Answers are grounded in this paper and cite their sources.
Monitoring numerous CCTV feeds requires concise descriptions of visible flood progression and criteria for identifying observations that warrant review. However, water annotations in selected keyframes provide limited temporal information, complicating comparisons across sequences and interpretation for monitoring. This study calculated water proportions within fixed traffic-surface regions of interest for road and mixed scenes. Four change features were extracted from 8,980 observations in 85 sequences from 2025. K-means, Gaussian mixture models, and Ward clustering were compared for two to six clusters. Four K-means types were adopted considering interpretability, separation, and stability: Limited change (21 sequences), Expansion and recession (38), Expansion and persistence (22), and Fluctuating (4). The silhouette score was 0.5104, compared with 0.5210 for six clusters; the four-type solution was therefore not considered globally optimal. Across 500 parent-case bootstrap resamples, median type-specific Jaccard indices ranged from 0.913 to 1.000, although the smallest type was sensitive to measurement assumptions. Fixed 2025 standardization parameters and centroids were applied to 2,562 observations in 40 sequences from 2024. Separately, an exploratory observation-level rule combined current water proportion, change from the initial observation, and change relative to the preceding three observations. Review calculations excluded future-informed water supplementation, and river-only observations used change criteria alone. The framework summarizes image-projected water-area progression and identifies observations for inspection. Its operational effectiveness and relationship to water depth or damage require validation using independent field data.
No takes yet. Share an insight, caveat, or question.
Kang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: