Using Sentinel 2 data for spatiotemporal mosaic of bare soils creation and mapping soil properties using machine learning
This is a modal window.
The media could not be loaded, either because the server or network failed or because the format is not supported.
Formal Metadata
Title |
| |
Title of Series | ||
Number of Parts | 24 | |
Author | ||
License | CC Attribution 3.0 Germany: You are free to use, adapt and copy, distribute and transmit the work or content in adapted or unchanged form for any legal purpose as long as the work is attributed to the author in the manner specified by the author or licensor. | |
Identifiers | 10.5446/59785 (DOI) | |
Publisher | ||
Release Date | ||
Language |
Content Metadata
Subject Area | |
Genre |
7
8
9
12
24
00:00
Computer animation
00:31
Computer animation
01:01
Computer animation
01:30
Computer animationProgram flowchart
02:12
Computer animation
04:19
Computer animation
07:29
Computer animation
09:40
Computer animation
10:15
Computer animation
10:49
Computer animation
11:53
Computer animation
12:08
Computer animation
13:19
Computer animation
14:27
Computer animationProgram flowchart
Transcript: English(auto-generated)
00:11
Okay, good afternoon ladies and gentlemen. I am Daniel Gijela and I will continue the soil issue and issue on mapping soil properties, specifically using Sentinel-2 data and using MOSAIC.
00:30
Me and my colleague, we are from Research Institute for Soil and Water Conservation from the Czech Republic, so I will talk about mapping soil properties only for the Czech Republic.
00:46
And in my presentation, I will talk about talk about BERSO composite creation, and then briefly about using this this composite for soil properties prediction.
01:03
So let's introduce a little bit. Traditionally, some maps are created in forms of analog maps or digital maps in form of polygons, which need
01:21
a lot of effort of sharing work and digitizing and so on, and these maps are a lot of shortcomings. On the other hand, nowadays, we have a lot of possibilities to use some new technologies, some new data, for example, remote sensing data or other spectral data.
01:44
Some new statistical methods, and we can use it for soil properties mapping, and so called digital soil mapping, when we use, when we use
02:02
point data with the ground truth information about soil properties, and a lot of covariates and to make a model for final prediction of map. So, how remote sensing data can help us.
02:22
Previously, the use of spectral data as proven is very useful for mapping soil properties, especially spectral active properties like organic carbon, clay content, carbonates or iron oxides.
02:44
The spectral data has been confirmed, the utilization has been confirmed, mainly on laboratory or airborne hyperspectral data. But, although it has been shown that also satellite multispectral data can be also used in anyways, but there are some,
03:10
there are some challenges, as Tom said, we can, we can use remote sensing data for mapping surface layer of soil.
03:22
And individual image, satellite image, we can use only for local mapping. Why? Because on every single image, there are a lot of crops on the fields.
03:45
So, as we can use the soil spectral information only from some field, we can, we can use individual image only for some fields, because the vegetation crops and crop residues mask the spectral information of the soil.
04:06
So, how to overcome this shortcoming, the possibilities is to use time series mosaic ink, as we can, we can call it bare soil, bare soil composite.
04:21
In bare soil, in creation of bare soil composite, we need to process all time series of satellite images, mask cloud. For this, we use, we use combination of S2 cloudless algorithm and synclassification layer from Sentinel-2 data.
04:48
And the other challenge is, is how to, how to mask crops and crop residues. And so, for this, we, we use some, some structural ink methods.
05:05
So, in this approach, we process all time series of data, we, we mask clouds, and then we, we mask bare soil, or we identify bare soil using, using two thresholds of spectral indices, specifically green, green vegetation.
05:32
We, we identify by threshold of the normalized difference vegetation index, and for
05:41
crop residues and sensitive vegetation, we use NBR2 index, normal bloom burn ratio index. For final creation of, of composites, we distinguish bare soil and permanent crops.
06:03
For each pixel, that this market more than five times as bare soil, we use average absorption reflectance from occurrence of bare soil. And if these pixels are identified less than five, we said that it's a permanent,
06:26
permanent vegetation, and we can compete with, with reflectance at the time of the highest NDVI. So, the results is, is bare soil composite or are of interest in this case for,
06:43
for Czech, for all Czech Republic. And here I can show you some, some detailed examples. For example, here is the example, the, the image, which show the contrast between some unique chernosamic soils and some ruby, ruby soils in eastern part of Bohemia.
07:06
On the next, the next image, you can see highly eroded, eroded soil in the south of Moravia, and the third example, some red-coloured can be solved on, on purple carbons in Potkal-Ganarsi Arabian, and it's, it's north, northeast of Bohemia.
