International Journal of Remote Sensing

ISSN: 0143-1161 (Print) 1366-5901 (Online) Journal homepage: http://www.tandfonline.com/loi/tres20

Remote sensing techniques to predict salinity

intrusion: application for a data-poor area of the
coastal Mekong Delta, Vietnam

Phuong T.B Nguyen, Werapong Koedsin, Donald McNeil & Tri P.D Van

To cite this article: Phuong T.B Nguyen, Werapong Koedsin, Donald McNeil & Tri P.D Van
(2018): Remote sensing techniques to predict salinity intrusion: application for a data-poor
area of the coastal Mekong Delta, Vietnam, International Journal of Remote Sensing, DOI:
10.1080/01431161.2018.1466071

To link to this article: https://doi.org/10.1080/01431161.2018.1466071

Published online: 26 Apr 2018.

Submit your article to this journal

Article views: 78

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=tres20
INTERNATIONAL JOURNAL OF REMOTE SENSING, 2018
https://doi.org/10.1080/01431161.2018.1466071

Remote sensing techniques to predict salinity intrusion:

application for a data-poor area of the coastal Mekong Delta,
Vietnam
Phuong T.B Nguyena, Werapong Koedsin a
, Donald McNeilb and Tri P.D Vanc
a
Remote Sensing & Geo-Spatial Science Research Unit, Faculty of Technology and Environment, Prince of
Songkla University, Phuket, Thailand; bDepartment of Statistics, Macquarie University, Sydney, New South
Wales, Australia; cDepartment of Water Resources, Can Tho University, Can Tho, Vietnam

ABSTRACT ARTICLE HISTORY

Salinity intrusion is a complex issue in coastal and estuarine areas. Received 29 July 2017
Currently, remote sensing techniques have been widely used to Accepted 5 April 2018
monitor water quality changes, ranging from inland river networks
to deep oceans. The Vietnamese Mekong Delta is an important
rice-growing area, and intrusion of saline water into irrigated
freshwater-based agriculture areas is one of the most crucial con-
straints for agriculture development. This study aimed at building
a numerical model to realize the salinity intrusion through the
relationship between reflectance from the Landsat-8 Operational
Land Imager images and salinity levels measured in situ. A total of
103 observed samples were divided into 50% training and 50%
test. Multiple Linear Regression, Decision Trees and Random Forest
(RF) approaches were applied in the study. The result showed that
the RF approach was the best model to estimate salinity along the
coastal river network in the study area. However, the large sam-
ples size needed was a significant challenge to circumscribe pre-
dicting ability of the RF model. The reflectance has a good
correlation with salinity when locations (latitude–longitude) of
salinity measured stations were added as a parameter of the
Step-wise model with R-square 77.48% in training and 74.16% in
test while Root Mean Square Error was smaller than 3.

1. Introduction
Saline intrusion is a difficult issue for freshwater supply in coastal areas (Trung et al. 2016; Himi
et al. 2017). Extraordinary saline and short-term salinity intrusion can cause huge damages but
are hard to detect, on the short and long timescale, and it acts as a slow-onset hazard
(Consultative Group on International Agricultural Research CGIAR Research Centers in
Southeast Asia 2016; Binh 2015). Much damage is done before the intrusion has been
identified. In the Vietnamese Mekong Delta (VMD), the salinity intrusion trend has historically
increased in terms of concentration and duration, leading to great threats to sustainable
freshwater-based agriculture development (Trung et al. 2016; Binh 2015). Agricultural

CONTACT Werapong Koedsin werapong.g@phuket.psu.ac.th Remote Sensing & Geo-Spatial Science Research
Unit, Faculty of Technology and Environment, Prince of Songkla University, Phuket, Thailand
© 2018 Informa UK Limited, trading as Taylor & Francis Group
2 P. T. B. NGUYEN ET AL.

