Can we take a photo in space

Create an exact picture of the landscape from space

With detailed information on land cover, we can improve our understanding of our environment - for example when it comes to important ecosystem services such as the pollination of plants or nitrate and nutrient inputs into bodies of water. This information is increasingly obtained from high-resolution satellite images. However, clouds often obstruct the view of the earth's surface from orbit. With the help of dynamic machine learning methods, these gaps, which inevitably occur in satellite observation, can be closed. We talk to Sebastian Preidl from the Helmholtz Center for Environmental Research in Leipzig (UFZ).

Mr. Preidl, what benefits does satellite observation of agricultural land bring to society? What can you observe with it?

Sebastian Preidl: Satellite observations have already been used frequently in the past to record land cover and to divide it into classes such as agricultural areas, forest areas, urban areas, bodies of water, etc. With newer satellites it is now possible to further differentiate individual classes thematically. In our study we distinguished 19 agricultural crops within Germany. This made it possible to generate a map that shows different types of grain, but also, for example, areas of wine-growing, orchards, rape and grassland.

With such a map it is possible to create a more detailed overall picture of the landscape. If we can better determine the habitats for flora and fauna in this way, ecological questions about biodiversity or the pollination performance of wild bees can also be answered more precisely. If the spatial distribution of agricultural crops is included in scientific models, the nutrient input - for example nitrogen - in the soil and water can be better estimated. In addition - in connection with EU agricultural subsidies, among other things - comprehensive monitoring of the agricultural landscape also arouses great interest at the political level.

First of all, we carried out the classification for 2016. In the future, however, we also want to work on documenting short and long-term changes in land cover. As a result of climate change, years of severe drought, as we have seen in the past two years, will pile up. In order to be able to take concrete measures, we need information about which crops or tree species in which regions are most severely affected by drought damage. Determining the vitality of plants with the help of satellite observations also plays a role here.

Which satellites do you use for your research? Why this one? What orbit are these satellites in?

Sebastian Preidl: For our study we used data from the Sentinel-2 satellite of the European Earth observation program Copernicus. These orbit the earth at an altitude of almost 790 kilometers. These satellites generate high-resolution images with a pixel edge length of 10 or 20 meters. In addition to the spatial resolution, the temporal resolution also speaks in favor of using the Sentinel-2. Over Germany, the same image section is taken in different spectral ranges every two to three days. For our study, we opted for a spatial resolution of 20 meters in order to be able to fall back on 9 spectral bands for image analysis. It is therefore the combination of high spatial, temporal and spectral resolution that makes the Sentinel-2 satellite attractive for observing dynamic processes on the land surface.

What problems do you have to contend with when observing remotely from orbit?

Sebastian Preidl: The problem with optical satellites is that they cannot see through clouds. The already mentioned high temporal resolution of the Sentinel-2 is a decisive advantage here because it increases the probability of obtaining a cloud-free view of the earth's surface. However, enormous amounts of data are generated by the many recordings. The aim was to develop a classification algorithm that takes into account the different cloud cover in numerous satellite images, taking into account the available training data.

Are you now combining satellite images with methods of dynamic machine learning? How can you imagine that exactly? What's new about it?

Sebastian Preidl: We used a classic machine learning method as a classifier. The novelty lies in the dynamic application of a large number of forecast models. First, the classifier has to learn which spectral properties are characteristic of a crop. Training data is therefore required with which an agricultural crop can be clearly assigned for some image pixels. In a first step, our algorithm now defines periods for which sufficient cloud-free training pixels are available. This process is dynamic and can differ from region to region due to different weather conditions.

In a second step, the algorithm looks at the time periods at which the pixels to be classified are free of clouds. Since we classify millions of pixels, there are many possible combinations. In order to avoid the generation of artificial data through interpolation, a new prediction model is calculated for each combination. What is new is that we have chosen a regionalized approach in which only available satellite data is used within dynamically generated time periods. Finally, individual predictive models are used to assign a crop to each pixel.

How accurate is this new method?

Sebastian Preidl: In our study, we achieved an overall accuracy of approx. 88 percent. A good result, especially because we have often achieved an even higher level of accuracy for the most common crops in Germany, such as wheat, rape, sugar beet and maize. However, other crops for which less training data were available or whose spectral signature does not clearly differ from those of other crops were classified with less accuracy. It is to be investigated to what extent the classification accuracy can be further increased, especially for spelled and some types of summer grain.

