cs.CVSep 29, 2026

Planetary Feature Fields are Scalable Earth Representations

Authors: Arjun Rao, Sebastian Loeschcke, Anthony Fuller, Isaac Corley, Nico Lang, Evan Shelhamer

Organizations: University of British Columbia · University of Copenhagen · Carleton University · Vector Institute · Taylor Geospatial

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Satellite observations, precomputed embeddings, and map products describe the same evolving Earth, yet are stored as independent, petabyte-scale data products. Their continued growth calls for compact representations of multiple products while preserving spatial and temporal detail. We introduce Planetary Feature Fields (PFFs), which exploit redundancy across data products by modeling them jointly as continuous functions of space and time at planetary scale. PFFs are spatially local explicit-implicit (hybrid) neural fields. Each field shares a factored feature volume---a decomposition of an explicit 3D grid with smaller factors---across products, while lightweight implicit decoders reconstruct individual products across multiple timesteps. PFFs reconstruct EO products over space and time more accurately than single-product fields at matched compression rates. At 1800×1800\times compression relative to the uncompressed source data, reconstructed features retain approximately 90%90\% or more of the performance achieved with the original features on pixel-level segmentation, change detection, and patch-level classification tasks. PFFs can add new timesteps by extending their factored feature volumes and add new products by attaching new decoders, while leaving existing outputs unchanged. PFFs reduce end-to-end feature access latency by an order of magnitude relative to evaluated API and cloud-storage pipelines.

