1. Introduction
Surface water plays a critical role in sustaining ecosystems and human activities, making it imperative for countries to establish comprehensive water monitoring and management systems. It is essential to ensure the availability and quality of this vital resource. Integrating multiple sources of water quality data and indicators into actionable information is a major challenge in water resource management and monitoring in the field of remote sensing. Researchers have been addressing this challenge, as evidenced by the works of Brown et al. (2015), El Serafy et al. (2021), and Yang et al. (2022), which have contributed to improving our understanding of the environmental, social, economic, and infrastructural aspects of water resources management
In recent years, the integration of remote sensing data with geographic information system (GIS) has revolutionized the automatic or semi-automatic extraction and mapping of water surfaces (Frey et al., 2010; Sarp, 2017). Remote sensing has emerged as a pivotal source of information for analyzing and providing data on changes in surface water (Feyisa et al., 2014). Compared to manual methods, remote sensing technology has significantly reduced the time and effort required for surface water mapping (Bijeesh and Narasimhamurthy, 2020). Water information extraction models typically use water’s spectral characteristics and imaging mechanisms directly (Luo et al., 2010). Various satellite sensors, with different spatial, temporal, and spectral resolutions, have been employed to extract and analyze surface water information (Feyisa et al., 2014).
Commonly used water extraction methods include multispectral water indices, such as the normalized difference water index (NDWI). NDWI utilizes data from the green- and near-infrared (NIR) band to accurately identify open water features (McFeeters, 1996). However, it is important to note that water classification based solely on water indices may not consistently distinguish between water pixels and non-watery dark surfaces (Verpoorter et al., 2014; Feyisa et al., 2014). Previous studies have shown that the main problems in water extraction research are about clearly making water edges and effectively distinguishing between water surfaces and shadows, particularly in urban areas or mountainous regions.
As a result, prior research has consistently employed the NDWI water index to address this concern. The NDWI index is used to highlight the characteristics of open water. This is because the green wavelength enhances water reflection while decreasing NIR reflectance from water features and the elevated NIR band reflectance from vegetation and soil features (Khalifeh Soltanian et al., 2019; Xu, 2006). In some cases, Xu (2006) replaced the NIR band with the shortwave infrared (SWIR) band, resulting in the development of the modified normalized difference water index (MNDWI) for improved analysis.
The MNDWI approach is because it delivers better results in extracting water features from satellite images compared to NDWI (Singh et al., 2015). This difference is due to NDWI’s lower sensitivity to vegetation that contains a lot of water, eg., paddy fields (Singh et al., 2015; Dennison et al., 2005). However, despite this, it has been observed that the NDWI index produces better outcomes and can effectively distinguish actual water areas and enhance water features in NDWI-derived images (Ghofrani et al., 2019; Khalifeh Soltanian et al., 2019). In our study focused on extracting water from open water regions such as reservoirs, lakes, and rivers, implementing the NDWI index is expected to yield positive results. As a result, using the MNDWI is not considered essential and is unlikely to have a significant impact on this study.
The Korean government has provided valuable spatial information about the land surface condition of Korea through digital maps and land cover maps. For the South Korea area, a digital map is available at a scale of 1:5,000, while for North Korea, it is provided at a scale of 1:25,000. These maps, developed for monitoring purposes, offer essential insights into the land surface, including water regions. Water regions on these maps encompass water surfaces, such as lakes, rivers, reservoir, banks, or oceans and their surrounding structures. Recognizing the significance of this comprehensive information, our study has chosen to utilize and compile the data from these maps to create a specialized database dedicated specifically to water bodies. The primary objective of this database is to support the water extraction process from satellite images, thereby enhancing our understanding of the water surface.
By leveraging the provided digital map and land cover map data, our research contributes to a deeper comprehension of the water surface, ultimately benefiting various applications related to water resource management and environmental studies. Through the utilization of the water regions database as our primary input, we have devised an automated water extraction method involving a two-stage process. This innovative approach holds the potential to enhance water extraction procedures. As a result, our future objectives are to apply this method to extract water from various satellite images. Furthermore, our study has achieved a remarkable enhancement in the precision of water surface extraction compared to earlier research. This is because our method effectively and automatically analyzes water extraction from satellite images, resulting in improved accuracy and efficiency within our study’s scope.
