Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 62 of the 62 references it cites were found in a registry or an index; 25 of the 34 sources its claims cite were checked for support.

Prithvi-EO-2.0: A Versatile Multitemporal Foundation Model for Earth Observation Applications

Open access

Abstract

This paper presents Prithvi-EO-2.0, a new geospatial foundation model that offers significant improvements over its predecessor, Prithvi-EO-1.0. Trained on 4.2 million global time series samples from NASA’s Harmonized Landsat and Sentinel-2 data archive at 30-m resolution, the new model incorporates temporal and location embeddings for enhanced performance across various geospatial tasks. Through extensive benchmarking with GEO-Bench, the model outperforms the previous Prithvi-EO model by 8% across a range of tasks. It also outperforms six other geospatial foundation models when benchmarked on remote sensing tasks from different domains and resolutions (i.e. from 0.1 m to 15 m). The results demonstrate the versatility of the model in both classical Earth observation and high-resolution applications. Early involvement of end-users and subject matter experts (SMEs) allowed constant feedback on model and dataset design, enabling customization across diverse SME-led applications in disaster response, land cover and crop mapping, and ecosystem dynamics monitoring. Prithvi-EO-2.0 is available as an open-source model on Hugging Face and IBM TerraTorch, with additional resources on GitHub. The project exemplifies the Trusted Open Science approach embraced by all involved organizations.

Overview

CoChat Summary

TL;DR

Prithvi-EO-2.0 is a 300M/600M-parameter geospatial foundation model pretrained with MAE on 4.2M global HLS 30-m multitemporal samples, adding temporal and location embeddings. It improves over Prithvi-EO-1.0 by up to 8% on GEO-Bench and outperforms six other GFMs across benchmark and SME-led downstream tasks.

Key findings

  • Pretrained on 4.2M training and 46k validation 256×256 samples from NASA HLS at 30-m resolution, using four timestamps per sample; released in 300M and 600M parameter versions.
  • On GEO-Bench, Prithvi-EO-2.0-600M-TL and Prithvi-EO-2.0-600M achieved the best combined performance across all 12 datasets; all Prithvi-EO-2.0 versions outperformed Prithvi-EO-1.0, with up to 8% improvement overall and a 3% overall gain for the 100M model over its predecessor.
  • Downstream gains include Sen1Floods11 water IoU 83.1 for Prithvi-EO-2.0-600M-TL vs. 79.6 for Prithvi-EO-1.0-100M; wildfire burn-scar IoU 83.2 vs. 76.8; L4S landslide mIoU 67.0 with only 50 training images for 600M vs. 59.7 for U-Net; US crop segmentation mIoU 50.7 vs. 42.6 for U-Net; and PASTIS 100% mIoU 53.4 vs. 52.9 for U-TAE.
  • For AGB estimation, the best Prithvi-EO-2.0-300M configuration reached test RMSE 33.40 with 12 Sentinel-2 timestamps and LoRA, while the baseline achieved 27.49; for GPP, Prithvi-EO-2.0-600M-TL reached R² up to 0.88 in 2018, outperforming random forest, XGBoost, and ResNet18 baselines.

Methods

  • Pretraining used a masked autoencoder with 3D patch and positional embeddings, plus learned time and location metadata biases with random dropping; models were ViT-L (300M) and ViT-H (600M), trained for 400 epochs on A100 GPUs, using 80 GPUs (~21,000 GPU-hours) for 300M and 240 GPUs (~58,000 GPU-hours) for 600M.
  • The pretraining dataset was built from HLS six-band reflectance data (Blue, Green, Red, NIR, SWIR1, SWIR2) spanning 2014–2023, sampled by LULC class and ecoregion, with sequences of four images separated by one to six months, filtered for clouds and missing data, and split into 256×256 patches.
  • Evaluation combined GEO-Bench benchmarking across 12 classification and segmentation datasets with 10 hyperparameter trials and 10 random seeds, plus SME-led downstream tasks in disaster response, land cover and crop mapping, and ecosystem dynamics; fine-tuning experiments required a single GPU.

