Machine Learning Method for Predicting the Depth of Shallow Lakes Using Multi-Band Remote Sensing Images

Journal of Soft Computing in Civil Engineering 3-2 (2019) 54-64
How to cite this article: Jalilzadeh A, Behzadi S. Machine learning method for predicting the depth of shallow lakes using multi-
band remote sensing images. J Soft Comput Civ Eng 2019;3(2):54–64. https://doi.org/10.22115/scce.2019.196533.1119.
2588-2872/ © 2019 The Authors. Published by Pouyan Press.
This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Contents lists available at SCCE
Journal of Soft Computing in Civil Engineering
Journal homepage: www.jsoftcivil.com
Machine Learning Method for Predicting the Depth of Shallow
Lakes Using Multi-Band Remote Sensing Images
A. Jalilzadeh1
, S. Behzadi2*
1. M.Sc. Student in Geographic Information Systems, Department of Civil Engineering, Shahid Rajaee Teacher
Training University, Tehran, Iran
2. Assistant Professor in Surveying Engineering, Department of Civil Engineering, Shahid Rajaee Teacher Training
University, Tehran, Iran
Corresponding author: behzadi.saeed@gmail.com
https://doi.org/10.22115/SCCE.2019.196533.1119
ARTICLE INFO ABSTRACT
Article history:
Received: 31 July 2019
Revised: 21 October 2019
Accepted: 25 October 2019
Knowing the lake’s characteristics information such as depth
is an essential requirement for the water managers; however,
conducting a comprehensive bathymetric survey is
considered as a difficult task. After the advent of remote
sensing, and satellite imagery, it has been recognized that
water depth can be estimated in some way over shallow
water. There are many models that can evaluate relationships
between multi-band images, and depth measurements;
however, artificial computation methods can be used as an
approximation tool for this issue. They are also considered as
fairly simple and practical models to estimate depth in
shallow waters. In this paper, different methods of artificial
computation are used to calculate the depth of shallow lake,
then these methods are compared. The results show that
Artificial Neural Network (ANN), Adaptive Neuro Fuzzy
Inference System (ANFIS), and regression learner are best
methods for this issue with RMSE 0.8, 1.47, and 0.96
respectively.
Keywords:
Remote sensing;
Geographic information systems
(GIS);
Artificial computation;
Bathymetry.
1. Introduction
The sea surface similar to the earth’s surface is very complex with full of hills and valleys [1].
Determination of the water depths, knowing the detailed structure of the bottom and estimating

A. Jalilzadeh, S. Behzadi/ Journal of Soft Computing in Civil Engineering 3-2 (2019) 54-64 55
the volume of the lakes are essential requirements for the majority of coastal engineering, coastal
science applications, managers and decision-makers of water management sector [2–4]. Many
attempts have been undertaken by researchers over the past few decades to assembly depth data.
These data are used for marine transportation. In addition, water bodies have important defects in
environmental studies such as climate change. Therefore, checking sea levels in different time
periods provides useful information. Moreover, these data are used to study sea life such as
plants and fish which live in water [5,6]. Different methods can be used to illustrate the shape of
underwater terrain for better understanding. Bathymetry is a common method to measure the
water depth to illustrate the land that lies underwater. Bathymetric maps are the most used data
for mapping the seafloor [7]. There are different sources of depth data. These sources can be
classified into three different classes. 1) Hydrologic surveys, 2) Echo sounding (sonar) and 3)
satellites. In the Hydrologic surveys which is an early technique, a long, heavy pre-measured
rope or plummet, is used to measure the depth of the points. However, today it is not applicable
due to many errors and limitations such as time-consuming and being expensive [8].
Sonar and LiDAR are two common equipment to measure the depth of seawater and lakes. Both
sonar and LiDAR are mounted on the boat; in sonar, a sound wave is used for depth measuring
while in the LiDAR sensor, the electromagnetic wave is widely used. Despite being high-
precision, the echo sounding method is time-consuming, costly and even is impractical in deep
depth. It is suitable for low and medium depths. Satellite Gravity data can also be useful in
estimating depth in the whole basin. For example, GRACE mission with twin satellites are
making detailed measurements of earth's gravity [9] and SWOT MISSION (Surface Water Ocean
Topography) [10] make a global survey of earth’s surface water. In addition, satellite radar
altimetry has been used successfully to derive water levels of continental surface water bodies
[8,11]. Although bathymetric information is an obvious need in many remote areas, applications
of bathymetry mapping in coastal areas are beneficial for a wide range of people and researchers.
Because of the limitations, the number of measuring points cannot be too large and cannot cover
the whole entire area. So the measured data have general limitations in representing the true
terrain of the lake bottom [12]. To overcome such challenges, it is essential to develop methods
for depth estimation in uncertain points. Numerous studies have proposed different techniques to
estimate terrain of water bodies from measured points, so mapping coastal water from optical
remote-sensing techniques has become an interesting method [13].
The theory of using remote sensing techniques for mapping water depth and bottom features was
developed by Lyzenga [2] for the first time after that various depth estimation methods are
developed and expanded by many researchers based on optical remote sensing. For instance,
Bramante, Raju et al. [14], Stumpf, Holderied et al. [15] and Clark, Fay et al. [16] used this
method to obtain depth data. In all these researches, they attempted to evaluate the relationships
between multi-band images and depth measurements. For instance, Karimi, Bagheri et al [17]
developed an equation for evaluating the relationships between Urmia lake depth and Landsat
image bands. Obtaining the best equation is not a simple work, so artificial intelligence can be a
great help for this purpose. In [3] artificial neural network is used to evaluate water depth in Foca
bay lake in Turkey.