2. Water DB Collection
Our water DB product was created by collecting water feature data from two sources: the 1:5,000 and 1:25,000 digital maps of Korea and the land cover map of Korea with special emphasis on oceanic regions. A comprehensive water DB product for the Korean region was developed through a three-step process. In the first step, essential properties of the water surface were extracted from the digital map and land cover map data sources. These properties were then merged in the next step. Finally, a rasterization process was employed to generate an accurate representation of the water regions across the entire Korean peninsula. It is important to highlight that the data collection is slightly different between the South and North Korean regions, with distinct digital maps used for each area. Fig. 1 provides a detailed illustration of the algorithmic process utilized to create the water DB, while Fig. 2 showcases the constructed water DB covering the entire Korean peninsula and zooms in on specific water features in the Korean area. The figure shows that rivers, lakes, and coastal water regions were collected successfully.
Fig. 1. Workflow process of making a water DB product from a digital map and land cover map.
Fig. 2. This is the result of water DB product. (a) Water DB in raster type. (b) Water DB in polygon type. (c) Detail of water DB.
The water properties integrated into our water DB product accurately depict the unique characteristics of each water surface on the Korean peninsula. This comprehensive data set includes over 3,000 water properties extracted from both maps, encompassing diverse areas such as rivers, lakes, reservoirs, and coastal regions. The output of our water DB product is represented in a single package, offering various water properties available in two file formats: polygon and raster files, as described in the following sections. This valuable resource provides essential information for water resource management, environmental analysis, and hydrological studies in the Korean region.
3. Experiment Data
In this study, we aim to use this database to enhance water index analysis and delineate water surface precisely. Our study addresses the challenge of removing errors in water analysis indices using Sentinel-2 satellite images. Our proposed method consists of two main stages. In the first stage, we used NDWI and k-means clustering analysis on satellite images to delineate the initial water surface. In the second stage, we combined the results from the first stage analysis with our water DB product which was created using the digital maps and the land cover maps of Korea.
We utilized Sentinel-2 Level-1C data for experiments, which provides orthorectified top-of-atmosphere (TOA) reflectance with subpixel multispectral registration (Meraner et al., 2020). The data used in this study was obtained from the Copernicus open-access hub, various data portals, and platforms. In this study, we selected 15 representative scenes from Sentinel-2 Level-1C imagery, covering different parts of South and North Korea. To ensure a comprehensive analysis, we included 8 scenes with cloud cover and 7 scenes without cloud cover. It should be noted that the primary focus of our study is on the development and comparison of methods for water extraction from water regions. Therefore, the temporal or seasonal aspects of image acquisition were not the determining factors in our sample selection. Instead, we aimed to represent the area of water regions in the regions of South and North Korea, which are characterized by substantial water areas. Given the above challenge of discriminating between water surfaces and shadows during the extraction process, we strategically included cloudy scenes. This allows us to evaluate the performance of our approach under such challenging conditions. By using this approach, we can make meaningful comparisons and validate the effectiveness of our water surface extraction methodology. As a result, Table 1 presents a summary of the image data utilized in this research, and Fig. 3 illustrates the geographic locations of the data, defining our designated study area.
Table 1. Detail of satellite images for sample data
Fig. 3. This is an overall sample for study area. (a) The study area of all the samples of this study. (b) The sample area for the North Korean region. (c) The sample area for the South Korean region.
To ensure the accuracy assessment of our study, we have created ground truth data for 15 sample images. These ground truth data points are strategically located within the study area covered by the 15 images. The process of generating the ground truth data involved two methods. Firstly, we utilized water features provided by digital maps. Secondly, we conducted a manual process to accurately identify and mark water surfaces within each of our sample locations. During the creation of ground truth data, we cross-referenced V-World satellite images as map backgrounds. This approach ensured that the ground truth data reflected the real conditions in the field, aligning with the water surface present in our sample locations. For a detailed illustration of our ground truth data production process supporting this research, please refer to Fig. 4. By employing these meticulous procedures for generating ground truth data, our study ensures a reliable basis for assessing the accuracy of our results and reinforces the credibility of our findings.