enterprises including paddy rice, fruit, vegetable and freshwater-based aquaculture are the
most vulnerable due to the lack of any alternative freshwater supplies. However, the lack of
agility in providing salinity information from manager to local people is a critical problem:
farmers are not able to keep in touch with current information on salinity (CGIAR Research
Centers in Southeast Asia 2016). Increasing uncertainties and complexities exacerbated by
climate change and abnormal weather guarantee accurate information for salinity forecasting
will become more difficult (CGIAR Research Centers in Southeast Asia 2016).
Salinity intrusion is a complex process depending on many factors in the VMD. These
include freshwater discharge from upstream, capacity and morphology of the channel,
configuration of the drainage network, tidal conditions and the presence of control structures
such as sluice gates (Hashimoto 2001; Nguyen et al. 2008). The magnitude of floods, summer–
autumn paddy production growing-stage and rainfall frequency and magnitude (Deltares,
Delta Alliance and Federation of Water Resources Planning and Investigation in the South,
Vietnam 2011) in combination with impacts of sea level rise also exacerbate the damage of
salinity intrusion. Ability to predict salinity intrusion was an issue of interest in previous studies.
Even though hydraulic models were considered to be successfully developed to simulate
salinity intrusion in the VMD (Khang et al. 2008; Trung and Tri 2014), they are highly complex
requiring large input data (e.g. detailed topographical information and the latest knowledge
about the infrastructure), which is in many cases not achievable in remote areas (Nguyen et al.
2008). Meanwhile, remote sensing techniques open new potential applications to monitor
freshwater resource, salinity in coastal area and ocean salinity (Wang and Xu 2012). The
resolution applications of available satellite imgaes are particularly suited to coastlines and
estuaries (Roy et al. 2014). The success of satellite images for detecting salinity intrusion was
demonstrated in different previous studies using the Landsat Thematic Mapper (Landsat TM;
Wang and Xu 2008; Baban 1997), the Moderate Resolution Imaging Spectroradiometer
(MODIS; Urquhart et al. 2012), the Earth Observing-1 (EO-1 ALI; Fang et al. 2010), the Sea-
viewing Wide Field-of-view Sensor (Sea WiFS; D’Sa et al. 2002) and recently by the Landsat-8
Operational Land Imager (OLI; Zhao, Temimi, and Ghedira 2017). However, none of these
studies were done in a tropical monsoon location.
Remote sensing may relate suspended solids and coloured dissolved organic matters
from the river discharge to salinity and water reflectance (Wang and Xu 2008). Thus,
salinity tends to precipitate Total Suspended Solids (TSS) and so has a good negative
correlation with salinity (Fang et al. 2010). Modelling salinity also was performed using
only single wavelength reflectance remote sensing (Urquhart et al. 2012; Zhao, Temimi,
and Ghedira 2017). On the other hand, it was found that the location was the most
significant predictor variable in surface salinity estimation models (Urquhart et al. 2012).
However, the locality issue is a major consideration when using the model for other
study areas that contains complex processing, freshwater inflow and seasonality.
Machine learning has been successfully applied for different studies in the remote sensing
field (Waske et al. 2009; Dev et al. 2016). Eight statistic models including a Categorical and
Regression Tree model (CART), a Generalized Linear Model (GLM), a Generalized Additive
Model (GAM), a Random Forest (RF) Model, a Mean model, an Artificial Neural Network (ANN),
a Multivariate Adaptive Regression Spline (MARS) and a Bayesian Additive Regression Tree
(BART) were applied to find out the best model for predicted salinity concentration in
Chesapeake Bay, in the United States (Urquhart et al. 2012). In another study, an ordinary
least square regression was performed to determine relationships between salinity and
 INTERNATIONAL JOURNAL OF REMOTE SENSING 3

reflectance in Lake Pontchartrain, in the US Gulf of Mexico (Wang and Xu 2008). A multi-
variable linear algorithm was employed for predicting sea surface salinity from remote sensing
reflectance in the hypersaline Arabian Gulf (Zhao, Temimi, and Ghedira 2017). Choosing a
suitable model for detecting salinity depends on the characteristics of each study area.
However, lack of in situ data for supervised modelling is one of the great challenges in the
way of developing realistic applications of salinity modelling (Urquhart et al. 2012; Zhou and
Zhang 2016). Small sample size is related to overfitting which causes serious limitations in
developing a useful model (Liu and Gillies 2016; Shcheglovitova and Anderson 2013). In such
cases, the model does good fitting on the small number of training data points available but
would not do well in predicting for a new task on new samples (Ratner 2011).
This study sought to identify risk of salinity intrusion and develop a model using
reflectance data from Landsat-8 OLI as a new approach to predict salinity in the
VMD. Two objectives of the study were: (i) to determine the most useful reflectance
wavelength of the Landsat-8 OLI images and choosing a suitable statistic model to
predict salinity intrusion with the limited salinity observations available; and (ii) to
examine trends of saline intrusion by combining wavelength and location factors.

2. Methodology
2.1. Study area
The VMD, located in the downstream of the Mekong River, is flat and low-lying area of highly
complex rivers and channels (Noh et al. 2013). The Tien River is one of the biggest
distributary river systems in the VMD. Before entering the East Sea, the Tien River splits
into four branches including the Cua Tieu, Cua Dai, Ham Luong and Co Chien–Cung Hau
(Figure 1). At a distance of 30 km from the South China Sea, the Co Chien River again splits
into two estuary branches: Co Chien–Cung Hau (Nguyen and Savenije 2006).

Figure 1. The study area and sampling locations of 17 salinity stations in Cua Dai, Cua Tieu, Ham
Luong and Co Chien–Cung Hau River.
4 P. T. B. NGUYEN ET AL.

Annually, declining discharge from the upstream from the onset of the dry
season, strong wind speed from the sea and high tide level cause incursions of
saltwater along the rivers up to about 40–50 km inland. Furthermore, the propor-
tion of discharge entering the delta river has significant roles to determine salinity
distribution (both in terms of time and magnitude) in the different branches. In
2005, the discharge contribution at Cua Tieu was about 10% of total discharge in
the Tien River while the discharge in Cua Dai and Ham Luong were about 20.8%
and 10.2%, respectively. In the Co Chien river, the Co Chien estuary received 10.5%
and Cung Hau received 3.3%, and the remainder of the discharge was through
other parts of the delta (Nguyen et al. 2008). Thus, about 55% of the total Mekong
discharge occurs through the study area. In the VMD, the standard salinity alert
value is 4 parts per thousand (ppt), equivalent to about 12% of seawater (Deltares,
Delta Alliance and Federation of Water Resources Planning and Investigation in the
South, Vietnam 2011).