Limitations

  • Stated: future work is needed to compare with multimodal GFMs and to identify the benefits and limitations of incorporating additional modalities at the fine-tuning stage.
  • Stated: incorporating SAR inputs did not improve biomass estimation in their setting, and adapting Prithvi-EO to SAR is challenging because of fundamental optical–SAR differences; fusing a separate SAR encoder with the pretrained Prithvi encoder could help.
  • Evident: pretraining and the main benchmark are optical multispectral HLS/30-m focused, some GEO-Bench datasets lack temporal and location information, and the combined GEO-Bench score averages heterogeneous metrics (accuracy, F1, mIoU) into a single aggregate.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/tgrs.2025.3642610

Cites

62 of 62 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Red, NIR, SWIR1, and SWIR2. This band combination will be referred to as the six-band HLS reflectance data.
Red, NIR, SWIR1, and SWIR2. This band combination will be referred to as the six-band HLS reflectance data.CC-BY · From the PDF
Fig. 2 Global HLS dataset distribution visualized on a tile-level. The number of training samples are color-coded in blue to green, while validation tiles are visualized in magenta.
Fig. 2 Global HLS dataset distribution visualized on a tile-level. The number of training samples are color-coded in blue to green, while validation tiles are visualized in magenta.CC-BY · From the PDF
Fig. 3 Prithvi-EO-2.0 architecture and general pretraining framework.
Fig. 3 Prithvi-EO-2.0 architecture and general pretraining framework.CC-BY · From the PDF
Fig. 4 Locations of the burn intensity dataset used in training and testing.
Fig. 4 Locations of the burn intensity dataset used in training and testing.CC-BY · From the PDF
Fig. 5 Workflow to fine-tune Prithvi-EO-2.0 for predicting GPP at flux observations sites. The model consists of two branches: the (frozen) Prithvi encoder generating HLS embeddings later processed by a simple decoder, and convolutional layers transforming MERRA-2 atmospheric and land surface variables. The decoder and convolved MERRA-2 outputs are concatenated and processed with a linear layer to predict GPP.
Fig. 5 Workflow to fine-tune Prithvi-EO-2.0 for predicting GPP at flux observations sites. The model consists of two branches: the (frozen) Prithvi encoder generating HLS embeddings later processed by a simple decoder, and convolutional layers transforming MERRA-2 atmospheric and land surface variables. The decoder and convolved MERRA-2 outputs are concatenated and processed with a linear layer to predict GPP.CC-BY · From the PDF
Fig. 6 Aggregated performance across (a) all GEO-Bench datasets, (b) all classification tasks, and (c) all segmentation tasks. The Prithvi-EO-2.0 models are highlighted in blue.
Fig. 6 Aggregated performance across (a) all GEO-Bench datasets, (b) all classification tasks, and (c) all segmentation tasks. The Prithvi-EO-2.0 models are highlighted in blue.CC-BY · From the PDF
Fig. 7 Distribution of performance across all models from 10 repeated runs for the GEO-Bench (a) classification and (b) segmentation tasks in terms of accuracy (F1 score for m-bigearthnet) and mIoU, respectively. We list input sensor and resolution for each dataset.
Fig. 7 Distribution of performance across all models from 10 repeated runs for the GEO-Bench (a) classification and (b) segmentation tasks in terms of accuracy (F1 score for m-bigearthnet) and mIoU, respectively. We list input sensor and resolution for each dataset.CC-BY · From the PDF
Image 8
CC-BY · From the PDF
Fig. 9 Comparison of AGB predictions versus observed AGB on all inputs in the test set for (a) fine-tuned Prithvi-EO-2.0300M and (b) baseline model. Prithvi-EO-2.0-300M is finetuned and tested using 12 timestamps of Sentinel-2 data. A log scale is used for visualization.
Fig. 9 Comparison of AGB predictions versus observed AGB on all inputs in the test set for (a) fine-tuned Prithvi-EO-2.0300M and (b) baseline model. Prithvi-EO-2.0-300M is finetuned and tested using 12 timestamps of Sentinel-2 data. A log scale is used for visualization.CC-BY · From the PDF
Fig. 10 Comparison of fine-tuned Prithvi and baseline model prediction for 4 images and their corresponding AGB target pairs from the test set. Columns from left to right: RGB visualization of Sentinel-2 scene from the same month as the AGB observation, target AGB, baseline model prediction, and fine-tuned Prithvi prediction. As high-quality Sentinel-2 data does not exist for all AGB targets, some S2 inputs, such as the one in row 4, are either obscured by clouds or contain partial data.
Fig. 10 Comparison of fine-tuned Prithvi and baseline model prediction for 4 images and their corresponding AGB target pairs from the test set. Columns from left to right: RGB visualization of Sentinel-2 scene from the same month as the AGB observation, target AGB, baseline model prediction, and fine-tuned Prithvi prediction. As high-quality Sentinel-2 data does not exist for all AGB targets, some S2 inputs, such as the one in row 4, are either obscured by clouds or contain partial data.CC-BY · From the PDF
Fig. 11 Train-test set R 2 for GPP flux estimation using leave-one-year-out cross-validation over 975 instances of HLS and MERRA-2 data using Prithvi-EO-2.0-600M-TL. Each row shows one of the leave-one-year-out train-test split.
Fig. 11 Train-test set R 2 for GPP flux estimation using leave-one-year-out cross-validation over 975 instances of HLS and MERRA-2 data using Prithvi-EO-2.0-600M-TL. Each row shows one of the leave-one-year-out train-test split.CC-BY · From the PDF

