Abstract

Abstract Mapping global forest biomass from NASA ’s GEDI mission requires interpolating sparse LiDAR observations across diverse landscapes. Standard machine learning approaches like Random Forest and XGBoost fail to produce calibrated uncertainty estimates, as they conflate ensemble variance with true predictive uncertainty and ignore spatial context. We introduce Attentive Neural Processes (ANPs), a probabilistic meta-learning framework that conditions predictions on local observations and geospatial foundation model embeddings. ANPs learn flexible spatial covariance functions, expanding uncertainty in complex landscapes and contracting it in homogeneous areas. Validated across five biomes from tropical Amazonian to boreal and alpine ecosystems, ANPs achieve competitive accuracy with near-ideal uncertainty calibration. The framework also enables few-shot adaptation, recovering most cross-region transfer performance with minimal local data. This provides a scalable, principled alternative to ensemble methods for continental-scale biomass mapping.

Bio

Robin Young is a first-year PhD student in Computer Science at the University of Cambridge.