Novel Index for Hydrological Drought Monitoring Using Remote Sensing Approach: Standardized Water Surface Index (SWSI)

Abstract

Most of the drought indices designed for hydrological drought monitoring use location-specific data, while there are only a handful of indices designed for hydrological drought monitoring using remote sensing data. This study revealed a novel drought index, Standardized Water Surface Index (SWSI), developed for hydrological drought monitoring. The water surface areas required to calculate the SWSI can be extracted from remote sensing data entirely using both the optical (Landsat 5, 7, and 8) and SAR (Sentinel-1). Furthermore, the developed index was applied to five major reservoirs/tanks; Iranamadu, Mahavilachchiya, Kantale, Senanayaka Samudhraya, and Udawalawa, located in Sri Lanka to monitor respective hydrological drought status for the period from 2000 to 2020. Cloud computing platform such as Google Earth Engine (GEE) provides a good basement to use this index effectively, as it can extract long-term water surface area covering a large geographical area efficiently and accurately. The surface water area extraction from satellite data of those tanks shows an accuracy of more than 95%, and in the event of a severe hydrological drought, the water surface area of the tanks is less than 25% of the total and lasts for more than three to four months. It was also determined that in some years, the surface water area of tanks dropped to as low as 7%. The strong correlation observed between the Standardized Precipitation Index (SPI) and SWSI is indicated by the Pearson correlation coefficient ranging from 0.58 to 0.67, while the correlation between the Vegetation Condition Index (VCI) and SWSI ranges from 0.75 to 0.81. Timely drought monitoring over large geographical areas can be more accurately performed with the SWSI index compared to existing hydrological drought monitoring indices. The SWSI could be more useful for areas that do not have measurable field data.

1. Introduction

Drought is a severe natural disaster that can have serious environmental and social consequences. It directly affects the growth of vegetation and accelerates the degradation of grasslands and the disappearance of wetlands [1,2]. Rapid depletion of rivers and lakes due to drought poses a greater risk to ecosystems. Accurate and timely drought assessment is especially important as the incidence of drought has increased due to the increasing trend of climate warming around the world [3]. Therefore, detailed studies on drought are very important to take necessary measures to mitigate the severe effects of drought. Thus, detailed studies of drought have attracted extensive research attention, and a large number of drought monitoring indices have been developed over the past few decades as a robust tool for qualitative and quantitative estimation of different types of drought [4,5,6].

Every drought is different because of the affected environment, the drought system, the population, the economy, and the attitudes of the people. Simply, the effects of drought are not limited to water resources and agriculture [7]. Despite technological advancement, the effects of the drought have been intensified by the expansion of hydropower generation, increasing demand for water, population growth, expansion of irrigated lands, and increasing per capita water use [8].

There are four types of droughts, namely meteorological, agricultural, hydrological, and socio-economic [9]. The occurrence of hydrological drought is usually indicated by a significant decrease in the surface water quality of rivers, lakes, reservoirs, and groundwater. The hydrological drought is especially observed after meteorological and agricultural droughts. The main reason for this is that groundwater, as well as surface water sources, are not recharged due to the long-term gradual decrease in rainfall [10]. Among other droughts, hydrological drought has been gaining more attention because water is an integral part of daily human life. However, a relatively small number of indicators have been introduced for hydrological drought monitoring compared to meteorological and agricultural drought monitoring indices [11].

Various researchers developed standardized indices for drought monitoring using a reliable and flexible method used for Standardized Precipitation Index (SPI) calculation [12]. Standardized Runoff Index (SRI), Standardized Reservoir Supply Index (SRSI), Standardized Streamflow Index (SSFI/SSI), and Standardized Snow Melt and Rain Index (SMRI) can be introduced as standardized indices designed for hydrological drought monitoring [13,14,15]. The most commonly used data for calculating these standardized indices were location-specific measurements such as river discharge, reservoir water level, groundwater level, and surface flow [16,17].

Since the onset of hydrological drought is well reflected by a continuous decrease in surface water resources, monitoring its long-term variation is more effective for studying hydrological drought. Previous studies suggested that the measurement and monitoring of surface water area using remote sensing technology provides valuable information for hydrological drought-related studies [18,19]. Although not directly relevant to drought studies, many studies have shown that different data and algorithms are used to map the surface water area at different spatial and temporal resolutions with greater accuracy [20,21]. However, the low cost and high spatial and temporal resolution of satellite data have led researchers to focus more on satellite-based methods than conventional methods for surface water area extraction.

Various remote sensing data captured from different satellite missions such as Landsat Multispectral Scanner (MSS), Thematic Mapper (TM), Enhanced Thematic Mapper Plus (ETM+) [22,23], Moderate Resolution Imaging Spectroradiometer (MODIS) [24,25,26], and National Ocean, Atmospheric Administration (NOAA), Advanced Very High Resolution Radiometer (AVHRR), and Sentinel-1 Synthetic Aperture Radar (SAR) have been used to detect and monitor the surface water bodies under various spatial and temporal resolutions from the 1970s to the present [27,28,29].

On the other hand, remote sensing (RS) based methods for surface water detection can be roughly divided into two classes; feature classification methods and thematic water surface extraction algorithms. Even though the most commonly used feature classification methods can be identified as spectral mixing models [30,31], maximum probability classification [32], and artificial intelligence methods [33,34,35], these methods are not easy to use for a large area and long-term use of multiple-temporal images to extract the water surface area. The thematic water surface extraction algorithm can be classified as row satellite imagery classification and the water indexes calculation [36]. Water indices created through multi-satellite imagery are more appropriate than single-satellite image bands because they take into account the reflective differences that exist between water and land [37]. It is more effective to extract surface water using remotely sensed data due to its accuracy, ease of use, speed, and reproducibility, as well as the ability to analyze time-series data [38].

