Evapotranspiration on Natural and Reclaimed Coral Islands in the South China Sea

Abstract

Studies of evapotranspiration on remote tropical coral islands are important to explore and sustain scarce freshwater resources. However, there is a significant knowledge gap between research to evaluate evapotranspiration based on remote sensing methods and the influences of different land use types on water dynamics on reclaimed coral islands. This study applied the remote-sensing-based Vegetation Interfaces Processes (VIP-RS) model to estimate actual evapotranspiration (ET_a) on Zhaoshu Island, Yongxing Island, and Yongshu Island in the South China Sea from 2016 to 2019. The results showed that the average annual ET_a of Zhaoshu Island, Yongxing Island, and Yongshu Island was 685 mm, 530 mm, and 210 mm, respectively. Annual transpiration (E_c) and soil evaporation (E_s) exhibited similar patterns on the natural islands; however, E_s controlled the water consumption on the reclaimed islands. Water dynamics exhibited seasonal fluctuations due to the uneven distribution of precipitation (PRP). However, ET_a of the natural islands was higher than PRP in the dry season, indicating vegetation has to absorb water from the groundwater to sustain growth. The results also agreed with the analysis of dominant driving factors based on partial correlation analysis, which demonstrated that the Normalized Difference Vegetation Index (NDVI) is the most important factor that influences ET_a, while relative humidity (RH) controlled the bare land or sparsely vegetated areas on the reclaimed islands. The setting of different land use types showed that vegetation and built-up or hardened roads took control of evapotranspiration and rainwater collection, respectively, which play important roles in water dynamics on corals islands. The evaluation of ET_a based on a remote-sensing-based model overcame the difficulty in fieldwork observation, which improves the certainty and accuracy at a spatial scale. In addition, it gave us a new reference to protect and manage scarce freshwater resources properly.

1. Introduction

Coral islands are widely distributed in the Pacific, Indian, and Atlantic Oceans between 30°N and 30°S (Figure 1a). Most of them have ecosystems vulnerable to both natural processes and anthropogenic activities [1,2]. Fresh water is one of the most important resources for the natural ecosystem and the people inhabiting the coral islands. In addition, it also attracts much attention for its sensitivity to climate change [3,4]. Freshwater resources on coral islands differ from those on the continent: (1) Freshwater resources on coral islands mostly come from precipitation. However, it is easy for precipitation to penetrate the thin, permeable unsaturated zone. There is nearly no ponding water on coral islands [5,6]. (2) The fresh water is mainly stored in the form of groundwater, whose availability is highly dependent on its resilience in drought periods. When droughts occur, the water of the coral islands gotten from precipitation decreases, and the vegetation may increase the consumption of water from groundwater [7,8]. (3) The low land surface elevation of atoll islands also makes fresh water vulnerable to inundation by seawater, such as storm surges, wave setup, extremely high tides, and tsunamis [2,9,10]. (4) The high population densities of some coral islands have led to unsustainable rates of extraction. Changes in land use types and contamination from human settlements can also significantly affect freshwater resources [3,11].

Actual evapotranspiration (ET_a), which represents the sum of water evaporation and transpiration from a surface area to the atmosphere, is the least quantified freshwater component of atoll islands’ water budget analysis because of the inherent difficulties in measuring it [1,5]. ET_a could be estimated from potential evapotranspiration (ET_p) by means of a water balance, taking account of evapotranspiration losses from interception and soil water stores [5], where ET_p has been investigated based on the Priestley–Taylor formula, the Penman formula, and pan evaporation [12,13,14]. The ET_a of coral islands could also be estimated by field observations such as eddy covariance, heat-dissipative sap flow, and monitor wells [15,16,17,18,19]. However, these studies only concentrated on a few coral islands, and it was hard to go further in such studies due to the difficulty in data collection and equipment maintenance. Satellite remote sensing data, which consider the vegetation and land use type changes at a long temporal scale [20,21], could be convenient to study hydrological processes on remote islands; however, it has rarely been used in ET_a evaluation on coral islands.

In this study, the remote-sensing-based Vegetation Interfaces Processes (VIP-RS) model [22,23,24] was used to evaluate the ET_a of coral islands in the South China Sea from 2016 to 2019. The VIP-RS model takes account of the detailed physical mechanism of the exchanges of energy and water between the land surface and atmosphere, assimilated with remote sensing information. We aimed to (1) explore the ET_a and water dynamics on natural and reclaimed coral islands in the South China Sea and dominant driving factors for ET_a using partial correlation analysis and (2) compare the differences in ET_a between natural and reclaimed coral islands and analyze the influences of ET_a on different land use types according to the scenario analysis of land use types. This method considers the spatial changes in vegetation and land use types and reduces the hard work of field measurements, helping us evaluate the characteristics of freshwater resources accurately and improve water resource management.

