IEEE MELECON 2004, May 12-15,2004, Dubrovnik, Croatia

Radar Remote Sensing for Oil Spill Classification (Optimization for Enhanced Classification)

Krishnaswamy Sankaran* and Joaquim Fortuny Guasch** * University of Karlsruhe TH / Institute for High Frequency Techniques & Electronics (IHE)

Karlsruhe, 761 3 1 - Germany Email: saykrish@rediffmail.com * * DG Joint Research Centre, Ispra (VA), 21020 - Italy Email: Joaquim.Fortuny@jrc.it

Abstruct- Oil spills are a major factor in the ocean pollution. The complications involved in detecting oil spills are due to varying wind and sea surface conditions. The main aim of this paper is to find the best combination of transmit and receive polarizations for optimal detection of oil spills which can extend the tolerance and validity ranges of single and dual polarization spaceborne missions. The optimization has to be handled in a careful way for the oil spill detection because the backscattering is very low from the oil spilled region due to high Fresnel reflection. This paper deals with an improved optimization technique for the detection and classification scheme using the Spaceborne Imaging Radar (SIR) data sets of the oil spilled regions. Both the theory and the experimental results obtained are discussed.

I. INTRODUCTION Ocean pollution in the recent years has become a major

issue due to the increasing number of illicit discharges. Apart from the increasing political pressure in handling these issues technical viability to give better solution in controlling and monitoring the illicit discharges is gaining momentum among various research organizations. Many research papers [ l , 2, 61 were published in the last two decades for more general cases but still a good solution to this problem is in search. In fact no robust technique has been developed to solve the problem of oil spill detection and classification over any sea surface conditions. Varying wind and ocean surface conditions, lack of credible data and technology are the main reasons for this impediment. In the advent of fully polarimetric system we are able to analyze the situation well beyond the limits which were imposed on the former single and dual polarization missions. Monitoring and automatic analysis of SAR images are not applied regularly [7] due to lack of a robust and reliable detection and classification scheme to determine the nature of the slick (real oil slicks, look- alikes, ocean currents etc).

A. Fully polarimetric radar systems Compared to the former single or dual polarization

systems, fully polarimetric SAR systems have a better control on improving the contrast of the radar image. Spaceborne Imaging Radar - C (SIR-C) provides increased capability over SEASAT, SIR-A, and SIR-B by acquiring digital images simultaneously at two microwave wavelengths (I): L- band (h = 23.5 cm) and C- band (h = 5.8 cm). The vertically and horizontally polarized transmitted waves can be received on two separate channels, to give images of the magnitude of

Figure 1. oil spills along with the ship

Processed SIR-C imagery showing a typical example of

radar backscatter for four polarization combinations: HH (Horizontally-transmitted, Horizontally-received), VV (Vertically-transmitted, Vertically-received), HV, and VH; and also data on the relative phase difference between the HH, VV, VH, and HV returns. This allows derivation of the complete scattering matrix of a scene on a pixel by pixel basis. From this scattering matrix, every polarization configuration (linear, circular or elliptical) can be generated during ground processing.

B. The data obtained from the SIR - C are specially coded

and compressed data set. For each pixel in a SIR-C image, we obtain a measurement of the "multi-look complex" (MLC) "scattering matrix" S, which represents the reflectivity of the area being observed at a given radar wavelength. Each of the four complex elements of this matrix is the amplitude and phase of the backscattered radiation as measured at one of four orthogonal transmit / receive polarizations: Horizontal transmit, Horizontal receive (HH); Horizontal transmit, Vertical receive (HV); Vertical transmit, Horizontal receive (VH); and Vertical transmit, Vertical receive (VV). These are denoted: Shh, Svh, Shv and Svv respectively. (Sometimes, only two -

polarizations are acquired, and this is known as the dual- pol mode. Occasionally, only one transmitheceive polarization measurement is made, and this is known as the single-pol mode. When all four transmitlreceiye polarizations are acquired, this is known as the quad-pol mode. The particular mode selected for any given data is determined by the scientific requirements at that site).

SIR - C digital data set

0-7803-827 1-4/04/$20.00 02004 IEEE 51 1

C. After appropriate signal processing we get the final

intensity image from the raw data obtained from the satellite system. This image (called as digital image) depicts the target phenomenon through the contour and varying brightness depending on the backscattering from the particular object. As a thumb rule, greater the difference in intensity of backscattering from the target and the surrounding, greater will be our understanding over the image. But when we analysis the image from the ocean area containing the oil spills this is not an easy task. The quality of image obtained strongly depends on various factors namely ocean surface conditions, ocean currents, wind speed, presence of oil look-alikes (algae, ocean surface currents etc) and also on the radar transmit and receive polarization and wavelength. There may be situations where we confirm the presence of oil spills in some areas by looking at the final image but in reality there may not be any oil spills but they may be some ocean currents or oil look-alikes and vice versa.

The general oil spill classification problem

D. Generic Contrast Enhancement of radar imagery The basic idea of contrast enhancement for single-pol

radar systems was introduced in many applications. But the scope of applying these generic schemes for the classification of oil spills is quite doubtful due to the aforesaid reasons. Hence the task is very specialized and in fact the most extreme case of detection. The algorithm which we implemented is built upon the physics of radar polarimetry. Consider any transmitted signal represented by T having both horizontal and vertical components as Th and Tv respectively. The Jones vector representation of the transmitted radar signal is given as follows.

Th * exp(i@) E] = [ Tv * exp(i p)] L J

The Jones vector representation of the transmitted wave gives basic information about the incident wave with the angles cph and cpv being the inclination of the transmitted wave to its horizontal and vertical axes respectively [2]. The difference angle is cp (where cp = cpv - cph) which gives an idea about the ellipticity x and orientation U, of the transmitted wave. By varying the two angles (cph and cpv) and one of the magnitudes (Th or Tv) we can generate any arbitrary wave polarization for transmission which follows the constraint given below [2,3,4].

This can be represented graphically using various locations on a Poincare's sphere [2]. Our main aim in running this algorithm is to find the best possible transmit and receive polarizations for the task of identifying the oil spills under worst conditions with best contrast difference between the oil spilled and clean sea areas. There are two bands of analysis possible namely the C and L band. Generally the C-band analysis gives a good result but this is not a thumb rule as the conditions of ocean are highly variable.

E. Optimized SOC Algorithm The Sea-Oil-Classifier (SOC) algorithm is an

extension of the techniques proposed in some research works on radar polarimetry [2 , 3, 41. Consider the two sets of covariance matrix [2] represented by C = Coil for oil (target) and G = C,,, for sea area (clutter). So the SOC is the ratio of the received power of the two classes of objects, which is given by

Backscatter power from Sea area SOC = (3)

Backscatter power from Oil area In mathematical terms it can be represented as

sea

(4) soc= OfiOl/ (Wr 3 X r 9 W , 9 x 1 )

o r t (Vr,Xr,Vt>Xt) Hence to maximize the contrast we have to look for the maximum SOC possible under certain constraints which are discussed below. The above problem can be formulated as a standard eigen-value problem as follows. The polarimetric feature vector X is expressed in a horizontal-vertical polarization basis as

x=[;] ( 5 )

In the task of finding the best SOC, we indeed fmd the best weighting function W which gives best possible contrast difference between the oil and the sea area. W is a function of the four angles given by

where wT is the transpose of W and (yr , E , u,~, K) gives the orientation and ellipticity angles of receive and transmit polarizations respectively. Hence task of finding the maximum SOC can be rewritten in a different way as

SOC = W'CseaW w 'Coil w

As [3] we can imagine of the inverse of SOC which gives the sea-to-oil received backscatter power ratio as

The best SOC classifier will be defined as shown below soc = max{max SOC(W>, min OSC(W>} (8) Now we can represent W as a fimction of our transmitted and received radar signal as