Figures & tables

Figure 1: Planetary Feature Fields (PFFs) jointly represent heterogeneous Earth observation (EO) products with high fidelity at high compression ratios ( ≈\approx 1800 ×\times ). We present PFFs as spatially local, explicit–implicit (hybrid) neural fields that share a factored feature volume across raw observations, precomputed embeddings, and map products. Queried by space–time coordinates, this shared representation exploits redundancy across products, preserves downstream utility, and supports adding new timesteps and products while leaving the existing representation unchanged.
Figure 3: PFFs are capable of continually learning (a) new timesteps and (b) new products. (a) New timesteps for one or more products requires a small addition to the shared field (0.2% params) and are trained with the rest of the PFF frozen. (b) New products are added by training only a decoder against the frozen shared field, and deprecated products can simply be discarded.
Figure 4: Reconstruction results on PFFBench. (a) One shared PFF beats compression rate-matched single-product neural fields on EO product reconstructions tested across 100 global PFFBench tiles (see Figure 10 ). (b) Rate-distortion curves on the mean reconstruction metric (10-tile subset). Sharing the explicit representation is uniquely beneficial to factored feature volumes (NGP is an explicit–implicit representation, but not a factored feature volume). Full table shown in Table 5 .
Figure 5: PFFs are robust to missing data. We train with products that may be observed at every timestep, miss entire timesteps, carry only some of its channels, miss spatial information in one or more years, or be entirely absent from the region.
Figure 6: Continual learning results. (a) Only training the small temporal slice (0.19% params) achieves ≈\approx 92% of joint pretraining’s new-year reconstruction while leaving existing years unchanged ( PFF-Append ). When the product decoders are trained with this slice ( FT-Dec ), or the entire PFF is finetuned ( FT-All ), existing-year retention collapses. (b) We add Galileo (GA) ( Tseng et al., 2025 ) 10m embeddings to PFFs covering the PASTIS segmentation benchmark. Only training a new decoder on a small subset of pixels significantly outperforms finetuning the PFF on GA’s PASTIS mIoU while fully retaining AlphaEarth (AE) and TESSERA’s (TE) reconstruction and performance on PASTIS. Expanded in Figure 13 .
(a) PASTISmIoU
 AETES2
Original features53.661.313.3
TensoRF(VM)67%69%92%
K-Planes69%59%91%
iNGP46%46%75%
PFF-VM90%90%136%
Table 1: PFF features retain most downstream performance and can even outperform the original features. We report retention % : the percentage recovery of the Original features ’ performance. (a) PASTIS crop-type semantic segmentation. (b) FTW field delineation in Brazil and South Africa. (c) SwissCrop25 crop-type segmentation from 2019–2025. (d) EuroMineNet annual change detection. Original features are AlphaEarth (AE), TESSERA (TE), and Sentinel-2 (S2).
Figure 7: PFFs materialize features an order of magnitude faster than API or cloud-based methods. Time (s) to materialize 1 million AE and TE embeddings visualized.
 ProductSourceddGSDYearsUnitsPrecision
Raw observationsSentinel-2 s2pcSentinel-2 L2A1010 m2017–2025surface reflectancefloat16
Sentinel-1 s1rtcSentinel-1 RTC, asc+desc410 m2017–2025γ0\gamma^{0} dBfloat16
Landsat landsatLandsat C2 L2, L8/L9630 m2017–2025surface reflectancefloat16
Temperature lstLandsat C2 L2 thermal130 m2017–2025∘\circ Cfloat16
PALSAR palsarALOS PALSAR-2 mosaic220 m2017–2021γ0\gamma^{0} dBfloat16
Precomputed embeddingsAlphaEarth AEAlphaEarth Foundations6410 m2017–2025embeddingint8
Table 2: EO products used to train PFFs.
(a) BEN-v2AlphaEarthTESSERASentinel-2Landsat
 Micro-F1mAPMicro-F1mAPMicro-F1mAPMicro-F1mAP
Original features74.968.874.268.361.551.858.647.3
TensoRF100%101%101%100%102%100%101%102%
K-Planes100%100%101%98%99%97%98%97%
iNGP101%101%100%95%99%97%101%101%
PFF-VM100%101%101%101%103%103%103%103%
Table 3: PFFs and most hybrid neural fields saturate performance on patch-level classification tasks. We report linear-probe transfer results using stats pooling. All values are percentages of the original features (gray). We also evaluate traditional geographic location encoders (fully implicit, global) SatCLIP ( Klemmer et al., 2025 ) , GeoCLIP ( Cepeda et al., 2023 ) , and Tessellating the Earth (TTE) ( Cher et al., 2026 ) on classification tasks and report their retention as a percentage of AlphaEarth’s original features.
Table 4: Probing shared and stacked PFF features. PASTIS test mIoU (%) from a pooled linear probe across 17 regional fields, using identical evaluation pixels and train-only standardization. Stack combine observations, embeddings, maps, or all 14 products. Bold / underline : best / second-best merging result per architecture. Gray rows show single-product controls; raw scores span both architectures.
