Spatio-temporal dynamics monitoring of surface water bodies in Nhat Le River Basin, Vietnam, by Google Earth Engine

Nowadays, remote sensing has become a powerful and widely used technique for a wide range of applications (from environmental monitoring to disaster management, agriculture, forestry, urban planning, and beyond) due to its numerous advantages, i.e., wide coverage, high revisit frequency, valuable and good historical data, clear and stable images, rich information with low cost and sometimes free access, and real-time monitoring. Satellite sensors can be broadly categorized into two main types, i.e., active sensors and passive sensors. Optical images from passive sensors are indeed well-suited for specific studies and conditions, where cloud cover and daylight availability are less of an issue. On the other hand, radar sensors have the ability to penetrate cloud cover, completely neglect the cloud shadow effect, and provide data in all weather conditions. Recently, Sentinel-1, a part of the European Space Agency's Copernicus program, has become a prominent source of Synthetic Aperture Radar (SAR) data for water body mapping. With a spatial resolution of 10 m and a repeat cycle of 6 days for most of Europe and 12 days for other parts of the world, Sentinel-1 offers the creation of time series data, enabling long-term monitoring and analysis of surface water changes (Schlund & Erasmi 2020). Sentinel-1 data have been widely used for various studies, from flood extent detection (Bioresita et al. 2018; Uddin et al. 2019; Huang & Jin 2020) to temporal dynamics of surface water (Gulácsi & Kovács 2020), groundwater change monitoring (Liu et al. 2019), and soil moisture mapping (Paloscia et al. 2013).

The automatic extraction methods of remote sensing water body information are often categorized into two main approaches: threshold segmentation and image classification (Li et al. 2022). While the first approach uses the spectrum, texture, and spatial characteristics of both images and the ground for water extraction, the latter approach primarily extracts water based on its particular spectral characteristics. Although image classification methods generally produce superior results for water detection and extraction, they would be overly complex and time consuming for detecting surface water over a long study period due to difficulties in developing classification criteria and recurrent collection of training samples. On the contrary, the threshold segmentation technique is well known for its ease of use in extracting water by simply applying a threshold value to separate water pixels from other land cover categories (Atta et al. 2018). For optical images, the threshold value is typically based on the reflectance properties of the water surface and surrounding terrain. Therefore, the threshold segmentation approaches are often integrated with water indices such as normalized difference water index (NDWI) (McFeeters 1996) and modified normalized difference water index (mNDWI) (Xu 2006), which help to highlight the presence of water in an image from other categories by taking the normalized difference between the near-infrared and green bands. Various researchers have adopted these indices for their studies on water body detection, wetland management, landcover change, river and coastal morphological change, and disaster management (Eid et al. 2020; Quang et al. 2021a, 2021b; Shashikant et al. 2021). In contrast, for SAR images, the threshold value is usually determined by analysing the histogram of SAR backscatter intensity and estimating the probability distributions of water and non-water pixels. Thresholding-based methods can be categorized into two main groups: global thresholding and local thresholding (Liang & Liu 2020). In fact, the first method is more common since it employs a single threshold for all the images, making them well-suited for processing images with varying grey ranges between objects and the background. Otsu's method is one of the most widely used approaches in global thresholding, which selects an optimum threshold by maximizing the between-class variance (Huang & Wang 2009). This approach assumes the histogram of pixel values is bimodal and then selects the optimal threshold at the separated points of two classes in the histogram. Hence, Otsu's method can provide a dynamic threshold for each image to separate water and non-water areas with a fairly good accuracy compared to other methods (Bangira et al. 2019). The bimodal distribution in the pixel value histogram mainly focuses on the flood season when the water presence significantly increases, leading to lower backscatter submerged vegetation cover area, creating a second histogram peak. In contrast, the ground surface in the dry season is less affected by flooding and likely has a unimodal distribution in the histogram, reducing the accuracy of Otsu's method. Moreover, applying Otsu's method for a large study area of more than 2,600 km² can require relatively high computation and, hence, be less effective in real-time applications or high-resolution images. Therefore, this study is interested in investigating the annual variation of surface water, and then a unique threshold for the whole year is ideally sufficient. Furthermore, the method of using a single threshold can reduce time and effort for data processing and simply apply in other case studies.

Long-term studies with huge data collection require a high demand for data storage and processing time, which can be an obstacle for conventional Geographic Information System software, e.g., QGIS, ArcGIS, or ENVI. Fortunately, cloud computing has significantly transformed the landscape of big data analysis, including remote sensing data, by offering scalability, accessibility, and cost-effectiveness. Google Earth Engine (GEE), a cloud-based computing platform, typically makes use of high-performance parallel computing capabilities, a wealth of ready-to-use geospatial products, machine learning algorithms, and a library of Application Programming Interfaces with development environments that support well-known coding languages like JavaScript and Python (Tamiminia et al. 2020). GEE has been successfully utilized in problems related to large-scale and long-term surface water monitoring (Pekel et al. 2016; Pickens et al. 2020).