Fig. 4. This is the ground truth data algorithm and process. (a) The flow process for making ground truth data. (b) Manual editing process with a base map layer. (c) The ground truth results.
4. Water Surface Extraction Method
Upon acquiring the water DB product, we employed it in our automated water extraction method to enhance water surface extraction from satellite imagery and increase overall accuracy (OA). Our method involved a two-stage process for our automated water extraction. In the first stage, we utilized the NDWI and k-means clustering analysis to delineate the water regions from sample images captured by the Sentinel-2 satellite imagery. For the NDWI analysis, we utilized the NIR and green band images. Subsequently, the results from the NDWI analysis were further processed using k-means clustering analysis as a second approach. As our study aimed to develop a water DB product for the entire Korean peninsula, we employed this product for the main stage of our automatic water extraction method. We combined the results from the first stage analysis with the water DB product as the primary input for the analysis. This was followed by image processing analysis, including image segmentation algorithms and binary mask analysis. Overall, our automated water extraction method generated three main results, the first approach (utilizing water DB product), the second approach (utilizing NDWI analysis), and the third approach process (using k-means clustering analysis). The overall algorithm process in this study is presented in Fig. 5, while the specific algorithm process flow for water extraction is presented in Fig. 6.
Fig. 5. Automated water extraction full workflow.
Fig. 6. Workflow for water extraction process. DN: digital number.
4.1. Water Extraction with Water DB
In our study, we adopted the first approach for our automated water extraction method, where we utilized our water DB product to delineate the water surface from satellite images. The sample data consisted of Sentinel-2 images as the first input, supplemented by our water DB product as the second input. To extract water areas, we employed image segmentation as the first stage in this first approach, followed by the generation of a binary mask. Image segmentation divides an image into contiguous and homogeneous regions, known as segments, serving as the basis for further analysis (Blaschke et al., 2004; Nussbaum and Menz, 2008; Kotaridis and Lazaridou, 2021).
These algorithms group pixels based on three essential criteria: homogeneity within a segment, differentiation from adjacent segments, and shape homogeneity (Nussbaum and Menz, 2008; Kotaridis and Lazaridou, 2021). In our study, we adopted spectrally based methods that analyze individual pixels (ThenKabail, 2015; Kotaridis and Lazaridou, 2021). Specifically, we utilized thresholding-based algorithms, a well-established and widely used technique for image segmentation. Employing a threshold value greater than 0, we delineated the boundaries for extracting water body areas. Hence, the final step in this stage involved binary mask image processing to distinguish non-water areas and water areas, assigning a digital number (DN) value of 0 and 1, respectively. These outcomes form the basis of primary approach findings.
4.2. Water Extraction with NDWI Analysis and Water DB
The NDWI plays a critical role in enhancing the local water signal. To optimize the spectral decision space for water discrimination, multispectral vegetation, and water indices exploit the differences in reflectance (ρ) between water, vegetation, and other types of land cover types in the visible and infrared wavelengths. The NDWI, introduced by McFeeters (1996), is as follows:
\(\begin{aligned}N D W I=\frac{\rho \text { Green }-\rho N I R}{\rho \text { Green }+\rho N I R}\end{aligned}\)
In our second approach, we incorporated this NDWI analysis as the first stage. The sample image from Sentinel-2 NIR and green band images were processed as the first input for the NDWI analysis, as mentioned earlier. This analysis played a crucial role in effectively delineating water bodies from water regions, in satellite imagery. After obtaining the results from the NDWI analysis, we utilized them as the first input, along with our water DB product as the second input, in our automated water extraction method. The application of NDWI analysis to the images before their integration with our water DB product for the water extraction process is expected to yield favorable results.
NDWI analysis is expected to effectively delineate the boundaries between water and non-water areas. Consequently, when this analysis is combined with our water DB, errors in water detection during the extraction process are likely to be reduced. Hence, in the water extraction process, we employed image segmentation and binary mask generation as a process in the second stage. The design of this second approach aimed to address any potential imperfection that may have appeared in the first approach. By combining the insights from NDWI analysis and the utilization of our water DB product, we sought to improve the accuracy and robustness of our automated water extraction process.