Table 5: Reconstruction scores ( ↑\uparrow ) on PFFBench . Shared PFFs are compared with rate-matched single-product neural fields across 100 globally distributed tiles ( Figure 10 ). Single-product fields use either uniform field budgets ( 1/141/14 per product) or bit-proportional budgets. Scores are means ×100\times 100 , with one standard error across tiles beneath each score. Mean (14) averages products within each tile before aggregation. Bold and underline denote the best and second-best scores in each column, including ties.
 AlphaEarthTESSERASentinel-2LandsatSentinel-1
 B-F1OF1B-F1OF1B-F1OF1B-F1OF1B-F1OF1
Original features66.578.067.477.363.970.756.464.849.645.7
TensoRF(VM)97%100%94%103%85%104%89%103%91%104%
K-Planes97%100%94%103%85%105%90%103%92%105%
iNGP88%98%83%100%82%102%86%101%85%100%
PFF-VM99%100%98%102%91%104%97%104%98%106%
Table 6: EuroMineNet footprint mapping. Gray rows show original-feature scores ( ×100\times 100 ); other rows show retention relative to those scores. B-F1 measures mining-class F1 within three pixels of a ground-truth boundary; OF1 measures mining-class F1 over reconstructed site scenes.
Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: We sample 10610^{6} points globally spanning 100 AlphaEarth COGs randomly across 2017–2025 to measure latency of feature materialization.
  Seconds to answer one query [-1.5pt] A query == one request for 10610^{6} embeddings 
HardwarePFFs held in memoryFirst query Loads all 100 PFFsEach later query Reloads what did not fitThroughput (emb / s)
Cascade Lake 2.8 GHz ⋅\cdot 2 cores ⋅\cdot 8 GB13 / 100178.1167.56.0 k
Sapphire Rapids ⋅\cdot 8 cores ⋅\cdot 31 GB60 / 10057.432.730.6 k
EPYC 7763 ⋅\cdot 32 threads ⋅\cdot 503 GB100 / 10026.514.569.1 k
NVIDIA L4 GPU ⋅\cdot 24 GB100 / 10079.50.911.10 M
Appendix
Table 7: Serving PFFs on hardware-constrained environments. We benchmark serving 10610^{6} AlphaEarth embeddings with PFFs spread over approximately 100 fully valid regions globally across a variety of hardware environments. PFFs are faster to materialize compared to the Earth Engine API, even when run on a 2-core, 8 GB RAM CPU machine, averaging 6k embeddings/second compared to max API throughput of ≈\approx 1800 embeddings/s.
Figure 9: PFF max throughput by hardware. CPUs use 8 threads. Multiplier comparisons are made against the fastest Google Earth Engine API throughput recorded on a virtual machine in us-west1-a .
Figure 10: Geographic coverage of PFFBench . Colored points show selected tile centers; gray points show the remaining eligible pool.
Figure 11: Entropy ranking shifts the distribution toward higher scores. Mean HH is 3.253 nats for PFFBench and 2.804 for constrained random selection (Cohen’s d=1.15d=1.15 ).
Figure 12: Sampling across space, time, and products. (a) We randomly shuffle spatial block–year pairs from Cloud-Optimized GeoTIFFs (COGs), sampling across locations and years without replacement within each pass. (b) CPU workers read all available products for each selected block and year. Each stack has one layer per product; pixels with no observations are excluded. Prefetched blocks fill a fixed GPU buffer, where new blocks replace the oldest. (c) We sample uniformly, with replacement, from all valid pixels in this buffer, using the same coordinates for all active products.
Figure 13: Continually learning new timesteps. Sentinel-2 observations show each region before and during the added year. (a, b) New-year reconstruction and old-year retention for PFF -VM and PFF -K-Planes . Labels report retention.
Figure 14: Continual learning (PFF-VM) qualitative results on the Pacific Palisades burn scar area. Visualized Sentinel-2 (top) and TESSERA (bottom). Time copy and other nonparametric baselines are unable to capture spatial change in 2025 caused by the burn scar. PFF-Append captures this 2025 change while leaving the existing year (2024) unchanged, matching the pretrained result at ≈\approx 2% of its training time. Finetuning the PFF (FT-All) overfits to the spatial change and exhibits catastrophic forgetting on 2024.
Figure 15: Selected regional reconstructions: embeddings and observations. Two examples per view, curated for source structure and geographic/year variety. Each inset covers 2.56 ×\times 2.56 km (256 ×\times 256 cells on the 10m product grid), north up. PFFs and compression rate-matched single-product neural fields occupy 64M parameters. We use per-crop PCA for embeddings, visible RGB for optical, co-pol, cross-pol, and their dB difference with per-crop stretches for SAR (Sentinel-1 ascending and descending orbits). AE scores are angular skill; other scores are channel-averaged R2R^{2} (two polarizations for each Sentinel-1 orbit).
Figure 16: Selected regional reconstructions: native-lattice maps. Two source-curated examples per product, each covering 2.56 ×\times 2.56 km. Original cells retain their native sizes and footprints; every model is queried at those same cell centers. Cells are drawn without value interpolation. Dates identify field queries.

