Abstract
Rapid disaster response is often hindered by the lack of labeled data for novel catastrophes, creating a "cold start" problem for traditional supervised AI models. We present Deep-Disaster, an unsupervised framework that detects and localizes disaster damage without requiring labeled damage data for training. By framing disaster detection as an anomaly detection task, the system utilizes Knowledge Distillation (KD) to learn the distribution of "normal" (no-damage) imagery. During inference, the discrepancy between a pre-trained teacher network and a specialized student network allows for robust detection and precise localization of destruction across diverse disaster types.
Introduction
In the immediate aftermath of a disaster, social media provides a wealth of visual data that could assist humanitarian organizations. However, supervised computer vision models face two critical hurdles: the inability to train on a new event due to lack of labels, and poor generalization across different disaster types (e.g., a model trained on floods failing to recognize earthquake rubble). Deep-Disaster addresses these by modeling what a city looks like under normal circumstances, treating any disaster-induced change as a detectable anomaly.
Methodology
The core of the Deep-Disaster framework is a Knowledge Distillation architecture featuring a Teacher and a Student network.
Knowledge Distillation
The Teacher is a powerful network pre-trained on a massive general dataset (ImageNet), providing a comprehensive understanding of visual features. The Student is a smaller, GAN-based network trained exclusively on "no-damage" images. By forcing the student to mimic the teacher’s behavior only on normal samples, the student learns a narrow manifold of the non-disaster world.
Crucially, the student network must be smaller than the teacher to prevent over-memorization, ensuring that it fails to replicate the teacher's features when encountering anomalous (damaged) data. Furthermore, the authors emphasize that the direction (cosine similarity) of activation vectors is as vital as Euclidean distance for reliable knowledge transfer.
Detection and Localization
When a test image is processed, the system performs two primary functions:
- Detection: A scoring function $S(\dot{x})$ measures the difference in both value and direction of the networks' activations. If the image contains damage, the student produces features that diverge significantly from the teacher's, flagging the image as an anomaly.
- Localization: To identify the specific location of damage, the framework employs gradient-based interpretability methods like Smooth Gradient. By identifying which pixels cause the highest gradient spikes in the loss function, the model generates a heatmap over the affected areas.
Experimental Results
The model was evaluated using the Damage Severity Assessment Dataset (DAD), which includes approximately 25,000 images from events such as the Nepal Earthquake and Hurricane Matthew.
- State-of-the-Art Performance: Deep-Disaster achieved an average AUC-ROC of 0.84, significantly outperforming unsupervised domain adaptation baselines.
- Generalization to Novel Events: When tested on disaster types entirely absent from the training set, the model maintained a high AUC-ROC of 0.8035, proving its utility for "unseen" catastrophes.
- Localization Quality: Qualitative analysis shows that the model produces cleaner, more precise damage heatmaps compared to traditional supervised methods like VGG-19 with Grad-CAM.
Conclusion
Deep-Disaster demonstrates that unsupervised anomaly detection is a powerful alternative to supervised learning for disaster response. By focusing on the manifold of "normal" imagery, the framework overcomes the cold-start problem and generalizes effectively to novel events. Future work will focus on increasing the diversity of the "normal" training data to reduce false positives in complex, non-damaged scenes.
Summary updated