4.3. Water Extraction with NDWI + K-means Clustering and Water DB
K-means is a numerical, unsupervised, and iterative method that has proven to be highly effective in producing reliable clustering results (Na et al., 2010). Clustering is used to classify the raw data in a meaningful way and uncover any hidden patterns that may be present in the dataset (Ng et al., 2007). It involves the grouping of data objects into different clusters, where objects within the same cluster have similarities, while the different clusters are distinct (Na et al., 2010). In our study, we applied k-means clustering as the basis for our third approach. We utilized the NDWI analysis results from the first stage and employed k-means clustering for further analysis.
As a result, it is expected that the results of the k-means analysis in the second stage will provide clear and accurate boundaries that effectively distinguish water from non-water areas. In addition, the integration of these results with our water DB product is expected to improve the accuracy of water region delineation and correct for the shortcomings of the first and second approaches. Consistent with the preceding methodology, the final analysis of the water extraction process involves image segmentation and binary mask processes. The objective of this approach was to effectively differentiate between water bodies, clouds, and terrain shadows, with a particular focus on areas with cloud cover in the sample data. By doing so, we gained a deeper understanding of the spatial distribution and characteristics of these different features within the dataset. We anticipate that this approach will help address any limitations from our first and second approaches, ultimately enhancing the overall quality and completeness of our results.
5. Results
We have obtained water masks through our approaches of automated water extraction methods. In this study, we present our results from the first, second, and third approaches obtained from all samples and compare them with the first approach from our automatic water extraction. However, it has been mentioned that we faced challenges in detecting smaller features such as streams, lakes, and reservoirs. In addition, noise due to misclassification of terrain shadows or building shadows, which share similar pixel values with water bodies, posed further difficulties. In contrast, our three models successfully addressed and improved upon these limitations. They effectively delineated more detailed water bodies while reducing noise from non-water areas.
Looking at Table 2, we can see the outcomes of our first approach, which involved using solely our water DB product to extract water bodies from satellite images. These results demonstrated improvement and consistently yielded better outcomes when contrasted with the outcomes of the conventional method that relied on only NDWI analysis and k-means analysis. To validate the quality of our resolution results, we employed a visual interpretation method to compare our water extraction results with the reference data obtained from the global surface water (GSW) mask, which offers a resolution of 60 m for the same location within our study area. Encouraged by these outcomes, in this table we also show our second and third approach results, incorporating NDWI analysis and k-means clustering, along with our water DB, which also yielded high-quality results. In summary, our study’s findings indicate a notable enhancement quality of water mask resolution. Our approach yielded superior results compared to both the conventional method and the reference data water mask outcomes. Furthermore, when visually comparing the results from our first to third approaches, we consistently achieved a high standard of quality in the water extraction process.
Table 2. Visual comparison results from 1st and 2nd approaches with reference data from GSW mask and conventional results
Furthermore, in line with our objectives, we wanted to show the improved accuracy achieved by our second and third approaches in contrast to the results of the first approach. To illustrate this, we performed a comparison of the three sets of results, and the results are presented in Table 3. Examining the results in Table 3, a notable observation emerges: while the results of the first approach failed to identify small areas in the middle of the river as land, both the second and third approaches demonstrated an advanced ability to accurately distinguish between water and land regions. In the first approach, the small area of land in the middle of the river was incorrectly identified as water. However, subsequent analyses performed in the second and third approaches corrected this by accurately delineating this small area of land as part of the water body. Therefore, these results decisively validate the achievement of our intended goals in this particular case.
Table 3. Detailed comparison results of 1st approach with 2nd and 3rd approaches
Hence, we conducted an accuracy assessment using an occurrence matrix to calculate the OA of our results that we present in Tables 4 and 5. In this table, we present a comparison of our accuracy result with reference data from GSW which obtained its water mask product with a 60 m resolution from collected water extraction results from Landsat satellite imagery. This approach provided a robust means of assessing the performance and reliability of our methods. The accuracy results demonstrated that our approaches to OA and kappa value have shown improvements compared to the reference data.