2. Materials and Methods

2.1. Study Area

The South China Sea (Figure 1b), situated between the Karimata Strait (~3°S) and the middle of the Taiwan Strait (~23.5°N), has numerous coral islands. These coral islands develop a wide variety of scenery and wildlife and provide convenience to people with fish, tourism, and energy. To study the water cycle in natural and reclaimed coral islands in the South China Sea, three typical islands were selected. Zhaoshu Island (latitude 16°58′N, longitude 112°16′E) and Yongxing Island (latitude 16°50′N, longitude 112°20′E) were chosen as typical natural islands with different island sizes. Yongshu Island (latitude 9°37′N, longitude 112°58′E), reclaimed in 2015, was chosen as a case of reclaimed islands. These three islands could represent most coral islands in the South China Sea in island size and type.

Zhaoshu Island (Figure 1c) has a natural area of 0.22 km², with an area of about 0.07 km² reclaimed in the west. Lush vegetation is distributed in a ring and band in the natural parts, and the reclaimed area is occupied with greens, a built-up area, and roads. The elevation is approximately 4.4 m, with the core of the island being somewhat low. The soil type is calcareous sand, which is marine sediment originating from coral and other creatures’ debris, with high permeability. It has a tropical monsoon climate that is warm all year, with dry and wet seasons, and the average annual temperature is above 26 ℃, while the daily temperature variation in a year is only about 6 ℃. The rainfall is mainly controlled by typhoons and convective rains, with an average annual rainfall of about 1500 mm, 85% of which is concentrated in June–November [25]. The water table is between 2.3 and 3.0 m and is influenced by the tides. Yongxing Island (Figure 1d) is the largest coral island of the Xisha Archipelago, which has an area of 3.58 km², with the reclaimed area in the northeast. It is 18 km away from Zhaoshu Island, with approximately 3000 people living there. It has similar climate conditions, morphology, and depositional environments as Zhaoshu Island. However, its altitude is as high as 8.2 m. The vegetation is more diverse than that on Zhaoshu Island, but it has suffered some destruction because of human activities.

Yongshu Island (Figure 1e) is a reclaimed island that is located in Yongshu Atoll. It was reclaimed from August 2014 to September 2015, with the land area increased from less than 0.2 km² to 2.23 km². The length is 3.71 km, and the maximum width is 0.99 km [26]. The topography is flat, with an elevation normally 3 to 4 m above the mean sea level. It was reclaimed stratified, which is similar to the formation process of natural coral islands who were deposited by coral debris under the influence of storms or tidal waves [14,16], but the reclaimed island has accelerated the process. Therefore, the coral sand properties of the reclaimed islands are similar to those of natural calcareous sands, but the vegetation is sparse and needs time to recover under natural and artificial processes [27].

2.2. Model Description

The VIP-RS model (Mo et al., 2015) is based on the VIP-distributed ecohydrological dynamic model (Mo and Liu, 2001, Mo et al., 2004). The model inputs include land use types, atmospheric forcing, Digital Elevation Model (DEM), and the Normalized Difference Vegetation Index (NDVI) data series. The model was run from 01 August 2015, to 31 December 2019, at a daily scale, with a spatial resolution of 10 m. The simulated period was chosen for two reasons: (1) the availability of image data from Sentinel-2 is from July 2015, and (2) the reclaimed construction of Yongshu Island was completed in July 2015.

In the VIP-RS model, actual ${ET}_a$ consists of vegetation transpiration ($E_c$), soil evaporation ($E_s$), and canopy interception evaporation ($E_i$). $E_c$ and $E_s$ based on the energy budgets of canopy and soil surface calculated separately with the Penman–Monteith equation. $E_c$ is estimated from potential transpiration ($E_{cp}$), mediated by plant physiological characteristics and environmental stresses, and expressed as follows:

$$E_c=E_{cp}{f}_w{f}_t$$

(1)

where $f_w$ and $f_t$ are stress functions of the atmospheric water vapor pressure deficit and air temperature, respectively, and they are calculated using algorithms proposed by Mu et al. [28] and Zhang et al. [29]. $E_{cp}$ is based on the Penman–Monteith equation:

$$E_{cp}=\frac{1}{\lambda }\left(\mathsf{\Delta }{R}_{nc}+f_c\rho c_{p}D/r_a\right)/\left(\mathsf{\Delta }+\gamma \eta \right)$$

(2)

where $R_{nc}$ is the net radiation absorbed by the canopy (MJ d⁻¹), $f_c$ is the fractional cover of vegetation, $c_p$ is the specific heat capacity of air (J kg⁻¹ °C⁻¹), $\rho$ is the air density (kg m⁻³), $\lambda$ is the latent heat of vaporization of water (J kg⁻¹), $r_a$ is the aerodynamic resistance between the canopy and the reference height (s m⁻¹), and D is the saturated water vapor pressure deficit of air (hPa). Δ is the slope of the curve relating saturated water vapor pressure to temperature (hPa °C⁻¹), and γ is the psychrometric constant (hPa °C⁻¹).

