EarthBridge: A Solution for 4th Multi-modal Aerial View Image Challenge Translation Track

This paper introduces EarthBridge, a high-fidelity cross-modal translation framework combining Diffusion Bridge Implicit Models and Contrastive Unpaired Translation to achieve second place in the 4th Multi-modal Aerial View Image Challenge by effectively translating between SAR, EO, and IR aerial imagery.

The Gist

Imagine you are a detective trying to solve a mystery, but you only have a blurry, black-and-white sketch of the crime scene (SAR images). You know the scene also has a vibrant, full-color photo (RGB), a heat map showing body temperatures (IR), and a clear daylight photo (EO). But you don't have those photos; you only have the sketch.

EarthBridge is a super-smart AI tool built by a team of researchers to turn that blurry sketch into all those other missing photos. It won second place in a high-stakes international competition called the "4th Multi-modal Aerial View Image Challenge."

Here is how they did it, explained simply:

1. The Problem: Speaking Different Languages

Think of the different camera types (SAR, RGB, IR) as people speaking different languages.

• SAR (Radar) is like a person speaking in "static and echoes." It can see through clouds and darkness, but the picture looks like a messy scribble to humans.
• RGB (Visible Light) is like a person speaking "full color." It looks beautiful but disappears in the dark or fog.
• IR (Infrared) is like a person speaking "heat." It sees warm objects but lacks fine details.

The challenge was to build a universal translator that could take a "SAR scribble" and instantly translate it into a "full-color photo" or a "heat map" without losing the important details of the buildings, roads, and trees.

2. The Solution: The "Time-Traveling Bridge"

The team built a system called EarthBridge. Instead of just guessing what the photo should look like, they used a clever mathematical trick called a Diffusion Bridge.

The Analogy: The Sculptor and the Clay
Imagine you have a block of clay (the SAR image) and you want to turn it into a statue of a horse (the RGB image).

• Old AI methods were like trying to smash the clay into dust and hoping it magically reforms into a horse. It often resulted in a blob.
• EarthBridge is like a sculptor who knows exactly how to chip away the clay while keeping the horse's shape in mind the whole time. It creates a direct "bridge" from the clay to the horse, ensuring the legs, head, and tail end up in the right places.

They used two main tools to build this bridge:

1. The "Deterministic Bridge" (DBIM): This is their main tool. It's like a high-speed train that travels on a fixed track from the "SAR station" to the "RGB station." Because the track is fixed, the train is incredibly fast and precise. It can make the trip in just a few stops (steps) instead of hundreds, making it very efficient.
2. The "Contrastive Translator" (CUT): For one specific task, they used a different tool that acts like a strict art teacher. It looks at the original sketch and the new drawing side-by-side, constantly checking: "Is this roof in the same spot as the roof in the sketch?" This ensures the structure stays perfect.

3. The Secret Sauce: "Booting Noise"

One tricky part of translation is that one SAR image could theoretically become many different-looking RGB images (maybe it's a sunny day, or maybe it's cloudy).
To handle this, EarthBridge uses something called "Booting Noise."

• Analogy: Imagine you are writing a story based on a single sentence. To make the story interesting, you need a little bit of randomness at the very beginning to decide the plot twists. EarthBridge adds a tiny, controlled "spark" of randomness at the start of the process. This allows it to generate realistic, varied textures (like grass or water) without messing up the structure (like roads or buildings).

4. The Results: A Masterpiece

The team tested their system on four different translation tasks:

• Turning Radar into Optical photos.
• Turning Radar into Infrared heat maps.
• Turning Optical photos into Infrared.
• Turning Radar into full-color RGB.

The Outcome:

• Speed: Because of their "train track" method, the AI was incredibly fast, generating high-quality images in seconds.
• Quality: The images looked real. The AI didn't just guess colors; it understood the geometry of the city. It could turn a radar blob into a sharp image of a city block, preserving the exact shape of the buildings.
• Ranking: They scored a 0.38 (where lower is better) and took 2nd place in the world, beating out many other top teams.

Summary

EarthBridge is like a magical translator that can look at a messy, all-weather radar map and instantly "paint" a perfect, high-definition photo of the same scene, complete with colors, shadows, and heat signatures. It does this by building a direct, mathematically perfect bridge between the different types of camera data, ensuring that what you see is both beautiful and structurally accurate.

Technical Summary

Here is a detailed technical summary of the paper "EarthBridge: A Solution for 4th Multi-modal Aerial View Image Challenge Translation Track."

1. Problem Definition

The paper addresses the challenge of cross-modal image-to-image translation in the context of aerial remote sensing. The goal is to translate images between three distinct sensor modalities:

• Electro-Optical (EO/RGB): Standard visible light imagery.
• Infrared (IR): Thermal imagery.
• Synthetic Aperture Radar (SAR): Radar imagery, which offers all-weather capabilities but is difficult to interpret due to speckle noise and unique electromagnetic signatures.