Table 4. Accuracy assessment results of images with cloud cover for each approach
Table 5. Accuracy assessment results of images without cloud cover for each approach
The results show that the accuracy of our results in all approaches exceeds that of the GSW mask reference data. As mentioned above, our approaches have successfully improved the resolution compared to the GSW mask result. This improvement is due to the ability of our approaches to delineate water areas in detail, which is achieved through the use of our water DB product. As a direct consequence, our accuracy values also exceed those of the GSW reference data. Thus, the accuracy results of our first, second, and third approaches show a consistent trend of high accuracy. This can be attributed to the use of our water DB in all three approaches, resulting in increased accuracy levels for each approach. In addition, since the water DB is compiled from comprehensive water features data obtained from the Korean government’s digital and land cover maps, the inherent quality of the map contributes to its OA. However, it is worth noting that while the second and third approaches have demonstrated the ability to correct the deficiencies observed in the first approach, this refinement is not universally applicable and is limited to specific locations.
In addition, the results also showed large differences between the OA and the kappa value of the GSW mask accuracy results. This can be explained because the discrepancy between the very low kappa value in the GSW mask result compared to the OA is due to the nature of the occurrence matrix analysis. This analysis calculates values by dividing the total number of instances in the datasets. Consequently, in this scenario, the detection accuracy of non-water areas in the GSW mask in this analysis appears commendable, as it agrees well with the ground truth data. This accuracy is primarily due to the generally higher accuracy of non-water area detection. However, when it comes to water areas in the GSW mask, the accuracy is significantly lower compared to the ground truth data. This discrepancy between water detection and ground truth data contributes to the low kappa value. Despite the low kappa value, the OA value is still high.
To demonstrate the superiority of our proposed method over previous approaches that rely solely on water indices NDWI analysis or k-means clustering algorithm for water body extraction, we performed a comparative analysis. In Tables 6 and 7, we present our result comparison with the conventional method that involved water extraction using only NDWI and k-means. The accuracy assessment results clearly show how significant progress was achieved by our approach compared to the conventional method. Specifically, our method achieved an impressive OA of up to 90% and a kappa value of up to 0.8. These results underscore the remarkable improvements achieved by our first, second, and third approaches over the conventional method.
Table 6. Accuracy assessment results of images with cloud cover for comparison with conventional method
Table 7. Accuracy assessment results of images without cloud cover for comparison with conventional method
However, the misclassification of water bodies in the NDWI results and the significant noise in the k-means classification contribute to the reduced accuracy observed in the results of the conventional method. This inaccuracy is significantly lower when compared to the results of our approaches, as convincingly demonstrated in Table 2. The visual representation of the resolution of the NDWI and k-means results presented in Table 2 confirms the lower quality of these results compared to those obtained by our approach.
Table 8. Visual comparison of conventional method results with new approaches
To demonstrate the improvements in our results over the conventional, we present our figure in Table 7 for visual comparative analysis. Upon close examination of the results, it is clear that the NDWI analysis alone struggles to extract intricate details of the water bodies from satellite imagery. Although the combination of NDWI analysis and k-means clustering improves the results compared to NDWI alone, issues such as noise from clouds or shadows remain significant. Our method, on the other hand, effectively addresses these imperfections and yields significant improvements in the final results.
6. Discussion
Our study aims to develop an improved automated water extraction method by utilizing our water DB product to aid in extracting water bodies from satellite imagery. To achieve this, we evaluate our results against multiple approaches, including our first approach and our second and third approaches, which employed NDWI and the k-means clustering algorithm for the automated water extraction process. Additionally, we compared our method with the conventional method used in previous studies to demonstrate the advancements made by our proposed technique. This comparison allows us to assess the significant improvements our methods offer over our conventional methods from previous research. We acknowledged that the NDWI approach can encounter difficulties due to challenges in thresholding, primarily caused by similarities in pixel values between water bodies and terrain shadows. Similarly, the k-means clustering algorithm, based on pixel value distances, could also lead to misclassifications if thresholds were not carefully defined.
To further validate our findings, we integrated reference data from the GSW mask, generated using high-resolution Landsat imagery. Visual interpretation confirmed the high-resolution quality of our water mask. While our initial water DB product yielded satisfactory results, we recognized its imperfections and decided to combine it with additional analyses as part of our second and third approaches. Remarkably, the incorporation of these supplementary analyses did not significantly impact the delineation of water extraction using our water DB product.
