3:45 PM - 4:00 PM
[MIS04-08] Combining Multi-Sensor and Multi-Temporal Data with Machine Learning to Enhance Flood Detection in Sentinel-1 Images
★Invited Papers
Keywords:Synthetic Aperture Radars, Flood detection, Machine Learning, Flood risk management, Tropical Cyclone Seroja
Floods are among the most frequent and devastating natural disasters worldwide, with their occurrence and intensity expected to rise due to climate change and evolving socioeconomic patterns. In flood-prone regions such as Kupang Regency, Indonesia, recurrent floods have caused significant economic losses and nearly 100 casualties in recent years. Effective flood management depends on timely and accurate flood maps to support preparedness, response, and mitigation efforts.
Traditional flood mapping methods, such as ground surveys and aerial observations, are often constrained by high costs, logistical challenges, and adverse weather conditions. Satellite-based remote sensing, particularly synthetic aperture radar (SAR) imagery (e.g., Sentinel-1), provides a cost-effective alternative, offering consistent, all-weather, and day-and-night monitoring of flood events. However, SAR-based approaches frequently suffer from false flood detections over smooth surfaces such as roads, salt pans, and paddy fields.
To address these limitations, we propose a machine-learning-based approach that integrates multi-sensor and multi-temporal data to enhance flood detection in Sentinel-1 imagery. Our study focuses on Kupang Regency, where we developed a flood-hazard frequency map using 10 years of hourly rain gauge corrected precipitation data of the Global Satellite Mapping of Precipitation (GSMaP) product developed by JAXA, combined with hydrogeological factors such as elevation, slope, land cover/land use, soil type, Topographic Wetness Index, and river characteristics. Additionally, we compiled a flood event catalog for 2021–2025 based on field surveys, news reports, and Sentinel-2 imagery, resulting in over 500 flood and non-flood reference points, respectively.
We applied three machine learning models: Random Forest (RF), Support Vector Machine (SVM), and Naïve Bayes (NB), to identify flooded areas by incorporating accumulated daily precipitation of GSMAP data, average daily soil moisture from ERA-5, the flood-hazard frequency map, and multi-temporal Sentinel-1 and Sentinel-2 imagery. Sentinel-1 data underwent preprocessing, including geometric and radiometric corrections, noise removal, terrain effect correction, and the masking of permanent water bodies.
Preliminary results indicate that the RF model outperforms SVM and NB in Sentinel-1-based flood detection. Feature importance analysis reveals that precipitation and soil moisture are the most influential factors in the RF model, followed by post-flood Sentinel-1 imagery. Next, we will evaluate the models on flood events caused by Tropical Cyclone Seroja in 2021, which brought 246 mm of rainfall on April 4. The performance of the model is validated by the Critical Success Index (CSI) and Probability of Detection (POD) using a comparison of the detected flood extents with the results of a distributed hydrological model.
This study aims to advance flood risk management by providing a data-driven approach to flood detection, particularly in data-scarce and flood-prone regions.
Traditional flood mapping methods, such as ground surveys and aerial observations, are often constrained by high costs, logistical challenges, and adverse weather conditions. Satellite-based remote sensing, particularly synthetic aperture radar (SAR) imagery (e.g., Sentinel-1), provides a cost-effective alternative, offering consistent, all-weather, and day-and-night monitoring of flood events. However, SAR-based approaches frequently suffer from false flood detections over smooth surfaces such as roads, salt pans, and paddy fields.
To address these limitations, we propose a machine-learning-based approach that integrates multi-sensor and multi-temporal data to enhance flood detection in Sentinel-1 imagery. Our study focuses on Kupang Regency, where we developed a flood-hazard frequency map using 10 years of hourly rain gauge corrected precipitation data of the Global Satellite Mapping of Precipitation (GSMaP) product developed by JAXA, combined with hydrogeological factors such as elevation, slope, land cover/land use, soil type, Topographic Wetness Index, and river characteristics. Additionally, we compiled a flood event catalog for 2021–2025 based on field surveys, news reports, and Sentinel-2 imagery, resulting in over 500 flood and non-flood reference points, respectively.
We applied three machine learning models: Random Forest (RF), Support Vector Machine (SVM), and Naïve Bayes (NB), to identify flooded areas by incorporating accumulated daily precipitation of GSMAP data, average daily soil moisture from ERA-5, the flood-hazard frequency map, and multi-temporal Sentinel-1 and Sentinel-2 imagery. Sentinel-1 data underwent preprocessing, including geometric and radiometric corrections, noise removal, terrain effect correction, and the masking of permanent water bodies.
Preliminary results indicate that the RF model outperforms SVM and NB in Sentinel-1-based flood detection. Feature importance analysis reveals that precipitation and soil moisture are the most influential factors in the RF model, followed by post-flood Sentinel-1 imagery. Next, we will evaluate the models on flood events caused by Tropical Cyclone Seroja in 2021, which brought 246 mm of rainfall on April 4. The performance of the model is validated by the Critical Success Index (CSI) and Probability of Detection (POD) using a comparison of the detected flood extents with the results of a distributed hydrological model.
This study aims to advance flood risk management by providing a data-driven approach to flood detection, particularly in data-scarce and flood-prone regions.