🤖 AI

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

This paper introduces HighFM, a foundation model leveraging over 2 TB of high-temporal-resolution SEVIRI satellite imagery and an enhanced masked autoencoding framework to achieve superior performance in real-time cloud masking and active fire detection compared to existing geospatial models.

Original authors: Stella Girtsou, Konstantinos Alexis, Giorgos Giannopoulos, Harris Kontoes

Published 2026-04-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Stella Girtsou, Konstantinos Alexis, Giorgos Giannopoulos, Harris Kontoes

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Earth is a giant, busy stage, and sometimes, dramatic events like wildfires or sudden storms happen in the blink of an eye. To catch these events, we need a security camera system that doesn't just take a photo once a day, but records a video every 15 minutes, 24 hours a day.

This paper introduces HighFM, a new "super-brain" (an Artificial Intelligence) designed specifically to watch this continuous video feed and spot disasters as they happen.

Here is the story of how they built it, explained simply:

1. The Problem: The "Slow-Motion" Camera

For years, scientists have used AI to look at Earth from space. But most of these AI models are like tourists with a camera. They take beautiful, high-definition photos of a mountain or a forest, but they only visit once every few days.

The Issue: If a fire starts and spreads in 30 minutes, a tourist camera misses it entirely. By the time the next photo is taken, the fire might be huge, or the smoke might have cleared.
The Need: We need a security guard who never blinks, watching the same spot every 15 minutes to catch fast-moving dangers.

2. The Solution: The "24-Hour Security Guard" (HighFM)

The authors built HighFM using data from a special satellite called MSG (Meteosat Second Generation).

The Data: This satellite sits in a fixed spot above the Earth and takes pictures of Europe, Africa, and the Middle East every 15 minutes. It's like a security camera that never sleeps.
The Challenge: These pictures aren't super sharp (they are a bit blurry compared to tourist photos), but they are fast. The AI needed to learn how to understand "blurry but fast" instead of "sharp but slow."

3. How They Taught the AI: "The Fill-in-the-Blanks Game"

You can't teach an AI by showing it a million labeled pictures of fires (because there aren't enough labeled fires). Instead, they used a clever trick called Self-Supervised Learning.

Imagine you have a 1,000-page comic book, but someone ripped out random pages.

The Game: The AI looks at the pages that are still there and tries to guess what the ripped-out pages should look like.
The Result: By playing this "Fill-in-the-Blanks" game with 2 Terabytes of satellite data (that's like a library of millions of hours of video), the AI learned the "grammar" of the sky. It learned how clouds move, how smoke looks, and how a fire starts to glow, all without a human teacher telling it what to look for.

4. The Secret Sauce: "Time Travel"

Most AI models treat time like a still photo. HighFM is different. The authors realized that for fast events, time is the most important clue.

The Analogy: If you see a single frame of a runner, you don't know if they are walking or sprinting. But if you see three frames in a row, you know exactly how fast they are moving.
The Innovation: HighFM was taught to look at three snapshots in a row (15 minutes apart). This allowed it to learn the speed and direction of clouds and fires, making it much better at predicting what will happen next.

5. The Test: Can It Spot Fires and Clouds?

Once the AI finished its "training," they tested it on two real-world jobs:

Finding Clouds: To help solar power companies know when the sun will be blocked.
Finding Fires: To help firefighters know where a wildfire is starting.

They compared HighFM against other smart models (the "tourists" and the "generic AI").

The Result: HighFM won. It was better at spotting tiny, hidden fires and tracking fast-moving clouds.
Why? Because it was trained on the right kind of data (fast, continuous video) rather than just high-quality photos.

6. Why This Matters

Think of HighFM as a super-early warning system.

For Firefighters: It can spot a tiny spark before it becomes a massive inferno, giving them precious extra minutes to act.
For Solar Power: It can predict exactly when clouds will block the sun, helping energy companies manage the grid.
For Everyone: It shows us that to protect our planet from fast-changing disasters, we don't just need sharper eyes; we need faster ones.

In a nutshell: The authors built a specialized AI that learned to watch the Earth's "live stream" instead of its "photo album," proving that speed and continuity are the keys to catching disasters before they get out of hand.

1. Problem Statement

The increasing frequency of climate-related disasters (wildfires, floods, storms) demands real-time, near-real-time Earth Observation (EO) systems. While Foundation Models (FMs) have revolutionized EO by enabling self-supervised pretraining on large datasets, existing models suffer from a critical limitation:

Temporal Sparsity: Most current EO FMs (e.g., SatMAE, Prithvi) are trained on high-resolution, low-revisit satellite imagery (e.g., Sentinel-2, Landsat) with revisit times of days.
Inapplicability to Fast Phenomena: These models are unsuitable for rapidly evolving events like wildfire spread or storm development, which require minute-level monitoring.
Gap: There is a lack of foundation models specifically designed for high-temporal-resolution, geostationary data (e.g., Meteosat Second Generation/SEVIRI), which offers 5–15 minute revisit rates but lower spatial resolution.

