Adaptive Neuro-Fuzzy Inference System for Echoes Classification in

Radar Images

Leila Sadouki

1,2

and Boualem Haddad

Institute of Electrical and Electronic Engineering, University M'Hamed Bougara of Boumerdes (UMBB),

Independence Avenue, 35000 Boumerdes, Algeria

Laboratory of Image Processing and Radiation, Faculty of Computer Science and Electronics (USTHB),

PObox 32 El Alia. Bab Ezzouar, Algiers, Algeria

Keywords: ANFIS, Image, Radar, Precipitations, Clutter.

Abstract: In order to remove the undesirable clutter which reduces the radar performances and causes significant

errors in the rainfall estimation, we implemented in this paper an algorithm deals with the classification of

radar echoes. The radar images studied are those recorded in Sétif (Algeria) every 15 minutes, we used a

combination of textural approach, with the grey-level co-occurrence matrices, and a grid partition based

fuzzy inference system, named ANFIS-GRID. We have used two parameters, namely Energy and local

homogeneity that are considered to be the most effective in discriminating between precipitation echoes and

clutter. Those parameters are used as inputs for the ANFIS-GRID, while the output of this system is the

radar echo types. In function of the best mean rate of correct recognition and using two different

optimization methods, the structure with 2 inputs, 4 membership functions, 16 rules and 1 output was

selected as the most efficient ANFIS-GRID. This method gives a mean rate of correct recognition of echoes

to over 93.52% (91.30% for precipitation echoes and 95.60% for clutter). In addition, the proposed

approach gives a process maximum time of less than 90 seconds, which allows the filtering of the images in

real time.

1 INTRODUCTION

The demands for finer scale meteorological services

have more and more required higher resolution

observations to initialize and evaluate weather and

climate models, applications, and products. In

response to these demands smart techniques are

increasingly used in the design, classification,

modeling and control of complex systems, such as

neural networks, fuzzy logic and genetic

algorithms…

For weather radar, the presence of echoes

coming from the earth's surface, or clutter, mixed

with precipitation echoes, making the hydrological

measurement very difficult (Sauvageot and

Despaux, 1990), a good way to eliminate clutter is to

compare the statistical properties of the ground

echoes to those of precipitations echoes, such as

textural features (Haddad et al., 2004). The most

common techniques used are Doppler filtering

(Doviak and Zrnic, 1993), or dual polarization

filtering (Islam et al., 2012; Chandrasekar et al.,

2013). Clutter can be removed by analyzing in real

time the coefficient of the autocorrelation function

of the radar signal (Hamuzu and Wakabayashi,

1991). Others applied the fuzzy logic technique, to

classify the Doppler radar echoes types (Hubbert et

al., 2009) or for identifying non-precipitating echoes

in radar scans (Berenguer et al., 2006; Cho et al.,

2006). In (Sadouki and Haddad, 2013), they

combined the textural properties of the echoes with

the fuzzy approach in order de classify the echoes.

Furthermore, (Xiang, 2010) used the neuro-fuzzy

approach to eliminate noise in Doppler radar signals.

Depending on the on above literature survey, it’s

interesting to implement an algorithm which

combines the textural features, based on the method

of grey-level co-occurrence matrices, and an

Adaptive Neuro-Fuzzy System (ANFIS) with grid

partition. The input variables for our system are two

textural parameters considered as effective elements

for distinguishing between precipitation echoes and

clutter (Sadouki and Haddad, 2013). This method

was applied to the images taken in the region of

Sétif (Algeria).

Sadouki, L. and Haddad, B.

Adaptive Neuro-Fuzzy Inference System for Echoes Classiﬁcation in Radar Images.

DOI: 10.5220/0005717401590166

In Proceedings of the 11th Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2016) - Volume 4: VISAPP, pages 159-166

ISBN: 978-989-758-175-5

159

The remainder of the paper is organized as

follows: Section 2 focuses on database used in this

work, sections 3 and 4 deal respectively with the

used concepts and the data processing. We illustrate

discuss and validate the different results in sections

5 and 6. Finally, and in section 7, we give our

conclusion.

