Deep Evidential Regression quantifies tree height uncertainty from Sentinel satellite data
New model estimates forest height at 10m resolution while predicting confidence—key for carbon accounting.
A team led by Laura Bader at LMU Munich introduced Deep Evidential Regression (DER) to tackle a persistent problem in Earth observation: predicting forest structure without knowing how confident the AI is. Their paper, posted on arXiv (2608.06406), applies DER to the TreeUQ benchmark, a large-scale dataset covering Bavaria at 10m resolution with Sentinel-1 (radar) and Sentinel-2 (optical) imagery, plus sparse tree inventory labels. The approach trains a U-Net to jointly output tree height and an uncertainty estimate in one forward pass, avoiding the cost of ensemble or Bayesian methods.
To handle the extreme sparsity of ground-truth tree inventory data, the researchers introduced a masked evidential loss that prevents empty pixels from skewing training. Results show DER achieves predictive performance comparable to a deterministic U-Net, while additionally providing well-calibrated uncertainty scores. This allows analysts to know where predictions are reliable and where more field data is needed—making the framework practical for operational monitoring of carbon stocks, biodiversity, and ecosystem health from satellite imagery.
- Deep Evidential Regression (DER) jointly predicts tree height and uncertainty in a single forward pass versus costly ensembles
- Masked evidential loss handles extreme label sparsity in the TreeUQ benchmark (10m resolution, Bavaria)
- Matches deterministic U-Net accuracy while delivering well-calibrated confidence metrics for Sentinel-1/-2 fusion
Why It Matters
Enables trustable, uncertainty-aware forest monitoring from satellites—critical for carbon credits and climate policy.