(ThR h ) * 1 W = I (ThTv)* + (TvRh)* I ( 9 4

5 12

L

where TC and RC are the transmit and receive classifiers. These two parameters control the transmit and receive polarizations. Hence for a given polarimetric radar system there exists only one best possible combination of the four angles ( w r , K , yt , E ) which gives the best contrast for detection. Hence by varying the four angles we can analyze the various polarization's impact for the best detection possibility in a filly polarimetric system using (8). At frst we vary the ph and pv values along with the Th with the constraint given in (2). Each combination of ph, pv and Th defines a particular polarization ( T o . For each combination of TC we find the best possible RC using generic eigen-value analysis as below. From (9b), we can rewrite (7a) as

Polarization

RCT *C., * RC RCT *Czl * RC

' soc=

Sea Oil SOC Spill (dB)

0 where Corllsea = TCT * Corllsea * TC

R T (dB)

Hence we can modify the problem into a generalized eigen-value problem as

C,,, * RC = SOC * Cor, * RC (1 1) Hence by solving the generic eigen-value problem (1 1) we get the eigen values given by SOC and the corresponding eigen vectors given by RC [2, 31. From (8) we get the optimized SOC. The eigen vector corresponding to optimized SOC is the optimized RC. We have used the above analysis for the fully polarimetric SIR-C data sets.

(&)

0 50 1M 150 200 No 300 350 400 450 500 Pixel CO Jrnn n-mtar

-25 I

L

R

Figure 2. Received power for each column pixel in dB along the row pixel 235. I t is showing a dip in the received power for those pixels corresponding to the oil spills in the original power image for the North Sea data set. We can see the SOC - Optimized Polarization shows better difference in oil to clean sea contrast.

L -23.7 -26.6 2.9

L I -16.8 -21.1 4.3

F. Effect of Local Characteristics of Sea Surface on the Optimisation Technique The SOC - Optimized detection algorithm tries to

extend the tolerance and validity ranges of standard

R

TABLE I . EXPERIMENTAL RESULTS -NORTH SEA C-BAND DATA SET

R -23.8 -27.8 4.0 SOC-ODtirnized

41 -19.1 -23.5 4.4

I

Oil

(dB)

Sea Polarization

R T (dB)

SOC (dB)

(yr, y, ) = (-j4.35., -34.99) 1 Ir (-43'2:121'72) -16.7 1 -21.4 1 4.7 1 ( y r , , ~ )=(44.21 ,21.72)

(40.11 ,21.52) an les in de ree

H

H

TABLE 11. EXPERIMENTAL RESULTS - ENGLISH CHANNEL L -BAND DATA SET

H -17.9 -20.0 2.1

V -34.8 I -35.9 1.1

V V -24.7 -27.0 2.3

R

R

513

L -20.9 -23.2 2.3

R -27.8 -28.6 0.8

or (13.10,22.63) angles in degree

single-pol and dual-pol missions. But in this process, the algorithm inherently assumes that the variation of local characteristics of sea surface (surface wind speed, etc) are more uniform inside the processing window of the algorithm. We also assume that the averaged value of covariance matrix inside the processing window as a representative covariance matrix of each pixel inside this window. Hence, the method is strongly dependent on the local sea surface variations.

G. Results and Discussions The optimization technique was applied to the North

Sea and English Channel data. The results of standard polarization were compared with the optimized polarization. The received power (represented in dB scale) for the oil spilled and clean sea areas for a sample row in the total power image of the North Sea (C-Band) is shown in Fig. 2. The results for the North Sea (C-Band) HV polarization and optimized polarization are shown in Fig. 3 and Fig. 4 respectively. The results of the English Channel (L-Band) for HH polarization and optimized

polarization respectively are shown in Fig. 5 and Fig. 6. Table I and Table I1 give the results for the North Sea and English Channel respectively. The optimized angles are given in degree. Note that in Table I & 11, the contrast difference (in dB) between the oil spilled and clean sea areas for the SOC-Optimized polarization is slightly higher than in VV and RL polarization. Also to highlight, RL and VV polarizations can be used particularly for oil spill detection which have comparable results with that of the SOC - Optimized polarizations.

