, calibration target on the ground, we've known reflectance or we measure its reflectance in the field and afterwards we will calibrate our remote sense data and convert into reflectance.

1:01:19 Speaker 1: In other words, we will try to minimize the impact of atmospheric effects on our data. That being said, sometimes people use multispectral data from drones without pre-processing, but personally I don't recommend doing that. Question 26, what part of the presentation was the ecological events that helped distinguish between species covered? So in this talk, we focused on trait-based mapping. We focused at one point in time. In other words, we didn't use phenology to do our mapping and the reason for it, there were several reasons for that.

1:02:14 Speaker 1: The first reason was that we didn't have access to remotely sensed data with fine temporal and spatial resolution. So we went with data with fine spectral and spatial resolution. But that being said, you can definitely use remotely sensitive fine-temp resolution to map invasive plants. And that's one of the projects that we have been working on, and we have published a few papers on that, where we use multispectral data from planets to see where we can, when we can map invasive species during their growing season.

1:02:56 Speaker 1: But in this talk, in this specific presentation, we focus on one point in time using hyperspectral data. Are there any three hyperspectral data sources I assume my colleague, Dr. Aaron Hester, answered this question. I really appreciate that. Yes, there are many free hyperspectral data sources. There is, of course, MS. It is a space-borne hyperspectral data. It provides data from 400 to 2,500 nanometers with a spatial resolution of 60 meters. We also have DASIS data, which is a sensor mounted on International Space Station.

1:03:37 Speaker 1: It also has a spatial resolution of 30 meters, but it doesn't provide, it provides limited hyperspectral data compared to EMIS. There is also AVERIS and AVERIS-NG. These are NASA sensors. These are specifically airborne hyperspectral data. And you can, you know, there are data portals where you can freely download AVERIS and AVERIS-NG airborne data. So there are several options. You have space-borne hyperspecial data, like EMID. Also, PACE will be another good option, but they often have a coarse spatial resolution.

1:04:20 Speaker 1: There are also airborne hyperspecial data that you can use, like AVERIS or AVERIS-NG. And both of them are really applicable. Are there, question 28, are there any computational imaging techniques that allow us to get data from multi-spec of drones that would be competitive enough with hyperspecial imagery to infect invasive species? I am not sure if I fully understand this question. So, Yeah, if we have a multispectral sensor on a drone, the advantage we have is that the data will be, the volume of data would be way smaller compared to hyperspectral data.

1:05:11 Speaker 1: Also, with multispectral drones, you can fly at low altitudes and collect data with really fine spatial resolution or pixel size or grain size. And as we discussed today, if you have data with really fine spatial resolution, then the odds, the likelihood of detecting invasive plants will be much higher. But in terms of a spectral resolution, multispectral centers have much fewer spectral bands and therefore coarser spectral resolution compared to hyperspectral data. But to answer your question in a sentence, yes, we can use multi-specular data from drones to map and detect invasive species because they provide data with really fine spatial resolution, which would help us to map invasive species more reliably.

1:06:10 Speaker 1: With question 29, with, for example, 320 bands, how do we go about choosing the correct one? What composites are useful? So this is a great question. So 320 bands refers to the number of spectral bands that we have in a microspectral data set. And the question is asking, are we selecting the specific bands, and if so, how? No, we are using all spectral bands. The only exception is we are removing water vapor absorption bands. These are the bands that are affected by atmospheric water vapor, we remove them because they are often noisy and not reliable.

1:06:53 Speaker 1: But after removing these unreliable or bad bands, we use all these bands in our data-driven model, in our case, PLSR. So what we did, we pre-processed our remote distance data, we removed unnecessary bands, the ones that have been impacted by water vapor, And then afterwards, we use PLSR. We use spectral bands or reflectance as input as independent variables and our traits that we measure in the field as our dependent variables and develop our models. I just want to mention that PLSR is one example of a data-driven model.

1:07:35 Speaker 1: There are many other machine learning approaches that you can use instead of PLSR to do what we did. But to answer this question, we use all the bands, except those bands that have been affected by the atmosphere. Question 30, how is the CCD gamma handled here? Is it differentiated over different oceans and land, or simple spectrography? I don't have a clear answer to this question, unfortunately. Question 31, do hyperspectral images help with DNA identification also? What are the best methods and instruments for biochemical sampling?