$\eta$ is the ratio of the minimum stomatal resistance (s m⁻¹) of a natural plant functional type to that of a reference crop. It was proposed by Mo et al. (2015) to avoid bias from considering the crop reference transpiration for other vegetation types, as done by most remote sensing evapotranspiration models. The minimum stomatal resistance was adopted from Leuning et al. (2008) and Bastiaanssen et al. (2008) [30,31].

Soil evaporation ($E_s$) is restricted by surface potential evaporation ($E_{sp}$) and soil moisture exfiltration ($E_{ex}$),

$$E_s=min\left(E_{sp},E_{ex}\right)$$

(3)

$$E_{sp}=\frac{1}{\lambda }\left[\mathsf{\Delta }\left(R_{ns}-G\right)+\left(1-f_c\right)\rho c_{p}D/r_{as}\right]/\left(\mathsf{\Delta }+\gamma \right)$$

(4)

$$E_{ex}=S\left(t^{0.5}-{(t-1)}^{0.5}\right)$$

(5)

where $R_{ns}$ is the net radiation absorbed by the soil surface, G is the soil heat flux, $r_{as}$ is the aerodynamic resistance between reference height and soil surface, S is the soil-controlled exfiltration volume and usually falls by 3–5 mm d^−0.5 (in this study, a value of 4 mm d^−0.5 is set), and t is the time (days) elapsed since the day following rainfall [32].

Net radiation $R_n$ is partitioned into energy for the soil and the canopy ($R_{ns}$ for the soil and $R_{nc}$ for the canopy):

$$R_{nc}=f_c{R}_n$$

(6)

$$R_{ns}=R_n-R_{nc}$$

(7)

The vegetation fraction ($f_c$) is estimated using an empirical formula based on the remote sensing vegetation index [33]:

$$f_c=1-{(\frac{{NDVI}_{max}-NDVI}{{NDVI}_{max}-{NDVI}_{min}})}^{\beta }$$

(8)

where $\beta$ is an empirical constant (from 0.6 to 1.2), which was 0.9 in this study; ${\mathrm{NDVI}}_{max}$ represents NDVI values of the full vegetation cover, and ${\mathrm{NDVI}}_{min}$ represents NDVI values of bare soil.

2.3. Data

The land use classification was partitioned through artificial interpretation according to field investigation and images of Google Earth. First, we investigated different plant communities around the whole islands and created the land use type classification according to plant growth characteristics and stand structure. The basis of the land use type classification and field investigation numbers is shown in

Table 1. Basis of land use types classification and field investigation numbers.

Land Use Types	Basis of Classification	Sample Numbers
Evergreen forest	Mainly the community of Cocos nucifera on Yongxing Island	15
Evergreen shrubs	Mainly the community of Scaevola sericea on Zhaoshu Island	12
Mixed forests	Mainly a mixture of different vegetation species (common vegetation species, such as Calophylluminophyllum, Terminalia catappa, Cocos nucifera, Clerodendrum inerme, Scaevola sericea, Pandanus tectorius, Ceodes grandis, Guettarda speciosa, Canavalia maritima, etc.)	31
Grassland	The lawn beside the airport, roads, or built-up area	12
Built-up or hardened roads	Construction land or concrete roads	12
Barren or sparsely vegetated area	Bare land or sparsely vegetated areas	15

. Second, we confirmed different plant species’ distributions and their land use types and drew the maps of land use types from 2016 to 2019, manually combing images of Google Earth. The map was made at an annual scale for rare changes in land use types within a year. Figure 1c–e showed the map of the land use types on Zhaoshu Island, Yongxing Island, and Yongshu Island, respectively, in 2019.

Meteorological data were obtained from the China Meteorological Administration at a daily scale at the Yongxing Island meteorological station (ID: 59981) and the Yongshu Island meteorological station (ID: 59985). Atmospheric forcing variables included air temperature (TA), maximum air temperature (TMX), minimum air temperature (TMN), relative humidity (RH), wind speed (WIN), sunshine duration (SHIN), precipitation (PRP), and atmospheric pressure (AP). The missing SHIN data at the Yongxing Island meteorological station from 2018 to 2019 were replaced by the data at the Qionghai Island meteorological station (ID: 59985; 19°14′N, 110°28′E). SHIN is used to calculate solar radiation (Ra) in the VIP-RS model, so the missing SHIN data at the Yongshu Island meteorological station were replaced by Ra, which was obtained from the modeling results of Hargreaves’ radiation formula [34]. According to the station meteorological data from 2010 to 2019 (Figure 2), precipitation and temperatures are similar at both meteorological sites. The annual TA ranges from 25.8 °C to 28 °C in different years, while the average monthly TA ranges from 24 °C to 30 °C, with its maximum value in June and minimum value in January. The annual PRP varies from 900 mm to 1900 mm, with a large fluctuation in different years. Similar to TA, the average monthly PRP has seasonal differences, with the maximum value of 270 mm in July and the minimum value of less than 10 mm in February.

