Near--Real-Time Conflict-Related Fire Detection in Sudan Using Unsupervised Deep Learning

This study introduces a lightweight, unsupervised Variational Auto-Encoder model utilizing 3-meter 4-band Planet Labs imagery to detect conflict-related fires in Sudan within 24 to 30 hours, demonstrating superior performance in recall and F1-scores compared to traditional change detection methods for near-real-time war zone monitoring.

The Gist

Imagine Sudan as a house where a violent fight has broken out. In the chaos, people are throwing things, breaking windows, and setting fires. The people inside are scared, and the people outside (the world) want to know: Where is the fire? How bad is the damage? Can we help?

Usually, getting this information is like trying to watch a fight through a thick, foggy window from a mile away. You might see smoke, but you can't see the broken chairs or the small fires. Or, you have to wait for a messenger to run all the way there, which takes days. By the time they get back, the fire might be out, or the situation might have changed.

This paper introduces a new, super-fast way to "see" the damage using a special pair of eyes (satellites) and a smart brain (artificial intelligence).

Here is the breakdown of how they did it, using simple analogies:

1. The Eyes: PlanetScope Satellites

Think of the PlanetScope satellites as a swarm of tiny, hyper-active bees buzzing around the house every single day.

• The Old Way: Other satellites (like Sentinel-2) are like a slow-moving owl. It flies over the house once every 5 or 10 days. If a fire starts and ends in a day, the owl misses it completely.
• The New Way: The PlanetScope bees fly over every day. They take pictures so sharp (3 meters resolution) that you can see individual cars and small piles of rubble, not just the whole neighborhood.

2. The Brain: The "Unsupervised" AI

The researchers built a special AI brain called a Variational Autoencoder (VAE). Here is the tricky part: In a war zone, you don't have a "teacher" to show the AI what a "burned house" looks like. There are no textbooks or labeled photos because the fighting is happening right now.

So, how does the AI learn?

• The Analogy: Imagine you have a friend who knows your house perfectly. They know exactly where the sofa is, the color of the rug, and where the lamp stands.
• The "Normal" Baseline: The AI is trained to memorize what a "normal," peaceful house looks like. It learns the "vibe" of the neighborhood.
• The "Anomaly" Detection: When the AI looks at a new photo taken after a fight, it doesn't need to know what "fire" looks like specifically. It just asks: "Does this look like the normal house I memorized?"
• If it sees a charred roof or a smoke plume, it says, "Whoa, this doesn't match my memory! Something is wrong here!" This is called unsupervised learning—the AI figures out what is "weird" on its own without being told what to look for.

3. The Speed: Near-Real-Time

The goal was to get this information to humanitarian workers within 24 to 30 hours.

• The Process:
  1. Take the photo: The satellite snaps a picture.
  2. Clean the data: The computer fixes the colors (like adjusting the white balance on a camera).
  3. The AI checks: The brain compares the "before" photo and the "after" photo.
  4. The Alert: Within a day, a map pops up showing exactly where the "weird" (burned) spots are.

4. The Results: Did it Work?

The researchers tested this in five different places in Sudan where fighting had happened. They compared their "Smart Brain" method against older, simpler methods (like just measuring the difference in pixel colors).

• The Winner: The Smart Brain (VAE) was much better. It found more fires (high Recall) and didn't get confused by shadows or clouds as often as the old methods.
• The "Extra" Bands: They also tried feeding the AI more colors (8 bands instead of 4) and watching the house over a few days (time-series).
  • The Surprise: It didn't help much! The simple 4-color version (Red, Green, Blue, Near-Infrared) was already doing a great job. It's like realizing you don't need a 4K TV to see a fire; a good standard TV works just fine and is much faster to set up.

Why This Matters

In a war zone, time is life.

• Old Way: Wait for news, wait for slow satellites, wait for ground reports. By then, the damage is done, and aid is late.
• New Way: A map appears the next day showing exactly where the fires are. Humanitarian groups can send food, water, and doctors to the exact right spots immediately.

The Bottom Line

This paper proves that we can use a "smart, unsupervised brain" combined with "daily, high-definition satellite photos" to act like a super-fast emergency responder. It doesn't need a teacher to tell it what a fire looks like; it just knows what "peace" looks like, and when that peace is broken by fire, it raises the alarm instantly.

It's a powerful tool for turning the chaos of war into clear, actionable data, helping the world see the invisible scars of conflict.

Technical Summary

1. Problem Statement

The ongoing armed conflict in Sudan (since April 2023) has caused widespread destruction, displacement, and civilian harm. Monitoring conflict-related fires and burn scars in active war zones is critical for humanitarian assessment but faces significant challenges:

• Data Scarcity: Ground-based reporting is often delayed, incomplete, or impossible due to insecurity.
• Labeling Limitations: Supervised deep learning models require high-quality labeled training data, which is scarce in active conflict zones.
• Sensor Constraints: Existing methods often rely on coarse-resolution sensors (e.g., MODIS, VIIRS) that miss small-scale, fragmented fires in dense urban environments, or they use computationally intensive models that are not scalable for rapid deployment.
• Timeliness: There is a need for near–real-time (NRT) monitoring (within 24–30 hours) to support operational decision-making.

Objective: To develop a scalable, unsupervised, near–real-time framework for detecting fire-affected areas in Sudan using high-resolution (3m), high-frequency commercial satellite imagery (PlanetScope) without relying on ground-truth labels for training.

