we are dealing with the grassland recognition with usage of thermal weights. What are thermal weights and how we define them is going to be shown in a few minutes. So as I know, I think I'm the last presenter of this block before lunch, so hopefully you are not so hungry and you can enjoy the next few minutes without so much pain.

So what is the con? Yeah, I apologize if there is some broken parts of the presentation, but as you know, it's like this, Murphy never sleeps. OK, what is the concept of our work? Is that there are different land usage, land cover

with similar vegetation indexes. So when we talk about recognition of land cover and land usage based on just NDVI and similar types, we get to meet some problematic areas, which are in this example, grasslands and grasslands alike crops. So time series can lead to these answers

without so much problems, but they also are prone to hide those answer if you use only optical data, because when there is a plugging process that happens, if it happens in a period of, for example, spring, which has a lot of clouds, we can easily

have our solutions and our results hidden from those. And those can be quite easily also seen by combining other data. For example, a great approach is to combine radar data, which is cloud penetrating, but it's all processing is quite painful and time demanding.

And sometimes we can just skip some steps and use other data to resolve our problems. So our hypothesis was based on the fact that the thermal imagery can give us a better insight of what's happening on the land cover. And we said that the usage of the thermal imagery

with visible infrared bands can improve the accuracy of our classification. So in this case, we were talking about a problem that was seen in other projects that dealt with classification of land cover and land use that saw that grasslands as seen as a simple field

are not simple at all. And we tried to combine the thermal infrared data that is derived from Landsat 7 and 8 to the other mentioned data that is open and distributed by ESA, by the Sentinel-2 platform.

So this is a diagram of all the processes. So to make a long story short, here is the pre-processing. We pre-process the Sentinel data, the Landsat data. We fill the gaps if we talk about the Sentinel-7 derived Landsat thermal images.

And then we calculate the NDVI and from it the proportional vegetation. I will show you later in the formulas what does it really mean so you can see these steps in a more clear way. Then we calculated the emissivity, which is a parameter for calculating the land surface

temperature from the NDVI derived from Sentinel, mixed it with the thermal data from Landsat, then evaluated everything to see if it's worthy or not, and then made the time series analysis from that. So in general, our hypothesis was that the NDVI calculated from Landsat and Sentinel-2.

If you talk about a small gap, it should not be really different. So the NDVI of some plants today is pretty similar as the NDVI of a plant tomorrow or two days after. So these are the formulas that we used to calculate.

This is the classic formula to transform the temperature to the top of atmosphere. So these are the constants, which are the thermal constants. They are contained in the Landsat-8 metadata. If we talk about Landsat-7, we can calculate them from the wavelength central plants.

In QG, there is a really nice plug-in that can convert them directly. It's the semi-automatic plug-in that I think each one that deals with thermal sensing knows. And the CNC radiation constants are just constants in this other formula from the Planck flow.

And in this case, before, as shown, we were calculating the emissivity from the NDVI of Sentinel. We calculated it because the emissivity is a parameter that is contained here. It's under a logarithmic function to calculate the land surface temperature. So after converting it in the first step,

we just recalculated the land surface temperature with Sentinel data. So this emissivity was calculated by a threshold method that was presented for Landsat by Sobrino and others. In 2008, one of the conclusions in the end

will be also that we could also develop a better threshold method for that. But this one, for now, seems to be really cool. This is the proportion of vegetation that we calculated from the minimal and maximal NDVI. This was used to calculate the second threshold values. So we reclassified in GRASS all the values of the NDVI

with the values of the emissivity calculated like that. This is the red channel. Then we continued with the other ones. This is the data that we used. So this is the area. It's located in Slovenia in the Styrian region. This is the Sentinel-2, Sentinel-8 data

that we used to calculate. In this scenario, we used just the eight data because the Sentinel-7 data, as some know and use, have that SLC gap problem. But if you are willing to work on it, I really recommend you to use the Landsat-7 data because its thermal bands is based on a 70-meter raster that

is resampled to 30 meters. And the Landsat-8 is 100 meters. So we talk about quite a big difference. And these are the time gaps within days. So it means that between this and these days, we have two times gaps. This is why also we hypothesized that those NDVI are practically similar, which we also checked.