The NDVI was extracted using Sentinel-2 images by Google Earth Engine. The data spanned from July 2015 to December 2019, with a monthly temporal composite at 10 m spatial resolution. The temporal changes in the NDVI in Figure 3 show that Zhaoshu Island has the largest spatial average NDVI for its large area of vegetation, followed by Yongxing Island and Yongshu Island. The NDVI in these islands exhibits a seasonal change and a slow fluctuated increase tendency during the simulated period. We noticed that a rapid decline in the NDVI occurred between September 2018 and May 2019 in Zhaoshu Island, which may be related to drought and vegetation destruction caused by insufficient precipitation. Figure 4 demonstrates that an NDVI above 0.4 has a larger ratio to the entire island area and a lower degree of fragmentation in Zhaoshu Island than in Yongxing Island. The reason for the differences between the two natural islands is the high-intensity human activities in Yongxing Island. The NDVI in reclaimed areas ranges from 0.0 to 0.2, except the grassland beside the airport and scattered green areas ranging between 0.2 and 0.4.

The altitude was set as a 2 m height for the whole area of islands because the small altitude differences on coral islands have a rare influence on simulated results.

2.4. Model Validation

The model validation could be divided into four parts:

(1) The potential evaporation was validated by pan evaporation measured with pan Φ20 at a monthly scale at the Yongxing Island meteorological station from 2008 to 2013 because pan evaporation data at the Yongxing Island meteorological station after 2014 are missing. Potential evapotranspiration was only driven by meteorological data, and the changes in vegetation and land use types have no influences on potential evapotranspiration. The validation of the simulated potential evapotranspiration from 2008 to 2013 confirmed the applicability of the model under tropical coral island meteorological conditions, although the validated periods had different time spans than our study periods.

(2) The daily simulated ET_a of grasses and E_s of bare land in July 2018 and August 2018 were validated by field-collected ET_a using a specially constructed micro-lysimeter in July 2018 and August 2018 [35]. The micro-lysimeter was constructed from a polyvinyl chloride (PVC) pipe with a height of 150 mm and a diameter of 101.6 mm, and the sampling processes could refer to Plauborg, (1995) [36]. The E_s of bare land was measured from 22 June 2018 to 6 July 2018 during the periods with no precipitation, with the ET_a of grasses being measured between 22 July 2018 and 28 July 2018 [35].

(3) The daily simulated E_c of evergreen forests, evergreen shrubs, and mixed forests was validated with the E_c measured by Jie et al. (2015) [37], who measured the leaf E_c of Calophylluminophyllum, Terminalia catappa, Cocos nucifera, Clerodendrum inerme, Scaevola sericea, Pandanus tectorius, and Canavalia maritima with the Li-Cor 6400 Portable Photosynthesis System on Shanhu Island of the Xisha Archipelago. In this study, evergreen forests were mainly the Cocos nucifera community on Yongxing Island, evergreen shrubs represented the community of Scaevola sericea on Zhaoshu Island, and mixed forests were mixtures of different vegetation species (Table 1). We compared the simulated average E_c of evergreen forests, evergreen shrubs, and mixed forests from 31 August 2015 to 2 September 2015 with the E_c of Cocos nucifera and Scaevola sericea and the average E_c of all vegetation species, as measured by Jie et al. (2015) [37] on 1 September 2015. The measured branch E_c at a daily scale was transferred from the leaf E_c using the leaf area index (LAI) [38] and the sinusoidal distributional shape of ET_p [38,39].

(4) To validate the model performance on coral islands at a temporal scale, we simulated the monthly ET_a and E_c of the Cocos nucifera plantation on Vanuatu in the South Pacific to compare with the works of Roupsard et al. (2007) [19], who monitored the monthly ET_a by eddy covariance and the monthly E_c by heat-dissipative sap flow, soil heat flux and, stand heat storage from 2001 to 2004. Vanuatu (15°26′N, 167°11′E) is a coral island having a similar distance to the equator as Yongxing Island (16°50′N, 112°20′E), and the climate of Vanuatu is tropical monsoons with seasonally fluctuating precipitation, which is similar to the climate in the South China Sea [19]. Although the meteorology of Vanuatu has some differences in temporal distribution, the validation of evapotranspiration in Vanuatu could also illustrate the applicability of the VIP-RS model in tropical coral islands. The meteorological data for simulating ET_a in Vanuatu were obtained from the website of the National Oceanic and Atmospheric Administration (ID: 915510-99999/915540-99999), and the LAI was set as 3.0, as given by Roupsard et al. (2007) [19].

2.5. Dominant Driving Factor Analysis

Dominant driving factor analysis can explore the main driving factors that control ET_a. The analysis process includes simple linear regression, the first difference de-trending method, normalization, and partial correlation analysis, which are all performed in each grid over the whole region. The dominant driving factor is set as the factor that has the highest partial correlation coefficient. The specific methods used in this study are as follows:

First, we used the first-difference de-trending method (i.e., the difference in values from one month to the previous month) to get rid of the trend influences on evapotranspiration [40,41].

Second, we normalized each variable to facilitate the comparison between variables of different units and orders of magnitude:

$$X_s=\frac{x_i-x_{min}}{x_{max}-x_{min}}$$

(9)

Third, since climatic factors have correlations with each other, the relationships between the de-trending climatic factors and the de-trending ET_a are quantified by partial correlation analysis. Partial correlation analysis explores the relationship between two variables independent of the influences of other factors [40,42], which is shown as follows:

$$Y_s=a_1{X}_{1s}+a_2{X}_{2s}+a_3{X}_{3s}+\cdots \cdots$$

(10)

The ET_a of the coral islands was simulated from 2016 to 2019 with a 10 m spatial resolution and a daily time step. Partial correlation analysis was used to investigate the relationships between ET_a, the NDVI, and climatic factors at the monthly scale. The factor that had the highest partial correlation coefficient was identified as the dominant climatic factor.

3. Results

3.1. Evaluation of the Model

The simulated water ET_p and pan evaporation (pan coefficient of 0.7) agreed well as they demonstrated similar inter-annual and seasonal variations with a maximum value of about 170 mm/month in July and a minimum value of about 80 mm/month in January (Figure 5a). The comparison (Figure 5b) between pan evaporation and simulated ET_p exhibited a determination coefficient of 0.82 within the significance of 0.01, with the root-mean-square error (RMSE) being 10.54. The results indicated that the model is sensitive to meteorological conditions and can simulate the potential evaporation well.

Figure 6 exhibits the evapotranspiration comparison of different land use types. The average daily E_c of evergreen forests, evergreen shrubs, and mixed forests was about 1.9 mm/day, 2.6 mm/day, and 1.51 mm/day, respectively. The average daily ET_a of grassland was 2.1 mm/day, with the E_s of bare land being approximately 0.9 mm/day. Evergreen shrubs showed the largest E_c because the dominant species of Scaevola sericea on Zhaoshu Island had a high transpiration capacity, as shown by Jie et al. (2015) [37], and its distribution is on a natural area with almost no human activities. Error bars showed that simulated values had a smaller fluctuation than measured values, the reason is that the measured evapotranspiration may show some differences at the inner-daily scale, versus the simulated values showed stable values. The comparison results demonstrated that the model can simulate evapotranspiration of different land use types well.

To validate the performance of the model at the temporal scale, we compared ET_a and E_c between simulated and measured values at the Cocos nucifera plantation on Vanuatu from 2001 to 2004 (Figure 7). The measured ET_a of Cocos nucifera had an average value of 2.6 mm/day, with the highest value of approximately 3.4 mm/day in December and the lowest value of 1.8 mm/day in June. The measured E_c of Cocos nucifera had an average value of 1.7 mm/day, with the largest value of approximately 2.3 mm/day in November versus the lowest value of 1.3 mm/day in June. The simulated ET_a and E_c showed a similar fluctuation with measured values in different months, with the relative errors of ET_a and E_c being less than 0.1 mm/day. The results demonstrated that the model can simulate ET_a and E_c at a monthly scale well.

3.2. Evapotranspiration Characteristics on Different Islands

Figure 8 shows the spatial patterns of the average annual ET_a from 2016 to 2019, which was similar to the distribution of the NDVI. ET_a in the natural area showed higher values than that in reclaimed areas. Zhaoshu Island had the largest average annual ET_a of 685 mm for its largest proportion of the natural vegetation area. The average annual ET_a of Yongxing Island was 530 mm. Although Yongxing Island also had a large area of natural vegetation, the vegetation was somewhat destroyed because of the inhabitants’ activities. Yongshu Island was a reclaimed island that had the lowest ET_a of 210 mm because of the high proportion of built-up or hardened roads.

The average annual ET_a of different land use types is shown in the bar figures in Figure 8. We found that evergreen forests on Yongxing Island and evergreen shrubs on Zhaoshu Island had the largest average annual ET_a, ranging from 1000 mm/year to 1200 mm/year. The ET_a of grassland and mixed forests ranged from 600 mm/year to 800 mm/year versus the ET_a of bare land or sparsely vegetated areas, which ranging from 400 mm/year to 600 mm/year on Yongxing Island and Zhaoshu Island. The ET_a of grassland, mixed forests, and bare land or sparsely vegetated land ranged from 300 mm/year to 600 mm/year on Yongshu Island, which was smaller than that on Yongxing Island and Zhaoshu Island. The differences may be caused by the somewhat small vegetation cover in the reclaimed island. The ET_a of bare land or sparsely vegetated land was similar to the ET_a of grassland and mixed forests for high values of E_s in bare land. The ET_a of built-up or hardened roads was less than 50 mm/year on all three islands.