What are the advantages of this method, for example when it comes to wild bee populations or the entry of nitrate into water?

Sebastian Preidl: With the presented method, we first classify the land cover. Our study has shown what level of thematic detail can be achieved with the new satellite data. This opens up new possibilities in spatial analysis and scientific modeling. If we know where fertilizer-intensive crops are grown, the effects of land management practices and thus the nitrate pollution in the soil and adjacent waters can be better quantified. The designation of grassland areas or agricultural crops that can serve as a source of nectar for wild bee populations is no less important. Only by determining the size, spatial distribution and arrangement of such areas can one investigate whether the food supply is sufficient for the bees and whether the flight distances are not too great. The created land cover map can serve as a basis for answering such pressing ecological questions

You have now created a map of Germany of the agriculturally used areas. How can you use this map?

Sebastian Preidl: The map, which was initially created for 2016, can be viewed in a WebGIS at It is also possible to download the map from the PANGEA data platform and use it for your own evaluations under the specified license.

Are there other possible uses for the new observation and evaluation method, for example in forestry or in wine and hops cultivation?

Sebastian Preidl: Yes, the method developed has already been used elsewhere at the UFZ. The Federal Agency for Nature Conservation was interested in evaluating forests from a nature conservation point of view, largely on the basis of remote sensing data. So I used the method for the classification of the main tree species in Germany. In principle, the approach can be applied to any type of land cover. It is crucial that the training data set is as extensive as possible. However, sometimes it is the case that plant species can hardly be distinguished spectrally due to their similar physiology and structure.

We may be heading for a multi-hazard situation this year. Covid-19 crisis and another drought summer can overlap. On the one hand, we would then have interrupted supraregional supply chains, and on the other hand, we would have problems with domestic food production. Do you see possible uses of your method to create reliable forecasts on the supply situation in good time?

Sebastian Preidl: The entire Sentinel-2 recordings of one year were used for the classification. Thus, species-specific reflectance profiles could be created according to the individual growth behavior. When using shorter time series, which would be a prerequisite for the situation mentioned, I expect greater inaccuracies in the differentiation of individual crops. For a forecast of the supply situation, yield estimates would also be necessary, which we did not make in our study.

However, climate change can affect the supply situation of individual foods in the long term. It should now be investigated at an early stage how strongly agricultural crops react to persistent drought in accordance with their cultivation area. In addition, the forest is increasingly suffering from the drought of recent years. The weakened trees can hardly withstand fungal diseases or the bark beetle in the following year. We would like to pursue research on these topics at the UFZ. A submitted project aims to document changes in agriculture and in the forest. Other artificial intelligence approaches, namely deep learning methods, should also be used for the evaluation. A close exchange with environmental and nature conservation authorities as well as agricultural and forestry interest groups is sought.

Mr. Preidl, thank you for the interview.

The questions were asked by Oliver Jorzik and Dierk Spreen (Earth System Knowledge Platform | ESKP)

  Helmholtz Center for Environmental Research - UFZ. (2020, April 21). Corridor view from space. Machine learning methods provide detailed information on land cover [press release,]. Accessed on April 21, 2020.

  Lausch, A., Pause, M., Doktor, D., Preidl, S. & Schulz, K. (2013). Monitoring and assessing of landscape heterogeneity at different scales. Environmental Monitoring and Assessment, 185(11), 9419-9434. doi: 10.1007 / s10661-013-3262-8

  Preidl, S., Lange, M. & Doktor, D. (2020). Introducing APiC for regionalized land cover mapping on the national scale using Sentinel-2A imagery. Remote sensing of the environment, 240: 111673. doi: 10.1016 / j.rse.2020.111673


Published: 04/29/2020, Volume 7

Citation reference: Preidl, S. (2020, April 29). Create an exact picture of the landscape from space (interview).Earth System Knowledge Platform [],7. doi: 10.2312 / eskp.021

Text, photos and graphics unless other licenses are concerned: | CC BY 4.0 | Earth System Knowledge Platform - the knowledge platform of the research area Earth and Environment of the Helmholtz Association