Identification of Flash Flood-Prone Areas in Arid and Semi-arid Regions Using Optical and Radar Imagery (Case Study: Semnan Province)

Extended Abstract

Introduction

Flash floods are among the most destructive natural disasters, causing substantial financial and human losses worldwide each year. These phenomena primarily occur in arid and semi-arid regions and have exhibited increased frequency and intensity in recent years due to climate change and intensified human activities. The application of satellite data and imagery within advanced analytical platforms such as Google Earth Engine provides precise, near-real-time information and robust cloud computing capabilities for rapidly processing large datasets. Sentinel satellite imagery represents one of the most advanced sources of Earth observation data available. Despite the effectiveness of radar imagery, accurately distinguishing between water and sandy areas in arid regions remains challenging due to the similar backscatter signatures of these surfaces in radar wavelengths. This research aims to identify flood-prone areas in Semnan Province, located within Iran's central desert basin, using Sentinel-1, -2, and -3 imagery combined with multiple images processing techniques, including composite analysis, NDWI (Normalized Difference Water Index), automated thresholding, and band differencing, to determine the optimal approach for delineating flood zones in arid mountainous environments.



Materials and Methods

This study focuses on identifying and evaluating flood-affected areas following the flash flood events recorded in May 2021, using an integrated remote sensing and hydrological analysis framework. The approach was initiated with the acquisition of Sentinel-1, Sentinel-2, and Sentinel-3 satellite imagery, which facilitated the high-precision delineation of flood-prone zones. A suite of image processing techniques was subsequently applied, including Automatic Thresholding using OTSU’s method to optimize pixel classification by maximizing inter-class variance within intensity histograms. To further enhance water body detection, the Normalized Difference Water Index (NDWI) was employed, leveraging the reflectance contrast between green (560 nm) and near-infrared (842 nm) bands. Additionally, spectral band composite techniques were utilized—particularly combinations such as RGB, NIR-SWIR-Red—to improve the differentiation of land surface features including vegetation, soil moisture, water bodies, and mineralogical attributes. Complementary to the remote sensing analysis, ground-based precipitation and peak instantaneous discharge data from hydrometric stations were extracted and analyzed for validation purposes. The flood mapping results were then compared with an existing flood susceptibility map generated through machine learning models, enabling an assessment of spatial accuracy. Finally, the effectiveness of each method was evaluated to identify the most suitable approach for mapping flood extents in arid and semi-arid mountainous regions.



Results and Discussion

Utilizing multi-sensor satellite imagery from Sentinel-1, Sentinel-2, and Sentinel-3, this study delineated flood-affected areas within Semnan Province with high spatial and temporal precision. Sentinel-1, operating with C-band microwave radar in VH polarization, proved particularly effective for flood mapping under persistent cloud cover. Through band differencing and thresholding techniques, this sensor enabled the identification of approximately 431,835 hectares of inundated land, primarily concentrated in the northern, central, and southern parts of the province. In parallel, Sentinel-2 optical imagery, processed through rigorous cloud masking and NDWI analysis, detected around 268,000 hectares of flooding, predominantly located within depressions and low-lying terrain. Although Sentinel-3, equipped with advanced multispectral sensors such as SLSTR and OLCI, offered extensive spatial coverage—encompassing an estimated 1,005,045 hectares—its lower spatial resolution (300 m) and heightened sensitivity to clouds and atmospheric interference limited its ability to capture smaller, discrete flood patches. To enhance delineation accuracy, automated OTSU thresholding was applied to the Sentinel-1 dataset, which improved boundary detection and reduced classification noise, resulting in a refined flood estimate of approximately 467,379 hectares. Validation against hydrometric station data revealed that over 60% of flash flood events recorded during the 2020–2021 hydrological year occurred in May 2021, showing strong temporal alignment with satellite-derived flood extents. Moreover, spatial comparison with flood susceptibility maps generated from machine learning models demonstrated substantial overlap in high-risk zones, reinforcing the reliability of the remote sensing-based mapping approach. Collectively, these findings underscore the complementary strengths and inherent limitations of both radar and optical sensors in capturing flood dynamics across arid and topographically complex regions. The integration of multi-source satellite data with advanced image processing techniques significantly enhances the accuracy and credibility of flash flood detection, thereby contributing valuable insights for disaster risk mitigation in data-scarce environments.





Conclusion

Exacerbated extreme events such as flash floods, driven by climate variability and human activities, pose serious hazards in arid and semi-arid regions where complex environmental, climatic, and geological conditions—combined with low infiltration capacities and limited historical data—complicate hydrological assessments. In dry mountainous zones, the prevalence of gypsum and limestone soils further reduces permeability, accelerating surface runoff and intensifying flood risks. This study demonstrated that advanced cloud computing platforms such as Google Earth Engine, integrated with multispectral and radar data from Sentinel satellites, provide accurate, rapid, and cost-effective means for identifying flood-prone areas in data-scarce environments. Among the evaluated techniques, automated Otsu thresholding applied to Sentinel-1 imagery proved the most effective for delineating flood extents, yielding refined estimates of approximately 467379 hectares.



Keywords: Sentinel; Image Processing; NDWI; OTSU; Semnan