However, in specific cases, the results from the second and third approaches demonstrated improvements over the first approach. To comprehensively evaluate our method’s performance, we utilized two different sample sets, one with cloud cover and the other without cloud cover. Our approach successfully delineated the area of water bodies even under cloud cover, demonstrating its effectiveness. Overall, our study highlights the effectiveness of our approach, which combines the strengths of our water DB product with other analyses, resulting in enhanced accuracy and quality in water body delineation from satellite images.
7. Conclusions
Given the significance of our water DB product in the development of our automatic detection method, our overall findings demonstrate its effectiveness in accurately delineating water bodies from satellite images. The incorporation of water properties from our water DB product greatly aided the water extraction process, enabling precise and detailed delineation that closely matches real conditions. A comparative analysis between our water extraction results and conventional methods such as NDWI analysis and k-means clustering revealed substantial improvements in terms of quality and resolution.
This improvement was evident in our accuracy assessment results, with an OA of 90% and kappa values exceeding 0.8. In addition, our visual interpretation involved a comparison with the GSW mask, which confirmed that our results exhibited superior resolution compared to the reference data. To address any imperfections in our first approach, we devised a strategy to combine our first approach with additional analyses as second and third approaches.
Although these supplementary methods did not significantly impact the water extraction results overall, there were notable improvements observed in specific sample areas, aligning with our objective of refining the initial outcomes. Overall, our study underscored the effectiveness of our water DB product in enhancing the delineating of water bodies from satellite images. The utilization of water properties and the integration of other analysis approaches contributed to the improved quality and resolution of our results. This demonstrated the value of our approach in achieving more accurate and detailed water extraction, particularly in the Korean region.
Acknowledgments
This study was conducted as part of the task of the Korea Forest Service’s Nation Institute of Forest Science, “Reception, Processing, ARD Standardization, and Development of Intelligent Forest Information Platform” (Task number: FM0103-2021-01).
Conflict of Interest
No potential conflict of interest relevant to this article was reported.
References
- Bijeesh, T. V., and Narasimhamurthy, K. N., 2020. Surface water detection and delineation using remote sensing images: A review of methods and algorithms. Sustainable Water Resources Management, 6, 68. https://doi.org/10.1007/s40899-020-00425-4
- Blaschke, T., Burnett, C., and Pekkarinen, A., 2004. Image segmentation methods for object-based analysis and classification. In: Jong, S. M. D., Meer, F. D. V. (eds.), Remote sensing image analysis: Including the spatial domain, Springer, pp.211-236. https://doi.org/10.1007/978-1-4020-2560-0_12
- Brown, C. M., Lund, J. R., Cai, X., Reed, P. M., Zagona, E. A., Ostfeld, A. et al., 2015. The future of water resources systems analysis: Toward a scientific framework for sustainable water management. Water Resources Research, 51(8), 6110-6124. https://doi.org/10.1002/2015wr017114
- Dennison, P. E., Roberts, D. A., Peterson, S. H., and Rechel, J., 2005. Use of normalized difference water index for monitoring live fuel moisture. International Journal of Remote Sensing, 26(5), 1035-1042. https://doi.org/10.1080/0143116042000273998
- El Serafy, G.Y., Schaeffer,B.A., Neely, M.B., Spinosa, A., Odermatt, D., Weathers, K. C. et al., 2021. Integrating inland and coastal water quality data for actionable knowledge. Remote Sensing, 13(15), 2899. https://doi.org/10.3390/rs13152899
