New paper in Biogeosciences!

To build resilience against climate change, we need a better understanding of our ecosystems. Remote sensing, in combination with deep learning, can tell us a lot about plant traits that are essential for environmental applications. Hyperspectral satellite missions collect data on biophysical, biochemical, and structural variations within plant canopies, while deep learning models help predict plant traits based on these remote sensing data.

But here’s the problem: when deep learning models are applied to data that differ from the training material (out-of-domain: OOD), for example from different regions, biomes, or sensors, the predictions become less accurate. It is therefore important to effectively quantify the uncertainty of the predicted values.

Because current methods for measuring uncertainty are often overly optimistic, Cherif et al. propose a distance-based uncertainty estimation method (Dis_UN) in their latest article published in Biogeosciences. The method, supported amongst others by multi-year data collected at LMU Munich in Tobias Hank’s research group, quantifies uncertainty by measuring how (dis)similar the training data and the new data are, based on quantile regressions that capture worst-case scenarios (i.e. the highest possible error).

Key results

• The method performed especially well for mixed pixels in satellite imagery: At 30 m resolution, one pixel often contains multiple land cover types, making predictions with conventional methods difficult. Dis_UN still identified these pixels as uncertain because their spectral signatures differed from the vegetation samples used during training.
• Dis_UN was both more interpretable and computationally efficient: Instead of relying on many repeated model runs, the method estimated uncertainty from data similarity. This made the results easier to understand and significantly faster to compute for large-scale remote sensing applications.

Assessing uncertainties is an important step when applying models beyond their training domain. Only then can we make reliable and robust statements about our ecosystems that can support resilience-building in a warming climate.

Read the paper