2.2. Data collection

2.2.1. Landsat-8 OLI collecting and processing
The Landsat-8 OLI has an entire Earth coverage every 16 days and 30 m × 30 m of
spatial resolution. Twenty-five satellite images were used consisting of mostly cloud-
free scenes in the dry season (January–June) from 2013 to 2016. Scenes of images
were between 124/53 and 125/53 paths/rows. These images were used to extract
reflectance data for a single pixel that also has data on salinity measured in the field
at the time. ‘Ground truth’ from 17 salinity stations was applied in the Region of
Interest (ROI) function in ENVI to identify 103 samples from these images.
First, radiometric correction was performed to normalize satellite images for
factors such as sensor degradation, Earth-Sun distance variation, incidence angle,
view angle and time of data gathering were applied to the image data. This process
involved converting Digital Number (DN) into radiance using calibration parameters
that accompanied with the images metadata. The image was then atmospherically
corrected and transformed to reflectance using the Moderate Resolution
Atmospheric Transmission-based Fast Line-of-sight Atmospheric Analysis of
Spectral Hypercubes (MOD-TRAN-based FLAASH) algorithm under the ENVI 5.3
program.
Second, the Function of mask (Fmask) algorithm was applied to detect clouds and
cloud shadows on the created cloudy maps. Finally, the mask of the study area prepared
for every single satellite image was created by overlapping the river network map and
cloudy maps. The reflectance-image prepared as inputs was cleared of cloud, cloud
shadows and focused on the water area.
Depending on each object (water, soil), the spectral reflectance has specified wavelengths
but the range should be in the range 0–1 (Peddle et al. 2001). Normally, the spectral
reflectance of clear water is quite low. The wavelength reflectance of seawater was small
between 400 nm and 850 nm, ranging from 0.01 to 0.14 (Xiong et al. 2012). In this study, mean
reflectance of water was around 0.03–0.13 (Figure 2).
 INTERNATIONAL JOURNAL OF REMOTE SENSING 5

Figure 2. Maximum, minimum and mean of reflectance wavelength water area based on the actual
sample size in the whole study river reaches.

2.2.2. Salinity measurement

Saltwater data were collected at the same time with satellite imagery at 14 hydrology
stations: Vam Kenh 1 (VK1), Long Hai (LH), Hoa Binh (HB), Hoa Dinh (HD), My Thoa (MT), Binh
Dai (BD), Loc Thuan (LT), An Thuan (AT), Son Doc (SD), Giong Chom (GC), My Hoa (MH), Ben
Trai (BT), Huong My (HM) and Khanh Thanh Tan (KTT) and 3 sluice gates: Vam Kenh 2 (VK2),
Vam Giong (VG) and Xuan Hoa (XH) along the rivers: Tieu River, Dai River, Ham Luong River
and Co Chien River (Figure 1). Saltwater data were provided by Tien Giang and Ben Tre
Hydrology and Metrology Centres and Department of Agriculture and Rural Development
Tien Giang during the dry seasons (January to June) from 2013 to 2016. The hydrology
stations measured salinity using Electrical Conductivity (EC) metres manufactured by Yellow
Springs Instruments (YSI); along the transects, the measured saltwater samples were col-
lected by mixing water at 0.5 m depth and surface water for every 2 hours.

2.3. Model development

To assess effective band compositions of seven bands, multi-linear regression methods
were employed. The Multiple Linear Regression (MLR), Decision Trees (DTs) and Random
Forest (RF) model were applied to develop the relationship between salinity and
reflectance from Landsat-8 OLI using R sorfware version 3.3.2. Moreover, this study
used the Step-wise model to develop reflectance locations for predicting based on the
real data. The input for these models was prepared using 103 samples as the real data
set and then divided into 50:50 for training and test purpose. The best model was used
to predict salinity intrusion for the whole study area using R software and ArcGIS. The
best model was employed to detect salinity intrusion for the whole study area in the
early dry season: on 24 January 2015 and 9 February 2015.
6 P. T. B. NGUYEN ET AL.

2.3.1. Bands selection for modelling the salinity intrusion

To choose suitable bands for the model development, the reflectance wavelength data from
7 bands: Band 1 – Coastal/Aerosol (0.433–0.453 μm), Band 2 – Bule (0.450–0.515 μm), Band 3
– Green (0.525–0.600 μm), Band 4 – Red (0.630–0.680 μm), Band 5 – Near Infrared (0.845–
0.885 μm), Band 6 – Shortwave Infrared 1 (1.560–1.660 μm), Band 7 – Shortwave Infrared 2
(2.08–2.35 μm) were considered to investigate the correlation both single band and multiple
bands with salinity observation. The MLR was employed to analyse these relationships. The
probability value (p-value) and coefficient of determination (R2) were factors to assess which
were the effective bands to use as an input to apply statistical model.

2.3.2. Modelling salinity intrusion with reflectance wavelength

The MLR analysis is the most commonly applied statistical method (Owen 2001) and is
highly flexible for examining the relationship between independent variables and a
single dependent variable (Aiken, West and Pitts 2003). Meanwhile, the DTs is a non-
parametric and highly nonlinear method that can deal with high-dimensional problems
where the number of features is much larger than the number of samples. The RF
regression algorithm is an ensemble learning algorithm that combines and grows trees
from a large set of the regression trees (Epifanio 2017; Breiman 2001).
The MLR, DTs and RF were considered to find out the best model to explore the
relationship between salinity and reflectance wavelength. The DTs model was integrated
in the R software by the Recursive partitioning function (Rpart). The Rpart’s Complexity
Parameter (CP) coefficient was set as 0.01. In the RF analysis, two parameters need to be
optimized including the number of regression tree and the number of input variables
per node. This study chose the number of regression tree with a high number of 1000
trees and the number of input variables per node was set at the default value (i.e. 0.333).

2.3.3. Modelling salinity intrusion with reflectance wavelength and location

The Step-wise MLR was employed for predicting salinity using variuos reflectance
wavelengths of Landsat-8 OLI and the location (latitude and longitude) where salinity
was measured in situ.
Root Mean Square Error (RMSE), R2 and p-value were used to evaluate the goodness
of fit and significance of these models.

