Abstract
Automatic post-disaster damage detection using aerial imagery is crucial for quick assessment of damage caused by disaster and development of a recovery plan. The main problem preventing us from creating an applicable model in practice is that damaged (positive) examples we are trying to detect are much harder to obtain than undamaged (negative) examples, especially in short time. In this paper, we revisit the classical bootstrap aggregating approach in the context of modern transfer learning for data-efficient disaster damage detection. Unlike previous classical ensemble learning articles, our work points out the effectiveness of simple bagging in deep transfer learning that has been underestimated in the context of imbalanced classification. Benchmark results on the AIST Building Change Detection dataset show that our approach significantly outperforms existing methodologies, including the recently proposed disentanglement learning.
Introduction
When natural disasters strike, emergency responders require immediate damage assessments. While satellite and aerial imagery are plentiful, the "ground truth" labels needed to train AI models—specifically examples of damaged buildings—are incredibly scarce in the first few hours or days. This creates two major hurdles: extreme class imbalance and the need for few-shot learning. Previous approaches often relied on complex "disentanglement learning" or simple under-sampling, which can be either too complex for rapid deployment or too wasteful of available information.
Methodology
The proposed approach combines modern transfer learning with classical bagging. A ResNet34 architecture pre-trained on ImageNet provides a strong feature extractor baseline. To address the data imbalance, the large set of undamaged (negative) examples is partitioned into chunks matching the size of the small damaged (positive) set. An ensemble of models is trained, where each base model sees all available damaged examples but only one specific chunk of undamaged examples. Final predictions are an average of the ensemble's scores. Additionally, label smoothing is employed to prevent overconfidence and improve ensemble calibration.
Experimental Results
The methodology was benchmarked using the AIST Building Change Detection (ABCD) dataset, simulating real-world disaster scenarios with "5-shot" and "50-shot" training sets.
In the 5-shot scenario, the ensemble achieved 89.96% accuracy, representing a 53.3% reduction in the average error rate compared to previous state-of-the-art methods (Hamaguchi et al.). As shown in Figure 3, the ensemble approach not only improves mean accuracy but also significantly reduces the standard deviation, leading to more stable and reliable predictions.
In the 50-shot scenario, the model achieved 92.56% accuracy, outperforming the previous best of 89.70%. Figure 6 illustrates that while performance gains begin to saturate as the number of ensembles increases, the stability of the model continues to be a key advantage of the bagging approach.
Discussion and Conclusion
The power of this approach lies in its simplicity and the effective use of calibration. Pre-trained models are better calibrated than those trained from scratch, making them ideal for ensembles. By partitioning rather than discarding negative data, the system learns the full diversity of undamaged structures. While training multiple models introduces computational overhead, the significant gains in accuracy and stability make this a superior choice for on-the-fly disaster response compared to more complex generative models.
Summary updated