This is the NDVI for this region from the Sentinel-2 area. As you see, those are the two histograms. The first one is made from the Sentinel-2 data. We can see two nice peaks. One is the peak from the land cover from agricultural crops.

The other is water and other. So if you want to continue with the reclassification, we need to divide it at least in two parts. To simulate if it's really cool, we checked the kappa validation error within GRASS. And we used the Sentinel and the Landsat images

as classification results. And the results were pretty good. To also evaluate if kappa was wrong or not, we made the correlation coefficient. And as our hypothesis was that the NDVI are not so different,

it was true. So these are the reclassified values from the NDVI to the emissions. If we see, there are quite differences. This is the Landsat-8, and this is the Landsat Sentinel-2. For example, in some parts, like the parts under the circles, we see that there is a big loss of data in the Landsat-8 derived emission

This is, of course, prone to the resolution of the different resolutions. But after, when we combine it with the thermal data, the difference is quite clear. So this is the land surface temperature that is derived from it.

So from the originally 100 meter re-sampled thermal band from Sentinel-8. And after that, we developed the thermal weights, which for us just means that we used the thermal factors to weight our results. So in this case, we were talking about time series

that deal with grasslands and grasslands alike crop fields diversification. And we saw that the possible solution was to calculate raster representing the range of the values in this time series, and the raster represented the maximum land surface emissivity

values. And how we did it? Because as we saw before, the values for the emissivity are really, really near and really narrow. We just normalized those time series on a scale from 1 to 100 and then made their sum.

The emissivity values, in general, already in the first step filtered the majority of the water bodies, the woods, and the urban areas, which left us raster representing other crops and significantly improved their diversification in the further step.

So why we did it? We did it because the land crops, in general, have bigger junior temperature variation. When you calculate the land surface temperature, the difference between the plant and the soil is not so big. But a classical grassland has not such a big rate of growth.

So that difference remains constant through time. An agricultural field changes a lot. And especially in the plug-in time or in the biggest vegetation time, the differences are quite clear. So this is one of the examples of those sums.

So this is the sum of two normalized raster values by 200. As we see, all the values practically that are in this peak in the last quantile are croplands, and the others are other.

We were calculating also the correlation between those by checking the in-situ data with the experiments rank coefficient. And we saw that there is a strong negative correlation

between those for a big number of polygons, which means that the increase of this weight represents the decrease of possibility that we talk about an agricultural crop, about a grassland. So the grasslands are practically in this part here.

So what we concluded in the end, because we wanted to see if the thermal impact can be used as a variable for assessing that. So the answer is yes. There is a string of direction of association between those variables, and we can use it to differentiate. We made a quite simple usage of them.

So any major and further development in some robust statistical methods will for sure improve. A big improvement would be also to re-tweak the NDVI threshold method. Probably we can find some inspirations in some tesselled-cup formulas that deal with the band that

are within the sentinel to domain. And in general, when we deal with photo interpretation of some areas which are not clear, at one point, we can try to find a solution within this approach.

So our further steps will be dedicating our time to build a robust statistical model for evaluating this, and sooner or later, a development of a plugin that lets us just to insert the adequate bands and get some results without too much effort.

So this was all from me. And if anybody has some questions, please shoot. Do we have any questions? Then, so did you think about to use any other method

to compare your results as a kind of a verification about your precision of the grassland detection? So we had the in-situ data from the state, and the comparison was made by creating statistical samples

of those. So one was the kappa value, one was the correlation between them to see how much those rasters danced between themselves. And after that, we used the Spareman's rank of correlation that we built within our script, within QG's. I see. Thank you. All right.

Then, thanks for your presentation. And thanks again for all the presenters in this session. They were really interesting. And we also managed to work down almost all of our delay,

so we are really good. And as far as I know, we will have a lunch break now, and the next track will continue at 2 o'clock.