1:08:33 Speaker 1: How can I get chronology for classifying open-ended species? Yes, hyperspectral data potentially can help us with mapping different species. Again, it depends on several factors that we discussed earlier during this Q & A. And the methodology for classifying elk or any other species, we have provided details on our methodology.

1:09:01 Speaker 1: in the papers that we have published, and I will make sure to provide the links for these papers after the talk. Question number 32, what are the most effective spectral bands or indices for differentiating ISPPs from native grassland species? We did another great question. We partially covered it in question number 29, I think. So all the spectral bands can potentially help us to differentiate invasive plants from native plants. But it also depends if we are, for example, following an approach similar to ours based on trait-based methoding.

1:09:51 Speaker 1: Specific traits affect specific spectral bands. So although we use all the spectral bands in our analysis, there are specific bands that are more important than others in identifying or mapping invasive plants and distinguishing them from native plants. Question 33, did you use points or polygons for accuracy assessment? What are your thoughts on the controversy of using overall accuracy metrics? It's another good question. We basically used two different approaches to do classification accuracy assessment.

1:10:33 Speaker 1: The first scenario was when an invasive plant called Lespinida cunata is the common species. In other words, when it forms homogeneous and large patches. And in order to do that, basically use the high-accuracy GPS in the field and map where this invasive plant is. Basically, we map the polygon and we overlay that with our remote defense data and data and accuracy assessment. But then, this is not the case all the time in the real world. When you are dealing with invasive plants, there are cases where this invasive plant can form sparse and small patches of small patches.

1:11:17 Speaker 1: And to address this, to test the accuracy of our model in situations like this, we also visited a large number of data points, several hundreds.

1:11:29 Speaker 3: Of small plots, 10 meter by 10 meter plots, and documented the abundance of invasive species in our study site.

1:11:43 Speaker 1: And then we compared the estimated abundance in the field with the estimated abundance from our remotely sensed data to see how reliable our models are. Question 34. Can you explain the type of challenges that arise when detecting invasive species in their eugenics graph environments using hyperspectral data? We partly covered this question earlier, but the main challenges from a remote sensing perspective, if you are using hyperspectral data, is data availability, pre-processing, which is extremely important when you're using hyperspectral data, and also the volume of the data.

1:12:27 Speaker 1: If you're using hyperspectral data, especially fine resolution hyperspectral data, the data volume is pretty high. Also, in the field, we have a lot of challenges. It's often remote places. We need to collect data in remote places. If you want to measure specific traits like chlorophyll, for example, you need to flash freeze your samples in the field. So you need to have liquid nitrogen, dry ice, and all of that in the field with you. So there are many, many challenges there. So it's been quite a learning process for all of us doing projects like this.

1:13:08 Speaker 1: Question 35, do we require image fusion if you have multi-spectral sensor drone imagery if the output scale is due to medium resolution Sentinel data? If you have fine resolution multi-spectral data and if you have map invasive plans, then you can directly use your drone data. Probably you don't need to fuse it with Sentinel data. So multi-spectral sensor-bound drones usually have several spectral bands. So they have seven, eight, nine, 11, 12, many spectral bands. So I would recommend using drone data directly without using it with Sentinel data.

1:13:55 Speaker 1: if you want to use it for mapping invasive plants. The reason we've used planets and DASIS in our project was that DASIS has fine spectral resolution. It is hyperspectral, but its spatial resolution is coarse. It is about 30 meters, which is really coarse for our work mapping invasive plants in grasslands. And as a result of that, we decided to use this data, this data with PlanetScope data, which have finer spectral resolution, spatial resolution. And at the end, we had to use data that defines spectral and spatial resolution.

1:14:38 Speaker 1: I hope my answer made sense. Thank you.

1:14:43 Speaker 2: So much, Ahmed, for taking the time to add additional questions. Information there, your expertise is much appreciated. And everybody else working behind the scenes, thank you all for making this webinar possible. And thank you, participants, for being here and your interest in the topic. As mentioned, homework is due on September 11th. Please get that in. And we will post this Q & A in about a week. Thank you so, so much.

1:15:12 Speaker 1: Thank you very much.

1:15:13 Speaker 2: We hope you all have a fabulous day.