Above-ground Biomass Estimation with Geospatial Foundation Models

Technical Summary: Above-ground Biomass Estimation with Geospatial Foundation Models

Problem Statement
Accurate estimation of Above-Ground Biomass (AGB) from satellite imagery is critical for global carbon stock monitoring and climate policy. However, generating high-resolution, global, and regularly updated AGB maps remains a challenging regression task. Current approaches face three primary limitations: (1) persistent biases where satellite signals saturate over dense canopies, leading to underestimation of high biomass; (2) uneven distribution of reference labels (e.g., GEDI footprints are confined to latitudes below ~51.6°, excluding boreal forests); and (3) prohibitive computational costs for processing petabytes of Earth Observation (EO) data at 10m resolution. While Geospatial Foundation Models (GFMs) have emerged as a paradigm for learning general-purpose representations from unlabeled EO data, their utility for quantitative regression tasks like biomass estimation remains largely unexplored, as most existing benchmarks focus on classification and segmentation.

Methodology
The authors present a comprehensive benchmark of GFMs for global-scale AGB estimation using the AGBD dataset, a machine-learning-ready benchmark spanning diverse biomes with ~16 million samples pairing GEDI L4A biomass values with multi-source satellite data (Sentinel-2, ALOS-2 PALSAR-2, and ancillary features).

The study distinguishes two distinct modes of GFM deployment and evaluates them against a fully supervised state-of-the-art (SOTA) baseline:

1. Frozen Encoders (Weight-Distributed): 11 GFMs distributed as model weights (e.g., CROMA, Prithvi, SSL4EO-MoCo) were evaluated within the PANGAEA framework. These models were run as frozen encoders, with only a task-specific decoder (UPerNet) trained on top. Crucially, these models were restricted to the subset of input features they support (primarily Sentinel-2 optical bands), lacking access to the L-band SAR and ancillary variables available in the full AGBD dataset.
2. Pre-computed Embeddings (Ready-to-Use): Two embedding products, AlphaEarth Foundations (AEF) and TESSERA, were evaluated. These are per-pixel, pre-computed representations derived from multi-modal, multi-temporal EO data. The authors tested these embeddings using Linear Probes (LP), Multi-Layer Perceptrons (MLP), and the SOTA fully convolutional network (fcn_film).

The evaluation included:

• Standard Benchmarking: Comparing RMSE on the AGBD dataset.
• Generalization Tests: Assessing performance in zero-shot cross-region (training on all regions except the target) and cross-year (training on 2019, testing on 2020) scenarios.
• Operational Validation: Comparing results against the independent AGBref dataset and the operational ESA CCI Biomass product.

Key Contributions

1. Comprehensive Benchmark: The authors benchmarked 11 weight-distributed GFMs and two embedding products on the AGBD dataset, the largest and most geographically diverse ML-ready biomass dataset available.
2. Comparative Analysis: They established a direct comparison between frozen GFM encoders, pre-computed embeddings, and a fully supervised SOTA model trained on raw AGBD features.
3. Generalization Experiments: The study designed and analyzed specific experiments to test geographical (cross-continent) and temporal (cross-year) generalization capabilities.
4. Operational Validation: Results were validated against an independent reference dataset (AGBref) and compared with the widely used ESA CCI biomass maps.

Results

• Frozen Encoders Underperform: When run as frozen encoders, the 11 weight-distributed GFMs substantially underperformed the fully supervised SOTA model trained on raw AGBD features. The best GFM (SSL4EO-MoCo) achieved an RMSE of 60.56 Mg/ha, compared to 53.73 Mg/ha for the supervised baseline. This gap is attributed to the models' inability to ingest L-band SAR and ancillary data, and their restriction to single-date or specific band inputs.
• Pre-computed Embeddings Excel: In contrast, pre-computed embedding products proved highly effective. An MLP trained solely on AEF embeddings (52.22 Mg/ha) outperformed the supervised baseline trained on raw AGBD features. The best overall result was achieved by the fcn_film model trained on AEF embeddings augmented with selected raw features (land cover, slope, aspect, coordinates), achieving an RMSE of 50.79 Mg/ha.
• Generalization: AEF-based models demonstrated superior geographical and temporal generalization. In zero-shot cross-region transfers (e.g., to Africa and South America), AEF+ reduced RMSE by 17–25% compared to raw features. However, in South Asia, AEF alone suffered a performance drop in zero-shot settings, which was mitigated by the AEF+ configuration (retaining raw features). Temporally, AEF embeddings remained stable across year gaps where raw spectral features degraded significantly.
• Operational Parity: When evaluated against the independent AGBref dataset, the model trained on AEF embeddings (trained on only 11 regions) achieved near-parity with the globally calibrated ESA CCI product in terms of R² and RMSE, despite the latter's use of globally distributed reference data.

Significance and Claims
The paper claims that the promise of foundation models for quantitative environmental monitoring is best realized when a model trained on rich, multi-modal, multi-temporal data is distributed as an accessible, pre-computed embedding layer, rather than as weights for users to run as frozen encoders.

The authors argue that:

• Input Richness is Critical: The superior performance of AEF and TESSERA stems not just from pre-training scale, but from the richness of the input context (multi-modal, multi-temporal) distilled into the embeddings. Weight-distributed GFMs often lack the architectural flexibility to ingest the specific sensors (e.g., L-band SAR) required for accurate biomass estimation.
• Efficiency and Accessibility: Pre-computed embeddings eliminate the need for GPU-intensive feature extraction, allowing lightweight models (like a simple MLP) to outperform complex, fully supervised networks trained on raw data.
• Robustness: Embeddings provide transferable ecological representations that generalize better across space and time, though retaining raw features (AEF+) offers a safeguard against out-of-distribution failures in ecologically distinct regions.

The study concludes that while frozen GFM features can be highly informative, their utility is currently limited by input restrictions and engineering frictions. The most effective path forward involves either making weight-distributed models flexible enough to handle diverse inputs and resolutions, or continuing to serve representations as pre-computed, analysis-ready embedding layers.