The purposes of this study are as follows: (1) to identify the optimal threshold value for surface water extraction in Nhat Le river basin from Sentinel-1 SAR image; (2) to develop a free GEE Tool for dynamic monitoring of surface water area; and (3) to assess the spatial and temporal dynamics of surface water area in Nhat Le river basin over the period 2016–2022.

Nhat Le is the second largest basin in Quang Binh province with an area of 2,622 km² (Figure 1). Two main tributaries, called the Long Dai and Kien Giang rivers, meet at Tran Xa village to form the Nhat Le River, which flows to the sea through the Nhat Le estuary. The topography of the Nhat Le River Basin is quite complex and gradually lowers from the west to the east. Separated by mountainous terrain in the west and high sand strips running along the coast in the east, Nhat Le delta is known as a flood-prone area with the lowest elevation of 0.7 m below sea level. Tung & Quang (2022) pointed out four main reasons for prolonged flooding in Nhat Le River Basin during the historic flood in 2020: (i) the low-lying and concave basin; (ii) the deposition of Nhat Le estuary – the unique drainage outlet of the basin; (iii) the construction of structures and infrastructures along the river and within the basin; and (iv) abnormally heavy rainfall coincides with the occurrence of high tide.

Several studies on surface water mapping using a threshold segmentation method recommended a selection of sub-case study, e.g., a lake or a reservoir, to represent the whole study areas like a basin, a province, and even a country (Feyisa et al. 2014; Zhou et al. 2020; Jiang et al. 2021). In this study, Phu Hoa reservoir, which was constructed in 2000 with a total capacity of 8.64 million m³ and a maximum water level of 30.86 m (Figure 1), is chosen as a sub-case study to identify the most suitable threshold range for surface water extraction in Nhat Le River Basin.

Optical images have been widely applied in water detection and delineation in previous studies (Huang et al. 2018); therefore, it is necessary to compare the performance of surface water extraction between radar and optical images. In this study, three images of Landsat 8, Sentinel-2, and Sentinel-1 are carefully chosen to represent the surface water conditions at Phu Hoa reservoir, the sub-case study, as closely as possible to the available date and approximate water level for the fair comparison (Table 1).

Being constructed in 2000, the bathymetry of Phu Hoa reservoir has been significantly altered due to sedimentation after 23 years of operation. Therefore, its updated topographic map has been collected from the survey campaign in 2019 to reconstruct the new water level – water surface area curve. Besides surveyed bathymetry, the water level of Phu Hoa reservoir for the period 2016–2022 is also collected to identify its surface water body, which will be used later to evaluate the accuracy of extracted results.

Several studies have pointed out that the optimal threshold value for separating water and non-water pixels may vary depending on the specific dataset, the time of acquisition, atmospheric conditions, and the nature of the water bodies being analysed. Jiang et al. (2021) stated that the suitable threshold values for surface water extraction in the spring, summer, fall, and winter are −14.45, −13.97, −14.78, and −16.32 dB, respectively. Tran et al. (2022) also suggested the values of −20 dB in the autumn, −24 dB in the winter, and −22 dB in the flood season for VH polarization images. Therefore, the collected images should cover the whole calendar year to propose the most suitable threshold range. Hence, in the second stage, 12 Sentinel-1 images from September 2020 to August 2021 that correspond to 12 months of a year are gathered (Table 2). The duration of collected images coincides with the rainy season in 2020 and the dry season in 2021, as well as the historic flood in October 2020. It is significant to convert the sensing time of collected images into Indochina Time, i.e., GMT + 7, to acquire the observed water level of Phu Hoa reservoir at the same time with the captured satellite image. By combining this water level with reservoir bathymetry, the ‘actual’ surface water can be derived and then will be used to validate the extracted results.

After applying the threshold, two land cover features named Water and Non-Water are included in the classified output with assigned values equal to 1 and 0, respectively. The accuracy assessment is then performed based on two accuracy metrics, i.e., overall accuracy (OA) and Cohen's Kappa coefficient (Kappa), to determine the best threshold range. Finally, the optimal threshold for surface water extraction for the study area is investigated based on surveyed flood marks and flood extent during the historical flood event on 18 October 2020 as follows:

• (i)

  Sentinel-1 image captured on 18 October 2020 is collected to extract surface water using different thresholds within the proposed threshold range.