3. Results
3.1. The relationship between reflectance and salinity
The composite spectral of 7 bands including Band 1 – Ultra Blue (Coastal/Aerosol),
Band 2 – Blue, Band 3 – Green, Band 4 – Red, Band 5 – Near Infrared (NIR), Band 6 –
Shortwave Infrared 1 (SWIR-1) and Band 7 – Shortwave Infrared 2 (SWIR-2) provided
highly effective prediction significant for the salinity model (Figure 3). The composi-
tion of Band 2, Band 3, Band 4 and Band 7 was highly significant with a p-value
(approximately, 3 × 10–9) which was greater than the combination of 7 bands. Use of
a single band was unlikely to be productive, and only Band 3 presented a high
correlation salinity (see Figure 3). Therefore, four bands (Band 2, Band 3, Band 4 and
Band 7) were selected as inputs for developing the further salinity intrusion models.
 INTERNATIONAL JOURNAL OF REMOTE SENSING 7

Figure 3. Multiple linear analysed the relationship between salinity and composite bands.

3.2. Modelling salinity intrusion

The selected bands (i.e. Band 2, Band 3, Band 4 and Band 7) were used as the input of each
procedure. Figures 4-6 show the observations and prediction of the MLR, DTs and RF
regression, respectively. The MLR model presented fairly good R2 in training when the R2
and RMSE were 50.23% and 4.15, respectively, while the test showed a very poor non-
significant correlation with an R2 of 22.61% and RMSE of 4.12. The range of predicted salinity
value was 0–20 ppt in training which was higher than in the test runs (Figure 4(a,b)). While
the DTs model indicated less significant correlation in both training and test, the R2 values

Figure 4. Scatter plot of the MLR based on (a) the real data in training with 51 training samples and (b) test
with 51 test samples.
8 P. T. B. NGUYEN ET AL.

Figure 5. Scatter plot of the DTs regression based on (a) the real data in training with 51 training
samples and (b) test with 51 test samples.

Figure 6. Scatter plot of the RF based on (a) the real data in training with 51 training samples and
(b) test with 51 test samples.

were 25.41% and 13.09% while the RMSE values were 2.72 and 2.83 for training and test,
respectively. Predicted salinity’s range was around 5–15 ppt in the training as well as the test
(Figure 5(a,b)). However, RMSE of the DTs model was better than that of the MLR model for
all processing; the p-value was 9.7 × 10–5 in the DTs and 2.0 × 10–7 in the MLR.
The RF model was highly correlated to salinity in the training (i.e. R2 and RMSE were
90.41% and 1.1, respectively). However, the test also showed the limitation of the model (R2
and RMSE were 22.55% and 2.13, respectively). The range of predicted salinity was between
4–18 ppt in the training and 6–18 ppt in the test (Figure 6(a,b)). The RF model turn out to the
best salinity intrusion model. The p-value of the RF had a very high significance level with
p-value < 2.0 × 10–16 in training and p-value <0.0002 in test. Moreover, based on the result
 INTERNATIONAL JOURNAL OF REMOTE SENSING 9

shown in this study, the RF performed well in the training but the testing of the model
exposed weaknesses that maybe caused by inadequate number of observations.

3.3. Reflectance-location salinity modelling and mapping salinity intrusion

The reflectance and location data were combined as parameters in the Step-wise model that
showed a good relationship with salinity. Figure 7(a,b) presents the regression between
predicted values and observed salinity, R2 of training and test ordered 77.48% and 74.16%,
respectively, while RMSE was 2.57 and 2.96. However, some extraordinary predicted values
were lower than 0 ppt in both training and test where 0 ppt was exceptional in the actual
salinity observations. Thus, locality issues need to verify the model if it was used in other
study areas. The formula (1) can only be used in this particular study area, and if it is applied
to a different area, the weight and perhaps also the number of coefficients in the model
would have to be changed. Furthermore, location played as a strong factor in the formula.
Even though the location was helpful to illustrate a general trend but it was a cause of
unrecovered essential information from reflectance wavelength data.



Salinity ¼ ð3:741  104 Þx  8:235  105 y þ 1:575  102 B2



 2:92  102 B3 þ 40:51B4 þ 1:048  102 B7  1:366  102 (1)

where x is latitude; y is longitude; B2, B3, B4 and B7 are the Landsat-8 OLI spectral bands
(i.e. Band 2, Band 3, Band 4 and Band 7, respectively).
Mapping salinity intrusion based on the reflectance-location model showed that the
salinity level clearly differentiated between estuaries and the upstream of the river. The
Cua Tieu and Cua Dai rivers presented high salinity intrusion while the Co Chien–Cung
Hau River had the lowest salinity. The results demonstrated a change in saltwater
concentration and intrusion on different days. Mean of salinity level was 7.12 ppt and
standard deviation was 9.88 ppt on 24 January 2015 (Figure 8(a)). It was 6.44 ppt at
mean and 9.48 ppt at Standard Deviation (SD) on 9 February 2015 (Figure 8(b)).

Figure 7. Scatter plot of the stepwise regression model (reflectance-location) between predicted salinity
and observed salinity (a) training with 51 training samples and (b) test with 51 test samples.
10 P. T. B. NGUYEN ET AL.

Figure 8. Salinity intrusion from the reflectance-location model; (a) 24 January 2015 – mean: 7.12;
SD: 9.88 with 51 training samples; (b) 9 February 2015 – mean: 6.44; SD: 9.48 with test samples.

