Near-real-time flood inundation monitoring by Bayesian analysis for change point problems for Sentinel-1 time series

2025-04-28

AI Summary

Existing near-real-time flood monitoring methods suffer from heavy reliance on labeled training data, susceptibility to cloud contamination, and difficulty distinguishing floodwater from permanent water bodies. To address these limitations, this paper proposes an unsupervised Bayesian changepoint detection framework that directly models abrupt changes in Sentinel-1 SAR time-series backscatter coefficients for end-to-end flood inundation mapping. It is the first work to apply Bayesian Changepoint Analysis (BCP) to SAR-based flood detection—requiring no labeled samples, cloud masks, or prior water body maps. Evaluated on the UrbanSARFloods dataset across three urban sites, the method achieves F1-scores of 0.41–0.76 (IoU: 0.25–0.61), substantially outperforming Otsu thresholding and Siamese CNN baselines. In open-area scenes, the F1-score reaches up to 0.81, demonstrating suitability for rapid flood early warning in agricultural regions.

Technology Category

• Machine Learning: Bayesian Learning
• Computer Vision: Object Detection & Categorization
• Reasoning under Uncertainty: Relational Probabilistic Models

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Data management and stream processing for Web, mobile and wireless applications
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

Near real-time flood monitoring is crucial for disaster response, yet existing methods face significant limitations in training data requirements and cloud cover interference. Here we present a novel approach using Bayesian analysis for change point problems (BCP) applied to Sentinel-1 SAR time series data, which automatically detects temporal discontinuities in backscatter patterns to distinguish flood inundation from permanent water bodies without requiring training data or ancillary information. We validate our method using the UrbanSARFloods benchmark dataset across three diverse geographical contexts (Weihui, China; Jubba, Somalia; and NovaKakhovka, Ukraine). Our BCP approach achieves F1 scores ranging from 0.41 to 0.76 (IoU: 0.25-0.61), significantly outperforming both OTSU thresholding (F1: 0.03-0.12, IoU: 0.02-0.08) and Siamese convolutional neural network approaches (F1: 0.08-0.34, IoU: 0.05-0.24). Further analysis reveals exceptional performance in open areas with F1 scores of 0.47-0.81 (IoU: 0.31-0.68) and high recall (0.36-0.84), contrasted with substantially lower performance in urban areas (F1: 0.00-0.01, IoU: 0.00-0.01), indicating a common challenge across current flood detection methods in urban environments. The proposed method's ability to process raw SAR data directly with minimal preprocessing enables integration into operational early warning systems for rapid flood mapping, particularly in agricultural and open landscapes where it demonstrates the strongest performance.

Problem

• Detects flood inundation in real-time without training data
• Overcomes cloud cover interference in flood monitoring
• Addresses poor urban performance in current flood detection

Innovation

• Bayesian change point analysis for flood detection
• Sentinel-1 SAR time series without training data
• Automated inundation mapping outperforms existing methods