Researchers warn about inaccurate forest estimates

Remote sensing data are playing an increasingly important role in providing information for decision-making about forests. According to researchers at Mistra Digital Forest, this development highlights a bias-related problem that is difficult for many to grasp. “The results of standard model-based estimates using remote sensing data are often distorted. We need to take this into account so that we don't make decisions based on false premises”, says Göran Ståhl, Professor of Forest Inventory at SLU.

Using remote sensing data to assess the conditions of forests is nothing new, but it is becoming increasingly common. There are plans within the EU to use optical satellite data to map Europe’s forests, as part of climate reporting. National laser scanning has long been providing important information about the Swedish forests.

This development highlights an issue of bias, that may cause problems in many areas of application. This is the view of Professor Göran Ståhl at SLU, who has studied the phenomenon within Mistra Digital Forest, and has published a scientific article on the subject together with his colleagues.

– Remote sensing predictions describing forests at pixel level are typically biased. Low true values tend to be overestimated and high true values underestimated. This has been known for some time, but the cause has been unclear. We have shown that it is due to inadequate combining of statistical methods, says Göran Ståhl.

Laser data can be corrected – this is much more difficult with optical satellite data 

Researchers are currently investigating various strategies to remedy the problem, with a particular focus on optical satellite data and laser data, such as the data from the Swedish Mapping, Cadastral and Land Registration Authority’s national laser scanning programme.

– We compare both established and new methods to see if they can reduce the systematic errors. Ideally, we want to find methods that lead to small systematic errors as well as small random errors, says Magnus Ekström, Professor of Forest Mathematical Statistics at SLU.

The results so far show that correcting for discrepancies works reasonably well in cases where the remote sensing data are strongly correlated with the variable being estimated, such as with timber volume.

– Laser scanning data have this strong correlation. However, it appears to be far more difficult to correct for the bias when using optical satellite data. In this case, the relationships are weaker, making it harder to make a good correction, says Magnus Ekström.

Back to Göran Ståhl, who emphasises that we need to understand, and to deal with this problem:

- Otherwise, we will make errors in several applications, for example when using this type of data for planning or for estimating changes. There is also a risk that we will not find hotspots. These are extreme values of various kind which are often the most interesting ones. As both forestry and political stakeholders are using remote sensing data to an ever-increasing extent, there is an increased risk of making decisions on the wrong basis. We need new ways of thinking regarding data analysis and modelling.