Despite some data gaps, Landsat data represent the world’s longest-running terrestrial satellite record, making it more successful for the study of long-term variability in hydrological activity and climate change [39]. Even though Landsat data have been available since 1972, it is more challenging to analyze them over the entire period (1972 to 2020) with traditional Geographical Information System (GIS) and RS methods [40]. As a result of significant technological improvement, various cloud computing platforms around the world have been set up to facilitate the easy analysis of large-scale geographical data [41]. The Google Earth Engine (GEE) is a high-performance cloud-based platform that incorporates free multi-petabyte remote sensing data freely available worldwide [42]. The GEE has successfully been used to map regional, continental and global level land cover [43,44], cropland areas, forests [45], and open surface water bodies [46,47].

It is not easy to capture historical droughts through more successful hydrological drought monitoring using such data as it is common in most reservoirs to have large data gaps and irregular data in the depth of manually measured reservoirs. Although more successful and complex methods have been introduced to calculate surface water extent using both optical and SAR satellite data, as discussed above, this study demonstrates the possibility of defining an index for hydrological drought monitoring using reservoir water surface area as a reasonable substitute for depth measurements in shallow and gently sloping reservoirs. Thus, the main objective of this study was to introduce a novel index, the Standardized Water Surface Index (SWSI), for hydrological drought monitoring, by measuring the time-series water surface area (WSA) using both optical and SAR satellite data. The GEE was employed to calculate the long-term water surface area on a monthly basis from 2001 to 2020. On the other hand, since both optical and SAR data are used to calculate the surface water area, the accuracy of the area calculated from them was studied using a correlation coefficient. Furthermore, the relationship of the novel index, from SWSI to the SPI and Vegetation Condition Index (VCI), was also analyzed to investigate the performance of the SWSI index and determine its validity using the available information, including the surface water area of the five tanks in Sri Lanka by considering the specific hydrological setting.

2. Materials and Methods

2.1. Study Area

It is preferable to use tank areas for SWSI calculations rather than geographical areas such as districts or provinces, as SWSI calculation directly uses monthly surface water area changes. Therefore, as indicated in Figure 1, the five tanks in Sri Lanka, Tank A (Iranamadu), Tank B (Mahavilachchiya), Tank C (Kantale), Tank D (Senanayaka Samudhraya), and Tank E (Udawalawa), were selected as case studies to apply and investigate the performance of the developed index.

Sri Lanka is an island, which is located at latitudes 5–10°N and longitudes 79–82°E below the Indian subcontinent [48]. Rainfall in Sri Lanka is controlled by the monsoon winds in the Indian Ocean and the Bay of Bengal. The unique characteristics of those five tanks are given in

Table 1. Maximum reservoir capacity, surface water area, and maximum depth for selected five tanks. (MCM—Million Cubic Meters) source: Department of Irrigation Sri Lanka.

Reservoir/Tank Name	Maximum Capacity (MCM)	Maximum Surface Area (km2)	Depth (m)
Tank A—Iranamadu	111	23	10
Tank B—Mahavilachchiya tank	40	11.5	6.75
Tank C—Kantale tank	135	25	15
Tank D—Senanayak Samudhraya	950	91	43
Tank E—Udawalawa	267	41	36

, and those characteristics indicate that those tanks are shallow and gentle slopes. In order to investigate the relevance of the SWSI index, the characteristics of the tanks, such as the shape, surface water area, and spatial distribution, were considered.

Depending on the months of rainy receives, the rainy seasons are the First Inter-Monsoon (FIM) from March to April; the Second Inter-Monsoon (SIM) from October to November; the North-East Monsoon (NEM) from December to February; and the South-West Monsoon (SWM) from May to September. Major rivers in Sri Lanka start from the central hills, and all the rivers have a length of about 4500 km. Moreover, there are 103 natural river basins in total [49]. The Mahaweli Basin is the longest river and the largest natural river basin in Sri Lanka, which is 335 km long and 10,448 square kilometers in size [50]. In addition, Sri Lanka has a significant number of tanks/reservoirs, including ancient irrigation reservoirs and the recently constructed multipurpose reservoirs that are significantly contributing to increasing the total surface water area of Sri Lanka. The total surface water area of Sri Lanka is about 169,941 hectares, and about 41.7% of that is in major irrigation, and 23.1% is in minor irrigation [50]. Medium tanks cover about 10% of the total area, and both the Mahaweli multipurpose reservoir system and the highland hydropower reservoirs cover about 12.8% of the area [50]. It is estimated that about 7800 million cubic meters of groundwater are recharged annually in Sri Lanka [50,51,52]. Groundwater has been identified as the primary source of water for about 72% of the rural population [52].

2.2. Data

In this study, satellite data were used to quantify the surface water area, primarily covering both the optical and microwave remote sensing data from 2001 to 2020. Landsat 5, 7, and 8 satellite data were used to cover the optical remote sensing, while Sentinel-1 data were used for microwave remote sensing. Climate Hazard Group InfraRed Precipitation with Station data (CHIRPS) rainfall product was used as the main rainfall input for calculation.

2.2.1. Landsat Data

From January 2001 to December 2020, Landsat 5 TM, Landsat 7 ETM, and Landsat 8 OLI with level 1 satellite images were used to calculate the surface water areas with 30 m spatial resolution. Landsat data are represented by Paths/Rows, and there are nine such combinations to cover Sri Lanka. Although Landsat 5 data have been available since 1987, it has several consecutive years as well as months of data gaps in some years. However, this study is more suitable for the study of hydrological drought in Sri Lanka; at least monthly water surface area data are essential for calculating drought indices. There are nine path/row combinations in Landsat images covering the whole of Sri Lanka, and the five tanks used for this study are covered in only three paths/rows, such as 141/54, 141/55, and 140/55 (ee.ImageCollection (“LANDSAT/LM05/C01/T1”), https://www.usgs.gov/core-science-systems, (accessed on 14 September 2022)).