0

50

100

150

200

250

300

350

400

4%

-5

.IO

zoom,channell cdp (Power for HH)

Figure 5. English Channel area in L-Band and HH polarization

Received power image in dB showing oil spills in

Figure 3. North Sea in C-Band and HV polarization

Received power image in dB showing oil spills in

Figure 6. Received power image in dB showing oil spills in English Channel area in L-Band and SOC - Optimized polarization

Figure 4. North Sea area in C-Band and SOC - Optimized Polarization

Received power image in dB showing oil spills in

H. Future Developments and Research Presently various other SIR-C images are being

analyzed to better the optimization scheme. We are also working on novel algorithms to get the optimal polarizations with less computational effort. Application of neural networks and genetic algorithms for an effective detection and classification are the areas of future interest. Regular monitoring and automatic analysis of SAR images for effective real time surveillance is the future developmental goal in illicit oil spill monitoring.

REFERENCES

A. A. Swartz, H. A. Yeuh, J. A. Kong, L. M. Novak and R. T. Shin, “Optimal polarizations for achieving maximum contrast in radar images,” J. Geophys. Res, vol. 93, pp. 15252-15260, December 1988. J. J. Van Zyl, H. A. Zebker and C. Elachi, Radar Polarimetry for Geoscience Applications, Artech House, 1990, pp.324-338. J. Yang, Y. Yamaguchi, W. M. Boerner and S. Lin, “Numerical Methods for Solving the Optimal Problem of Contrast Enhancement”, IEEE Trans. Geosc. and Remote Sensing, vol. 38 no. 2, March 2002, pp. 965-971. L. M. Novak, M. C. Burl, W. W. Irving, “Optimal Polarimetric Processing for Enhanced Target Detection”, IEEE Trans. Aero. andElec. Sys, vol. 29, no. 1, January 1993,pp. 234-244. S. R. Cloude and E. Pottier, “An Entropy Based Classification Scheme for Land Applications of Polarimetric S A R , IEEE Trans. Geosc. andRemole Sensing, vol. 35, no.1, January 1997,pp. 68-78. V. Wismann, “Radar signatures of mineral oil spills measured by an airborne multi-frequency radar and the ERS-I SAR’, IGARSS’93 symposium, vol. 3, pp. 940-942. H. Espedal, “Detection of oil spill and natural film in the marine environment by spaceborne S A R , IGARSS’99 proceedings, vol. 3, pp. 1478-1480.

5 14

IEEE MELECON 2004, May 12-15,2004, Dubrovnik, Croatia

Modeling and Performance Analysis of IP Access Interface in Optical Packet Switched Networks

M. Lackovic, C. Bungarzeanu Telecommunication Laboratory

EPFL, CH-1015 Lausanne, Suisse -' (marko.lackovic; cristian.bungarzeanu) @epfl.ch

Abstract-The article analyzes the influence of the IP access interface on the packet loss probability and delay times in the optical packet switched network. The network and node model have been proposed, and the assembler and hold features have been included in the analysis. It has been shown that the increase of the maximum optical packet sizes, as well as holding feature as contention resolution mechanism, decrease the packet loss ratio, but introduce delays at the optical network access points. The modeling and analysis were based on discrete event simulation assuming self-similar traffic sources.

I. INTRODUCTION The WDM boosted optical fiber transmission by

solving the ever increasing capacity demands, but optical transmission networks made a small step from the first point-to-point systems in terms of reconfigurability and manageability. Photonic layer still serves as a sole transmission medium controlled by the inefficient and unscalaeable protocol stack. The routing, protection and restoration and traffic/QoS management is spread across several stack layers resulting in overlapping protocol functionalities and waste of resources due to large overhead.

In the context of decreasing operators' revenues due to cheap pPice per best-effort bit, and increasing user demands for quality of service guaranties along with capacity, the idea of intelligent optical layer started to emerge. The user stack should be consolidated to provide a simple and robust communication between the photonic layer and the IP as its primary client. Thus, the intermediate protocol functionalities can migrate either to upper IP layer, or to lower photonic layer. The resulting paradigm, where IP entities communicate directly using photonic layer as dynamical service provider for packet based communication, is often denoted as the optical packet switching (OPS) [l]. The interface between the IP and the photonic layer is thus a main issue in designing the future optical network. This includes the transformation of ingress IP packets in optical packets, as well as contention resolution schemes in the IP/photonic interfaces.

