🤖 machine learning

When Earth Foundation Models Meet Diffusion: An Application to Land Surface Temperature Super-Resolution

This paper proposes Earth Foundation Model-guided Diffusion (EFDiff), a novel framework that leverages the Prithvi-EO-2.0 model to guide diffusion-based super-resolution of Land Surface Temperature, achieving superior performance over baseline methods in reconstructing fine-scale structures from highly degraded thermal observations.

Original authors: Yiheng Chen, Zihui Ma, Peishi Jiang, Yilong Dai, Qikai Hu, Xinyue Ye, Lingyao Li, Rita Sousa, Runlong Yu

Published 2026-04-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yiheng Chen, Zihui Ma, Peishi Jiang, Yilong Dai, Qikai Hu, Xinyue Ye, Lingyao Li, Rita Sousa, Runlong Yu

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 you are trying to restore an old, blurry family photo. The photo is so pixelated that you can't tell if the person in the background is holding a dog or a cat, or if the grass is green or brown. You know the general shape of the person, but the details are lost.

This is exactly the problem scientists face with Land Surface Temperature (LST) data from space.

The Problem: The "Blurry Satellite"

Satellites like MODIS give us a global view of Earth's temperature every day, but the images are very "chunky" (about 1 kilometer per pixel). It's like looking at a city from a plane: you see the general heat of the neighborhood, but you can't tell if a specific house is hot because of an air conditioner or a barbecue.

Other satellites, like Landsat, give us incredibly sharp, detailed views (30 meters per pixel), but they only fly over the same spot once every 16 days.

The Challenge: We need a way to take the "chunky" daily heat map and magically fill in the missing details to look like the sharp, 16-day map, but every day. This is called Super-Resolution.

The Old Way: Guessing the Average

Previous methods tried to solve this by acting like a strict accountant. They looked at the blurry data and calculated the "average" temperature for every tiny spot.

The Result: The image looked smooth and safe, but it was boring and blurry. It was like trying to paint a forest by mixing all the greens into one muddy color. You lost the sharp edges between a river and a forest, or a city street and a park.

The New Solution: EFDiff (The "Smart Artist")

The authors of this paper created a new system called EFDiff. Think of it as a team of two experts working together to restore that blurry photo:

The "Earth Foundation Model" (The Knowledgeable Librarian):
This is a super-smart AI that has read millions of satellite images of Earth. It knows that "if I see a dark green patch next to a blue patch, that's usually a forest next to a lake." It doesn't just see pixels; it understands geography.
- Analogy: Imagine a librarian who has memorized every map of the world. When you show them a blurry spot, they instantly know, "Ah, that's a mountain range, so the temperature should drop sharply here."
The "Diffusion Model" (The Creative Artist):
This is a type of AI that is great at imagining possibilities. Instead of just calculating an average, it knows that there are many ways a scene could look. It can "hallucinate" (in a good way) realistic details like the texture of a roof or the ripple of a lake.
- Analogy: Imagine an artist who is great at painting, but needs a prompt. If you just say "paint a forest," they might paint a generic one. But if you say "paint a forest next to a lake with a specific rock formation," they can create something incredibly realistic.

How They Work Together

In the past, the "Artist" (Diffusion) and the "Librarian" (Foundation Model) worked in separate rooms. The Artist tried to guess the details based on the blurry photo alone, often getting it wrong.

EFDiff puts them in the same room.

The Librarian looks at the sharp, high-resolution satellite photos (which show trees, water, and buildings) and whispers to the Artist: "Hey, this blurry spot is actually a city park next to a highway. Make sure you paint the sharp temperature difference between the hot asphalt and the cool grass."
The Artist uses this advice to generate a highly detailed, realistic temperature map that isn't just a guess, but a smart reconstruction based on real-world geography.

The Two Versions: "The Realist" vs. "The Accurate One"

The researchers built two slightly different versions of this system to handle different needs:

EFDiff-ϵ (The Realist):
This version takes its time. It generates many possible versions of the scene and picks the one that looks the most "real" and textured. It's great for seeing the feel of the landscape, even if a single pixel's temperature is slightly off.
- Metaphor: Like a photographer who takes a slightly noisy but incredibly sharp and detailed photo. You can see the individual leaves on the trees.
EFDiff-x0 (The Accurate One):
This version is faster and focuses on getting the exact numbers right. It's better at matching the precise temperature values, making it great for scientific calculations where exact numbers matter.
- Metaphor: Like a surveyor who measures the temperature with a laser. It might look a bit smoother, but the numbers are spot-on.

Why This Matters

The paper tested this on a massive scale (32 times sharper than the original blurry image) across the whole globe.

The Result: EFDiff was much better than the old methods. It didn't just smooth things out; it correctly identified where the temperature changes sharply (like the edge of a lake or a city block).
The Big Win: It works best in complex places (like cities or forests) where the old methods failed. It proved that giving AI "geographic common sense" (via the Foundation Model) helps it solve problems that pure math cannot.

In short: This paper teaches computers how to use their "geographic brain" to fill in the missing details of Earth's temperature, turning a blurry, low-res map into a crystal-clear, daily weather forecast for the entire planet.

1. Problem Definition

The paper addresses the challenge of Land Surface Temperature (LST) Super-Resolution (SR), specifically under extreme spatial degradation (a 32× scale factor).