Since Sri Lanka is a tropical country, the sky is often covered with clouds, so the accuracy of the extraction of the water surface area through Landsat data depends on the extent of the clouds above the reservoir and the amount of cloud shadow.

Table 2. Image percentage with different cloud cover percentages.

Reservoir/Tank Name	Cloud Cover in Percentage (%) over the Tanks
	100	100–75	75–50	50–25	25–10	Less than 10
Tank A—Iranamadu	10	6	6	9	9	59
Tank B—Mahavilacchchiya	8	12	12	6	13	49
Tank C—Kantale	11	5	6	8	15	55
Tank D—Senanayak Samudraya	10	5	5	7	8	65
Tank E—Udawalava	10	12	6	14	18	41

shows the percentage of cloud and the number of image percentages for the respective cloud cover that each reservoir covers. It is shown that about 10% of the total Landsat images are completely covered by clouds, and on the other hand, 41% to 65% of images fall into the 0–10% cloud category, which indicates that water surface area can be extracted more accurately.

Tank D covers the highest number of Landsat images with the lowest cloud cover, which is around 65% but the lowest percentage of the segment covering Tank E is about 41%. The main reason for bringing the least percentage of low cloud cover in Tank E is that this tank is more or less affected by all the rainy seasons of NEM, FIM, SIM, and SWM. Due to the high prevalence of drought in the dry zone [53], the study of hydrological drought mainly focused on the dry zone of Sri Lanka. Furthermore, the suitability of the SWSI index, which was introduced in this study analyzed using the five major tanks in Sri Lanka; Tank A (Iranamadu), Tank B (Mahavilachchiya), Tank C (Kantale), Tank D (Senanayaka Samudhraya), and Tank E (Udawalawa).

2.2.2. Sentinel-1 Data

The Sentinel-1 was the first satellite to be built by the European Space Agency to capture SAR data and could be seen as the most significant leap in the Copernican program. The Sentinel-1A satellite was the starting point for this space mission, and with the launch of the Sentinel-1B satellite, data were obtained with greater temporal resolution. The pair of satellites have a C-band (5.7 cm wavelength beam generator) that produces data under single (HH or VV) and dual (HH VH or VV VH) polarization with 10 m spatial resolution. The Sentinel-1 SAR data are available from November 2014 to the present, and in order to generate the monthly water surface area, it was used more than 1710 images received from Sentinel-1. When considering both Landsat and Sentinel-1, there were 2915 images in total. Therefore, the water surface area dynamics were generated from both optical and SAR data using the GEE platform (ee.ImageCollection (“COPERNICUS/S1_GRD”), https://scihub.copernicus.eu/, (accessed on 14 September 2022)).

2.2.3. Rainfall Data (CHIRPS)

CHIRPS data were used as the source of the input rainfall data for this study, which are available in various temporal dimensions such as daily, monthly and annual [54]. Working beyond the traditional methods of data processing via downloading, monthly CHIRPS data were generated to cover the relevant watersheds using the Google Earth Engine (GEE) platform [54]. Validation of CHIRPS data in Sri Lanka was performed in previous studies, so they were used for this study [48]. The CHIRPS data were generated by integrating satellite estimated rainfall product with gauge rainfall data, and the final product provided with 5 km spatial resolution (ee.ImageCollection (“UCSB-CHG/CHIRPS/DAILY”), https://www.chc.ucsb.edu/data/chirps, (accessed on 14 September 2022)).

2.3. Methodology

Rainfall mainly controls the variability of surface water sources, and it can be used effectively for hydrological drought monitoring as the surface water area of lakes varies with the amount of water received by the lake. Therefore, the strategic approach to this study was carried out in four main steps, and a detailed representation of the flow chart that reflects the overall methodology followed in this study is given in Figure 2. The detailed overall process used for this study can be divided into four main sections: surface water area extraction, SPI calculation, Standardized Water Surface Index (SWSI) calculation, and its validation.

2.3.1. The Novel Hydrological Drought Index: Standardized Water Surface Index (SWSI)

The novel Standardized Water Surface Index (SWSI) revealed in this study is designed to determine hydrological drought at multiple time intervals. Long-term water surface area (WSA) data are used to calculate the SWSI, and those data are then subjected to a probability distribution followed by a normal distribution by keeping the median SWSI at zero for a desired location or period. SWSI values vary from −3 to +3, the same as SPI values, and if SWSI becomes positive, it indicates the wetness, and a hydrological drought represents if the SWSI value is negative. The respective gamma distribution of water surface area is defined by its frequency or probability function, as stated in Equation (1).

$$\mathrm{g}\left(w\right)=\frac{{\mathrm{w}}^{\propto -1}.{\mathrm{e}}^{-\frac{W}{\mathsf{\beta }}}}{{\mathsf{\beta }}^2.\mathsf{\Gamma }\left(\mathsf{\alpha }\right)}\mathrm{for}\mathrm{w}>0$$

(1)

where “α” and “β” are the parameters of shape and scale, w is the water surface area, and Γ (α) is the gamma distribution function. Equations (2) and (3) are used to calculate α and β parameters.

$$\mathsf{\alpha }=\frac{1}{4\mathrm{A}}\left[1+\sqrt{1+\frac{4\mathrm{A}}{3}}\right]$$

(2)

$$\mathsf{\beta }=\frac{W_a}{\mathsf{\alpha }}$$

(3)

where $\mathrm{A}=\left(W_a\right)-\frac{\mathsf{\Sigma }\ln \left(W\right)}{\mathrm{n}}$ W_a is the mean water surface area, n is the number of records, and cumulative probability is calculated using Equation (4).

$$\mathrm{H}\left(w\right)=\mathrm{q}+\left(1-\mathrm{q}\right)\mathrm{G}\left(w\right)$$

(4)

where the probability of zero WSA is given as q, G (w) is the gamma function’s cumulative probability, and H(w) is the cumulative probability.