4. Discussion
Identification bands for setting the model were very important to the accuracy of the
predictive salinity map: it means that not only does R2 and RMSE of the modelling need
to be considered but also the composite of wavelength data set as inputs for predicting
salinity in the whole study area. On the other hand, using too many input variables can
 INTERNATIONAL JOURNAL OF REMOTE SENSING 11

lead to fittings quickly which become computationally infeasible if attempts are made to
use a higher number of potential predictor variables (Messner et al. 2016). Thus, proper
selection of the identification bands set-up for use in the model is very important for the
predictive accuracy of the salinity map.
The salinity of the sea surface can be represented by a function combining several
remotely sensed ocean colour bands (Urquhart et al. 2012). This study found that the
combination of spectral wavelength of Band 2 (Blue: 0.45–0.51 μm), Band 3 (Green:
0.53–0.59 μm), Band 4 (Red: 0.63–0.67 μm) and Band 7 (SWIR 2: 2.11–2.29 μm) was
essential to detect salinity intrusion on the Tien River. Likewise, Wang and Xu (2008)
reported that the TM Bands 1–5 in Landsat-5 TM were strongly correlated with salinity.
Castillo (2005), where the study on the salinity at the Gulf of Mexico, reported that
higher salinity levels resulted in escalated reflectance of coloured dissolved organic
matter (CDOM) at wavelengths less than 600 nm,which is in agreement with our results
for Band 2 and Band 3. Urquhart et al. (2012) reported the reflectance wavelength of
MODIS at 488, 443 and 667 nm associated with salinity. In addition, Zhao, Temimi, and
Ghedira (2017) reported that the Landsat-8 OLI’s Bands 1–4 were selected to use in the
algorithm development which is in agreement with our findings for Landsat-8 OLI’s
Bands 2–4. Nevertheless, there have been contradictory results reported by some
researchers. Baban (1997) and Wang and Xu (2008) reported that the TM Band 3 (Red:
0.63–0.69 μm) was the most correlative with salinity. Conversely, this study showed
that Landsat-8 OLI Band 3 (Green: 0.53–0.59 μm) was the most significant with the
salinity (see Figure 3). In a study on salinity in southwestern Australian estuaries, Lavery
et al. (1993) reported the linear negative relationship of salinity with the reflectance of
TM Band 4 (NIR: 0.76–0.90 μm) and of TM Band 7 (SWIR: 2.08–2.35 μm) that agreement
with our study results with combination of Landsat-8 OLI Band 7 (SWIR 2: 2.11–
2.29 μm); however, the NIR spectral bands were not selected in this study. The dis-
crepancy in relationships between salinity and water reflectance were derived from
satellite data and the differences of R2 may have been caused by environmental of the
site-specific that makes differences of some parameter in the water such as the CDOM,
TSS properties and salinity level. Moreover, the biogeochemical processes on land
divergent origins of CDOM and various concentrations of TSS in freshwater discharges
could be responsible for changes in the optical properties of CDOM (Castillo 2005) and
TSS in coastal environments (Wang and Xu 2008).
The VMD is located in the tropical monsoon region where season has a strong impact
on discharge, and tidal incursion of saltwater is characteristic during the low-flow season
(dry season), the estuarine being in partially well-mixed conditions, with saltwater
intrusion around 40 km inland, bringing fine-sediment up-river to a turbidity maximum
area. The majority of the sediment may have been deposited in shallow coastal waters
(Wolanski et al. 1996) and the current situation of estuarine sediments and open bay
muddy sediments expand more and more in the Mekong Delta region (Oanh et al.
2001). It is a challenge to detect salinity level when the number of observations might
not be enough to represent a representative sample of the study area. Meanwhile, the
study of Urquhart et al. (2012) was successful in the application of MODIS to detect
salinity intrusion in Chesapeake Bay (USA) using a huge set of data collected from 2003
to 2010 at 67 monitoring stations (620 samples). On the other hand, the data obtained
for the present study were collected from 17 stations which were measured and
12 P. T. B. NGUYEN ET AL.

provided by different offices which may lead to some inherent inconsistencies and
uncertainties out of the control in the study.
Choosing a suitable model seriously effects on the accuracy of the model and is
almost as equally important as variables and is more important than sample size
(Fassnacht et al. 2014). The MLR method clearly showed the structure of model but in
cases where the relationship between variables is complex, linear regression may lack
the ability to achieve a confidence level useful for predictive purposes. The DTs showed
very low potential for useful predictions of salinity. Nevertheless, the RF model is a
popular method in many fields since they can be successfully applied to complex data,
with a small sample size, complex interactions and correlations, with mixed type
predictors (Epifanio 2017). The RF model also exhibits the highest predictive capability
compared with the MLR and DTs models (Chen et al. 2017). This study showed that the
RF model was the best model to predict salinity intrusion. Unfortunately, the sample size
is not enough to present a sufficiently detailed profile of the study area.
The overfitting problem was a big issue, which showed even in training tasks, the R2
was rarely good enough to make useful predictions (Figure 5). The quality of satellite
images, the schedules used for measuring salinity and the timing of satellite images
were factors which limit observation in this study area. If a remote sensing program is
set up to monitor salinity intrusion, then the data set would naturally increase over time
and its predictive value would steadily improve. Another source of error is that the
remote sensing data which were used represent the integrated information on the
surface water while salinity data from the monitoring stations were taken from a specific
depth. Furthermore, special conditions like high water turbidity may affect the accuracy
of the algorithm. It is suggested that the algorithm uncertainties were attributed to the
several parameters and more measurements were required for improving the perfor-
mance of the algorithm. The application of the developed algorithm to other sites with
difference in optical properties and environmental parameters must be cautious, and
recalibration with the local in situ measurement is necessary.
The model was acceptable when successfully applied for mapping salinity in the
whole study area. The sensitivity of the reflectance-location data was examined to
recognize salinity changes. Salinity mapping distinctly divided the low salinity-low
intrusion upstream areas and high salinity-high intrusion frequency areas downstream.
The merging of reflectance-location data improved upon the limitations of the single
reflectance model which is in agreement with Urquhart et al. (2012) where their study of
the salinity in the Chesapeake Bay showed that the latitude and longitude are the most
significant predictor variables in the surface salinity estimation model. This achievement
may be due to the fact that the salinity intrudes into the river as a function of the
distance from the mouth of the estuary.
Furthermore, combining remote sensing and geospatial interpolation techniques was
illustrated as significantly improving accuracy to address the issue of limited data
availability in the coastal area (Urquhart et al. 2013). Thus, to consider location as
variables the model should be set up and updated frequently to adapt to the changing
of salinity over time. However, this method contains uncertainties due to salinity intru-
sion being influenced by different factors (hydrology of rivers and weather conditions)
which cannot simply expressed by location criteria. Merging the location and reflectance
 INTERNATIONAL JOURNAL OF REMOTE SENSING 13

