Introduction
Rapid disaster response is critical for public safety, yet identifying specific disaster types from high-altitude aerial footage remains a significant technical challenge. The University of Technology Sydney (UTS) team addressed the TRECVID 2020 Disaster Scene Description and Indexing (DSDI) task by developing a robust framework to classify aerial imagery. The primary hurdles identified in this domain include context ambiguity (e.g., distinguishing natural water from flood zones), scale variance where critical details are minute relative to the frame, and the scarcity of high-quality labeled aerial data.
Methodology
The UTS approach centers on two pillars: Feature Extraction and Feature Enhancement.
Backbones and Feature Extraction
The team utilized a variety of state-of-the-art deep neural networks as backbones, including ResNet (18, 50, 152), Inception-v4, SENet, and PolyNet. These models were pre-trained on ImageNet and fine-tuned specifically for the disaster dataset. For video processing, the team extracted frames and applied mean pooling to the resulting scores to generate final video-level classifications.
Domain-Specific Enhancements
- 360-Degree Data Augmentation: Recognizing that aerial perspectives lack a fixed "up" orientation, the team expanded standard rotation augmentation from 20 degrees to a full 360 degrees. This adjustment improved the mean Average Precision (mAP) from 37.02 to 37.44.
- Mixture of Experts (MoE): Instead of a single classification layer, the team employed multiple classifiers on the same feature vector, averaging their scores. This "voting" mechanism provided a 0.8 mAP gain, enhancing model generalization on noisy data.
- Model Ensemble: The final system combined results from multiple architectures and pre-training regimes (including the Places365 dataset) to maximize performance.
Datasets
The study utilized the Low Altitude Disaster Imagery (LADI) Dataset, which contains aerial imagery from Atlantic Hurricane and spring flooding seasons. The dataset was split into training and validation sets using a 10:1 ratio to ensure robust evaluation.
Results and Analysis
Single Model Performance
Evaluation of individual backbones revealed that SENet was the top-performing single model, achieving a validation mAP of 38.39. ResNet-152 followed closely with 38.12 mAP. These results underscore the effectiveness of attention-based mechanisms (SENet) in identifying fine-grained features in aerial imagery.
Ensemble and Submission Results
By ensembling different architectures, the team reached a peak validation mAP of 40.03. For the official TRECVID submission, the "fully selected" ensemble (UTS.run.3) utilizing seven model variations achieved the highest score of 0.281.
Conclusion
The UTS team demonstrated that tailoring data augmentation to sensor physics (360-degree rotation) and employing a Mixture of Experts strategy significantly improves disaster scene classification. While the current framework achieves competitive results, future work will focus on advanced temporal modeling, such as Transformers or LSTMs, to better capture the evolution of disaster scenes in video sequences and resolve remaining contextual ambiguities.
Summary updated