2. Methodology

A. Data Source and Preprocessing

• Imagery: Utilized PlanetScope imagery (3m spatial resolution) with 4 spectral bands (Red, Green, Blue, Near-Infrared).
• Study Areas: Five case studies in Sudan (Gandahar Market, El Fasher, Muqrin, Jaranga, Sarafaya) covering incidents between April and May 2024.
• Preprocessing:
  • Spectral Alignment: Applied Major Axis (MA) Linear Regression to align 3m imagery with 10m reference data.
  • Normalization: Used logarithmic compression and robust min-max scaling (using 1st and 99th percentiles) to handle the high dynamic range of NIR bands and reduce sensor artifacts.
  • Tiling: Imagery was partitioned into $32 \times 32$ pixel tiles.

B. Model Architecture: Adapted RaVAEn

The authors adapted RaVAEn (a lightweight Convolutional Variational Autoencoder) originally designed for 10m/10-band imagery to work with 3m/4-band inputs.

• Architecture: An unsupervised encoder–decoder framework with four hierarchical stages.
• Key Adaptations:
  • Frequency Decompression: Input layers use a fixed Gaussian low-pass filter (coarse structure) and high-pass residual (fine texture) to handle resolution-dependent spatial frequencies.
  • Multi-scale Convolutions: Replaced standard strided convolutions with Atrous Spatial Pyramid-like blocks (parallel $3 \times 3$ convolutions with increasing dilation rates) to capture multi-scale context while preserving detail.
  • Anti-aliased Downsampling: Implemented BlurPool to mitigate high-frequency information loss during downsampling.
  • Latent Space: The encoder projects features into a 128-dimensional latent space parameterized by a diagonal Gaussian distribution, regularized with a $\beta$-VAE objective ($\beta=1$).
• Training Strategy:
  • Unsupervised: Trained on the WorldFloods dataset (diverse disaster types) to learn generalized nominal land-surface representations, avoiding overfitting to specific Sudanese biomes.
  • No Reconstruction Error: Unlike traditional VAE anomaly detection, the model does not use reconstruction error. Instead, it uses the latent space embeddings for change detection.

C. Change Detection and Inference

• Mechanism: The model embeds temporally paired images (pre-incident and post-incident) into the latent space.
• Anomaly Scoring: Anomaly scores are calculated using the Cosine Distance between the latent embeddings of the pre- and post-incident tiles.
• Logic: Significant deviations in the latent space indicate a change from nominal surface conditions (e.g., burn scars), independent of absolute radiometric differences.

D. Evaluation Framework

• Ground Truth: Labels were curated by the Sudan Conflict Observatory (SCO) using human rights reports, ACLED data, and field interviews, verified by multiple analysts. Labels were used only for evaluation, not training.
• Baselines: Compared against Cosine Distance (pixel-level), Canonical Variates Analysis (CVA), and Iteratively Reweighted Multivariate Alteration Detection (IR-MAD).
• Metrics: Precision, Recall, F1-score, and Area Under the Precision-Recall Curve (AUPRC) (preferred due to severe class imbalance where fire tiles < 10% of the scene).
• Statistical Validation: Used 1,000 bootstrap resamples and paired Wilcoxon signed-rank tests.

3. Key Contributions

1. Unsupervised NRT Framework: Demonstrated a viable unsupervised deep learning pipeline for detecting conflict-related fires without labeled training data.
2. Resolution Adaptation: Successfully adapted a VAE architecture from 10m/10-band to 3m/4-band inputs, proving that high-resolution, low-bandwidth data is sufficient for this task.
3. Latent Space Change Detection: Validated that comparing latent embeddings is more robust than pixel-wise reconstruction error or traditional spectral indices for detecting fragmented urban fires.
4. Operational Feasibility: Established a workflow capable of producing results within 24–30 hours of image acquisition, suitable for humanitarian response.

4. Results

• Performance Superiority: The proposed Latent Representation Change (LRC) approach consistently outperformed all baselines (Cosine, CVA, IR-MAD) across all five case studies.
  • AUPRC Improvements: LRC achieved AUPRC values ranging from 0.62 to 0.81, representing a 14% to 36% relative improvement over IR-MAD.
  • Recall: The model prioritized recall (detecting all fire-affected areas), achieving F1-scores up to 0.86 in favorable conditions (e.g., Sarafaya).
• Spectral vs. Temporal Data:
  • Experiments with 8-band imagery and time-series sequences yielded only marginal performance gains (3–7% relative AUPRC increase) over the standard 4-band single-snapshot input.
  • In some complex urban scenes (El Fasher), 8-band/time-series data actually performed slightly worse due to increased noise and misregistration sensitivity.
• Error Analysis:
  • False Positives: Often linked to residual debris or prior burn scars.
  • False Negatives: Primarily occurred with fires smaller than the 3m sensor resolution or in highly fragmented urban settings, an inherent hardware limitation.

5. Significance and Conclusion

• Humanitarian Impact: This study provides a practical tool for rapid, independent verification of conflict damage in data-scarce environments, supporting humanitarian aid and accountability efforts.
• Cost-Efficiency: By proving that lightweight 4-band models are sufficient, the approach reduces computational costs and data bandwidth requirements compared to multi-spectral or multi-temporal heavy models.
• Generalizability: The unsupervised nature of the model allows it to be applied to other conflict zones or disaster types without retraining on local labeled data.
• Limitations: The approach relies on cloud-free optical imagery and accurate co-registration. It is currently limited to specific case studies in Sudan, though the methodology is transferable.

Final Verdict: The paper successfully demonstrates that unsupervised deep learning on high-resolution 4-band satellite imagery is a highly effective, scalable, and operationally viable solution for near–real-time monitoring of conflict-related fires, offering a superior trade-off between accuracy and speed compared to traditional methods.