HighFM: Towards a Foundation Model for Learning Representations from High-Frequency Earth Observation Data

Stella Girtsou; Konstantinos Alexis; Giorgos Giannopoulos; Charalambos Kontoes

cs.CVApr 5, 2026

HighFM: Towards a Foundation Model for Learning Representations from High-Frequency Earth Observation Data

Authors: Stella Girtsou, Konstantinos Alexis, Giorgos Giannopoulos, Charalambos Kontoes

Organizations: National Observatory of Athens · National Technical University of Athens · National and Kapodistrian University of Athens · Athena Research Center

Abstract

The increasing frequency and severity of climate related disasters have intensified the need for real time monitoring, early warning, and informed decision-making. Earth Observation (EO), powered by satellite data and Machine Learning (ML), offers powerful tools to meet these challenges. Foundation Models (FMs) have revolutionized EO ML by enabling general-purpose pretraining on large scale remote sensing datasets. However most existing models rely on high-resolution satellite imagery with low revisit rates limiting their suitability for fast-evolving phenomena and time critical emergency response. In this work, we present HighFM, a first cut approach towards a FM for high temporal resolution, multispectral EO data. Leveraging over 2 TB of SEVIRI imagery from the Meteosat Second Generation (MSG) platform, we adapt the SatMAE masked autoencoding framework to learn robust spatiotemporal representations. To support real time monitoring, we enhance the original architecture with fine grained temporal encodings to capture short term variability. The pretrained models are then finetuned on cloud masking and active fire detection tasks. We benchmark our SEVIRI pretrained Vision Transformers against traditional baselines and recent geospatial FMs, demonstrating consistent gains across both balanced accuracy and IoU metrics. Our results highlight the potential of temporally dense geostationary data for real-time EO, offering a scalable path toward foundation models for disaster detection and tracking.

Figures & tables

Figure 1: High level flowchart of our methodology

Split	Images	Background Pixels	Target Pixels	Target Ratio
Pretraining dataset
Train	13.6M	–	–	–
Val	5.3M	-	–	–
Test	1.3M	–	–	–
Active Fires dataset
Train	2140	2.18M	7.77k	0.00355

Table 1: Dataset Statistics for the HighFM model

Models trained with cross-entropy loss
Method	Balanced Accuracy	IoU no-cloud	IoU cloud	Recall no-cloud	Recall cloud
U-Net scratch	$0.792\pm 0.002$	$0.618\pm 0.005$	$0.701\pm 0.003$	$0.743\pm 0.014$	$\mathbf{0.842\pm 0.010}$
ViT-B/4 scratch	$0.826\pm 0.001$	$0.677\pm 0.002$	$0.727\pm 0.008$	$\underline{0.827\pm 0.018}$	$0.826\pm 0.019$
ViT-B/4 ImageNet	$0.819\pm 0.002$	$0.667\pm 0.002$	$0.713\pm 0.008$	$\mathbf{0.829\pm 0.013}$	$0.808\pm 0.016$
Copernicus-FM	$0.825\pm 0.001$	$0.672\pm 0.002$	$0.730\pm 0.006$	$0.809\pm 0.014$	$\underline{0.840\pm 0.015}$
Panopticon	$0.826\pm 0.002$	$0.674\pm 0.003$	$0.730\pm 0.004$	$0.815\pm 0.013$	$0.837\pm 0.012$

Table 2: Test set performance of cloud segmentation models trained without data augmentation. Results are reported as mean

\pm

std over 5 runs.

Models trained with cross-entropy loss
Method	Balanced Accuracy	IoU no-fire	IoU fire	Recall no-fire	Recall fire
U-Net scratch	$0.834\pm 0.004$	$0.866\pm 0.007$	$0.022\pm 0.001$	$0.867\pm 0.007$	$0.802\pm 0.007$
ViT-B/4 scratch	$0.883\pm 0.006$	$0.937\pm 0.010$	$0.048\pm 0.006$	$0.937\pm 0.010$	$0.829\pm 0.019$
ViT-B/4 ImageNet	$0.856\pm 0.010$	$0.942\pm 0.010$	$0.049\pm 0.007$	$0.943\pm 0.010$	$0.769\pm 0.027$
Copernicus-FM	$0.894\pm 0.003$	$0.943\pm 0.005$	$0.053\pm 0.004$	$0.943\pm 0.005$	$0.845\pm 0.007$
Panopticon	$0.888\pm 0.019$	$0.928\pm 0.013$	$0.044\pm 0.008$	$0.929\pm 0.013$	$0.848\pm 0.028$

Table 3: Test set performance of fire detection models trained with data augmentation. Results are reported as mean

\pm

std over 5 runs.