Figure 9 shows the simulated daily ET_a of different land use types on the three islands. The daily ET_a exhibited seasonal fluctuations due to the conditions of energies and water. ET_a was high from July to September, while it was low from January to March, which was caused by the change in meteorological conditions (Figure 2). At a daily scale, ET_a rose rapidly after rainfall, especially for the bare land or sparsely vegetated areas, due to increases in E_s and E_i. After a period of no rainfall, the ET_a of bare land or sparsely vegetated areas dropped to approximately 1 mm/day, while evergreen forests and evergreen shrubs had a value of at least 2 mm/day, even in the dry season, and the ET_a of grassland or mixed forests fell in between. The ET_a of built-up or hardened roads exhibited small values during a few days after rainfall. ET_a was influenced by PRP on the reclaimed island more than that on the natural island because E_s and E_i are sensitive in a short time, which also indicated that the water deficit should be considered for ET_a of the reclaimed island.

We partially enlarged the detail of ET_a from 20 March 2019 to 30 April 2019 when these three islands experienced two periods of rainfall after more than 30 days of no rainfall. One rainfall level is 37 mm on 31 March 2019, and the other is 13 mm on 16 April 2019. When it rained, we could see that ET_a rose rapidly for all land use types. E_s and E_i displayed an obvious increase with rainfall; however, E_c somewhat decreased because of stomatal closure. After rainfall, the ET_a of the land use types with large vegetation cover had a high value, and that of bare land or sparsely vegetated areas dropped to a low level rapidly.

3.3. Dominant Driving Factors for ET_a

ET_a is mainly influenced by meteorological factors and NDVI conditions [43,44,45]. Gao et al. [46] found that PRP controlled ET_a changes in most parts of Chinese areas, especially in arid and semi-arid regions. Coral islands in the South China Sea have a tropical oceanic monsoon climate, where vegetation is somewhat similar to the tropical dry evergreen forests in Southeast Asia [47]. Pour et al. [48] demonstrated that the minimum temperature (31.5–48.2%) is the most influencing factor, followed by wind speed (15.1–32.8%), in defining ET_p in peninsular Malaysia. To identify the dominant driving factors for ET_a of the coral islands in the South China Sea, partial correlation analysis was used based on the detrending data series for each grid.

Table 2. Dominant impact variables of evapotranspiration of different land use types.

Land Use Types	Yongxing Island	Yongshu Island	Zhaoshu Island
EF/ES	NDVI	-	NDVI
MF	NDVI	RH	NDVI
GL	NDVI	NDVI	-
BS	NDVI	RH	NDVI

Note: EF: evergreen forests; ES: evergreen shrubs; MF: mixed forests; GL: grassland; BS: barren or sparse vegetation.

shows the spatial distribution of dominant driving factors within the significance of 0.01. The NDVI was the main influencing factor for ET_a of Zhaoshu Island and Yongxing Island. On Yongshu Island, RH controlled the ET_a of mixed forests and bare land or sparsely vegetated areas, while the NDVI controlled the ET_a of grassland. The results revealed that the NDVI is the dominant driving factor in the natural islands. However, for reclaimed islands, RH was another dominant driving factor, except for the NDVI. The reason for the difference is that RH controlled the diffusion of soil water to air in bare land and sparsely vegetated areas. Salvucci and Gentine [49] also demonstrated that ET_a is controlled by the gradient of humidity near the land surface, which reflects a balance of radiative heating and cooling, conversion to latent heat, canopy and soil heating, and the turbulent transport of sensible heat into the atmosphere.

3.4. Evapotranspiration Distribution on the Three Islands

PRP, the only source of fresh water on the coral islands, is mainly consumed by ET_a, infiltration, and rainwater collection or water flowing into the sea. Storage changes within the unsaturated zone and surface runoff can be neglected [1,5] on coral islands due to the low water-binding capacity of gravelly and sandy soil texture, both for natural and for reclaimed islands [26,50]. In this study, the water budget analysis could be written as ΔV = PRP − Eta, where PRP is rainfall (mm); ET_a is the actual evapotranspiration (mm), which contains E_c, E_s, and E_i; and ΔV is the total amount of infiltration, rainwater collection, and water flowing into the sea. Figure 10 shows the annual water balance from 2016 to 2019; the average PRP was about 1443 mm/year on the three islands. The average annual E_c on Zhaoshu Island and Yongxing Island was 319 mm/year and 246 mm/year, respectively, while the E_s on Zhaoshu Island and Yongxing Island was 293 mm/year and 306 mm/year, respectively. The E_c on Yongshu Island was 48 mm/year, which was far smaller than the E_s of 174 mm/year. The annual E_i on the three islands ranged from 20 mm to 70 mm, which is a low value. Figure 11 shows the ratio of E_c, E_s, E_i, and ΔV to precipitation on Zhaoshu Island, Yongxing Island, and Yongshu Island from 2016 to 2019. E_c and E_s occupied more than 20% of PRP on Zhaoshu Island and Yongxing Island. However, the ratio of E_c and E_s to PRP was 3% and 13%, respectively. PRP had a somewhat large fluctuation in different years, and ET_a had small fluctuations compared to PRP, indicating that ET_a is restricted more by vegetation or energy than by PRP. The results were consistent with the analysis of Table 2.