• (ii)

  Munich-Re (Munich-Re 1997) defined flood as a temporary condition of surface water, in which the water level and/or discharge exceeds a certain value, thereby escaping from its normal confines. Therefore, the permanent water bodies such as rivers and lakes should be removed from extracted surface flooded area. In this research, the permanent water is obtained from a free global surface water data source called JRC global surface water mapping (Pekel et al. 2016).

• (iii)

  Removing steep areas is also significant to avoid misunderstanding natural drainage flow and noises from permanent mountain shadows with flood layers (Lin et al. 2019). In this study, digital elevation model data obtained from HydroSHEDS is used to remove areas with slopes above 5% over the basin.

• (iv)

  To eliminate small inundations that confound the final floodplain results, flood areas with coverage below eight connected pixels are also excluded.

• (v)

  Performance of extracted results is finally assessed based on surveyed flood marks and its boundary using OA and Kappa to identify the optimal threshold.

This section aims to assess the temporal dynamics and trends of water surface area as well as construct a water occurrence map, which shows the frequency of the pixel presented as water for Nhat Le River Basin over the studied period. The first task is done by adopting the chosen optimal threshold to delineate surface water from 380 collected Sentinel-1 images using GEE. The second aim can be obtained by evaluating the spatial distribution of the intra- and inter-annual dynamics of surface water. More specifically, the pixel value of all classified images collected from the same month is summed and averaged for the total number of images in that month to generate a map called the monthly surface water occurrence map. For instance, the pixel value of classified images in January 2016 is summed up with the pixel value of classified images in January 2017 and so on before averaging for the total number of collected images in January from 2016 to 2022. Finally, the pixel value of all 12 monthly surface water occurrence maps is averaged to obtain the final surface water occurrence map. The occurrence map is presented in the unit of percentage (%) and is classified into five categories: 0–10, 10–30, 30–50, 50–80, and 80–100% represent permanent not water, low seasonal, medium seasonal, high seasonal, and permanent water, respectively, as the suggestion of Pekel et al. (2016).

After narrowing the range of threshold, the trial-and-error method is applied again for the values ranging from −20 to −19 dB with a 0.25 dB increment to select the optimal threshold using the validation data in Nhat Le River Basin from the flood event on 18 October 2020. Table 5 reveals that the optimal threshold for surface water extraction in Nhat Le River Basin is −19 dB with 183/216 corrected flood marks, and the OA and Kappa values are 0.934 and 0.777, respectively.

The performance of extracted results from Sentinel-1 images is compared with the ones obtained from optical images since these images have been widely applied in delineating water bodies. More specifically, the surface water of Phu Hoa reservoir is derived from two optical images (Landsat 8 captured on 25 February 2021 and Sentinel-2 captured on 20 January 2021; Table 1), by using the two most common water indices, i.e., NDWI and mNDWI, with conventional thresholds that were suggested by Minh et al. (2022). Table 6 shows the performance of extracted results for Phu Hoa reservoir for both optical and radar images. Apparently, the chosen optimal threshold significantly improves the quality of surface water extraction from Sentinel-1 images with the values of OA and Kappa up to 0.944 and 0.859, respectively.

This study analyses not only temporal patterns of surface water distribution within the target basin but also the spatial distribution patterns are being evaluated. A spatiotemporal approach is employed to:

• Elucidate temporal variation in surface water levels and transitions between wet and dry periods.

• Evaluate spatial distribution and areal extent of surface water across the landscape.

• Identify regions vulnerable to water-related hazards by integrating temporal flow dynamics with spatial water distribution patterns.

By considering both temporal variation and spatial distribution of surface water simultaneously, this study aims to improve understanding of hydrological processes governing water movement within the basin. Insights gained can help inform hazard mitigation and water resource management strategies.

Our study provides a fully automatic processing chain for monitoring surface water and mapping water dynamics towards a well management of surface water in Nhat Le River Basin. We emphasize that the temporal dynamics follow two typical trends corresponding to dry and flood seasons throughout the study period, which can be influenced in specific years with abnormal natural hazards. Furthermore, the spatial dynamics show the transition of surface water in Nhat Le River Basin throughout the study period, and the most flooding-vulnerable area is Kien Giang River delta. The analysis of seasonal and inter-annual variation of surface water in this study states the temporal and spatial risk in the data-scarce region and improves the understanding of flood pattern, which highly contributes to water-related disaster risk management. In addition, our findings emphasize the importance of determining an optimal threshold for distinguished study areas to enhance the accuracy of surface water extraction. This study also provides a free-access tool for publicity to increase the feasibility of replication in further studies. Aiming to mitigate the challenges of water-related disasters, it is recommended that policymakers and stakeholders integrate our study's results to monitor surface water and prepare response plans.

