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

Researchers developed a lightweight AI model using unsupervised deep learning to detect conflict-related fires in Sudan within 24-30 hours using commercially available satellite imagery. The Variational Auto-Encoder (VAE) approach outperformed traditional methods in identifying burn signatures from 4-band Planet Labs satellite data at 3-meter resolution.

• →AI-powered fire detection system can identify conflict-related burn areas in Sudan within 24-30 hours using commercial satellite data
• →Lightweight VAE model adapted for 4-band imagery outperforms traditional detection methods like cosine distance and CVA
• →System achieves higher recall and F1-scores while maintaining operationally viable precision in imbalanced fire-detection scenarios
• →Approach uses unsupervised learning to identify anomalies by comparing temporal changes in satellite image representations
• →Additional spectral bands and temporal sequences provide only marginal improvements over single 4-band inputs