The average monthly water balance analysis demonstrated a seasonal distribution over the simulated period (Figure 12). The wet season had PRP between 120 mm/month and 220 mm/month from June to November on Zhaoshu Island and Yongxing Island. In comparison, the dry season had PRP between 2 mm/month and 90 mm/month from December to May. Yongshu Island had a similar seasonal distribution, but PRP had a larger fluctuation than that on Zhaoshu Island and Yongxing Island. In the dry season, PRP on Zhaoshu Island and Yongxing Island was 10 mm/month to 30 mm/month less than ET_a. Vegetation had to absorb water from groundwater to sustain growth, which has been found in many studies on coral islands [4,5]. However, on the reclaimed island, the groundwater could be recharged by PRP through the entire year. Comparing the water balance at the annual scale, we found that a relatively large fluctuation of PRP mainly occurs in the wet season. In the wet season, PRP was sufficient for ET_a. In addition, ET_a could be limited by PPR in the dry season. However, the changes in ET_a due to fluctuations of rainfall in the dry season were small for the low values of ET_a. The results further proved that ET_a is controlled more by energy and the NDVI than by PRP.

4. Discussion

4.1. Evapotranspiration Differences in Natural and Reclaimed Areas of Coral Islands

Reclaimed coral islands have somewhat large areas of hardening and barren land or sparsely vegetated areas compared to natural islands, which indicates that reclaimed islands consume small volumes of water due to the low E_c caused by the lack of evergreen forests or shrubs. Figure 13 compares the differences in ET_a between natural areas and reclaimed areas. ET_a had similar seasonal fluctuations at temporal series on different islands. However, ET_a in natural areas was 25–50 mm/month higher than that in reclaimed areas. Natural and reclaimed coral islands face different water resource problems. The most important freshwater problem on natural islands is water contamination and sustainable exploitation of water resources [1]. However, for newly reclaimed coral islands that have no fresh groundwater, it will take 20 years to form a stable freshwater lens [26]. The most important thing is to evaluate the spatial and temporal distribution of the fresh water of coral islands and the available water resources [51]. Falkland and Woodroffe [4] predicted an approximately 50% increase in groundwater recharge if all coconut trees were removed from Tarawa Atoll and Kiritimati. However, it could bring new problems if we simply decrease the vegetation cover. For the people on coral islands, however, a low vegetation cover could increase the groundwater recharge, which have a positive influence on the freshwater storage [52,53]. It could also provide shelters to seabirds, which could promote soil-forming processes, and decrease the surface temperature to make the islands livable [54,55,56]. However, the proper way to manage the limited fresh water and how the different land use types influence the water dynamics on the reclaimed islands are still not clear.

4.2. Influences of Land Use Types on Water Dynamics in Reclaimed Islands

To study the influences of land use types on water dynamics in reclaimed coral islands, we attempted to set different land uses on Yongshu Island to study its possible water transport process. The land use types on Yongshu Island in 2019 were set as a reference. The ratio of trees and shrubs, grassland, barren or sparsely vegetated areas, and built-up or hardened roads to the whole island area was 0.08, 0.19, 0.22, and 0.51, respectively, with E_c, E_s, E_i, and ΔV in 2019 of 98 mm, 161 mm, 24 mm, and 1160 mm, respectively. The water distribution changes are shown in

Table 3. Influence of changes in land use types on water dynamics of coral islands.

Land Use Types	Ec (mm)/Ratio	Es (mm)/Ratio	Ei (mm)/Ratio	ΔV (mm)/Ratio
Trees and shrubs	520/430%	253/57%	118/389%	552/−50%
Grassland	193/97%	307/90%	46/89%	897/−23%
Barren or sparsely vegetated areas	89/−9%	319/98%	27/13%	1007/−13%
Built-up or hardened roads	0/−100%	24/−85%	0/−100%	1419/22%

Note: The ratio means the relative value of the difference between the evapotranspiration of different land use settings and evapotranspiration in 2019 to evapotranspiration in 2019.

when we covered one of these land use types over the entire island. If we cover the whole islands with trees and shrubs, E_c, E_s, and E_i increase to 520 mm, 253 mm, and 118 mm, respectively, with their values increasing from 57% to 430%. In such a setting, ET_a is larger than ΔV obviously and vegetation consumes a relatively large proportion of precipitation. If we cover the entire coral island with grassland, E_c, E_s, and E_i increase to 193 mm, 307 mm, and 46 mm, respectively, with their values increasing from 89% to 93%, and ΔV decreases to 897 mm. In this situation, E_s becomes the largest ET_a component, and ΔV is the largest consumption of precipitation. If we cover the entire coral island with barren or sparsely vegetated areas, the overall ET_a still exhibits a 57% increase, although the area of vegetation decreases. The reason for the increase in ET_a can be attributed to the reduction in the area of built-up or hardened roads. The results showed that ET_a of Yongshu Island in 2019 was still at a low value due to the restriction of nearly 51% built-up or hardened roads.