The Gap: Operational thermal sensors (e.g., MODIS) offer high temporal resolution but coarse spatial resolution (~1 km), while high-resolution sensors (e.g., Landsat) offer ~100 m resolution but with low temporal frequency. Reconstructing fine-scale (30 m) thermal fields from coarse observations is critical for applications like urban heat island monitoring and drought assessment.
The Challenge: This is a severely ill-posed inverse problem. A single coarse thermal pixel aggregates multiple land-cover types, moisture conditions, and microclimates, discarding fine-scale boundaries.
Limitations of Existing Methods:
- Deterministic Models: Tend to predict conditional means, resulting in blurred transitions and attenuated local extremes.
- Regional Bias: Most models are trained on specific regions, limiting transferability across diverse climates and biomes.
- Weak Conditioning: Current methods rely on hand-crafted variables or raw multispectral bands (HLS), which lack the high-level semantic understanding needed to disambiguate sub-pixel mixtures.

2. Methodology: EFDiff Framework

The authors propose Earth Foundation Model-guided Diffusion (EFDiff), a framework that couples generative diffusion models with pretrained Earth Foundation Models (EFMs).

A. Core Architecture

The framework operates in residual space ( $R = Y - \tilde{X}$ ), where $Y$ is the target HR LST and $\tilde{X}$ is the upsampled coarse LST. The goal is to predict the fine-scale residual.

Conditioning via Earth Foundation Models (EFM):
- Instead of feeding raw multispectral bands directly to the network, the authors use a frozen Prithvi-EO-2.0 encoder (a ViT-Large model pretrained on global HLS imagery).
- This encoder transforms high-resolution multispectral reflectance into geospatial embeddings ( $Z$ ). These embeddings capture high-level semantics (vegetation boundaries, urban fabric, water bodies) that correlate with thermal structures.
- Mechanism: These embeddings are injected into the diffusion denoising network via cross-attention layers. This allows the model to use global, semantically rich context to guide the reconstruction of fine details.
Denoising Network:
- A UNet architecture is used.
- The upsampled coarse thermal field ( $\tilde{X}$ ) is concatenated with the noisy residual state at the input.
- The geospatial embeddings ( $Z$ ) are fused into the decoder and bottleneck layers via cross-attention, allowing the network to attend to relevant land-surface features during the upsampling process.
Two Diffusion Variants:
The paper studies two formulations to trade off between perceptual realism and pixel-level fidelity:
- EFDiff-ϵ: Follows the standard DDPM (Variance-Preserving) framework. It predicts the noise ( $\epsilon$ ) added at each step. It uses stochastic sampling (up to 1000 steps) to generate highly realistic textures.
- EFDiff-x0: Follows the ResShift (Transition-based) framework. It directly predicts the clean residual ( $x_0$ ). It allows for much faster inference (fewer steps) and achieves superior pixel-level fidelity (lower RMSE).

3. Key Contributions

Novel Framework (EFDiff): The first integration of Earth Foundation Models with diffusion-based generative modeling for LST super-resolution. It demonstrates how frozen, pretrained geospatial representations can serve as strong priors for generative reconstruction.
Dual-Formulation Design: Introduction of EFDiff-ϵ and EFDiff-x0, providing practitioners with a choice between high perceptual realism (stochastic) and high pixel-level fidelity (deterministic/fast).
Global Benchmark: Creation of a diverse, globally distributed benchmark comprising 242,416 co-registered Landsat thermal-reflectance patches across 3,094 MGRS tiles, evaluated under a unified 32× protocol.
Superior Conditioning Strategy: Demonstration that EFM cross-attention conditioning significantly outperforms traditional HLS channel concatenation, particularly in spatially heterogeneous scenes.

4. Experimental Results

The models were evaluated against nine baselines (including CNNs, Transformers, GANs, and standard Diffusion models) across Forest, Low Vegetation, and Water land-cover types.

Performance Metrics:
- EFDiff-x0: Achieved the best distortion metrics (lowest RMSE and highest SSIM) across all categories, outperforming even the best deterministic baselines (e.g., RCAN). It reduced RMSE on Water by 16% compared to the best non-EFM baseline.
- EFDiff-ϵ: Achieved the best perceptual metrics (lowest LPIPS and FID), showing a ~4× improvement in FID over the best baseline. It generated more realistic spatial textures, though with slightly higher pixel-wise error.
Ablation Studies:
- Conditioning: EFM cross-attention provided consistent gains over HLS channel concatenation, with the advantage increasing as scene complexity (spatial heterogeneity) increased.
- Sampling Steps: EFDiff-x0 achieved optimal results with as few as 1–3 steps, while EFDiff-ϵ required more steps (up to 1000) to maximize perceptual quality.
Qualitative Analysis: Visual comparisons showed that EFDiff models successfully resolved thermal boundaries and temperature discontinuities that other methods blurred or failed to reconstruct.

5. Significance and Impact

Bridging the Gap: The paper successfully bridges the gap between discriminative remote sensing (using EFMs for classification/segmentation) and generative remote sensing. It proves that EFMs can act as powerful semantic guides for generative tasks.
Solving Ill-Posed Problems: By leveraging the "semantic prior" from EFMs, the model effectively resolves the ambiguity inherent in extreme super-resolution (32×), where raw spectral data is insufficient to determine fine-scale thermal structures.
Generalizability: While focused on LST, the framework is broadly applicable to other remote sensing problems where pretrained geospatial representations can guide the reconstruction of fine-scale details from degraded observations.
Practical Utility: The dual-variant approach allows users to select the model based on specific needs: EFDiff-x0 for applications requiring precise temperature values (e.g., energy balance modeling) and EFDiff-ϵ for applications requiring realistic spatial patterns (e.g., urban planning visualization).