This way new AI methods can turn satellite data into tools to achieve the global sustainability goals

Around the clock, everything that happens on Earth is followed by satellites. Data collected is used to monitor and control groundwater as well as beach protection and pastures. The National Space Data Lab is now developing new AI methods to automatically transform satellite data into tools for more sustainable development.  

European satellites are constantly working to collect data about the earth. Every day, millions of satellite images are taken, which give us a unique picture of both global and local changes in the earth's surface. Copernicus is the EU's Earth observation action program, where data is collected for comparisons and analysis. But Copernicus' library of space data is growing by about 20 terabytes every day, and new AI methods are needed to make efficient analyzes of that amount of data. At the National Space Data Lab, such methods are currently being investigated, including deep machine learning and deep neural networks in order to, for example, be able to automatically screen out information and show changes in the landscape over time. The lab is run in collaboration between the Swedish Space Agency, AI Sweden, RISE and Luleå University of Technology.

Important for Agenda 2030

The United Nations (UN), the European Space Agency (ESA) and the European Union (EU) have stated that space-based sensors and communication solutions are a prerequisite for achieving the sustainability goals in Agenda 2030. In order to provide the best benefit from the observations, a well-functioning chain of data collection for good decision bases. Here, the National Space Data Lab has an important task in Sweden's climate change.

Since 2019, the National Space Data Lab has been operating a calculation and storage infrastructure for space applications on a platform that is being developed at the ICE Datacenter at RISE in Luleå. There is access to open data from Copernicus as well as tools for making data searchable such as, Open Data Cube, OpenEO and STAC. The fact that the development platform is based on open source code is important to facilitate the use of data from space within more authorities and companies.

During the first two years, the lab has worked together with, among others, the Swedish Forest Agency, the Swedish Board of Agriculture, the Swedish Marine and Water Authority, the Swedish Environmental Protection Agency, SMHI and several county administrative boards. Collaboration increases the degree of use of space data by the authorities, and is facilitated by the fact that collection, tools and knowledge are gathered in one place.

Today, the National Space Data Lab is a knowledge node for storing, handling and analyzing large amounts of space data all the way from simple statistics to AI. Three projects that have shown good results are Drought in the Mälardalen Valley, AI & Coastal Zones / Shallow Water and AI & Agriculture.

Climate impact can be seen in satellite images

The first project carried out at the lab was Torka i Mälardalen, a collaboration between SMHI and the County Administrative Board of Västmanland.

Here, data from 2018 were compared with 2019 to find differences in vegetation in a selected area. The comparisons were based on the so-called spectral index, which makes it possible to distinguish vegetation from, among other things, water, soil or buildings, all depending on how much light is reflected.

The project shows the way to a new AI-based method for using satellite data to better understand the impact of climate change on vegetation.  

Two spectral indices were selected: NDVI (Normalized Difference Vegetation Index, see Figure 1) which is a good indicator of green vegetation and MSI (Moisture Stress Index, see Figure 2) which measures the moisture content. Seven sites near Västerås were selected for the study. First, a time series was compiled for the comparison, which was then processed using Gaussian Process Regression (GPR). The GPR method makes it possible to include occasions when data is missing because it uses machine learning to predict reasonable values ​​even for those times.

The project shows the way to a new AI-based method for using satellite data to better understand the impact of climate change on vegetation.

AI sees changes in the coastline

Within the AI ​​& Coastal Zones / Shallow Water project, new AI methods were developed to identify physical changes along the Swedish coast. In the part of the project that was carried out within the framework of RISE, the focus was on identifying changes to the coastline itself.

Data were used from the Copernicus Sentinel-2 satellite, which was analyzed at the ICE Datacenter using Phyton software libraries. Three different AI models were developed based on the satellite images. The models classify individual pixels in the images as water, vegetation or non-vegetation to detect physical changes. The first model compared images taken every few days to find changes in geography, while the others used particle filters based on so-called "Hidden Markov" methods which can better determine whether the changes that appeared were permanent or not.

By extension, the method can be used as a tool for automatic monitoring of beach protection throughout the country. Unique is that the developed AI method can use low-resolution satellite images.

Field visits from space 

The last example is about grazing monitoring. Every year, the Swedish Board of Agriculture, together with the county administrative boards, conducts thousands of field visits to find out which land is used for agriculture. Both the EU and the Swedish Board of Agriculture hope to be able to replace field visits with analyzes of satellite images. Already today, the authority works with manual analyzes of the images, but even this is a time-consuming job.

In the AI ​​& Agriculture project, unsupervised machine learning has been used to analyze the satellite images, which means that input data is not categorized and that there is no known or expected result of the data analysis. Satellite data from Copernicus Sentinel-2 has been used to generate time series based on NDVI, which gives an indication of whether the observed area is water, bare ground or vegetation, and which has then also been analyzed using Gaussian Process Regression.

With the help of classical cluster analysis, such as Hierarchical clustering, the project was able to show a correlation between the time series and the use of the pastures.

The result opens up the possibility of developing more advanced AI models to automatically monitor all agricultural land in the country. Among other things, deep machine learning can open up the possibility of finding better representations of pastures as a complement to statistical measures based on NDVI.

Continued development on behalf of the government

During the autumn of 2021, the Swedish Space Agency was commissioned by the government to continue investigating how the capacity of the National Space Data Lab can be strengthened. This so that Sweden, the EU and international partners can increase the use of data from satellites for a more sustainable societal development. A developed national space data lab is important for the implementation of the European Data Strategy and the forthcoming National Data Strategy. The Government decision stipulates that in carrying out the assignment, the Swedish Space Agency shall obtain information through dialogue with relevant stakeholders, including RISE. The Swedish Space Agency will also participate when the European Space Agency's Council at ministerial level meets next year to decide on the direction of ESA's long-term space policy and future programs.

Bild 1. Normalized Difference Vegetation Index

Bild 2. Moisture Stress Index