Built-up or hardened roads influence the water cycle in another way. They cut off the interchange of water between air and the land surface, which can benefit rainwater collection. For reclaimed coral islands, it will take a long time to form fresh groundwater, so a rainwater collection system is an important measure to collect the limited freshwater for the people living there [57]. The water of the rainwater collection system mainly comes from built-up or hardened roads. Setting Yongshu Island in 2019 as a reference, if 50% of the precipitation on built-up or hardened roads were collected by the rainwater collection system, the rainwater collection system could collect 332 mm of fresh water in a year, which is about 0.74 million squares of water. The effect of built-up or hardened roads on the rainwater collection system needs further research for reclaimed islands.

4.3. Implications and Limitations

To evaluate ET_a of coral islands conveniently, free and relatively high-resolution satellite products are the first reference. We compared the NDVI products that extracted Landsat series images (30 m spatial resolution) and Sentinel-2 images (10 m spatial resolution). However, the products of Landsat-5 (1985 to 2011) had a wrong geospatial calibration, which causes deviations in the coordinates of images at different time, sand products of Landsat-7 (1999 to now) had some Scan Line Corrector failure. Landsat-8 (2014 to now) can simulate ET_a well, but it has a similar on-orbit time with Sentinel-2 (2015 to now), which has a high temporal-spatial resolution. Another limitation is the shortage of data in the wet season due to the influence of cloud cover. In this study, we use 31 days of temporal-scale data to run the model, although the products can provide a sight image of approximately a week. Land use types were obtained using artificial interpretation methods because of the difficulty in distinguishing different land use types from the satellite remote sensing image due to limits of spatial resolution. We tried to classify land use types using supervised classification and unsupervised classification methods, but the results were unsatisfactory, especially for differences in build-ups and bare lands in reclaimed areas, so an effective algorithm should be developed for coral islands.

The simulated results were validated with data of field samplings and other literature; however, uncertainties still exist due to the limited hydrological observations and intrinsic uncertainties of the hydrological model. Some plants close pores to defend against drought on coral islands in the South China Sea [37], which is similar to the study of Pitman [47], who found that the stomatal conductance of planta may reduce by 50% when the soil water potential reaches −1.2 (±0.07) MPa. Stomatal conductance is an important parameter during simulated processes. However, the model does not differentiate stomatal conductance changes between the dry season and the wet season, which may lead to some errors for some vegetation species. So, it is necessary to adjust sensitive parameters for coral islands referring to dominant driving factors. Even so, the evaluation of ET_a based on the remote-sensing-based model makes up for the difficulty in fieldwork, which improves the certainty and accuracy at a spatial scale. In addition, it gives us a new reference to protect and manage scarce freshwater resources properly.

5. Conclusions

This study evaluated ET_a and water dynamics on natural and reclaimed coral islands in the South China Sea from 2016 to 2019 based on a simplified remote-sensing-based VIP model. The simulated results showed that the water dynamics on coral islands have a seasonal distribution. The average annual PRP was about 1443 mm/year on the three islands, with 1088 mm in the wet season and 355 mm in the dry season. The average annual ET_a of Zhaoshu Island, Yongxing Island, and Yongshu Island was 685 mm, 530 mm, 210 mm, respectively. E_c and E_s had a similar value, which ranged from 246 mm to 319 mm (approximately 20% of PRP) on Zhaoshu Island and Yongxing Island. However, E_s controls water consumption on Yongshu Island. ET_a is restricted more by vegetation or energy than by PRP, which is consistent with the partial correlation analysis result that the NDVI is the dominant factor for the whole area of the natural islands and grassland of the reclaimed island, RH is the dominant driving factor for mixed forests and barren land or sparsely vegetated areas on Yongshu Island. Approximately half of PRP is consumed by evapotranspiration on the natural islands, while approximately 20% of PRP can be consumed by evapotranspiration on the reclaimed island.

Under the background of global climate changes, more and more islands are constructed to defend against the possible impact of extreme climate events. Through setting different land uses on the entire islands, we quantitatively evaluated the effects of vegetation and built-up or hardened roads to water cycles on coral islands. Although further studies need to be carried out using the VIP-RS model to optimize parameters due to the specific condition of coral islands, the evaluation of ET_a based on a remote-sensing-based model makes up for the difficulty in fieldwork, improving the certainty and accuracy at a spatial scale. In addition, it gives us a new reference to protect and manage scarce freshwater resources properly.