2. Methodology

The authors propose HighFM, a novel approach to adapt the SatMAE (Satellite Masked Autoencoder) framework for high-frequency geostationary data.

A. Data Curation

Pretraining Dataset: Utilized >2 TB of SEVIRI imagery from the MSG platform (2014–2019) covering the Mediterranean region.
- Resolution: 11 spectral bands, 15-minute cadence.
- Preprocessing: Scenes were split into non-overlapping $32 \times 32$ patches. Patches containing only ocean or fully cloud-covered areas were discarded.
- Split: Strict time-based split (2014–2018 for pretraining; 2019 for validation/testing) to prevent temporal data leakage.
Fine-tuning Datasets: Curated datasets for two downstream tasks (2020–2024):
1. Cloud Segmentation: Paired with MODIS cloud masks (MOD35/MYD35).
2. Active Fire Detection: Paired with NASA FIRMS detections, cross-validated with EFFIS and FireHUB burned-area products.
- Note: Fire pixels are extremely rare (<0.4% of total pixels), creating severe class imbalance.

B. Model Architecture (Adapted SatMAE)

The authors modified the standard SatMAE architecture to handle the specific characteristics of geostationary data:

Enhanced Temporal Encoding: Unlike standard SatMAE which discards fine-grained time (minutes/seconds) as irrelevant for slow phenomena, HighFM retains minute-level temporal encodings. This is critical for capturing short-term variability in fast-evolving events.
Input Variants: Two pretraining strategies were tested:
1. Single-Timestep (HighFMST): One SEVIRI acquisition per sample.
2. Multi-Timestep (HighFMMT): Three SEVIRI acquisitions from the same patch within the same hour, allowing the model to learn intra-hour dynamics.
Tokenization: Due to the coarse resolution of SEVIRI, a $4 \times 4$ token embedding was used (smaller than standard) to preserve spatial detail and avoid oversmoothing.
Backbone: Vision Transformer (ViT-Base, ~90M parameters).

C. Fine-tuning Strategy

The pretrained encoders were fine-tuned for binary semantic segmentation using two distinct optimization objectives to address different operational needs:

High-Recall Detection (Cross-Entropy Loss): Prioritizes capturing all positive pixels (critical for early warning), tolerating false positives.
Precision-Oriented Localization (Dice Loss): Prioritizes spatial accuracy and low false positives (critical for operational trust and resource allocation).

3. Key Contributions

First High-Frequency EO FM: Introduces the first foundation model specifically tailored for high-frequency, geostationary EO data, addressing the gap between sparse temporal models and real-time monitoring needs.
Architectural Innovation: Adapts SatMAE by incorporating fine-grained temporal encodings and multi-timestep inputs, enabling the model to learn short-term atmospheric and fire dynamics.
Large-Scale Dataset: Curated a massive, self-supervised pretraining dataset (2.23 TB) and two specialized fine-tuning datasets for cloud and fire detection.
Operational Co-Design: Developed in collaboration with stakeholders (fire brigades, civil protection, meteorological institutes) to define realistic error trade-offs (recall vs. precision) and evaluation protocols.

4. Experimental Results

The model was benchmarked against strong baselines: UNet (from scratch), ViT (from scratch), ViT (ImageNet pretrained), Copernicus-FM, and Panopticon.

Cloud Segmentation:
- HighFMMT achieved the best overall performance with a Balanced Accuracy of 0.831 and superior IoU scores for both cloud and no-cloud classes compared to all baselines.
- The multi-timestep variant consistently outperformed the single-timestep variant, proving the value of temporal context.
Active Fire Detection:
- HighFMMT significantly outperformed all baselines.
- Cross-Entropy Setting: Achieved 0.925 Balanced Accuracy and 0.890 Fire Recall, surpassing the next best model (Copernicus-FM) by +0.045 in recall.
- Dice Loss Setting: Achieved the highest Fire IoU (0.352) and Fire Recall (0.497), demonstrating the ability to provide precise segmentation without sacrificing detection capability.
- Qualitative Analysis: HighFM produced compact, well-localized fire predictions with significantly fewer false positives and less "over-segmentation" compared to baselines, which tended to activate large irrelevant regions.

5. Significance and Impact

Real-Time Disaster Response: HighFM demonstrates that geostationary data, despite lower spatial resolution, is highly effective for tracking fast-evolving disasters when paired with a temporally dense foundation model.
Scalability: The framework is not limited to fire and clouds; it offers a scalable path for other near-real-time applications such as storm tracking, solar energy forecasting, and severe weather nowcasting.
Domain Adaptation: The results validate that domain-specific pretraining on relevant sensor data (SEVIRI) yields superior performance compared to generic FMs pretrained on high-resolution but temporally sparse data (Sentinel/Landsat).
Future Direction: While the current work focuses on a single sensor, the authors identify the need for future multimodal pretraining (integrating MODIS, VIIRS, SAR) to create a unified, resolution-agnostic real-time monitoring framework.