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

AI summary

Overview

Research area: Remote sensing / computer vision applied to conflict monitoring — specifically, detecting conflict-related fires and burn scars in Sudan from high-resolution commercial satellite imagery.

Technical level: Intermediate. The reader should be comfortable with machine learning concepts (autoencoders, latent spaces, precision/recall) and satellite remote sensing basics (spectral bands, revisit cadence, cloud masking).

Scope in one sentence: The paper adapts a lightweight, unsupervised Variational Autoencoder (VAE) called RaVAEn to 3 m, 4-band PlanetScope imagery and benchmarks it against five conventional change-detection baselines across five conflict-related fire incidents in El Fasher and Khartoum, Sudan.

What This Paper Is About

Ground reporting of attacks in active war zones like Sudan is delayed, incomplete, or impossible, so satellite imagery has become a key source of independent evidence. Active fires and burn scars are especially informative because they often accompany airstrikes, shelling, looting, and destruction of civilian structures. The problem is that existing fire-detection methods rely on coarse-resolution sensors, labeled training data that does not exist in war zones, or heavy models that are hard to deploy quickly — so this paper asks whether a small, label-free deep learning model fed with near-daily 3 m imagery can flag fire-affected areas fast enough to be operationally useful.

Key Contributions

1. Adaptation of RaVAEn to high-resolution 4-band imagery. The authors retrained a lightweight convolutional VAE originally built for 10-band, 10 m imagery so that it works on 3 m PlanetScope imagery with only RGB and near-infrared channels, adding a frequency-decomposition input layer, multi-scale dilated convolution blocks, BlurPool anti-aliased downsampling, and a scale-augmented training strategy.
2. A latent-space change detection formulation instead of reconstruction-error scoring. Rather than treating reconstruction error as the anomaly signal (which prior work found unreliable in fully unsupervised settings), the model embeds pre- and post-incident tiles into a 128-dimensional latent space and scores change with cosine distance between the paired embeddings. The authors label this approach "latent representation change" (LRC).
3. A quantitative benchmark across five Sudan case studies against cosine distance, Canonical Variates Analysis (CVA), Iteratively Reweighted Multivariate Alteration Detection (IR-MAD), and two task-specific spectral indices (dNDVI and dBAI), evaluated with AUPRC, precision, recall, F1-score, and commission/omission error rates.
4. An empirical test of the marginal value of extra data. Experiments with 8-band imagery and with short pre-incident temporal sequences (Pre-1 to Pre-3) are compared against a single 4-band snapshot to see whether added spectral or temporal information actually helps.

Main Findings

• The VAE-based approach (LRC) outperformed all baselines at every site and imagery configuration. Comparisons included cosine distance, CVA, IR-MAD, dNDVI, and dBAI.
• Relative AUPRC improvement over IR-MAD ranged from 14% to 36%. Sites with clearer fire signatures, such as Muqrin and Sarafaya, reached AUPRC values exceeding 0.80.
• Fire-affected tiles make up less than 10% of each scene, which is the stated reason the authors use AUPRC rather than ROC curves; the remaining tiles define the nominal unburned background distribution.
• Gandahar Market (median across folds, 4-band): LRC reached AUPRC 0.74, precision 0.68, recall 0.75, F1 0.71, commission error 0.32, omission error 0.25 — versus IR-MAD at AUPRC 0.65, precision 0.66, recall 0.59, F1 0.62.
• El Fasher showed a different pattern: the 4-band LRC run scored AUPRC 0.68, precision 0.62, recall 0.66, F1 0.64, while the 8-band run scored lower (AUPRC 0.62, precision 0.57, recall 0.63, F1 0.60) and the time-series run scored AUPRC 0.65, precision 0.60.
• Additional bands and temporal sequences produced only marginal gains on the clearest scenes. At Gandahar Market, LRC AUPRC moved from 0.74 (4-band) to 0.80 (8-band) to 0.81 (time-series), and F1 from 0.71 to 0.75 to 0.77 — improvements the authors characterize as marginal relative to the added complexity.
• Performance degrades in high-noise or degraded conditions. Spectral and structural ambiguity reduced accuracy, and the authors note that baseline behavior is scene-dependent: dNDVI was generally less effective in sparsely vegetated or urban environments, while dBAI detected burn signatures more consistently but remained sensitive to dark rooftops, soil, and shadows.
• Qualitative results aligned with verified ground truth. Predicted fire-affected regions showed spatial coherence across Jaranga (multiple simultaneous fires), Muqrin (previously impacted locations), Sarafaya (recently burned surfaces), El Fasher (extensive charred swaths after successive incidents), and Gandahar Market, where an active fire was detected alongside a visible smoke plume. The authors state smoke plumes are transient and are not treated as a primary or reliable indicator of fire activity.
• Near-real-time detection is feasible under favorable conditions, within approximately 24 to 30 hours. The stated latency budget is roughly 18–20 hours for PlanetScope image acquisition and delivery, 1–2 hours for preprocessing and tiling, and approximately 1–2 hours per scene for GPU-based inference and postprocessing.
• Operational availability is constrained by cloud and georegistration. Approximately 95% of ground-lock failures are associated with cloud cover, and a small proportion of scenes cannot be fully processed due to acquisition anomalies or atmospheric contamination. The authors state operational latency was not explicitly stress-tested.
• Statistical rigor: 95% confidence intervals were estimated with 1,000 bootstrap resamples of tiles within each scene, and paired Wilcoxon signed-rank tests were applied to AUPRC values from identical bootstrap resamples across methods.

Methodology in Plain English

The authors took an existing small neural network called RaVAEn — a Variational Autoencoder that compresses an image tile down to a compact numerical "fingerprint" and then tries to reconstruct it — and rebuilt its early layers so it works on 3 m resolution imagery with four bands instead of 10 m imagery with 10 bands. They feed it 32 × 32 pixel tiles drawn from PlanetScope ortho-analytic imagery.

