AUTOMATED DETECTION OF WATER SURFACES FROM SENTINEL-2 IMAGES AND PERFORMANCE ANALYSIS

초록

Water, which covers more than 50% of the Earth's surface, plays a crucial role in scientific research and is of immense importance in applications such as disaster prevention, urban planning, and water resource management. Precise and automatic detection of water bodies in remote sensing images has become a critical and challenging task in various studies. Therefore, to effectively address these various concerns, automated monitoring of water bodies is of paramount importance. To address the challenge of detecting water bodies in remotely sensed images, researchers have investigated various methods in previous studies. These approaches include the analysis of water indices, such as the Normalized Difference Water Index (NDWI), derived from the visible or infrared bands of satellite imagery. In addition, researchers have used k-means clustering analysis to identify patterns in land cover and segment regions with similar characteristics. Despite ongoing challenges such as distinguishing water spectral signatures from artifacts such as cloud and terrain shadows, our study aims to address these issues. The primary objective is to significantly improve the accuracy of water surface detection by constructing a comprehensive water database using existing digital and land cover maps. To achieve this, we used 1:5000 and 1:25000 digital maps of Korea to extract water properties, including rivers, lakes, and reservoirs. In addition, the inclusion of the 1:5000 and 1:50000 Land Cover Map of Korea was instrumental in extracting water properties from oceanic areas. Our research highlights the effectiveness of using the Water DB layer as the primary approach for efficiently extracting water surfaces from satellite imagery. This primary approach consists of two methods: the first involves the use of water indices through NDWI analysis, while the second employs unsupervised classification techniques, specifically the k-means clustering algorithm. During the water extraction process, we incorporated image segmentation and binary mask methods for image analysis. To evaluate the accuracy of our approach, we performed two evaluations using both reference data and our ground truth data. Visual interpretation involved comparing our results to the Global Surface Water (GSW) mask, which showed significant improvements in both quality and resolution. In addition, accuracy evaluation measures, including an overall accuracy (OA) of 90% and kappa values greater than 0.8, provide substantial evidence of the effectiveness of our methodology. In summary, our primary approach produced superior results compared to traditional methods used in previous studies and reference data, producing water mask results with improved resolution. Specifically, our first approach using (NDWI) analysis demonstrated higher accuracy than our second approach using k-means clustering for water detection. Overall, our approach consistently outperforms conventional methods. © 2023 ACRS. All Rights Reserved.

키워드