56 A. Jalilzadeh, S. Behzadi/ Journal of Soft Computing in Civil Engineering 3-2 (2019) 54-64
For over 40 years, the Landsat mission collects and provides information about earth, and it
helps to understand earth better. Landsat sensors receive natural light and thermal radiation from
the earth's surface. It can also measure water-leaving reflectance in water-bodies, which can be
used to estimate depth. There are several factors that affect the water leaving radiance especially
in shallow waters such as depth, the degree of transparency of the water, reflection from the
surrounding area, colored dissolved organic matter, nature and material of the bottom and water
turbidity. In a study area with a low operating radius, the depth factor can be more effective than
other factors [13,18,19].
There are different models and methods in machine learning algorithms including Artificial
Neural Network (ANN) and ANFIS and classify method. The purpose of this study is the
implementation of these methods and demonstrating their validity, and problems involved with
using these methods. The remainder of the paper is structured as follows: Section 2 presents the
available data. In section 3, the implementation of Artificial Neural Network, ANFIS and
Classify methods are discussed. Section 4 shows the results obtained with simulated and real
data, and finally, section 5 draws conclusions and Discussions.
2. Study area
The Caspian Sea, located in the west of Asia, is the world's largest inland water body and
bordered by 5 countries including: in the northeast by Kazakhstan, in the southeast by
Turkmenistan, in the south by Iran, in the southwest by Azerbaijan, and in the northwest by
Russia. The study area is located in Mazandaran province. The geographical location of the study
area is 36° 50′ 50” N to 36° 51′ 17” N latitude and 53° 16′ 17” E to 54° 50′ 49” E longitude.
Fig. 1. The Study area and collected points.