2 DATA BASE

Weather radar of Sétif is of the Type AWSR-81

(Algerian Weather Service Radar). It is non-coherent

pulsed radar which consists of a transmitter, a

receiver, a duplexer, antenna of 3 m in diameter and

is associated with a SANAGA chain (Système

d’Acquisition Numérique pour l’Analyse des Grains

Africains) which is a system of acquisition and

digitization of images (Sauvageot and Despaux,

1990). The main characteristics of this radar are:

 Transmission power: 250 kW;

 Transmission frequency: 5.6 GHz;

 Reception sensitivity to: -110 dBm;

 Pulse width: 2μs;

 Antenna gain: 30 dB;

 Return period: 4 ms;

 Beam width (3 dB): 1.1;

Our database consists of images taken during the

period of 1997-2001. Sétif radar is positioned at

latitude of 36°11'N, a longitude of 5°25' E and an

altitude of 1700 m above the sea level. It records

every fifteen minutes an image of 512×512 pixels

using the PPI (Plan Position Indicator) presentation,

with a resolution of 1 km per pixel.

As shown in the image of Figure 1, the images

recorded in this site using the C-band meteorological

radar, use a palette of sixteen colors. We find in

those images a lot of ground echoes coming from the

earth's surface. These echoes are, in particular, due

to the fact that Sétif region is a part of the Algerian

highlands and its Radar is surrounded by several

ground obstacles. The nearest ground echoes are

produced largely by the industrial area. Beyond the

horizon, the ground obstacles produce several

ground echoes in the radar images. For example, to

the southwest, 60 km away from the radar, there are

the mountains of Djurdjura, which reach an altitude

of 2300 m. In the same direction, 40 km away from

the radar, we find the mountains of Bibans with a

height of 1417 m. To the northeast, at a shorter

distance (about 30 km away from the radar), are

located the mountains of Babors, which reach an

altitude of 2004 m.

Figure 1: Radar images of Sétif.

3 USED CONCEPTS

The main methods that are considered in this paper

are textural approach, using the co-occurrence

matrices, and the Neuro-Fuzzy controller. The latter,

is the combination of the Neural networks and Fuzzy

logic, that takes advantages of both approaches.

Figure 2 summarizes the concepts used in this study.

Figure 2: Block diagram of the used concepts.

3.1 Co-occurrence Matrices

The grey-level co-occurrence matrices are among

the most frequently used statistical methods in the

field of the texture analysis of the radar and the

satellite images (Haralick, 1979). The gray-level Co-

occurrence matrix of an image is obtained by

estimating the joint conditional density of

probability functions of second-order P (i, j / d, θ ),

the latter represents the transition probability of a

pixel of gray level “i” to a pixel of gray level “j”.

Input: Features extraction

from the Radar images

ANFIS

Output: Class identification

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

160

This transition is controlled by: the distance “d”

between the two pixels and the orientation “θ”which

is defined by the angle between the direction of

transition and the image scanning direction.

The orientation “θ” can be determined also with

Cartesian coordinates (Δx, displacement in the

horizontal direction and Δy, displacement in the

vertical direction)

The elements P (i, j) denoted Pij of the Co-

occurrence matrix represent the frequency of

occurrence of the pair of gray levels (i, j) in the

processing window “W” of T1×T2 size, according to

a relationship represented by the pair (Δx, Δy). They

are defined as follows:

P (i , j /Δx ,Δy) = Card { (m ,n), ( m+Δx ,

n+Δy)∈ W/ I( m , n) = i and I ( m +Δx , n

+Δy) = j }

(1)

Where Card, is the cardinal or the number of

elements, and I (m, n) and I (m+Δx, n+Δy) represent

respectively the intensities of pixels “i” and “j”

located at (m,n) and (m+Δx, n+Δy) in the window

“W”.

The elements of the direction matrix C

(θ, d) are

written:

(θ,d) = P (i , j /

x , Δy) / r

(2)

Where “r” is the normalization parameter which is

equal to: (T1-|Δx|) × (T2-|Δy|).

There are eight Co-occurrence matrices C (θ, d)

for different directions (θ = 0°, 45°, 90°, 135°, 180°,

225°, 270°, 315°).

