Accounting for uncertainty in AI-based tree cover monitoring

No. 28.6 - 29.09.2026

Göttingen Researchers Win Best Paper Award at ACM GoodIT 2026

Researchers from the Institute of Computer Science and the Department of Economics at the University of Göttingen have received the Best Paper Award at the sixth ACM International Conference on Information Technology for Social Good (GoodIT 2026). The award, the conference's sole Best Paper Award this year, recognizes their research on making AI-based tree cover monitoring from satellite imagery more reliable. First author Fabian Wölk presented the work at the conference in Pisa.

The award-winning paper, entitled “Uncertainty-Aware Annotation for Reliable Tree Cover Monitoring from Satellite Imagery,” was authored by Fabian Wölk, Dr David Degenhardt, Tobias Hellmundt and Professor Xiaoming Fu. The researchers investigate how artificial intelligence can be trained to better distinguish trees from other vegetation in satellite images.

A central challenge is that objects in satellite images are not always clearly identifiable. A small tree, a shrub or another type of vegetation can look very similar in satellite imagery. When creating training examples for an AI system, people nevertheless usually have to make a clear decision when labeling the images – even when they are uncertain. Such uncertain decisions can introduce errors into the training data and later affect the maps of tree distribution and estimates of tree numbers produced by the AI. In applications such as monitoring forest development and reforestation, this can lead to misleading assessments of how an area is developing.

The Göttingen researchers therefore pursue a simple idea: people who label satellite images for AI training should be able to mark ambiguous objects as “uncertain” instead of being forced to assign them to a specific category. Additional auxiliary models then estimate how likely the different categories are for these uncertain areas. These probabilities are used as additional training information for the main AI model, which is later tasked with distinguishing trees from other vegetation in satellite images. This allows the model to better account for ambiguity in the image data.

The approach was tested in a case study on mapping tree cover in parts of Madagascar. The researchers examined not only how accurately the AI could distinguish different types of vegetation, but also whether independently trained models produced similar maps of tree distribution and similar estimates of tree numbers.

The results show that retaining information about uncertainty and incorporating it deliberately into training can contribute to more reliable results. The previously described methods, in which uncertain areas are incorporated into training with additional probability information, were particularly successful. In the experiments, they led to higher accuracy and more stable results.