Usually, in the case of water surface area, the zero (0) values are not present or very rare. The cumulative probability H(w) is then converted to the standard normal distribution (z) using the mean of zeros and the variance of one, and the result is SWSI. The input data for the SWSI calculation are based on the five main water tanks in Sri Lanka as described before for the period from January 2001 to December 2020 (240 months) and were calculated using WMO’s SPI generate tool [55]. Furthermore, the SWSI was calculated for different time sequences as 1, 3, 6, 9, 12, and 24 months so that the long-term and short-term variability of the drought could be identified. SWSI values can be categorized into drought classes in the same way as the SPI index. The SWSI values of 1–3 months can be used to identify short-term droughts, 6–9 months of monsoonal droughts, and 12–24 months of long-term as well as inter-annual droughts.

When generating the SWSI index in continuous reservoir water surface area, the calculation is complicated due to the presence of clouds in the optical remote sensing data. Therefore, it has a higher possibility of representing data gaps or abnormal values, especially when determining the surface area of water by visible satellite data. Therefore, gap-filling and smoothing techniques were used to make more accurate use of this data. When mapping the monthly surface water distribution with the help of satellite data, more satellite images can be obtained for one month, thereby calculating the surface water distribution with maximum accuracy. The surface water area, which is used to calculate the SWSI index, was extracted with the following approaches using optical and SAR satellite data.

2.3.2. Water Surface Area Extraction from Optical Satellite Data

The normalized difference water index (NDWI) is the most widely used and efficient indicator applied by various researchers worldwide to represent open water surfaces using green and shortwave infrared with optical remote sensing data [56]. However, using this indicator alone still makes it difficult to distinguish water surface from vegetation [57]. In order to avoid that drawback, it was shown that it is better to identify water surfaces by combining NDWI and NDVI [58]. Accordingly, Equations (5) and (6) were used to compute the NDVI and NDWI indices in this study. The two indices of NDVI and NDWI were combined by the logical criteria for water surface extraction (NDWI > 0.1), and the criteria for NDVI < 0.1 can be used to minimize wet vegetation interference on water pixels [59].

$$NDVI=\frac{R_{nir}-R_{red}}{R_{nir}+R_{red}}$$

(5)

$$NDWI=\frac{R_{nir}-R_{swir}}{R_{nir}+R_{swir}}$$

(6)

where R_nir is the reflectance of near-infrared, R_nir is the reflectance of red, and R_swir is the reflectance of shortwave infrared.

Landsat 5/7/8 surface reflectance data were used to cover selected 5 major tanks in Sri Lanka for the period from 2000 to 2020, and the “Fmask” function was used to remove most of the low-quality pixels, which were recorded due to clouds and their shadows. The green, red, and near-infrared-red bands of the Landsat data (5/7/8) were used to calculate the NDWI and NDVI based on Equations (5) and (6) in the GEE. Subsequently, NDVI and NDWI were calculated, and surface water was extracted using the criteria (NDWI > 0.1 and NDVI < 0.1). Pixels that meet the above criteria are classified as water, and conversely, other pixels are classified as non-water surfaces. The surface water area was then calculated for each image through the total water pixels thus classified, which was then converted to the maximum water area per month. In particular, the calculation area was limited to the reservoir area, which made it possible to increase the calculation accuracy. In order to further confirm the accuracy of the water surface area from which the extraction was performed, all the extracted water pixels were downloaded, and their accuracy was confirmed by visual interpretation using wet and drought years.

2.3.3. Water Surface Extraction with SAR Satellite Data

Sentinel-1 was used for SAR data and is available from 2014 to 2020 with 10 m spatial resolutions with different temporal resolutions. However, the SAR data do not contain cloud effects and can be used for 24 h irrespective of day and night, so a more accurate surface water estimation can be obtained. A speckle filter was used across the image before mapping the water surface, and the resulting image was again smoothed using the focal median. Water and non-water pixels were distinguished through threshold classification (backscatter ≤ −15) on Sentinel-1 VH-polarized satellite images and the water surface area was then calculated using a total number of water pixels in the reservoir [60]. Sentinel-1 data processing and water pixel extraction were performed completely through the GEE platform.

2.3.4. SPI Calculation and SWSI Validation

The average rainfall for the catchment of each reservoir was calculated using CHIRPS monthly rainfall data. Then, the SPI index was calculated using WMO’s SPI calculation tool in 1, 3, 6, 9, 12, and 24 months, the same as the steps followed for the SWSI index. The Pearson correlation coefficient was used to determine the correlation between the SPI and SWSI drought index. Similarly, 1, 3, 6, 9, 12, and 24 months intervals were analyzed for SPI and SWSI to determine which time interval provides the best correlation between those two parameters.