3. Dataset
In this research, several points with known depth are needed to extract depth information from
satellite images of shallow waters, and also for calibration. The obtained data about the lake’s
depth is scarce. Considering the available data of the Caspian Lake, the image obtained from the
LANDSAT satellite is used. The depth data is gathered with field observation and collected in
June 2013. The operation was designed in a way that the satellite's passage time is somewhat the
same as the data gathering time. The number of points is 456 and the minimum depth is -2.2800
and the maximum is -10.2100. Landsat satellite provides images with two instruments: the
Operational Land Imager (OLI) and the Thermal Infrared Sensor (TIRS). OLI consists of nine
spectral bands (Coastal, Blue, Green, Red, Near Infrared (NIR), - Shortwave Infrared (SWIR) 1,
Shortwave Infrared (SWIR) 2, Panchromatic, Cirrus) and TIRS consists of two thermal bands.
More information about Landsat bands is shown in Table 1.
Table 1
Landsat8 properties.
Landsat 8 Bands
Wavelength
(micrometers)
Resolution
(meters)
Operational Land
Imager (OLI)
Band 1 - coastal 0.435 - 0.451 30
Band 2 - Blue 0.452 - 0.512 30
Band 3 - Green 0.533 - 0.590 30
Band 4 - Red 0.636 - 0.673 30
Band 5 - Near Infrared (NIR) 0.851 - 0.879 30
Band 6 - Shortwave Infrared (SWIR) 1 1.566 - 1.651 30
Band 7 - Shortwave Infrared (SWIR) 2 2.107 - 2.294 30
Band 8 - Panchromatic 0.503 - 0.676 15
Band 9 - Cirrus 1.363 - 1.384 30
Thermal Infrared
Sensor (TIRS)
Band 10 - Thermal Infrared (TIRS) 1 10.60 - 11.19 100 * (30)
Band 11 - Thermal Infrared (TIRS) 2 11.50 - 12.51 100 * (30)
4. Methods
In this paper, attempt to estimate the structure of the sea bottom with different methods, then
these methods are compared. The used methods are respectively: 1) Artificial Neural Network 2)
ANFIS, and 3) Regression Learner. In these methods, input and output data play a main rule in
training and testing. Eleven bands of Landsat8 is considered as input data. Band No. 8 is not
involved in the model due to difference in pixel size. The surface temperature is also added as
input data. The formula for calculating the surface temperature is shown as Equatin.1:
𝑇𝐸 = (𝐾2 ÷ 𝑙𝑜𝑔(𝐾1 / 𝑅 + 1 )) − 273.5 (1)
Where TE is the temperature in Celsius, and K2 and K1 are constant value for thermal bands
(Band 10 and 11).

Band 1 Band 2 Band 3
Band 10 Band 11
Fig. 2. Image of study area in each band.
The depth points which is collected through field observation is used for training and testing. So,
10 bands from landsat8 and surface temperature is the input data. The input points are the pixels
with the same coordinates with collected depth points.

4.1. Neural network
Artificial Neural Network (ANN) as an approximation tool has found a fundamental role in
solving problems in various sciences. ANN can arrive at solutions by taking some data samples
rather than entire data. ANN has three interconnected layers: Input, hidden and output. The first
layer is for input data. The input data include water-leaving radiance in different bands. The
hidden layer may consist of numerous layers, and the output data include measured depth points.
Fig. 3. Structure of used neural network (Taken from Matlab).
In this research, ten bands among eleven bands of Landsat satellite are used (the data of band No.
8 or panchromatic band is not involved in input data because of its different structure). The test
areas include 456 points; these data are divided randomly into two groups: training and testing.
70% of the entire data set is utilized for training, and 30% remaining data is used for validating
and testing.
ANN is programmed with 1 hidden layer and trained with feed-forward backprop. The result is
shown in Table 2.
Table 2
Obtained result from neural network.
Network type Number of layers Number of neurons Iterations RMSE
feed-forward backprop 1 10 6 0.8
4.2. ANFIS
Fuzzy logic was first advanced by [20] and has come of age its applications have grown in
number and variety. The Adaptive neuro-fuzzy inference system or adaptive network-based
fuzzy inference system (ANFIS) is a subset of artificial intelligence, and it combines both neural
networks and fuzzy logic principles.

Fig. 4. ANFIS model structure for 2 bands (Taken from Matlab).
ANFIS Model generated with gauss2mf (combining two Gaussian membership functions for
computes fuzzy membership values) and 3 MFs (membership function (MF) represents the
membership value for each point) to each input. 350 points are used for training and 106 points
are also used for testing.
Fuzzy Inference System (FIS) structure is generated from data using grid partition. Input and
output data are the same as ANN. The data is divided into two groups: training and testing. 350
points are assigned for training, and 106 points for testing. For input data type of fuzzy
Membership Function (MF) Type is Triangular (trimf) and number of MFs is eleven as follow [2
2 1 1 1 2 1 2 1 1 1] and for output MF type is linear. In this paper, optimization method is hybrid
for Train FIS. The result of implementing ANFIS on data and some extra information is shown in
Table 3. Scatter plots against training and testing data are showed in figure 4 and 5.
Table 3
ANFIS info and Obtained result.
Number
of training
data pairs
Number
of nodes
Number of
linear
parameters
Number of
nonlinear
parameters
Total number
of
parameters
Number of
fuzzy rules
Epochs RMSE
350 76 192 45 237 16 10 1.47