Not only contributing to a clearer understanding of surface water dynamics in Nhat Le River Basin in both temporal and spatial scopes, but the study's method also has certain advantages compared to conventional methods in water-related disaster risk management. Firstly, regarding flood mapping, this method can determine potential flooding areas using free satellite data, while hydraulic models require multiple data such as elevation data, rainfall data, river flow data, and physical characteristic parameters. Furthermore, satellite data are globally covered and continuously updated on the GEE platform; hence, the effort to collect new data for different study areas or periods is mostly negligible. Secondly, thanks to the global scope of satellite data, the study approach can generate flood extent in a large study area. Moreover, the integration with GEE contributes to increasing time saving and reduces a lot of effort in model building and processing. Thirdly, this user-friendly approach allows users to map surface water dynamics at any time and region by simply changing input parameters, including the study area and the period. Finally, the water occurrence map is totally feasible to be used as an input to build a flood hazard map.

The study outcomes are also potentially applied in proposing a strategic regional plan and cultivation plan. According to the spatial analysis of the water occurrence map, Kien Giang river delta is the most vulnerable area to flooding in Nhat Le River Basin. Therefore, authorities and decision-makers should focus on this region in the strategic regional plan. For instance, the local government should invest more in applying structural and non-structural solutions for flood mitigation, such as anti-flood structures at the upstream Kien Giang River or low-impact development solutions. Not only focusing on proposing solutions for regional flooding situations but also the enhancement of local people's understanding and capacity plays a significant role in flood resilience. Particularly, local authorities can issue policies to support people living in Kien Giang, such as sending financial support to build typhoon-resilient low-income houses and raising human capacity building for resilience by holding special workshops or trainings. Furthermore, the spatiotemporal analysis has previously mentioned that floods might most likely occur in Kien Giang River delta, lasting from October to December. As Kien Giang River delta is the rice granary of the Central part of Vietnam, where the main source of income for most of the local people is from agricultural product trading, the impact of flooding can decrease rice productivity and threaten people's likelihood. Hence, the cultivation plan should be adjusted based on the study's results to reduce the negative impact of flooding on the locality. According to study outcomes, the most suitable time for starting the growing season should be at the beginning of February, when the flood season is going to end, and the surface water amount in the study area is abundant. Furthermore, shifting to plant short-duration rice varieties or floods and droughts-tolerant rice varieties might overcome the natural hazards in the study area.

As mentioned previously, this study's method is a fully automatic processing chain for detecting, monitoring surface water, and mapping water dynamics, which provides potentially rapid impact assessment of water-related disasters, i.e., floods and droughts, to study area. However, there are certain difficulties when replicating this approach in other study areas due to differences in regional characteristics, which require researchers to find a new optimal threshold for surface water detection. Therefore, it is necessary to gather sufficient data for processing and validating within at least one calendar year to ensure the objectivity of studies. Excluding the optimal threshold, the following stages of creating water dynamics mapping can be replicated without further adjustment.

The study aimed to automatically detect and monitor surface water dynamics in Nhat Le River Basin using Sentinel-1 satellite images. The threshold segmentation method is commonly used in surface water extraction from satellite data, and the proposal of an optimal threshold value of −19 dB for accurate surface water extraction in Nhat Le River Basin shows high reliability for both small and large areas. Comparisons with reference data from Phu Hoa reservoir confirmed the extracted surface water's accuracy, with OA and Kappa values surpassing 0.94 and 0.82, respectively. Applying the threshold on 18 October 2020 for flood extent mapping resulted in an OA of 0.93 and Kappa of 0.78.

Temporal analysis of water surface area in Nhat Le River Basin from 2016 to 2022 revealed two trends in surface water area: low levels during dry seasons, covering only 15% of the basin, and high levels during flood seasons, affecting nearly one-third of the basin. Floods were more likely between October and December, while severe droughts threatened the region from May to July.

Spatial analysis using a water occurrence map identified various water layers, including permanent not water, low seasonal, medium seasonal, high seasonal, and permanent water. The Kien Giang River delta emerged as the most vulnerable area to flooding, mainly covered by medium and high seasonal water layers.

The study's approach offered advantages over conventional hydraulic models for flood extent mapping, utilizing free and up-to-date data, saving time and effort in data processing, and being applicable to different study areas and periods. The surface water occurrence map proved valuable for flood hazard mapping and strategic regional planning by local authorities. However, the study acknowledged limitations, such as the need to improve accuracy and flood damage estimation. Future research should incorporate additional datasets like Digital Surface Model, Land Use/Land Cover data, and population density data to address these shortcomings and enhance the overall findings.

This research was supported by a Ministerial-level Research Project, namely, ‘Research on Digital Transformation in the Flood Warning Methods for the Community: An Experimental Flood Warning System for Nhat Le river basin, Quang Binh Province,’ funded by Ministry of Agricultural and Rural Development, Vietnam.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.