2.3.5. Vegetation Condition Index (VCI) Calculation

Agricultural drought has direct effects on plant health. Therefore, the NDVI index is very sensitive to the chlorophyll content of the crops; thus, it is possible to successfully identify healthy grown and stressed crops through the remote sensing tool. However, the interpretation of agricultural drought through NDVI alone can be problematic due to variability in climate, vegetation types, and environmental parameters of an area. Therefore, researchers introduced the Vegetation Condition Index (VCI) as a commonly used index for agricultural drought monitoring and assessment by identifying the impact on crops due to climate variability (Equation (7)) [61]. There, VCI values are in the range of 0–100, and when the value of VCI is close to 100, the crop shows good health condition, but when the value is 0, it indicates bad health condition of the crop.

$$VC{I}_{ijk}=\frac{NDV{I}_{ijk}-NDV{I}_{i,min}}{NDV{I}_{i,max}-NDV{I}_{i,min}}\ast 100$$

(7)

where $VC{I}_{ijk}$ is the VCI value for the given pixel (i) during week or month (j) for year (k), $NDV{I}_{ijk}$ is the NDVI value for the given pixel (i) during week or month (j) for year (k), and $NDV{I}_{i,min}$ and $NDV{I}_{i,max}$ are minimum and maximum multiyear NDVI for pixel (i), respectively.

3. Results

The results generated to achieve the objectives of the study are described in the following sections. The spatial and temporal changes in the water surface areas of the tanks extracted through optical and SAR satellite data are shown to vary for drought and non-drought years, and the numerical values of the surface areas of the tanks are also represented.

3.1. Landsat Base Reservior Water Dynamics

The spatial variation in the water surface area of five tanks called Tank A (Iranamadu), Tank B (Mahavilachchiya), Tank C (Kantale), Tank D (Senanayaka Samudraya), and Tank E (Udawalawa) was studied to analyze whether the surface water area of the tanks could be significantly changing during the drought. Figure 3 shows a spatial pattern of the monthly water surface area changes in Tank D during drought (2019) and non-drought (2015) years.

The Tank D catchment area received significantly low rainfall during the 2018–2019 NEM, which directly affected crop cultivation in the 2019 Yala season. Due to the low water inflow to the tank, the depth of the water decreased rapidly, and the water surface area also decreased accordingly. This suggests that changes in surface water area can be used effectively instead of the tank’s water depth data for hydrological drought monitoring. At the maximum capacity of Tank D, the surface water area is estimated to be 90.3 (sq. km), which dropped to 6.5 (sq. km.) during the 2019 drought, which is about 7.2% of the total area. This was the lowest value recorded in the last 20 years (Irrigation department).

According to the Disaster Management Center (DMC) reports on Sri Lanka’s drought, the dry zone of Sri Lanka was severely affected by the drought from August 2016 to November 2017. In contrast, from November 2019 to November 2020, significant rainfall was received, and more than 90% of Sri Lanka’s tanks reached their maximum capacity.

Table 3. Water surface area in square kilometers (sq.km) for drought (2017 and 2019) year, non-drought year (2015 and 2020), and maximum recorded water surface area from 2001 to 2020.

Tank/Month	Tank A		Tank B		Tank C		Tank D		Tank E
	2017	2020	2017	2020	2017	2020	2015	2019	2017	2020
January	7.9	21.2	4.5	12.2	11.5	19.5	83.8	48.8	19.1	39.3
February	7.7	21.1	4.6	12.9	13.9	19.1	81.2	36.7	11.6	38.9
March	9.5	21.0	4.3	11.1	12.9	17.5	81.7	47.3	21.8	36.9
April	12.7	22.0	4.0	11.0	13.8	23.7	90.1	50.4	25.2	32.9
May	11.8	20.5	3.2	10.1	13.6	21.3	79.9	36.1	26.9	28.6
June	13.4	20.1	3.4	8.8	11.9	17.8	84.2	5.2	22.6	20.8
July	13.0	16.9	3.4	8.5	11.5	15.2	79.6	9.3	5.2	17.9
August	10.5	16.6	2.9	7.3	10.2	15.0	74.4	4.4	8.3	18.8
September	8.1	15.8	3.0	6.3	10.3	14.2	79.3	6.5	10.4	19.2
October	9.2	15.9	3.0	6.6	11.9	14.0	76.4	6.2	17.0	19.5
November	13.4	13.7	2.7	7.7	11.2	13.8	74.7	26.1	21.5	23.3
December	15.4	20.9	3.9	10.7	13.3	16.2	78.7	56.2	33.3	27.7
Maximum recorded WSA (2001–2021)	23.1		13.1		25.2		90.3		40.2

shows the maximum monthly surface water area calculated through the RS data. Furthermore, the maximum water surface area calculated by RS data for the period 2000 to 2020 obtains values that are closer to the actual (measured) value when the maximum supply level of the relevant reservoir is reached. This implies that the water surface areas of the reservoir calculated from the satellite data are more accurate.

Figure 4 depicts the spatial variability in the water surface area of Tank A, Tank B, Tank C, and Tank E in 2017 (drought year) and 2020 (non-drought year). Analysis of the water surface area changes in the above four tanks shows that the water surface area values are very low in the drought years and very high in the wet years. The 2017 drought was identified as a severe drought that occurred in most parts of the dry zone, and the water surface area of Tank A, Tank C, Tank B, and Tank E dropped up to 17.2%, 19.8%, 36.5%, and 12.9%, respectively, to its full capacity. These numbers are also one of the lowest in the last twenty years.

3.2. SAR Base Reservior Water Dynamics

Figure 5 and Figure 6 show the monthly surface water dynamics successfully extracted for Tank D and Tank A, which are located in mountainous and plane topographic areas using Sentinel-1 images. Those figures also indicate the monthly variability in the water surface area in drought and non-drought years of both tanks. Analysis of that variability shows that during a hydrological drought, the water surface area of the tanks reached less than 25% of the total water surface area for more than three months or more. During the year 2017, the water surface area of Tank D was less than 25% of the total water surface area for more than five months. In 2017, the maximum water surface area of Tank D was about 43%, with a minimum of 15.6%, and for about six months (July to December), the percentage was less than 30%.