11. NODE AND NETWORK ARCHITECTURES

The aim was to create an optical packet switching model able to support the network planning procedure and the network performance analysis. The model should thus address the optical packet switching as well as the functionalities of network clients, primarily IP. Node and

network models were developed using the object-oriented paradigm of the Cosmos tool [2].

Network is structured using two node types namely IPPoP node (ZP Point of Presence) and WDMPoP node (WDM Point of Presence) as shown in figure 1. IPPoP and WDMPoP nodes are generic nodes with optical network (WDMPoP) and IP network (IPPoP) functionality.

Figure 1. Optical transmission network structure

A. IPPoP An IPPoP node implements the basic IP layer

functionality needed to transmit IP packets and thus enable the network performance analysis. The basic functionality includes initial definitions of traffic demands between each pair of nodes, which serve as the input data for the dimensioning procedure, accompanied by the complex traffic sources in the simulation phase.

IPPoP nodes represent communicating IP entities, such as traffic generators and IP routers, as clients of the optical layer. These clients are characterized in this work by the capacity required from the optical network. Thus an IPPoP node can represent a whole range of possible IP entities from one computer to a whole metropolitan network comprising large number of local area networks and users. Figure 2 clarifies the case where IPPoP node substitutes LANIMAN networks (e.g. a whole city).

The IPPoP node dimensioning is determined by the traffic demand capacity between the selected node (node A ) , and all other IPPoP nodes in the network (node B).

Optical &WO*

n

Figure 2. IPPoP node functionality

0-7803-8271-4/04/$20.00 02004 IEEE 515

Figure 2 illustrates the IPPoP node structure. It includes sources and destinations of IP traffic (denoted as IPSD) along with the IP router. Each IPSD module models one communication between this and some other IPPoP node. IPSD modules are connected to the optical network via routers using line cards (IPLC), whose structure is shown in Figure 3. They represent an electrical-optical interface converting the electrical (IP) signal to optical form. In this model router acts as a passive concentration point implementing an interface between IPSDs and WDMPoPs.

optical interface (POS)

Figure 3. Line card model

Line cards in the IPSDs and routers have PoS (Packet over Sonet) interfaces, which are currently a dominating solution for broadband optical communication. The number of IPSDs is determined by the demand capacity that has to be supported and the PoS interface capacity. The number of PoS line cards in the node i is equal to

T ( i , j ) Np<,S ( 9 = C N,, (i, j ) , NsD(i, j ) = -

OL,<,L.J#A i c p . s l * ( l ) where T(i, j ) denotes a capacity of traffic demands between IPPoP nodes i and j , Cpos denotes a PoS interface capacity and n stand for a number of IPPoP nodes. NsD ( i , j ) is the number of IPSD modules supporting the node i to j communication. Symmetrical communication capacities between all node pairs are assumed as a consequence of equal transmitter and receiver line card capacities.

B. WDMPoP

5 , primarily performs optical packet switching. WDMPoP node, whose architecture is shown in figure

wavelength amplifier demultiplexer c y e r t e r mult,;plexer h[Fkr

\ transponder

Figure 4. General WDMPoP node structure

The input traffic is the WDM multiplexed signal from the other WDMPoP node, or a baseband Sonet/SDH framed signal from IPPoP (add flows) as discussed in the previous chapter. Each WDMPoP node thus serves as an edge or core optical packet switch as it has ingress traffic flows from both core (multiplexed signal) and access (baseband signal) networks. The input multiplexed signals are demultiplexed and analyzed on the packet level in the control logic part. The header data along with the buffer state and routing table information (figure 6) determines the configuration of the coding and switching section.

Figure 5 . logic Figure 6 . Traffic aggregation in router structure

The coding section comprises tunable wavelength converters controlled by the control logic module. The output wavelength depends on the egress packet scheduling. This includes the choice of the output port determined by the routing table and the state of the port or its occupancy. If all the channels (wavelengths) on the output port are occupied, the packet has to be buffered or deflected as a part of the contention resolution scheme [3]. The choice of the wavelength and fiber delay line depends on the implemented buffer mechanism and/or QoS support.