Figure 3: Qualitative results on test samples. Rows 1 and 2 show models fine-tuned on the cloud segmentation task, while rows 3 and 4 correspond to the fire detection task. Rows 1 and 3 present outputs from models trained with cross-entropy loss, whereas rows 2 and 4 show results from models trained with Dice loss. SEVIRI input images are visualized using a false-color fire composite, consistent with EUMETSAT’s operational fire RGB products. In the prediction columns, true positives are shown in green, false positives in red, and false negatives in blue.

Geospatial Foundation Models

Geospatial foundation models turn satellite, map, and Earth-system data into reusable representations for spatial analysis and prediction.

See all 123 papers

Scalable and Trustworthy Earth Observation Foundation Models
Jul 8, 2026Syed Usama Imtiaz, Mitra Nasr Azadani, Nasrin AlamdariGeospatial Foundation Models Vision Foundation Model Adaptation
Department of Civil and Environmental Engineering, Florida State University
TerraFlow: Multimodal, Multitemporal Representation Learning for Earth Observation
Mar 13, 2026Nazar Puriy, Johannes Jakubik, Benedikt Blumenstiel +1Representation Learning Geospatial Foundation Models
IBM Research Europe, Zurich, Switzerland · ETH Zurich, Zurich, Switzerland
Earth System Foundation Model (ESFM): A unified framework for heterogeneous data integration and forecasting
Apr 20, 2026Firat Ozdemir, Yun Cheng, Salman Mohebi +11Geospatial Foundation Models Climate Science
Swiss Data Science Center (SDSC), ETH Zurich & EPFL, Zurich & Lausanne, Switzerland · ETH AI Center, ETH Zurich, Zurich, Switzerland · ETH Zurich, Zurich, Switzerland +1

Jul 8, 2026J/K move · Enter open · S save

Scalable and Trustworthy Earth Observation Foundation Models

Syed Usama Imtiaz, Mitra Nasr Azadani, Nasrin Alamdari

Department of Civil and Environmental Engineering, Florida State University

Geospatial Foundation Models Vision Foundation Model Adaptation Remote Sensing Remote Sensing Benchmarks

Open paper AskPDF arXiv

Foundation models (FMs) have transformed machine learning from isolated task-specific model development toward general-purpose models pretrained on broad data and adapted to multiple downstream tasks. Earth observation (EO) is an important domain for this paradigm because satellite and airborne archives are large, high-revisit, and increasingly multimodal, while reliable field labels are often sparse. Remote sensing foundation models (RSFMs) cannot be transferred reliably/optimally without domain-specific adaptation. This is because EO data are governed by measurement physics and operational decision constraints. This chapter reviews the design principles arising from these domain-specific constraints. It first defines the FMs paradigm in remote sensing (RS), then synthesizes the current model landscape, pretraining objectives, architecture designs, downstream adaptation and trustworthiness requirements. The chapter also incorporates recent benchmark evidence showing that no single geospatial foundation model is universally best and that inconsistent evaluation remains a major issue to fair comparison and reliable deployment. In addition, two brief environmental monitoring case studies; physics-informed spectral targeted masking for harmful algal bloom prediction and reinforcement learning for adaptive environmental monitoring station selection to illustrate the FMs domain-guided principles in practice. This chapter posits that next-generation RSFMs should be evaluated not only by benchmark accuracy, but also by modality-aware transfer and physically plausible representations for trustworthy EO decisions.

Abstract

Geospatial Foundation Models

Explore similar work

Scalable and Trustworthy Earth Observation Foundation Models

TerraFlow: Multimodal, Multitemporal Representation Learning for Earth Observation

Earth System Foundation Model (ESFM): A unified framework for heterogeneous data integration and forecasting