The water surface area of Tank A continued to be less than 30% of the maximum area from July to October 2016, with the lowest area of 20% being recorded in September. The specialty is that in 2019, the minimum water surface area was only 37%, and it was reported only in one month. In 2019 the surface water area was more than 80% for 7 months, and in 2016, only one month exceeded 80%; since January 2016, the surface water area has been gradually declining.

Table 4. Water surface area in square kilometers (sq.km) for drought year (2017), non-drought year (2019 and 2020), and maximum recorded WSA from 2015 to 2020.

Tank/Month	Tank A—Iranamadu		Tank D—Senanayak Samudhraya
	2016	2019	2017	2020
Jan	19.9	21.2	30.9	64.1
Feb	18.7	21.0	32.1	66.0
Mar	12.5	21.0	38.5	79.2
Apr	11.8	21.0	38.6	62.7
May	16.7	20.8	35.1	62.1
Jun	10.4	19.0	32.7	58.8
Jul	9.0	16.3	26.8	50.9
Aug	6.5	13.2	15.8	50.7
Sep	4.8	9.0	14.1	52.1
Oct	5.2	11.4	14.1	46.4
Nov	13.4	12.8	23.0	41.1
Dec	14.4	21.1	27.6	40.6
Maximum recorded WSA 2015 to 2020	23.8		89.8

shows the numerical values of the monthly water surface area of Tank A and Tank D.

Since the temporal resolution of S1 data in Sri Lanka range from 7 to 13 days, it is possible to calculate the surface water area using about 3–6 satellite images per month. The greatest advantage of using SAR (Sentinel-1) satellite data is that they provide more accurate values of water surface area consistently under any environmental conditions. However, long-term water area data can be easily calculated by combining Landsat data and Sentinel-1 data via GEE, making the SWSI index calculation even more efficient.

3.3. Water Surface Area and Rainfall Dynamics

The analysis of the tank’s water surface area dynamics clearly shows that changes in water surface area in the main tanks show strong seasonal patterns Figure 7. Although the timing of low and high-water surface area occurrence slightly varies more or less from tank to tank, the five tanks considered show more similarities in temporal patterns. For example, during the 2014 and 2015 North Eastern Monsoon (NEM) seasons, the maximum water surface areas of all tanks are shown, while the lowest water area levels from 2016 to 2019 are shown with the same increase shown in the NEM in 2019.

By considering the changes in the water surface of all five major tanks, it appears that more water is recharged to the tanks during the northeast monsoon season. Considering the main peaks in the surface water area variability graph of the tanks (Figure 7), it is implied that they correspond to the main rainy season, and the sub-peaks show the short-term fluctuations in rainfall during the respective rainy season. The recurrences show large peaks at the beginning of the year, with generally high rainfall and relatively high water areas from December to the end of April. Similarly, the decrease in the surface area of the tanks can be detected early from May to the end of August/September. The main reason for this is that the dry zone of Sri Lanka receives less rainfall during the period from May to August of the year. Although four tanks out of five have shown the same behavior as described above but the water surface dynamics of Tank E contradict as it receives rainfall from both NEM and SIM of Sri Lanka.

Furthermore, the variability in rainfall and water surface area reflects the fact that not only one month of heavy rainfall but also several months of continuous rainfall contribute significantly to the rising water level in tanks (2014 and 2019). For example, in the case of Tank C, it reached its maximum capacity with a significant speed due to the heavy rainfall experienced in December 2014, and the gradual increase in water surface areas can be identified in 2019 due to the effect of accumulated rainfall that occurred during December 2018 and January and February 2019. The variability of rainfall and water surface area in the years 2016–2018 is particularly indicative that if two or three consecutive rainfall periods show a significant decrease, there is a high possibility to occur a hydrological drought. From all the facts identified above, it is clear that the WSA is directly related to rainfall. Therefore, a WSA-based index could be more appropriately used to identify the occurrence, existence, and severity of hydrological drought. Therefore, satellite technology provides a more accurate and convenient basis for calculating the water surface area of a reservoir.

The primary focus of this study was on the development of a new hydrological drought indicator called the Standardized Water Surface Index (SWSI) using optical and SAR satellite data. This section introduces the variability of the SWSI index shown in the five tanks used in this study from 2001 to 2020 and details the validation process of the SWSI index performed using the SPI index in more detail.

3.4. Water Surface Area and Rainfall Dynamics

SWSI values were calculated using 240 months (20 years) of water surface area data from January 2001 to 2020, using five tanks of different sizes, shapes, and climatological and geographical settings of Sri Lanka (Figure 1). For that calculation, the water surface area of tanks extracted from satellite imagery was used. The WMO’s SPI Calculation Tool was used to calculate the SWSI Index, and values are calculated according to different time series of 1 month, 3 months, 6 months, 9 months, 12 months, 24 months, and 48 months, as per the last 20 years coverage. Figure 8 shows how the 3-month SWSI values for all five tanks have changed over the last 20 years. The SWSI index values change from −3 to +3, and using those values the severity of the hydrological drought can be determined. The classification scale for SWSI values and the corresponding event probabilities and cumulative probabilities are given in

Table 5. SWSI Drought classes, value ranges, and the event probability.

Values	SWSI Category	Event Probability (%)	Cumulative Probability
+2.0 < SWSI ≤ MAX	Extreme wet	2.5	0.975–1.000
+1.5 < SWSI ≤ +2.0	Severe wet	4.3	0.932–0.975
+1.0 < SWSI ≤ +1.5	Moderate wet	10.3	0.829–0.932
−1.0 < SWSI ≤ +1.0	Normal	65.8	0.171–0.829
−1.5 < SWSI ≤ −1.0	Moderate drought	10.5	0.066–0.171
−2.0 < SWSI ≤ −1.5	Severe drought	4.1	0.025–0.066
MIN ≤ SWSI ≤ −2.0	Extreme drought	2.5	0.000–0.025

.