wavelength demonstrated a good correlation with R2 and RMSE. It is likely the best way
to develop an interpolation model.
Field observations of salinity cannot cover enough of any real-world study area.
However, the usefulness of the model may face limitations due to locality issues
(Urquhart et al. 2012). Thus, using location parameters can lead to the neglect of
important information on satellite images if the location plays too strong role in
predicting salinity.

5. Conclusion
The Landsat-8 OLI offers a convenient approach to enhance prediction capacity of saline
intrusion. This study showed that Band 2, Band 3, Band 4 and Band 7 were vital for
developing a successful reflectance wavelength salinity model. The RF model was demon-
strated to maximize capacity for exploring the relationship between reflectance and
salinity. However, the small sample size was a cause of overfitting and limits the applica-
tion of these models for predictive purposes. Improving sample size is the most important
priority work for further study to explore more thoroughly the relationship between
salinity level and single reflectance wavelength from satellite images. Ongoing use of
the model incorporating new data as it is collected will progressively improve the model.
The combination of reflectance-locations for predicting salinity intrusion illustrated a
good correlation, which could be used to predict trends of salinity intrusion, when field
observations of salinity cannot cover enough of any real-world study area. However,
locality issues should be kept in mind which can negate the reflectance information in
the satellite images and limit the application of remote sensing when there is a lack of
sufficient ground observational data. Even though mapping salinity of reflectance-loca-
tion model clearly showed the salinity dynamic in study area, it also faced inherent
limitations due to lack of observation in upstream areas (where economic losses due to
unexpected salinity incursions would be likely to be the most severe) and locality issues
such as the suitability of monitoring station sites due to the latitude and longitude play
too strong a role for predicting salinity.
Due to the revisit frequency of 16 days for the satellite used in the present study and
the requirements of cloud-free conditions for satellite measurements especially in the
tropical area, there are very limited field measurements acquired during OLI acquisition
times. Salinity alerts may be too late to be useful in preventing salinity damage to
Mekong agriculture.

Acknowledgements
This work was supported by the Higher Education Research Promotion and the Thailand’s
Education Hub for Southern Region of ASEAN Countries Project Office of the Higher Education
Commission, under Prince of Songkla University. The authors are thankful to Dr. Raymond J Ritchie
who gave a lot of useful advice for this manuscript.

Disclosure statement
No potential conflict of interest was reported by the authors.
14 P. T. B. NGUYEN ET AL.

Funding
This work was supported by the Higher Education Research Promotion and the Thailand’s
Education Hub for Southern Region of ASEAN Countries Project Office of the Higher Education
Commission, under Prince of Songkla University.

ORCID
Werapong Koedsin http://orcid.org/0000-0002-4411-9148

References
Aiken, L. S., S. G. West, and S. C. Pitts. 2003. “Multiple Linear Regression.” In Handbook of
Psychology, edited by I. B. Weiner. New York, NY: John Wiley & Sons. doi:10.1002/0471264385.
wei0219
Baban, S. M. J. 1997. “Environmental Monitoring of Estuaries; Estimating and Mapping Various
Environmental Indicators in Breydon Water Estuary, U.K., Using Landsat TM Imagery.” Estuarine,
Coastal and Shelf Science 44: 589–598. doi:10.1006/ecss.1996.0142.
Binh, N. T. 2015. “Vulnerability and Adaptation to Climate Change.”[PhD thesis.], The Faculty of
Agriculture, Bonn University and Institute for Environment and Human Security, United Nations
University (UNU-EHS), Germany. http://hss.ulb.uni-bonn.de/2015/3924/3924.pdf.
Breiman, L. 2001. “Random Forests.” Machine Learning 45: 5–32. doi:10.1023/A:1010933404324.
Castillo, C. E. D. 2005. “Remote Sensing of Organic Matter in Coastal Waters. Chap.7.” In Remote
Sensing of Coastal Aquatic Environments, edited by R. L. Miller, C. E. D. Castillo, and B. A. McKee,
157–180. The Netherlands: Springer.
CGIAR Research Centers in Southeast Asia. 2016. “The Drought and Salinity Intrusion in the
Mekong River Delta of Vietnam.” Accessed June 25 2016. https://cgspace.cgiar.org/rest/bit
streams/78534/retrieve/.
Chen, W., X. Xie, J. Wang, B. Pradhan, H. Hong, D. T. Bui, Z. Duan, and J. Ma. 2017. “A Comparative
Study of Logistic Model Tree, Random Forest, and Classification and Regression Tree Models for
Spatial Prediction of Landslide Susceptibility.” Catena 151: 147–160. doi:10.1016/j.
catena.2016.11.032.
D’Sa, E. J., J. B. Zaitzeff, C. S. Yentsch, J. L. Miller, and R. Ives. 2002. “Rapid Remote Assessments of
Salinity and Ocean Color in Florida Bay.” In The Everglades, Florida Bay, and Coral Reefs of the
Florida Keys: An Ecosystem Sourcebook, edited by J. Porter. Boca Raton, FL: CRC Press.
Deltares, Delta Alliance and Federation of Water Resources Planning and Investigation in the
South Vietnam.2011. “Vietnam-Netherlands Mekong Delta Masterplan Project Mekong Delta
Water Resources Assesment Studies, Viet nam-Neetherlands Mekong Delta Master
Plan.” Accessed 15 January 2017. https://www.wur.nl/upload_mm/2/c/3/b5f2e669-cb48-4ed7-
afb6-682f5216fe7d_mekong.pdf
Dev, S., B. Wen, Y. H. Lee, and S. Winkler. 2016. “Machine Learning Techniques and Applications for
Ground-Based Image Analysis.” IEEE Geoscience and Remote Sensing 4: 79–93. doi:10.1109/
MGRS.2015.2510448.
Epifanio, I. 2017. “Intervention in Prediction Measure: A New Approach to Assessing Variable Importance
for Random Forests.” Epifanio BMC Bioinforma 18: 1–16. doi:10.1186/s12859-017-1650-8.
Fang, L., S. Chen, H. Wang, J. Qian, and L. Zhang. 2010. “Detecting Marine Intrusion into Rivers
Using EO-1 ALI Satellite Imagery: Modaomen Waterway, Pearl River Estuary, China.” International
Journal of Remote Sensing 31 (15): 4125–4146. doi:10.1080/01431160903229218.
Fassnacht, F. E., F. Hartig, H. Latifi, C. Berger, J. Hernández, P. Corvalán, and B. Koch. 2014.
“Importance of Sample Size, Data Type and Prediction Method for Remote Sensing-Based
Estimations of Aboveground Forest Biomass.” Remote Sensing of Environment 154: 102–114.
doi:10.1016/j.rse.2014.07.028.
 INTERNATIONAL JOURNAL OF REMOTE SENSING 15

