NRK enlisted the help of NINA to shed light on a dark reality: while Norwegian nature is being lost bit-fot-bit, no one is watching.
Created using Dall-E.
The investigative journalism published over the weekend by NRK has left many shocked. Norway is losing at least 79 square meters of nature per minute when averaged over the past five years. This is 207 square kilometers in total. The story behind the numbers is also intriguing.
We have released the data publicly via this web app: nedbygging (earthengine.app) (best to open on PC, but also possible on mobile)
Lack of data on nature loss in Norway
During 2022, Zander Venter, a geospatial scientist at NINA, was contacted by Mads Nyborg Støstad to ask a very simple question: do we have data to show us how much nature our society consumes in Norway? Mads had approached SSB, NIBIO and Miljødirektoratet and they could not give him a straightforward answer. Venter also pointed out that there are no readily available maps of natural habitat loss in Norway
Venter had recently published a paper reviewing state-of-the-art satellite-based maps of land use (Venter et al 2022). It was coincidental that he was in the process of incorporating the maps into NINA’s projects on ecosystem accounting when Mads contacted him. He was able to produce a map* of nature loss over Norway between 2017 and 2022.
* The map will be made available via a link published here. Follow us on social media to get notified.
NINA uses eyes from the sky and AI
Venter used Google’s dataset called Dynamic World. It uses publicly available images from the two satellites Sentinel-2A and Sentinel-2B which are operated by the European Space Agency. These satellites fly continuously over the globe and take pictures of the whole world, including Norway, with a resolution of 10 metres.
The images are quite blurry, but Google uses artificial intelligence (AI) to analyze such satellite images and recognize different types of land use. It can recognize nine different categories: water, trees, grass, flooded vegetation, snow and ice, bushes and scrub, bare ground, crops and a final category called "built-up".
Zander used the raw outputs from the AI model to detect changes in built-up surfaces that signal nature loss. Specifically, the AI produces a probability that every single pixel on the Norwegian map (10 x 10 meters) is built-up or not. After using time series analysis, Venter could identify pixels, and ultimately patches of pixels, that went from natural to built-up cover.
Journalists acting as scientists
It is well-established that AI models are biased and not without error. Unfortunately, many end-users of AI products do not account for biases before the drawing conclusions from them. Venter asked the journalists if they were willing to follow scientific best practice and quantify error in the map and estimate uncertainty. They said yes!
Fortunately, NINA has a number of ongoing ecosystem accounting projects which are piloting methods for quantifying uncertainty in maps of ecosystem extent, condition and services (see Naturregnskap web page). Therefore, Venter had the statistical tools to guide the journalists through the map validation and uncertainty estimation.
It went down like this: Venter sent them thousands of random locations to verify if nature loss had in fact occurred or not. Using services such as norgeskart.no, Norge I bilder and Google Earth Pro, they verified each and every point and sent the data back to Zander. He then ran it through a statistical model and calculated that the margin of error was 18 percent – in other words, the AI missed just under one in five times when a pixel was flagged as nature loss. The validation points were also used in what is called a stratified area estimator to correct for the bias in the AI map and produce a bias-corrected estimate of nature loss.
The journalists went on to use many other datasets which NINA has helped produce to explore what types of nature we are destroying in Norway. For example, Vegar Bakkestuen, another NINA researcher, had also used AI and remote sensing to map wetlands over southern Norway (Bakkestuen et al 2023). The journalists could use this to explore how many nature loss events occurred in wetland ecosystems – important ecosystems for carbon and biodiversity.
A vision for the future
Venter notes that it is important to remember we only calculated how often the AI incorrectly identifies nature loss, not how often it overlooks nature loss (so-called “false-negatives”). To do the latter, we would require more manual verification work. This means that our area estimates are minimum figures, and that the real nature loss is greater.
Due to the coarse resolution of the satellite images, we also know that it does not detect smaller nature losses forestry roads, wind power roads, and small buildings like cabins. We have also not mapped other forms of habitat loss like forest harvesting, cropland expansion, and all the subtle impacts humans have through sound, light and chemical pollution.
On the bright side, there is hope for the future. NINA aims to use higher resolution satellites and perform more field work to detect all types of nature loss over Norway. “Although the technology is ready, the authorities are not investing money into the research. At the same time, nature is being lost bit-for-bit while no one is watching. Let us start watching, Venter says”.
Data available in this app: nedbygging (earthengine.app)
Contact: Zander Venter