Abstract
Humanitarian crisis response relies on rapid, accurate mapping, yet deep learning models often fail when applied to new geographic regions. This paper introduces a "smart" sampling strategy to prioritize image tiles for manual labeling during fine-tuning. By leveraging latent space analysis of a U-Net bottleneck layer, the proposed method identifies novel and challenging samples. Results demonstrate that this prioritization significantly outperforms random sampling in data-scarce scenarios, providing a stable and efficient path for model adaptation in time-critical environments.
Introduction
In the wake of humanitarian crises such as earthquakes or disease outbreaks, remote sensing and automated building detection are essential for coordinating aid. While deep learning models like U-Net can automate these tasks, they frequently suffer from poor generalization across different geographic regions. A model trained on European urban architecture may struggle with rural African terrain. Fine-tuning on local data is the standard solution, but obtaining high-quality labels from volunteers is time-consuming. This research addresses the "data scarcity" problem: identifying which small subset of a new area should be labeled first to achieve the highest model performance.
Methodology
The authors propose a multi-stage approach to guide data collection using interpretability and representation learning.
Latent Space Analysis and Clustering
The core of the method involves extracting activations from the bottleneck layer of a pre-trained U-Net. This layer contains the most compressed features of the input imagery. By applying Principal Component Analysis (PCA), the researchers visualized this latent space, discovering a "bipolar" distribution that naturally separates undeveloped areas from urban environments.
Priority Scoring
New, unlabeled images from the target region are mapped into this latent space. Two formulas are proposed to rank samples for labeling:
- Basic Priority Score (BPS): A weighted combination of distance from known training clusters (novelty) and predicted Intersection over Union (IoU), targeting areas where the model is likely to fail.
- Multiparty Priority Score (MPS): A more advanced version accounting for "orphaned" clusters (entirely new terrain types), historical "error" clusters, and Local Outlier Probability (LoOP) to filter out noisy outliers.
Results and Key Insights
The study utilized the Inria Aerial Labelling Dataset, holding out Kitsap County to simulate a difficult, new "crisis" region where the initial model performed poorly.
Stability and Performance
A primary finding is the volatility of random sampling. In data-scarce regimes, random selection results in highly variable model performance. In contrast, the prioritization strategy provides a stable, reliable improvement path. The advantage of "smart" sampling is most pronounced when only 10% to 40% of the potential labels are available.
As illustrated in the comparative analysis, the priority strategy consistently leads in the early stages of data collection. The initial model, which achieved 83.8% IoU on its original validation set, saw significantly faster recovery on the new region when using prioritized samples compared to random ones.
Conclusion
This research demonstrates that prioritizing data for labeling based on model uncertainty and feature novelty is a critical "force multiplier" in humanitarian contexts. While the Basic Priority Score (BPS) showed strong empirical results, future work remains to validate the more complex Multiparty Priority Score (MPS) and test these strategies across different architectures like Transformers or diverse tasks such as flood mapping.
Summary updated