07:30
So now, let's do the prediction mapping. Yeah, we use this, this bare soil composite as one input to do, to do prediction mapping.
07:44
And as, as input data, we also use some, some other covariates, like climate covariates, some legacy, legacy soil maps, and, for example, also coordinates, like buffer, buffer distance map.
08:06
We, we use, we used for, for the prediction mapping quantum random forest method. So we can, we can be able to, to produce some final map of prediction and also the answer uncertainties of prediction and every pixels.
08:27
So now it's about the input soil data. We use two, two basic data sets, and as Tom said, we have some dynamic soil properties and some not so dynamic.
08:44
So for, for, for these properties, which you can say that it's not so dynamic, we use the data from systematic soil survey from, from 60s. This, this database consists about the 300s, thousands of, of, of, of probes, of information from 300,000 probes.
09:15
And for, for more dynamic soil properties like organic carbon or pH, we, we use up to date database.
09:27
For, for, in this database, we have less data, but for example, for organic carbon, about, about 8,000 points.
09:41
So here you can see, you can see some, some metrics of, of prediction ability of models. And yeah, with some, with some exception, we reach a good, good results. And we predicted, predicted final maps and the maps of prediction confidence interval on
10:06
which we can, we can assess the, the uncertainty on, on, on each point. So this, this maps are in resolution 20 meters per pixel and are available for the Czech Republic.
10:25
And if you want to, you can see the results on our geo-portal on this, on this address. And some details about what methodology you can find in this paper in Caffeine Azure now, or, or our methodology, but, but sorry, it's all in Czech.
10:48
So yeah, it's, it's all from me. Thank you for attention. And as you can see, even though we, we use modern, modern methods, the method of collection of data is the, almost the same throughout the history.
11:05
Thank you. Thank you very much for interesting presentation. And please questions. Did you do a 3D model?
11:24
No, it's a, it's a separate models, but the four different, the different depth, we model three different depths, zero to 30, 30 to 60 and 60 to 100 centimeters. Yeah, but then you map also the ports.
11:44
No, no, no, we, we know agriculture. So, okay, you can try mask out. Yeah. I'm good.
12:01
Can you see the top for leads? Because you see, I see only one aspect. Yes. There is only some results. Yeah. You can see the results for, for two days. Yeah. Okay. But for example, for pH, we have not enough data for, for deeper layers.
12:24
So, so we predicted only, only the topsoil. So they can read a paper about reaching the inspector and it seems to be used for inputs. What's the, we use the, we use original, original bands, not, not spectral indices.
12:46
Have you considered using the index? It's called solving. And if they actually exist, actually, even if not the index, there's also a ratio between some bands.
13:08
You mean, yeah, yeah, we can, we can also use some, some spectral indices for, for predictions, but we use only, only the spectral bands and for, for thresholding of the, of the bare soil,
13:23
we use only this to do spectral indices. And from our analysis, it's enough for, for detection. But, but of course there are, there are space for, for some improvement. For example, we have planned to, to introduce, introduce data from Sentinel-1 data, because,
13:44
because in this approach, we can deal with the information about soil roughness or, or soil moisture, which is also the, some, some properties which affect the, the reflectance of soil.
14:02
But the results are already quite promising, given that the, the Elbere engine that's moving by, and the Moravia, the source around the Moravia river out in France. We don't need to do much more about it.
14:22
And which properties do you, to all points? Sorry? Which properties do you use? Yeah. I went to this, to this question, because we use, we use your idea from one of your paper. And to, we, we distinguish the,
14:43
the training data to, to intake quantiles. And we use the distance from points to the, to the nearest point in the, in the quantile. Like quantile. Okay. Yeah. So we, we use, we use 20, 20 layers.
15:06
Yeah. Yeah. Yeah. We discuss it in our paper. And this, this program is almost in the location where,
15:20
where are not good covered by, by soil samples. Yeah. Yeah. Yeah. No, not yet, but I read about it.
15:43
It's a very good idea. Any other questions? Please. Yes. I understand that we are on the open, open data,
16:01
but it's a little bit hard question for me, because our institution policy is not so open. And the data, the data of soil is the owner of the ministry of agriculture. So in Germany, when the same thing,
16:21
we are not allowed to share the point data, we signed the NDA and, and it's an ideal because we help them with their projects and they help us with ours. So we, we signed the bunch of agreement. And they're just allowed to data mining. We don't, we're not distributed data. Yeah. I think in this way, there is no problem,
16:42
but the data are not open. Unfortunate. No, no, I understand it. It doesn't have to be open, but at least it could go up. Yeah, it's possible. Okay. Thank you. I think it will work. Thank you so much for the presentation.