As shown in Figure 8, in the context of Tank A, it is well observed that hydrological droughts occurred in 2004, 2012 to 2014, and 2016 to 2017 in the extreme to severe, and those droughts correspond exactly to historical sources (DMC reports 2014, 2017). The other specialty is the hydrological drought that lasted for about three years for both Tank D and Tank C occurred from 2016 to 2019. On the other hand, Tank A, Tank B, and Tank C are located in the Northern and North-central Provinces, Tank D is located in the Eastern Province, and all the tanks belong to the dry and semi-arid zones of the country. The common indicator of the SWSI index is that all tanks are more prone to long-term hydrological drought after 2010. It is also important to note that from 2016 to 2018, there was an extreme to severe hydrological drought covering the dry zone of Sri Lanka. The hydrological drought of 2016–2017 is well represented by SWSI as the worst drought in the last 20 years.

3.5. Standardized Precipitation Index (SPI)

Since the water capacity of a reservoir is proportional to the precipitation received by the catchment area of the study tanks, the average rainfall in the catchment area of each tank was used to calculate the SPI. Figure 9 shows the variability of 3-month SWSI values for the last 20 years. When comparing the results of SPI calculated from different time frames (1, 3, 6, 9, 12, 24), it was shown that increase in the SPI time frame increase the duration of drought or wet condition, as the frequency of alternating positive and negative values decreases.

Analysis of the positive and negative values of SPI-3 for all three tanks above reveals that alternating periods can be identified. Tank D has 27 alternating periods, of which 14 are negative, and 13 are positive. In Tank C and Tank D, the alternate periods positive and negative were identified as 59 and 62, respectively, and those alternate periods are close to 50% between positive and negative. Furthermore, the 3-month SPI represents short-term droughts with a high frequency of changing positive and negative, while the 6–24 SPIs represent seasonal and annual droughts. However, the specialty is to better illustrate the long-term occurrences of the 2001–2002, 2013–2014, and 2016–2017 droughts through all the SPI timeframes, and drought severity becomes the highest in the 2016–2017 drought.

Table 6. Number of months for different drought classes from 2001 to 2020 at different reservoir catchments.

Drought Class/SPI Time Frame/Tank	Tank A—Iranamadu				Tank C—Kantale				Tank D—Senanayaka Samudraya
	3 M	6 M	12 M	24 M	3 M	6 M	12 M	24 M	3 M	6 M	12 M	24 M
Extreme drought	7	8	4	10	11	8	8	11	4	3	3	3
Severe drought	17	11	15	4	19	21	24	5	8	10	10	5
Moderate drought	22	20	15	13	18	20	10	13	23	22	28	29
Normal	143	159	155	145	167	166	151	136	165	163	153	149
Moderate Wet	27	17	23	22	21	16	36	52	18	16	22	8
Severe wet	16	16	15	22	2	4	0	0	11	14	2	13
Extreme wet	6	4	2	1	0	0	0	0	9	7	11	10

can be used to obtain a detailed understanding of the number of months of drought in the last 20 years, from 2001 to 2020, depending on the severity of the drought. When comparing the number of months corresponding to the different drought classes in the three tanks shows that no significant increase or decrease in the number of months of the drought was observed during the increase in SPI time frame from 3 months to 24 months. Compared to Tank A, Tank C, and Tank D, a maximum number of months of both extreme and severe droughts were recorded in the Kantale tank as 11 and 24 months, while the lowest number was recorded in Tank D with 3 and 5 months, respectively. Meteorological drought occurrences in the three tanks in the extreme, severe, and moderate classes averaged 6.6, 12.4, and 19.4 months, respectively. Furthermore, when analyzing the maximum number of months of drought events in the last 20 years by drought classes, the extreme drought events are 4.6%, severe events are 10%, and moderate events are 12.1%.

4. Discussion

4.1. SPI and SWSI Relationship

By analyzing the variability of the SPI and SWSI indices, despite the differences in the parameters used to calculate the SPI and SWSI, it appears that the positive and negative faces of both indices coincide very well over time (Figure 10). However, there are some differences between the values of the two indices. For example, the SPI values of Tank A, Tank B, and Tank D are higher negative than SWSI during the 2016–2017 drought. However, the values of SPI and SWSI might not be much suitable approach to compare the impact of meteorological and hydrological drought as their characteristics are much different. However, it seems more appropriate to analyze the validity of the two indices using the positive and negative faces instead of the values.

Analysis of the relationship between SPI and SWSI was taken beyond the visual interpretation of the three months, as shown in Figure 10 and Figure 11, which show the variability of the Pearson correlation coefficient in the time frames 1, 3, 9, 12, and 24. For monthly time scales, the SPI and SWSI values show a weak correlation of about less than 0.36.

As the SPI and SWSI calculation timeframe increases, the correlation coefficients gradually increase, and the highest values are detected at the three-month, six-month, and nine-month timeframe, and then the correlation coefficients decrease as the timeframe further increases. The maximum correlation coefficients of Tank A and Tank C can be identified in the three-month time frame, and they are 0.58 and 0.61, respectively, while Tank B and Tank D show 0.67 and 0.62 coefficients at the six-month timeframe as the maximum. However, the maximum correlation value of Tank E can be identified by a nine-month timeframe, and that value is 0.59. The most important point is that the maximum correlation is observed between 3 and 9 months, and no significant correlation can be detected at the upper and lower levels. Since the correlation values between SWLI and SPI reflect moderate to strong correlations from the study, it is important to note that the SWSI index is an effective indicator for hydrological drought monitoring.

