Calibrated Probabilistic Interpolation for GEDI Biomass

TL;DR

This work tackles the challenge of producing wall-to-wall, uncertainty-aware biomass maps from sparse GEDI observations. By combining Attentive Neural Processes with remote-sensing foundation-model embeddings, it delivers calibrated, context-aware prediction intervals that adapt to local landscape heterogeneity, outperforming standard ensemble baselines in both accuracy and uncertainty calibration. Across five biomes, ANPs demonstrate robust cross-regional generalization, with few-shot adaptation recovering much of the performance gap in transfer scenarios. The approach offers a scalable alternative to geostatistics for continental-scale land-cover mapping, with practical implications for conservation planning and carbon accounting where reliable uncertainty quantification is essential.

Abstract

Reliable wall-to-wall biomass mapping from NASA's GEDI mission requires interpolating sparse LiDAR observations across heterogeneous landscapes. While machine learning approaches like Random Forest and XGBoost are standard for this task, they treat spatial predictions of GEDI observations from multispectral or SAR remote sensing data as independent without adapting to the varying difficulty of heterogeneous landscapes. We demonstrate these approaches generally fail to produce calibrated prediction intervals. We identify that this stems from conflating ensemble variance with aleatoric uncertainty and ignoring local spatial context. To resolve this, we introduce Attentive Neural Processes (ANPs), a probabilistic meta-learning framework that explicitly conditions predictions on local observation sets and geospatial foundation model embeddings. Unlike static ensembles, ANPs learn a flexible spatial covariance function, allowing uncertainty estimates to expand in complex landscapes and contract in homogeneous areas. We validate this approach across five distinct biomes ranging from Tropical Amazonian forests to Boreal and Alpine ecosystems, demonstrating that ANPs achieve competitive accuracy while maintaining near-ideal uncertainty calibration. We demonstrate the operational utility of the method through few-shot adaptation, where the model recovers most of the performance gap in cross-region transfer using minimal local data. This work provides a scalable, theoretically rigorous alternative to ensemble variance for continental scale earth observation.

Figures (8)

• Figure 1: Performance-efficiency and accuracy-calibration limitations in baseline methods. Each point represents a hyperparameter configuration evaluated across 5 random seeds. Size indicates training time, where larger is longer training. Dashed lines show Pareto frontiers. Dotted horizontal line indicates ideal calibration (Z-score std = 1.0).
• Figure 2: Zero-shot spatial extrapolation performance across biomes. Models trained on one region (rows) are tested on all four regions (columns) without adaptation. Top row: 1$\sigma$ coverage for ANP (left) and XGBoost (right). Middle row: Z-score standard deviation, where 1.0 indicates perfect calibration. Bottom row: Accuracy (Log $R^2$). Diagonal elements represent within-region performance. Values show mean $\pm$ std.
• Figure 3: ANP few-shot adaptation results across region pairs. Each model is fine-tuned on 10 randomly sampled tiles from the target region (5 epochs). Top: 1$\sigma$ coverage. Middle: Z-score standard deviation. Bottom: Accuracy (Log $R^2$). Within-region entries (diagonal) are not shown. Values show mean $\pm$ std.
• Figure 4: Calibration diagnostics for ANP on Guaviare test set (representative run). Top left: Distribution of standardized residuals closely follows theoretical N(0,1). Top right: Prediction interval coverage shows slight conservatism. Bottom left: Binned uncertainty vs error plot shows positive correlation. Higher predicted uncertainty corresponds to larger actual errors. Bottom right: Quantile-quantile plot shows good agreement in central quantiles ($\pm$2$\sigma$), with slight divergence in tails.
• Figure 5: Spatial predictions and uncertainty for a 0.1$^\circ \times$ 0.1$^\circ$ tile in Guaviare, Colombia. Left: Mean AGBD prediction shows dendritic patterns of high biomass (green) corresponding to remaining forest corridors, with low biomass in cleared areas (yellow). Right: Predicted standard deviation (uncertainty) is spatially heterogeneous, with highest uncertainty in structurally complex forested areas and lowest uncertainty in homogeneous cleared land. The spatial pattern of uncertainty adapts to land cover configuration, increasing in regions with complex forest structure and decreasing in simple agricultural landscapes.