Fig. 5. Plot against training data.
Fig. 6. Plot against training data.
4.3. Regression learner
The Regression Learner app is used to predict data with trained regression models. These models
are including linear regression models, regression trees, Gaussian process regression models,
support vector machines, and ensembles of regression trees. After training models, their RMSE is
compared, which represented in Table 4 to search for the best regression model type. The best
answer belongs to Gaussian Process Regression Model and Exponential type with RMSE equal
to 0.96.
Table 4
Obtained results from regression learner models.
Linear Regression
Models
RMSE
Support Vector
Machines
RMSE Ensembles of Trees RMSE
Linear 1.404 Linear SVM 1.62 Boosted Trees 1.04
Interactions Linear 5.63 Quadratic SVM 2.15 Bagged Trees 1.04
Robust Linear 1.43 Cubic SVM 10.34
Stepwise Linear 1.48 Fine Gaussian SVM 1.03
Gaussian Process
Regression Models
RMSE
Medium Gaussian SVM 1.28 Rational Quadratic .97
Regression Trees RMSE Coarse Gaussian SVM 1.55 Squared Exponential .99
Fine Tree 1.05 Matern 5/2 .98
Medium Tree 1.13 Exponential .96
Coarse Tree 1.29

5. Results comparison
As seen, the neural network with RMSE Of 0.8 is the best answer in compare to ANFIS with
RMSE equal to 1.45, and regression learning with RMSE Of 0.96. The length of processing time
is another issue that can be used to compare these algorithms. The neural network responds with
the shortest time and can easily apply and test different structures. Although ANFIS takes the
advantages of fuzzy logic and neural network, its processing time is so long and even the
hardware of computer cannot respond with increasing the volume of data or complexity of
structure. However, the obtained result through ANFIS is not satisfactory. Regression learning
responds at a high-speed when only one model is running, but in order to find the best regression
model, all models need to be tested, which reduces the processing speed slightly. However, the
obtained result and processing time is satisfactory. Table 5 provides a comparison of these
models.
Table 5
Comparing used estimation methods.
Processing time RMSE Simplicity Easily at structure changing
Neural Network 1 1 1 1
ANFIS 3 3 3 2
Regression learning 2 2 2 3
The results show that ANN performs better than the others, so it has been tried to create a
bathymetry map and 3D view of the whole study area.
In the study area, there are some pixels which are not water, so these data needs to be filtered.
Water Ratio Index (WRI) is selected as a proper index to detect water pixels (Eq.2). in this
Equation, the range bigger than 1 is considered as water area [21].
𝑊𝑅𝐼 =
(𝑇3+𝑇4)
(𝑇5 +𝑇6)
(2)
The right image in figure 4 gives information of depth in the study area. In this figure, lighter
blue pixels present shallow areas and darker ones are deeper.
Fig 7. The study area and counters and bathymetry map.