Verification

Claims the paper makes

17 Valid · 9 Not valid · 3 Not checked · 125 Own results and interpretations

69 more claims · 6 with their cited sources checked

Create a free account to see all 154 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 38
  • It may state two separate claims as one. 56
  • The quote only partly states this claim. 3

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 34 references

Research works printed but not linked: 1

  1. Napper, 2006, Burned Area Emergency Response Treatments Catalog

Cited by

See all 100

Metadata

Authors
Szwarcman, D., Roy, S., Fraccaro, P., Gislason, T. E., Blumenstiel, B., Ghosal, R., de Oliveira, P. H., de Sousa Almeida, J. L., Sedona, R., Kang, Y., Chakraborty, S., Wang, S., Gomes, C., Kumar, A., Gaur, V., Truong, M., Godwin, D., Khallaghi, S., Lee, H., Hsu, C.-Y., Asanjan, A. A., Mujeci, B., Shidham, D., Balogun, R. O., Kolluru, V., Keenan, T., Arevalo, P., Li, W., Alemohammad, H., Olofsson, P., Mayer, T., Hain, C., Kennedy, R., Zadrozny, B., Bell, D., Cavallaro, G., Watson, C., Maskey, M., Ramachandran, R., & Moreno, J. B.
Affiliations
  • IBM (United States) (US)
  • Marshall Space Flight Center (US)
  • University of Iceland
  • University of Alabama in Huntsville (US)
  • IBM Research
  • Forschungszentrum Jülich (DE)
  • Virginia Tech (US)
  • Universities Space Research Association (US)
  • Arizona State University (US)
  • Oregon State University (US)
  • Clark University (US)
  • University of California, Berkeley (US)
  • Boston University
  • University of Iceland (IS)
Venue
IEEE Transactions on Geoscience and Remote Sensing
Published
2026
Volume
64
Pages
1-20
ISSN
  • 0196-2892 (ISSN-L)
  • 1558-0644
OpenAlex
W4417248811
Cited by
100
References
34
Citations by year
2026: 81
FWCI
40.73
Type
journal article
Language
en
Open access
Locations
  • IEEE Transactions on Geoscience and Remote Sensing (publishedVersion, cc-by)
  • JuSER (Forschungszentrum Jülich) (publishedVersion, other-oa)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Funders
  • National Aeronautics and Space Administration (80MSFC22M004)
  • Gauss Centre for Supercomputing e.V
  • NASA (80NSSC24K1562)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Remote-Sensing Image Classification
  • Physical Sciences › Environmental Science › Ecology › Remote Sensing in Agriculture
  • Social Sciences › Geography, Planning and Development › Remote Sensing and Land Use
Keywords
geospatial foundation models, Landsat, Sentinel-2, temporal embeddings, location embeddings, high-resolution remote sensing, disaster response, land cover mapping, crop mapping, open-source models

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Prithvi-EO-2.0: A Versatile Multitemporal Foundation Model for Earth Observation Applications. https://app.cochat.ai/papers/10.1109/tgrs.2025.3642610
The paper
Szwarcman, D., Roy, S., Fraccaro, P., Gislason, T. E., Blumenstiel, B., Ghosal, R., de Oliveira, P. H., de Sousa Almeida, J. L., Sedona, R., Kang, Y., Chakraborty, S., Wang, S., Gomes, C., Kumar, A., Gaur, V., Truong, M., Godwin, D., Khallaghi, S., Lee, H., . . . Moreno, J. B. (2026). Prithvi-EO-2.0: A Versatile Multitemporal Foundation Model for Earth Observation Applications. IEEE Transactions on Geoscience and Remote Sensing. https://doi.org/10.1109/tgrs.2025.3642610

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.