Improving Interpretability of Deep Active Learning for Flood Inundation Mapping Through Class Ambiguity Indices Using Multi-spectral Satellite Imagery

TL;DR

This work tackles the interpretability gap in deep active learning for flood inundation mapping by introducing the IDAL-FIM framework, which anchors DAL interpretation on two class-ambiguity indices, $BPR$ and $MDF$. Using a U-Net with MC-dropout on the Sen1Floods11 dataset, the study shows statistically significant correlations between these indices and predictive uncertainty at the tile level, enabling meaningful visualization through 2D density plots. The results demonstrate that uncertainty-based acquisitions (notably Margin and Entropy) can be interpreted in terms of sensor- and scene-driven ambiguities, while density-based approaches are more sensitive to the unlabeled pool distribution. Overall, IDAL-FIM provides a practical, interpretable lens for guiding data labeling in flood mapping and highlights how sensor characteristics shape active learning behavior and data selection strategies.

Abstract

Flood inundation mapping is a critical task for responding to the increasing risk of flooding linked to global warming. Significant advancements of deep learning in recent years have triggered its extensive applications, including flood inundation mapping. To cope with the time-consuming and labor-intensive data labeling process in supervised learning, deep active learning strategies are one of the feasible approaches. However, there remains limited exploration into the interpretability of how deep active learning strategies operate, with a specific focus on flood inundation mapping in the field of remote sensing. In this study, we introduce a novel framework of Interpretable Deep Active Learning for Flood inundation Mapping (IDAL-FIM), specifically in terms of class ambiguity of multi-spectral satellite images. In the experiments, we utilize Sen1Floods11 dataset, and adopt U-Net with MC-dropout. In addition, we employ five acquisition functions, which are the random, K-means, BALD, entropy, and margin acquisition functions. Based on the experimental results, we demonstrate that two proposed class ambiguity indices are effective variables to interpret the deep active learning by establishing statistically significant correlation with the predictive uncertainty of the deep learning model at the tile level. Then, we illustrate the behaviors of deep active learning through visualizing two-dimensional density plots and providing interpretations regarding the operation of deep active learning, in flood inundation mapping.

Figures (11)

• Figure 1: The process of the IDAL-FIM framework.
• Figure 2: The architecture of U-Net with MC-dropout. The input image assumes uniform width and height (I).
• Figure 3: Example images in Sen1Floods11. (left) Sentinel-2 false color composite image, and (right) corresponding labeled data with color codes: blue for flood, green for non-flood, and gray for no data.
• Figure 4: The geographical regions that make up the unlabeled data pool and the target regions. The target regions were selected, one from each of the continents of South America, Africa, and Asia. The remaining 8 regions were utilized for the unlabeled data pool.
• Figure 5: The comparison of the mean F1-score in different five acquisition functions: (left) Bolivia, (middle) Nigeria, (right) Vietnam. The horizontal blue dashed line (mF1-score$_{\text{Full}}$) represents the mean F1-score from models trained on the entire 1,532 data points in the pool. The horizontal black dashed line (mF1-score$_{\text{Random-500}}$) displays the mean F1-score from models trained on a selection of 500 data points using the random acquisition function.