Explore similar work

CardsList
Aug 4, 2026cs.CV

Earth Embeddings

Earth observation is moving from foundation models that users must run themselves toward embedding products that package model feature outputs as reusable data without needing to download and process the imagery used to generate them. Earth embeddings are vectors that summarize locations, image patches, or pixels, letting users analyze compact features instead of repeatedly training or running large models on raw satellite imagery. This chapter explains the main types of Earth embeddings, from implicit location encoders to explicit patch and pixel products, and compares their coverage, resolution, dimensionality, storage cost, licenses, and reproducibility. We review their use in land cover and crop mapping, ecological and hazard modeling, socioeconomic prediction, and semantic search, with evidence on when embeddings improve on conventional features and when pooling, fusion, or spatial transfer limit performance. Two case studies show practical workflows for similarity search and land cover mapping. We close with guidance for choosing, evaluating, storing, compressing, and publishing embeddings, and with open problems in oceanic and atmospheric coverage, uncertainty, and benchmarking.

Adam J. Stewart, Heng Fang, Isaac A. Corley +1

Chair of Data Science in Earth Observation, Technical University of Munich, Munich, Germany · Munich Center for Machine Learning, Munich, Germany · KTH Royal Institute of Technology, Stockholm, Sweden +1

Earth ObservationLand CoverSatellite Imagery
AskPDF
Jul 5, 2026cs.CV

Fields of the Planet: Field Boundary Mapping Beyond 10m

Field-boundary maps support crop monitoring, irrigation planning, and yield estimation, but many smallholder parcels span only a few 10 m Sentinel-2 pixels. We introduce Fields of the Planet (FTP), a 3 m PlanetScope companion to Fields of The World (FTW) that pairs the same polygons, seasonal windows, and train/test splits with 133,168 co-registered PlanetScope patch-window targets across 24 countries. FTP evaluates field delineation as parcel recovery by vectorizing predictions before scoring panoptic quality (PQ), object F1, size-stratified PQ, and meter-scale matched-boundary error. Under matched architectures and training recipes, 3 m imagery raises PQ from 21.0 to 35.5, raises PQ on sub-0.5 ha fields from 5.8 to 15.7, and cuts matched-boundary error from 18.6 m to 7.4 m.

Isaac Corley, Caleb Robinson, Jennifer Marcus +1

Taylor Geospatial · Microsoft AI for Good Research Lab · Arizona State University

Agricultural MonitoringSentinel-2DelineationPanoptic SegmentationBoundariesEarth
AskPDF
Jun 6, 2026cs.LG

Neural Field Tokenizations with Hierarchy and Spatial Locality Priors

Neural fields parameterize data as functions from coordinates to values, providing a unified framework for representation learning across modalities. Existing approaches are dominated by per-sample meta-learning, which scales poorly due to memory-intensive inner-loop optimization. The natural alternative -- feed-forward encoding -- typically introduces modality-specific assumptions, sacrificing the generality that makes learning with neural fields attractive. We argue that locality and hierarchy are useful priors for learning field representations that can be injected without compromising modality-agnosticism. We propose LH-NeF, a framework to learn general-purpose tokenized representations of continuous signals. A locality-preserving hierarchical encoder maps raw coordinate-value field observations to structured tokens, from which the field is reconstructed during training. By replacing meta-learning's inner loop with a single forward pass, LH-NeF uses 42×\times less memory and supports 133×\times larger batches than the strongest modality-agnostic baseline. Across images, 3D shapes, and climate fields, our learned representations match or exceed performance of modality-agnostic, modality-specific, and specialized generative neural field baselines on both reconstruction and downstream tasks.

Alonso Urbano, David W. Romero, Max Zimmer +1

Department for AI in Society, Science, and Technology, Zuse Institute Berlin (ZIB), Germany · Cartesia AI, San Francisco, CA, USA · Institute of Mathematics, Technische Universität Berlin, Germany

Neural FieldsSpatial PriorsHierarchical
AskPDF