The Challenge:

• Data Scarcity & Constraints: SAR data is hard to interpret and scarce; EO data is limited by lighting and weather.
• Domain Gap: The physical mechanisms generating these images differ vastly (electromagnetic reflection vs. thermal emission), making direct translation difficult.
• Specific Tasks: The paper focuses on four specific translation tasks defined by the MAVIC-T (Multi-modal Aerial View Image Challenge) 2025:
  1. SAR $\to$ EO (1 channel $\to$ 1 channel, 256$\times$256)
  2. SAR $\to$ RGB (1 channel $\to$ 3 channels, 1024$\times$1024)
  3. SAR $\to$ IR (1 channel $\to$ 1 channel, 1024$\times$1024)
  4. RGB $\to$ IR (3 channels $\to$ 1 channel, 1024$\times$1024)

2. Methodology: EarthBridge Framework

The authors propose EarthBridge, a multi-method framework that primarily utilizes Diffusion Bridge Implicit Models (DBIM) and supplements it with Contrastive Unpaired Translation (CUT) for specific tasks.

A. Core Approach: Diffusion Bridge Implicit Models (DBIM)

Unlike standard diffusion models that generate images from pure noise, EarthBridge uses a Diffusion Bridge formulation. This learns a stochastic process that directly connects a source distribution ($x_T$) to a target distribution ($x_0$).

• Non-Markovian Bridge Processes: The authors generalize DBIM using non-Markovian processes over discretized timesteps. This allows for deterministic sampling, significantly reducing the number of inference steps required (e.g., down to 5 steps for RGB$\to$IR) while maintaining high fidelity.
• Stochastic Bridges via h-transform: The method employs Doob's h-transform to condition the diffusion process on a fixed endpoint (the target image), ensuring the generative trajectory is constrained by both the source and target.
• Architecture:
  • UNet Denoiser: A channel-concatenated UNet is used. The source image $x$ is concatenated with the noisy intermediate state $z_t$ as input channels.
  • Karras-weighted Scaling: The model is trained using Karras-weighted bridge scalings to handle the specific noise schedules of remote sensing data.
  • Booting Noise: To handle the "one-to-many" ambiguity in cross-modal mapping (e.g., one SAR image could correspond to multiple plausible IR images), the authors introduce a "booting noise" ($\epsilon_{boot}$) at the initial step ($t=T$). This acts as a latent variable to encode the stochasticity of the translation.
• Channel Handling: For tasks with mismatched channel counts (e.g., SAR$\to$RGB), single-band modalities are expanded via channel repetition to match the model's shared channel space (typically 3 channels) before concatenation.

B. Supplementary Approach: Contrastive Unpaired Translation (CUT)

For the SAR$\to$IR task, the authors employ CUT as a baseline/alternative.

• Mechanism: A feed-forward GAN that uses PatchNCE (InfoNCE) loss.
• Goal: To preserve structural correspondence between the source and generated image by ensuring that patch embeddings in the generated image are closer to their corresponding source patches than to negative patches.
• Usage: Used specifically where the authors found contrastive learning beneficial for structural consistency in that specific modality pair.

3. Key Contributions

1. EarthBridge Framework: A novel multi-method framework combining DBIM and CUT, providing a robust comparison between diffusion-based and contrastive-learning-based approaches for aerial image translation.
2. Specialized Training Protocol:
   • Implementation of Karras-weighted bridge scalings and pixel-level modeling to handle the high dynamic range and resolution (up to 1024$\times$1024) of remote sensing data.
   • Introduction of Booting Noise to resolve ambiguity in cross-modal mappings during deterministic sampling.
3. Efficient Deterministic Sampling: Demonstrated that non-Markovian DBIM sampling can achieve high-quality results with significantly fewer steps (as few as 5) compared to standard diffusion models, drastically improving inference speed.
4. Comprehensive Evaluation: Conducted extensive experiments on all four MAVIC-T tasks, providing detailed quantitative and qualitative analysis.

4. Experimental Results

The solution was evaluated on the MAVIC-T 2025 leaderboard using a composite score derived from FID (distributional similarity), LPIPS (perceptual fidelity), and L1 (pixel-level accuracy).

• Overall Performance: EarthBridge achieved a composite score of 0.38, securing the 2nd place on the leaderboard.
• Task-Specific Scores:
  • SAR$\to$EO: 0.27 (DBIM)
  • SAR$\to$RGB: 0.58 (DBIM)
  • SAR$\to$IR: 0.46 (CUT)
  • RGB$\to$IR: 0.20 (DBIM)
• Efficiency:
  • RGB$\to$IR: Achieved high quality in just 5 sampling steps (1.5s latency).
  • SAR$\to$IR: Used CUT for 1-step inference (0.47s latency).
• Qualitative Findings: The model demonstrated superior ability to reconstruct complex urban layouts and fine-grained textures, effectively bridging the domain gap between SAR and optical/IR modalities.

5. Significance

• Advancing Sensor Interoperability: The framework enables the synthesis of missing modalities (e.g., generating IR from SAR), which is crucial for all-weather monitoring and data fusion in remote sensing.
• Efficiency in High-Resolution Translation: By optimizing diffusion bridges for deterministic sampling, the paper addresses the computational bottleneck of diffusion models, making them viable for large-scale aerial image processing (1024$\times$1024).
• Handling Ambiguity: The "booting noise" mechanism offers a principled way to handle the inherent ambiguity in translating between physically distinct sensor types, ensuring diverse and plausible outputs.
• Benchmarking: The work sets a new state-of-the-art baseline for the MAVIC-T challenge, demonstrating that combining diffusion bridges with contrastive learning yields robust results across diverse remote sensing tasks.

Code Availability: The authors have made their code publicly available at https://github.com/Bili-Sakura/EarthBridge-Preview.