We can calculate, using the Co-occurrence

matrix, a set of statistical properties (i.e. Mean,

Variance, Inertia, Local homogeneity, Energy,

Correlation, Entropy, Nuances grouping, and

Predominance grouping), which allow us to reveal

the particular characteristics of image texture. In this

paper, we will use only two parameters (i.e. Local

homogeneity and Energy) that give the results the

most uncorrelated, with the direction θ =0° and the

distance d=1, which correspond to the Cartesian

coordinates (Δx=0, Δy=1) (Sadouki and Haddad,

2013).

It’s worth noting that, among the values of d=

{1,2,3,4}, the distance d = 1, was chosen, by

experience, as the best value in terms of

effectiveness (in distinguishing between

precipitation echoes and clutter) and the

proportionality with the size of the study windows

(5×5 pixels window).

Equations of those parameters are: (Haralick,

1979; Unser, 1986; Peckinpaugh, 1991).

 Energy which measures textural uniformity:



−

(3)

 The local homogeneity that gives greater weight

to the occurrence frequencies of homogeneous

zones:

()

[]



−

−+=

ijL

CH ji

(4)

3.2 Neuro-Fuzzy Concept

A Neuro-Fuzzy (NF) system is a combination of

Artificial Neural Network (ANN) and Fuzzy

Inference System (FIS) in such a way that ANN

learning algorithms are used to determine the

parameters of FIS (Kurian et al., 2006).

ANFIS (Adaptive Neuro-Fuzzy Interface

System) is the fuzzy Sugeno model based paradigm

that grasps the learning abilities of ANN to enhance

the intelligent system’s performance using a priori

Knowledge.

Using a given input/output data set, ANFIS

constructs a fuzzy inference system (FIS) whose

membership function parameters are tuned using

either a back-propagation algorithm alone, or in

combination with a least squares method. This

allows your fuzzy systems to learn from the data

they are modeling. The learning method works

similarly to that of neural networks (Chaudhari et

al., 2012). In fact, ANFIS cancels out the

interference and gives better performance even if the

complexity of the signal is very high.

We used in this paper, the ANFIS-GRID fuzzy

inference system which is the combination of grid

partition and ANFIS. Grid partition divides the data

space into rectangular sub-spaces using axis-

paralleled partition based on predefined number of

membership functions and their types in each

dimension as shown in Figure 3.

Figure 3: Grid partition of an input domain with 2 input

variables and 2 membership functions for each input.

Adaptive Neuro-Fuzzy Inference System for Echoes Classiﬁcation in Radar Images

161

The number of fuzzy rules increases

exponentially when the number of input variables

increases. For example, if there are averagely “m”

membership functions (MF) for every input variable

and a total of “n” input variables for the problem, the

total number of fuzzy rules is “m

” (Wei et al.,

2007).

For a first-order Sugeno fuzzy model, a common

rule set with “k” fuzzy “if-then” rules is given by:

(Bhavani et al., 2012)

Rule k:

If X is Ai and Y is Bj,

Then fk = pk X + qk Y + rk

Where: k=1...i*j

To present the ANFIS architecture, let us

consider the example of the Figure 4 which has two

inputs (X , Y), and one output “f”.

Layer 1: Calculate Membership Value for Premise

Parameter: Every node “i” in this layer is an

adaptive node

Layer 2: Firing strength of rule: The nodes in this

layer are fixed (not adaptive).These are labeled “

Π”

to indicate that they play the role of a simple

multiplier.

Layer 3: Normalize firing strength: Nodes in this

layer are also fixed nodes. These are labeled “N” to

indicate that these perform a normalization of the

firing strength from previous layer.

Layer 4: Consequent Parameters: All the nodes in

this layer are adaptive nodes.

Layer 5: Overall output: This layer has only one

node labeled “Σ” indicated that is performs the

function of a simple summer.

Figure 4: ANFIS Architecture (an example with 2 inputs

and 4 rules).

4 DATA PROCESSING

To classify the types of echoes in radar images, a

variety of samples (sub-image), carefully selected

from images, where precipitation echoes (class 1)

are distinctly separated from the ground echoes

(class 2). It’s important to note that each sub-image,

of a maximum size of about 15×15 pixels, illustrates

a different meteorological situation.