- Feyisa, G. L., Meilby, H., Fensholt, R., and Proud, S. R., 2014.Automated water extraction index:A new technique forsurface water mapping using Landsat imagery. Remote Sensing of Environment, 140, 23-35. https://doi.org/10.1016/j.rse.2013.08.029
- Frey, H., Huggel, C., Paul, F., and Haeberli, W., 2010. Automated detection of glacier lakes based on remote sensing in view of assessing associated hazard potentials. Grazer Schriften der Geographie und Raumforschung, 45, 261-272. https://doi.org/10.5167/uzh-128917
- Ghofrani, Z., Sposito, V., and Faggian, R., 2019. Improving flood monitoring in rural areas using remote sensing. Water Practice and Technology, 14(1), 160-171. https://doi.org/10.2166/wpt.2018.118
- Khalifeh Soltanian, F.,Abbasi, M., and Riyahi Bakhtyari, H. R., 2019. Flood monitoring using NDWI and MNDWI spectral indices: A case study of Aghqala flood-2019, Golestan Province, Iran. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 42,605-607. https://doi.org/10.5194/isprs-archivesxlii-4-w18-605-2019
- Kotaridis,I., and Lazaridou, M., 2021.Remote sensing image segmentation advances: Ameta-analysis. ISPRS Journal of Photogrammetry and Remote Sensing, 173, 309-322. https://doi.org/10.1016/j.isprsjprs.2021.01.020
- Luo, J., Sheng, Y., Shen, Z., and Li, J., 2010. High-precise water extraction based on spectral-spatial coupled remote sensing information. In Proceedings of the 2010 IEEE International Geoscience and Remote Sensing Symposium, Honolulu, HI, USA,July 25-30, pp. 2840-2843. https://doi.org/10.1109/IGARSS.2010.5648978
- McFeeters, S. K., 1996. The use of the normalized difference water index (NDWI) in the delineation of open water features. International Journal of Remote Sensing, 17(7), 1425-1432. https://doi.org/10.1080/01431169608948714
- Meraner,A., Ebel, P., Zhu, X. X., and Schmitt, M., 2020. Cloud removal in Sentinel-2 imagery using a deep residual neural network and SAR-optical data fusion. ISPRS Journal of Photogrammetry and Remote Sensing, 166, 333-346. https://doi.org/10.1016/j.isprsjprs.2020.05.013
- Na, S., Xumin, L., and Yong, G., 2010. Research on k-means clustering algorithm: An improved k-means clustering algorithm. In Proceedings of the 2010 3rd International Symposium on Intelligent Information Technology and Security Informatics, Jian, China, Apr. 2-4, pp. 63-67. https://doi.org/10.1109/iitsi.2010.74
- Ng, M. K., Li, M. J., Huang, J. Z., and He, Z., 2007. On the impact of dissimilarity measure in k-modes clustering algorithm. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(3), 503-507. https://doi.org/10.1109/tpami.2007.53.
- Nussbaum, S., and Menz, G., 2008. Object-based image analysis and treaty verification: New approachesin remote sensing-applied to nuclear facilities in Iran. Springer Science & Business Media.
- Sarp, G., and Ozcelik, M., 2017.Water body extraction and change detection using time series: A case study of Lake Burdur, Turkey. Journal of Taibah University for Science, 11(3), 381-391. https://doi.org/10.1016/j.jtusci.2016.04.005
- Singh, K. V., Setia, R., Sahoo, S., Prasad, A., and Pateriya, B., 2015. Evaluation of NDWI and MNDWI for assessment of waterlogging by integrating digital elevation model and groundwater level. Geocarto International, 30(6), 650-661. https://doi.org/10.1080/10106049.2014.965757
- Thenkabail, P. S., 2015. Remotely sensed data characterization, classification, and accuracies. CRC Press. https://doi.org/10.1201/b19294
- Verpoorter, C., Kutser, T., Seekell, D. A., and Tranvik, L. J., 2014. A global inventory of lakes based on high-resolution satellite imagery. Geophysical Research Letters, 41(18), 6396-6402. https://doi.org/10.1002/2014gl060641
- Xu, H., 2006. Modification of normalised difference water index (NDWI) to enhance open water features in remotely sensed imagery. International Journal of Remote Sensing, 27(14), 3025-3033. https://doi.org/10.1080/01431160600589179
- Yang, L., Driscol,J., Sarigai, S.,Wu, Q., Lippitt, C. D., and Morgan, M., 2022. Towards synoptic water monitoring systems: A review of AI methods for automating water body detection and water quality monitoring using remote sensing. Sensors, 22(6), 2416. https://doi.org/10.3390/s22062416.