Training is unsupervised, meaning no fire labels are used at all. The model is trained on the WorldFloods dataset (233 scenes from Sentinel-2, with a 19-scene held-out validation set; validation images with more than 20% cloud cover were discarded), which contains several kinds of disaster-related surface change including fire burn scars. The idea is that even though Sudan's arid landscape differs from the training data, the model learns general structural and textural patterns of "normal" surfaces and can therefore flag anything that deviates. Training ran for 200 epochs on a machine with one NVIDIA H100 GPU, 112 Intel CPU cores, and 2 TB of system memory, using the Adam optimizer at a learning rate of 0.001 and batch size 128.

At inference, the model does not look at reconstruction error. Instead, it takes a pre-incident tile and the matching post-incident tile from the same location, pushes both through the encoder, and measures the cosine distance between the two latent fingerprints. A big jump in latent space means something changed on the ground.

Preprocessing involves two normalization steps: a spectral alignment using Major Axis linear regression to bring the 3 m reflectance in line with the 10 m reference, followed by logarithmic compression and robust min-max scaling to the [-1, +1] range using the 1st and 99th percentiles per band.

For evaluation, the authors compared against cosine distance in pixel space, CVA, IR-MAD, and the differenced spectral indices dNDVI and dBAI, thresholding every method at the 95th percentile of its own score distribution so that preprocessing, tile geometry, and temporal pairing stayed identical. Ground-truth labels came from the Sudan Conflict Observatory, drawing on human rights organizations, the ACLED catalog, and field interviews; each scene was reviewed by at least three analysts with inter-annotator agreement above 90%, and incidents were cross-validated with FIRMS and VIIRS data. Labels were used only for retrospective evaluation — never for training, threshold tuning, or model selection.

Why This Matters

Impact on research. This work extends a label-free anomaly detection framework from a generic multi-hazard setting into an active war zone, and it does so at 3 m resolution — a scale where small, fragmented, short-lived fires in dense settlements are actually visible. It also supplies a systematic head-to-head comparison showing that extra spectral bands and short temporal sequences buy relatively little, which is a useful negative result for anyone tempted to build heavier pipelines.

Real-world applications:

• Humanitarian response planning, where knowing which structures burned and when can guide aid delivery and needs assessments.
• International humanitarian law documentation, since burn signatures frequently accompany airstrikes, shelling, and destruction of civilian structures — evidence that ground access cannot gather.
• Conflict early warning and situational awareness, using the approximately 24–30 hour detection window to flag incidents for analyst follow-up.
• Scalable monitoring over large areas, because the model requires no ground-truth labels and runs on GPU-enabled hardware in 1–2 hours per scene.

Industry relevance. Commercial satellite operators and geospatial analytics firms have an obvious interest in turning near-daily revisit imagery into automated products. A lightweight unsupervised model matters commercially because it avoids the cost of building labeled training sets for every new conflict or disaster, and because inference cost — the paper's stated 1–2 hours per scene — is part of what determines whether a service is viable.

Future Directions

• Stress-testing operational latency. The authors explicitly state that actual operational latency may be longer in conflict zones due to cloud cover, image misregistration, platform-specific delivery policies, or tasking constraints, and that these were not explicitly stress-tested.
• Handling cloud cover and image availability. Dense cloud cover remains a stated limitation of optical remote sensing, roughly 95% of ground-lock failures are tied to cloud cover, and the near-real-time definition is framed as an ideal benchmark under cloud-free, well-registered conditions.
• Extending beyond detection to damage assessment and attribution. The paper deliberately scopes out retrospective damage assessment and causal attribution, and states that label-independent methods for rapidly evolving conflicts remain a gap.
• Testing transfer to other conflict settings. The model was trained on Sentinel-2-derived WorldFloods data and applied to Sudan, raising the open question of how well it generalizes to other ecological and urban contexts without retraining.

Target Audience

This paper is most valuable to remote sensing and geospatial researchers working on change detection and anomaly detection, machine learning practitioners interested in unsupervised methods that operate without labels, and analysts at conflict-monitoring organizations, human rights groups, and humanitarian agencies who need rapid, independent evidence from satellite imagery. Readers with a background in deep learning and multispectral imagery will get the most from the methodology sections.

Authors’ abstract

Ongoing armed conflict in Sudan highlights the need for rapid monitoring of conflict-related fire-affected areas. Recent advances in deep learning and high-frequency satellite imagery enable near--real-time assessment of active fires and burn scars in war zones. This study presents a near--real-time monitoring approach using a lightweight Variational Auto-Encoder (VAE)--based model integrated with 4-band Planet Labs imagery at 3 m spatial resolution. We demonstrate that these impacted regions can be detected within approximately 24 to 30 hours under favorable observational conditions using accessible, commercially available satellite data. To achieve this, we adapt a VAE--based model, originally designed for 10-band imagery, to operate effectively on high-resolution 4-band inputs. The model is trained in an unsupervised manner to learn compact latent representations of nominal land-surface conditions and identify burn signatures by quantifying changes between temporally paired latent embeddings. Performance is evaluated across five case studies in Sudan and compared against cosine distance, CVA, and IR-MAD using precision, recall, F1-score, and the area under the precision-recall curve (AUPRC) computed between temporally paired image tiles. Results show that the proposed approach consistently outperforms the other methods, achieving higher recall and F1-scores while maintaining viable precision in highly imbalanced fire-detection scenarios. Experiments with 8-band imagery and temporal image sequences yield only marginal performance gains over single 4-band inputs, underscoring the effectiveness of the proposed lightweight approach for scalable, near--real-time conflict monitoring.