6. Conclusion and discussion
Although the answer is somewhat satisfactory, these responses are greatly improvable. The most
important issue in improving the final result is the number of depth data as a trainer to the
algorithms. The satellite image as the input data is selected based on available depth data. For
example, Landsat satellite is used in this article, but there are other satellites that can be used,
such as Sentinel which is free and have a better resolution in comparison with Landsat. Depth
data is scarce in literature or available data is considered classified in many countries. Although
there is depth information on the sites, most of them are for deep points and this method does not
respond at very deep points when the bottom influence becomes negligible. Given that the pixel
resolution of the Landsat satellite is 30 meters, so some measured points were in the same pixel,
which can partly affect the training and response. Despite all these issues, artificial intelligence
has shown that it can overcome these difficulties, and this method can conveniently reduce the
labor needed and saves both time and money and provides a fast and practical solution for depth
estimation in shallow waters. The main advantage of using this method is it estimates depth
without removing the many troublemaker factors that exist.
References
[1] Cazenave A, Altimetry S. Reference module in earth systems and environmental sciences 2018.
[2] Lyzenga DR. Passive remote sensing techniques for mapping water depth and bottom features.
Appl Opt 1978;17:379. doi:10.1364/AO.17.000379.
[3] Ceyhun Ö, Yalçın A. Remote sensing of water depths in shallow waters via artificial neural
networks. Estuar Coast Shelf Sci 2010;89:89–96.
[4] Elsahabi M, Makboul O, Negm A. Lake Nubia sediment capacity estimation based on satellite
remotely sensed detected bathymetry. Procedia Manuf 2018;22:567–74.
[5] Dörnhöfer K, Oppelt N. Remote sensing for lake research and monitoring–Recent advances. Ecol
Indic 2016;64:105–22.
[6] https://oceanservice.noaa.gov/facts/bathyuses.html n.d.
[7] https://sealevel.jpl.nasa.gov/overview/oceansurfacetopo/ n.d.
[8] Duan Z, Bastiaanssen WGM. Estimating water volume variations in lakes and reservoirs from four
operational satellite altimetry databases and satellite imagery data. Remote Sens Environ
2013;134:403–16.
[9] http://www2.csr.utexas.edu/grace/. n.d.
[10] https://swot.jpl.nasa.gov/home.htm. n.d.
[11] Calmant S, Seyler F, Cretaux JF. Monitoring Continental Surface Waters by Satellite Altimetry.
Surv Geophys 2008;29:247–69. doi:10.1007/s10712-008-9051-1.
[12] Li C, Sun B, Jia K, Zhang S, Li W, Shi X, et al. Multi-band remote sensing based retrieval model
and 3D analysis of water depth in Hulun Lake, China. Math Comput Model 2013;58:771–81.
[13] Jay S, Guillaume M. A novel maximum likelihood based method for mapping depth and water
quality from hyperspectral remote-sensing data. Remote Sens Environ 2014;147:121–32.

[14] Bramante JF, Raju DK, Tsai Min S. Derivation of bathymetry from multispectral imagery in the
highly turbid waters of Singapore’s south islands: A comparative study. Digit 8-Band Res Chall
2010 2013:701–10.
[15] Stumpf RP, Holderied K, Sinclair M. Determination of water depth with high-resolution satellite
imagery over variable bottom types. Limnol Oceanogr 2003;48:547–56.
doi:10.4319/lo.2003.48.1_part_2.0547.
[16] Clark RK, Fay TH, Walker CL. Bathymetry calculations with Landsat 4 TM imagery under a
generalized ratio assumption. Appl Opt 1987;26:4036_1. doi:10.1364/AO.26.4036_1.
[17] Karimi N, Bagheri MH, Hooshyaripor F, Farokhnia A, Sheshangosht S. Deriving and Evaluating
Bathymetry Maps and Stage Curves for Shallow Lakes Using Remote Sensing Data. Water Resour
Manag 2016;30:5003–20. doi:10.1007/s11269-016-1465-9.
[18] Werdell PJ, McKinna LIW, Boss E, Ackleson SG, Craig SE, Gregg WW, et al. An overview of
approaches and challenges for retrieving marine inherent optical properties from ocean color
remote sensing. Prog Oceanogr 2018;160:186–212.
[19] Traganos D, Reinartz P. Mapping Mediterranean seagrasses with Sentinel-2 imagery. Mar Pollut
Bull 2018;134:197–209.
[20] Zadeh LA. Fuzzy logic= computing with words.-IEEE Transactions on Fuzzy Systems, vol. 4,№ 2,
pp. 103-111 n.d.
[21] Zhang FF, Zhang B, Li JS, Shen Q, Wu Y, Song Y. Comparative analysis of automatic water
identification method based on multispectral remote sensing. Procedia Environ Sci 2011;11:1482–
7.

Machine Learning Method for Predicting the Depth of Shallow Lakes Using Multi-Band Remote Sensing Images

More Related Content

Similar to Machine Learning Method for Predicting the Depth of Shallow Lakes Using Multi-Band Remote Sensing Images (20)

More from Journal of Soft Computing in Civil Engineering (20)

Recently uploaded (20)

Machine Learning Method for Predicting the Depth of Shallow Lakes Using Multi-Band Remote Sensing Images