For each 5×5 pixels window in a given sub-

image, we calculated the statistical parameters

Energy and Local homogeneity that have been found

to be useful in discriminating between precipitations

and clutter, and have been chosen as inputs of our

classifier. (Sadouki and Haddad, 2013)

As result to the previous process, we were

capable to construct our database of 1000 vectors

that corresponds to our two classes, 500 for clutter

and 500 for precipitations. Each vector is composed

of 3 elements, Energy, Local homogeneity and class.

In fact, we used MATLAB commands for

learning process with 1000 epochs and 600 training

sample from the two classes. In addition, the

optimization methods applied, in order to train the

membership function’s parameters to emulate the

training data, are the back-propagation and the

hybrid methods, where the second method is a

combination of least-squares and back-propagation

gradient descent method.

It’s worth mentioning that the others 400 vectors

of our data base were use in the testing process.

The output of the ANFIS classifier will be used

later to find an appropriate approach, which will

allow us to separate the ground echoes from the

precipitation echoes in order to eliminate the

undesirable echoes.

5 RESULTS AND DISCUSSION

After creating different architectures of ANFIS

classifier, and using triangular membership

functions for each Input, we perform the training of

each classifier using our database. Since the errors,

obtained during training and testing, seem to be of

the same level, so, we were obliged to validate our

classifier. Thus, we classified 20 images using 8

different topologies, after that, we calculated the rate

of correct recognition for each image (denoted

RCR). This rate is calculated by the following

expression:

100)N

(RCR

×=



(5)

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

162

Where:

c : Is the number of classes.

N : Is the total number of pixels.

Xi : Is the number of pixels correctly classified to

the class i.

Table 1 collects the results obtained, with

different topologies and by changing:

 The number of Membership Functions for each

input (MFin1-MFin2),

 Rules.

 Optimization Method.

Where, MFin1 and MFin2 are the number of

membership functions for the parameters Energy

and local homogeneity respectively.

Table 1: Rate of correct recognition (RCR in %) and the

associated time (in Seconds) for different topologies and

different optimization methods (average of 20 images).

Topology

(MFin1-MFin2)

Back propagation

Opt. Method

Hybrid Opt. Method

RCR Time RCR Time

(2-2) 92,85 71.30 92,65 72.14

(3-3) 92,96 71.61 92,22 73.13

(4-4) 93,52 72.14 92,57 74.93

(5-5) 90,60 73.25 91,95 75.78

(6-6) 83,89 73.78 92,30 76.70

(7-7) 79,42 74.84 92,88 77.19

(10-10) 75,37 76.28 88,52 80.40

(15-15) 73,11 78.67 87,16 81.23

According to the results of the Table 1, it’s clear

that from the topology (5-5) and when we use the

back propagation method, the RCR decreases with

the increasing of the number of the membership

functions, but for the hybrid method, the same rate

decreases from the level (7-7). Also, we can see that

the most adequate topology, which was fixed after

several trials based on the best rate of correct

recognition, is the (4-4) network with the back

propagation optimization method, with 2 inputs, 4

membership functions, 16 rules and 1 output. In

addition the processing time is less than 90 seconds,

which is a relatively small time comparing with the

image acquisition time.

The model of this ANFIS is shown in Figure 5.

Rules:

 Rules 1 to 4: with i,j=1..4

If (Input1 is In1MF1) and (Input2 is In2MFi)

then (Output is OutMFj)

 Rules 5 to 8: with i=1..4 and j=5..8

If (Input1 is In1MF2) and (Input2 is In2MFi)

Figure 5: ANFIS Structure.

then (Output is OutMFj)

 Rules 9 to 12: with i=1..4 and j=9..12

If (Input1 is In1MF3) and (Input2 is In2MFi)

then (Output is OutMFj)

 Rules 13 to 16: with i=1..4 and j=13..16

If (Input1 is In1MF4) and (Input2 is In2MFi)

then (Output is OutMFj)

Where, OutMFj (j=1..16) are the membership

functions of the outputs.

Figure 6: ANFIS Surface View.

Figure 7: Membership functions plot (final forms).