Hashimoto, T. R. 2001. Environmental Issues and Recent Infrastructure Development in the Mekong
Delta: Review, Analysis and Recommendations with Particular Reference to Largescale Water
Control Projects and the Development of Coastal Areas (No. 4). Sydney, Australia: Australian
Mekong Resource Centre, University of Sydney. Accessed 15 January 2017. http://sydney.edu.
au/mekong/documents/wp4.pdf.
Himi, M., S. Benabdelouahab, A. Salhi, L. Rivero, M. Elgetta, A. El, J. Stitou, and A. Casas. 2017.
“Geophysical Characterization of Saltwater Intrusion in a Coastal Aquifer : The Case of Martil-
Alila Plain (North Morocco).” Journal of African Earth Sciences African Earth Sciences 126: 136–
147. doi:10.1016/j.jafrearsci.2016.11.011.
Khang, N. D., A. Kotera, T. Sakamoto, and M. Yokozawa. 2008. “Sensitivity of Salinity Intrusion to
Sea Level Rise and River Flow Change in Vietnamese Mekong Delta-Impacts on Availabilityof
Irrigation Water for Rice Cropping.” Journal of Agricultural Meteorology 64: 167–176. doi:10.2480/
agrmet.64.3.4.
Lavery, P., C. Pattiaratchi, A. Wyllie, and P. Hick. 1993. “Water Quality Monitoring in Estuarine
Waters Using the Landsat Thematic Mapper.” Remote Sensing of Environment 46: 268–280.
doi:10.1016/0034-4257(93)90047-2.
Liu, R., and D. F. Gillies. 2016. “Overfitting in Linear Feature Extraction for Classification of High-
Dimensional Image Data.” Pattern Recognition 53: 73–86. doi:10.1016/j.patcog.2015.11.015.
Messner, J. W., G. J. Mayr, A. Zeileis, J. W. Messner, G. J. Mayr, and A. Zeileis. 2016.
“Nonhomogeneous Boosting for Predictor Selection in Ensemble Postprocessing.” American
Meteorological Society 145: 137–147. doi:10.1175/MWR-D-16-0088.1.
Nguyen, A. D., and H. H. G. Savenije. 2006. “Salt Intrusion in Multi-Channel Estuaries: A Case Study
in the Mekong Delta, Vietnam.” Hydrology and Earth System Sciences Sci 3 (2): 499–527.
doi:10.5194/hessd-3-499-2006.
Nguyen, A. D., H. H. G. Savenije, D. N. Pham, and D. T. Tang. 2008. “Using Salt Intrusion
Measurements to Determine the Freshwater Discharge Distribution over the Branches of a
Multi-Channel Estuary: The Mekong Delta Case.” Estuarine, Coastal and Shelf Science 77: 433–
445. doi:10.1016/j.ecss.2007.10.010.
Noh, S., M. Choi, E. Kim, N. P. Dan, B. X. Thanh, N. T. V. Ha, S. Sthiannopkao, and S. Han. 2013.
“Influence of Salinity Intrusion on the Speciation and Partitioning of Mercury in the Mekong
River Delta.” Geochimica Et Cosmochimica Acta 106: 379–390. doi:10.1016/j.gca.2012.12.018.
Oanh, T. K. T., N. V. Lap, M. Tateishi, I. Kobayashi, and Y. Saito. 2001. “Sedimentary Facies, Diatom
and Foraminifer Assemblages in a Late Pleistocene–Holocene Incised-Valley Sequence from the
Mekong River Delta, Bentre Province, Southern Vietnam. The BT2 Core.” Journal of Asian Earth
Sciences 20: 83–94. doi:10.1016/S1367-9120(01)00028-1.
Owen, A. B. 2001. ““Regression and Modelling.” Chap. 4.” In Empirical Likelihood. Boca Raton, FL:
Chapman and Hall.
Peddle, D. R., H. P. White, R. J. Soffer, J. R. Miller, and E. F. Ledrew. 2001. “Reflectance Processing of
Remote Sensing Spectroradiometer Data.” Computers & Geosciences 27 (2011): 203–213.
doi:10.1016/S0098-3004(00)00096-0.
Ratner, B. 2011. “Overfitting: Old Problem, New Solution. Chap.27.” In Statistical and Machine-
Learning Data Mining Techniques for Better Predictive Modeling and Analysis of Big Data. 2nd ed.
Taylor & Francis Group. Boca Raton, FL: CRC Press.
Roy, P. K., S. S. Roy, A. Giri, G. Banerjee, A. Majumder, and A. Mazumdar. 2014. “Study of Impact on
Surface Water and Groundwater around Flow Fields Due to Changes in River Stage Using
Groundwater Modeling System.” Clean Technologies and Environmental Policy 17: 145–154.
doi:10.1007/s10098-014-0769-9.
Shcheglovitova, M., and R. P. Anderson. 2013. “Estimating Optimal Complexity for Ecological Niche
Models: A Jackknife Approach for Species with Small Sample Sizes.” Ecological Modelling 269: 9–
17. doi:10.1016/j.ecolmodel.2013.08.011.
Trung, N. H., C. H. Hoanh, T. P. Tuong, X. H. Hien, L. Q. Tri, V. Q. Minh, D. K. Nhan, P. T. Vu, and V. P.
D. Tri. 2016. “Climate Change Affecting Land Use in the Mekong Delta: Adaptation of Rice-Based
Cropping Systems (CLUES) Theme 5: Integrated Adaptation Assessment of Bac Lieu Province
and Development of Adaptation Master Plan.” Accessed 3 January 2017 https://www.research
16 P. T. B. NGUYEN ET AL.