4.2. VCI and SWSI Relationship

In order to determine the suitability of the SWSI for drought monitoring, not only based on the meteorological drought index but also agricultural drought indices were used. Therefore, the correlation coefficient between monthly VCI and 3-month SWSI was used to represent the relationship between agricultural drought and hydrological drought conditions in the area. A buffer zone of 10 km from the reservoir/tank margin was used to generate the VCI value for the reservoir/tank studied using GEE. The correlation coefficient between the calculated VCI and SWSI for the five reservoirs is shown in Figure 12. There, the maximum correlation coefficient of 0.76 is shown in Udawalawa Lake, and the lowest value of 0.71 can be identified in Iranamadu Lake. Overall, the SWSI introduced through this study shows a significant correlation with agricultural drought.

5. Conclusions

This study introduced a novel drought index named Standardized Water Surface Index (SWSI) because of the lack of attention paid by researchers to develop hydrological drought indices and the fact that there is only a handful of hydrological drought monitoring compared to meteorological and agricultural drought indicators. The uniqueness of this novel index is the use of both optical and microwave (SAR) remote sensing data for calculations compared to other hydrological indices.

The calculation of SWSI used the changes in the water surface area at the monthly timeframe in a particular water-bearing formation such as a reservoir, lake, or tank. Thus, accurate and long-term water surface area extractions are essential for the calculation of the SWSI. Therefore, calculation of the surface water area using satellite data is more effective as the traditional methods used for that are more complex, inefficient, and slow process. It can be concluded that the use of a cloud computing platform, such as Google Earth Engine (GEE), made it possible to extract water surface area efficiently and quickly. On the other hand, water surface area extraction was more successful because GEE enabled to process of a large volume of satellite data through the new algorithm. In order to calculate the SWSI index, five tanks, Tank A (Iranamadu), Tank B (Mahavilachchiya), Tank C (Kantale), Tank D (Senanayaka Samudraya), and Tank E (Udawalawa), were selected. The maximum water surface area extracted from both the optical and the Synthetic Aperture Radar (SAR) satellite show greater accuracy (more than 95%) than the ground-measured maximum water surface area. It can be concluded without any doubt that the higher spatial resolution (30 m and 10 m) of both the Landsat and Sentinel-1 data is the main reason for such greater accuracy.

Various methods were introduced in water surface extraction with SAR satellite data, from simple thresholding [62] to more complex methods [63]. However, in these study areas, the simple threshold method gives better results when calculating the water area, while more complex indicators can be used for other areas based on the behavior. Moreover, by limiting surface water extraction to the maximum extent of the reservoir, impacts from other wetlands can be limited. The study also revealed that the difference in water surface area for severe drought years was less than 25% of the total and, in some cases, as low as 7%. Furthermore, the study observed that in the event of extreme hydrological drought, the surface water area decreases by more than 25%, which gradually decreases and persists for more than three to four months or even more. Furthermore, while the water level of a reservoir decreases, the change in watershed area is determined by the nature of the slope of the area where the reservoir was created, so the applicability and usability of this index in steep slope reservoirs can be studied using Digital Elevation Model (DEM) as a factor for future studies. When considering satellite data-based drought indices, the Temperature–Vegetation–soil Moisture–Precipitation Drought Index (TVMPDI) was developed by combining the parameters rainfall, temperature, soil moisture, and vegetation, which is much more successful for meteorological and agricultural drought monitoring [64] while the SWSI index is much useful for hydrological drought monitoring.

Since there is a direct correlation between precipitation and changes in water surface area in the tanks, it can be said that the SPI index, which was generated from the rainfall data, is more suitable for determining the validity of the SWSI index. Therefore, in order to study the validity of the SWSI index, a comparison analysis was carried out between the SWSI and SPI indices by using the Pearson correlation coefficient for the period 2001–2020 for five water tanks in Sri Lanka. The implication is that the maximum value of the Pearson correlation coefficient for the five tanks ranges from 0.58 to 0.67, indicating a strong correlation between SPI and SWSI. Even though there are differences between the values of SPI and SWSI, the visual interpretation clearly shows that the positive and negative faces of both indices are well coincide with each other. Moreover, event probabilities and cumulative probabilities of the SWSI classification suggest that it is much closer to the SPI classification, as reported in previous studies [12].

These results further prove that the new drought index (SWSI) can be used effectively to monitor hydrological drought by integrating multi-sensor (optical and SAR) satellite data. The study is based entirely on the freely available Landsat and Sentinel-1 time series data, which have high spatial resolution and can be easily applied to different areas due to their simplicity of use. Furthermore, hydrological drought monitoring does not easily apply to monitor drought covering a large area as more than 90% of the existing indicators require location-specific data. Therefore, timely drought monitoring of relatively large geographical areas can be performed using remote sensor data with greater accuracy via the SWSI index. Moreover, the SWSI index should be more useful for areas where there is a lack of field-measured data.

However, the important factors on which the surface water level change depending on the water depletion with drought in the considered reservoirs/lakes are the depth of the reservoir and the topography of the area. In reservoirs with a gentle slope, the change in water surface area is significant, but in steep or near vertical slope reservoirs, that change is insignificant. Considering the parameters of volume, surface water area and depth of the studied reservoirs (Table 1), it is clear that this index can be applied more successfully to shallow and gentle slope reservoirs. Therefore, a major limitation is the use of SWSI in reservoirs with vertical slopes. Another limitation that can be introduced is the unavailability of SAR data with all the polarization for more successful water surface area detection.