Adaptive Neuro-Fuzzy Inference System for Echoes Classiﬁcation in Radar Images

163

The ANFIS surface view and the final forms of

the 4 membership functions of each input are shown

in Figure 6 and Figure 7 respectively. Where,

In1MFi (i=1..4) are the membership functions of the

first input Energy or “E”, while In2MFi are the

membership functions of the second input Local

Homogeneity or “HL”.

As illustration of the output of our classifier, the

image of Figure 8 recorded in 09 November 2001 is

considered. It provides the case were the

precipitations partially cover the ground echoes.

Figure 8: Radar image of Sétif region recorded in

09/11/2001.

Figure 9: Radar image of Sétif region (Classified image).

In order to eliminate the undesirable echoes or

clutter, we used our approach, to classify and filter

this image with the (4-4) ANFIS network. To do

that, we performed the following process:

Figure 10: Radar image of Sétif region (Filtered image).

For each pixel in the image and using the surrounded

pixels, we computed the parameters Energy and

Local homogeneity, after that, we applied them to

the ANFIS network to get, as result, the appropriate

class for that pixel. With this classifier, we can

observe clearly through Figure 9 that the two classes

are well classified.

Whereas for the filtering, and for each pixel in

the image, we performed the following test: If the

current pixel is evaluated in the class of clutter, we

assign the black colour for that pixel, otherwise, we

maintain the initial colour. Figure 10 shows that the

ground echoes appearing on the considered image of

Sétif are eliminated and the precipitation fields are a

little bit affected by the filtering.

6 VALIDATION

It’s very important to note that for the case of

images, where the precipitations cover partially the

ground echoes, the estimations of the

RCR

in the

images, are very difficult. Consequently, we were

obliged to use another way to validate our technique,

which is the estimation of the intensity of rainfall

using the radar relationship for temperate climates:

(Sauvageot, 1992)

Z = 300 R

1.5

(6)

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

164

Where Z and R are, respectively, the radar

reflectivity factor (expressed in mm

−3

) and the

precipitation rate (expressed in mm h

−1

To verify that the filtering of clutter does not

affect the reflectivity of precipitation echoes, we

compared the intensity of rainfall collected and

measured by pluviometer, and that estimated by

radar images during the extreme rain event,

observed on November 09-10, 2001 in the region of

Algiers, which was at the origin of a natural disaster.

We recorded in the day of November 10

amount of rain equals to 132 mm in 6 hours duration

(6:00 to 12:00). (Haddad et al., 2003)

Since we have a chronological set of 25 images

recorded from 6:00 am to 12:00 pm, we were

capable to find the intensity of rainfall estimated by

the filtered radar images which is 121.6 mm. Thus,

the estimation error is about 7.87%.

7 CONCLUSIONS

The method described in this paper shows that the

combination of the textural features, using Co-

occurrence matrices, and Adaptive Neuro-Fuzzy

Interface System, with the utilization of grid

partition, allows an efficient radar echoes

classification. In function of two factors which are

filtering rate and computation time, the structure 2

inputs with 4 membership functions for each and 16

(or 4

) rules was selected as the most efficient

network. The application of this approach gives a

mean rate of correct recognition of echoes to over

93.52% (91.30% for precipitation echoes and

95.60% for clutter) for the images recorded in the

site of Sétif. In addition, time of processing is about

90s which is less than 2 minutes. It would be

interesting to extend this study to other sites of

different climates to check the effectiveness of the

technique and if the thresholds and membership

functions always stay invariant.

ACKNOWLEDGEMENTS

The authors would like to thank the National

Meteorology Office of Algeria for providing the

radar data base used in this study. We would also

like to thank the reviewers for their valuable

comments and suggestions.

REFERENCES

Berenguer, M., Sempere-Torres, D., Corral, C. and

Sanchez-Diezma, R., 2006. A fuzzy logic technique

for identifying nonprecipitating echoes in radar scans.

Journal of Atmospheric and Oceanic Technology, vol.

23, pp. 1157-1180.

Bhavani Sankar, A., Kumar, D., and Seethalakshmi, K.,

2012. A New Self-Adaptive Neuro Fuzzy Inference

System for the Removal of Non-Linear Artifacts from

the Respiratory Signal. Journal of Computer Science.

vol. 8 (5), 621-631.