gate.net/publication/301612048_Climate_change_affecting_land_use_in_the_Mekong_Delta_
Adaptation_of_rice-based_cropping_systems_CLUES_ISBN_978-1-925436-36-5.
Trung, N. H., and V. P. D. Tri. 2014. “Possible Impacts of Seawater Intrusion and Strategies for Water
Management in Coastal Areas in the Vietnamese Mekong Delta in the Context of Climate
Change.” In Coastal Disasters and Climate Change in Vietnam, Waltham, edited by Nguyen
Danh Thao Hiroshi Takagi Miguel Esteban. USA: Elsevier. doi:10.1016/B978-0-12-800007-
6.00010-1.
Urquhart, E. A., B. F. Zaitchik, M. J. Hoffman, R. R. Murphy, and B. F. Zaitchik. 2013. “Geospatial
Interpolation of MODIS-derived Salinity and Temperature in the Chesapeake Bay.” Remote
Sensing of Environment 135: 167–177. doi:10.1016/j.rse.2013.03.034.
Urquhart, E. A., B. F. Zaitchik, M. J. Hoffman, S. D. Guikema, and E. F. Geiger. 2012. “Remote Sensing
of Environment Remotely Sensed Estimates of Surface Salinity in the Chesapeake Bay : A
Statistical Approach.” Remote Sensing of Environment 123: 522–531. doi:10.1016/j.
rse.2012.04.008.
Wang, F., and Y. Xu. 2012. “Remote Sensing to Predict Estuarine Water Salinity.” In Environmental
Remote Sensing and Systems Analysis, edited by N. B. Chang, 85-107. Boca Raton, FL: Taylor &
Francis Group/CRC Press.
Wang, F., and Y. J. Xu. 2008. “Development and Application of a Remote Sensing-Based Salinity
Prediction Model for a Large Estuarine Lake in the US Gulf of Mexico Coast.” Journal of
Hydrology 360: 184–194. doi:10.1016/j.jhydrol.2008.07.036.
Waske, B., M. Fauvel, J. A. Benediktsson, and J. Chanussot. 2009. “Machine Learning Techniques in
Remote Sensing Data Analysis.” In Kernel Methods for Remote Sensing Data Analysis, edited by G.
C. Valls and L. Bruzzone. Chichester, UK: John Wiley & Sons.
Wolanski, E., N. N. Huan, L. T. Dao, N. H. Nhan, and N. N. Thuy. 1996. “Fine-Sediment Dynamics in
the Mekong River Estuary, Vietnam.” Estuarine, Coastal and Shelf Science 43: 565–582.
doi:10.1006/ecss.1996.0088.
Xiong, Y. J., G. Y. Qiu, X. H. Chen, S. L. Tan, and H. X. Feng. 2012. “Hyperspectral Characteristics of
Seawater Intrision in Pearl River Delta, China Based on Laboratory Experiments.” In Geoscience
and Remote Sensing Symposium (IGARSS), Proceeding of the IEEE InternationalConference on
Munich, Germany, July 22-27.
Zhao, J., M. Temimi, and H. Ghedira. 2017. “Remotely Sensed Sea Surface Salinity in the Hyper-
Saline Arabian Gulf: Application to Landsat 8 OLI Data.” Estuarine, Coastal and Shelf Science 187:
168–177. doi:10.1016/j.ecss.2017.01.008.
Zhou, F., and A. Zhang. 2016. “Optimal Subset Selection of Time-Series MODIS Images and Sample
Data Transfer with Random Forests for Supervised Classification Modelling.” Sensors(Switzerland)
16: 1–18. doi:10.3390/s16111783.