Chandrasekar, V., Keränen, R., Lim, S., and Moisseev, D.,

2013. Recent advances in classification of

observations from dual polarization weather radars.

Atmospheric Research, vol. 119, pp. 97-111.

Chaudhari, O. K., Khot, P. G., Deshmukh, K. C., and

Bawne, N. G., 2012. ANFIS based model in decision

making to optimize the profit in farm cultivation.

International Journal of Engineering Science and

Technology (IJEST). Vol. 4 (2), 442-448.

Cho, Y. H., Lee, G., Kim, K. E. and Zawadzki, I., 2006.

Identification and removal of ground echoes and

anomalous propagation using the characteristics of

radar echoes. Journal of Atmospheric and Oceanic

Technology. vol. 23, pp. 1206-1222.

Doviak, R. J., and Zrnic, D. S., 1993. Doppler radar and

weather observations, Academic Press., pp. 562.

Haddad, B., Sadouki, L., Naili, R, Adane, A., and

Sauvageot, H., 2003. Analyse De La Dimension

Fractale Des Echos De Precipitations: Cas Des

Inondations D'Alger. Publication de l 'Association

Internationale de Climatologie. vol. 15, pp.386-392.

Haddad, B., Adane, A., Sauvageot, H., and Sadouki, L.,

2004. Identification and filtering of rainfall and ground

radar echoes using textural features. International

Journal of Remote Sensing. vol. 25(21), pp. 4641–

4656.

Hamuzu, K., and Wakabayashi, M., 1991. Ground clutter

rejection. In Hydological applications of Weather

Radar, Clukie and Collier. Ed Ellis Horwood Ltd, pp.

131–142.

Haralick, R. M., 1979. Statistical and structural

approaches to textures. Proceedings of the IEEE on

Image Processes, vol. 67, pp. 786–804.

Hubbert, J. C., Dixon, M., and Ellis, S. M., 2009. Weather

Radar Clutter. Part II: Real-Time identification and

filtering. Journal of Atmospheric and Oceanic

Technology, vol. 26, pp. 1181–1197.

Islam, T., Rico-Ramirez, M. A., Han, D. and Srivastava,

PK., 2012. Artificial Intelligence Techniques for

Clutter Identification with Polarimetric Radar

Signatures. Atmospheric Research, 109-110, pp. 95-

113.

Kurian, C. P., George, V. I., Jayadev, B., and

Radhakrishna, S. A., 2006. ANFIS Model For The

Time Series Prediction of Interior Daylight

illuminance. AIML Journal. Vol. 6 (3).

Peckinpaugh, S. H., 1991. An Improved Method for

Computing Grey-Level Co-Occurrence Matrix Based

Adaptive Neuro-Fuzzy Inference System for Echoes Classiﬁcation in Radar Images

165

Texture Measures. CVGIP: Graphical Models and

Image Processing. vol. 53, 574–580.

Sadouki, L., and Haddad, B., 2013. Classification of radar

echoes with a textural–fuzzy approach: an application

for the removal of ground clutter observed in Sétif

(Algeria) and Bordeaux (France) sites. Int. J. of

Remote Sensing, vol. 34(21), 7447-7463.

Sauvageot, H., and Despaux, G., 1990. SANAGA: Un

système d’acquisition numérique et de visualisation

des données radar pour la validité des estimations

satellitaires de précipitations. Veille Climatique

Satellitaire, vol. 31, pp. 51–55.

Sauvageot, H., 1992. Radar Meteorology. Norwood:

Artech House., pp. 361.

Unser, M., 1986. Sum and difference histograms for

texture classification. IEEE Transactions on Pattern

Analysis and Machine Intelligence. vol. 8(1), pp. 118–

125.

Wei, M., Bai, B., Sung, A. H., Liu, Q., Wang, J., and

Cather, M. E., 2007. Predicting injection profiles using

ANFIS. Information Sciences. vol. 177, 4445–4461.

Xiang, L., 2010. Adaptive Network Fuzzy Inference

System Used in Interference Cancellation of Radar

Seeker. IEEE International Conference on Intelligent

Computing and Intelligent Systems (ICIS). vol. 2, pp.

93–97.

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

166