III. IPPOP - WDMPOP INTERFACE

A baseband optical fiber, with Sonet/SDH framing, serves as a connection between an IPPoP and a WDMPoP node. This enables the same type of interface for communication between IPSDs (e.g. LAN egress switch) and router, as well as between router and WDMPoP.

The PoS interfaced cards for the IPPoP to WDMPoP communication is denoted as PoSW. The number of PoSW card number depends on the total egress traffic from one IPPoP node as one assigned WDMPoP per IPPoP is assumed. For node A this equals to (symmetric communication assumption)

Npc,sw (A) = o i J < n ~ J t A Zcp:;Apj)l> (2)

where T(A, j ) denotes the communication capacity between IPPoP nodes A and j , and Cposw the capacity of a PoS interfaced IPPoP to a WDMPoP line card.

The egress traffic from IPSDs to WDMPoP is aggregated on IPPoP to WDMPoP links as shown in figure 7. The aggregation is done in the electrical domain on the packet level after the ingress OE conversion and before the egress EO conversion.

I PpOP WDMPoP

Figure 7. IPPoP/WDMPoP communication

A packet arrives from the PoS interfaced card to the forward engine module (denoted as FE) which forwards it to a correct backplane input and eventually to one of the egress FIFO organized electrical buffers. Each buffer corresponds to one PoSW card, or one baseband IPPoP - WDMPoP connection. The forwarding choice is based on

516

the buffer states which are examined on each packet arrival. The buffer with the shortest queue (smallest number of buffered packets) is chosen. This algorithm assures that the input traffic do not overflow the IPPoP- WDMPoP connections, and balances the load among all access links. Figure 8 depicts the interconnection of described modules in the router module.

IV. ASSEMBLER A sort of aggregation mechanism is needed to store and

classify input IP packets in order to form payloads for optical packets. Beside the payload formation, the aggregation also shapes input traffic and eliminates high variability introduced by the self-similarity [4]. Optical burst switching (OBS) assumes the input aggregation to form bursts [5] which are, in general, longer than optical packets. The difference between the OPS and OBS thus arises more in the core than in access nodes of the optical network as the OBS network requires preconfiguration of core switches by a signaling protocol.

The assembler module implements the aggregation along with the contention resolution mechanism based on delaying locally generated packets [6]. General assembler structure is shown in figure 8. It includes line cards, aggregator, deaggregator and holder module.

POS %OW

....... ~ ......,

- h.- optical electrical I optical

Figure 8. Assembler structure

The communication with the IPPoP node is done using the same number of PoSW line cards as in the IPPoP node. The cards supporting the communication between assembler and the other parts of the WDMPoP node do not utilize the PoS interfaces, but the packet over wavelength (POW) interfaces. From the broader functional perspective on a WDMPoP level, assembler performs the conversion. between input PoS optical signal (used in optical circuit switched network) to packet over wavelength optical signal (used in optical packet switched networks). The POW interfaces do not utilize the layer 2 protocol (like SoneVSDH), but just a simple framing protocol used for synchronization and packet delimitation.

in

I destination # add port #

buffers Athens

Figure 9. Aggregator and holder Figure 10. Analyzed network

Figure 9 depicts a simplified aggregator and holder scheme. Aggregator stores ingress IP packets to collect

enough data for the optical packet payload. The input packets are sorted according to their final destination as one optical packet can transport only several IP packets with the same destination (IPPoP node). This is implemented in the sub-module denoted as sorter.

The classification and buffering is done in the electrical domain as input packets arrive after OE conversion and deframing in PoSW cards. The FIFO organized buffers are time and length constrained. Data is collected until some maximum data size for one optical packets is acquired (MPS - Maximum Packet Size) or until the buffered data reaches some upper time. In that case the output packet load is shorter than the MPS.

Aggregated data serves as the payload for optical packet which is created in POW interfaced cards. Each card is connected to a transponder which adjusts the power level and assigns a correct wavelength. The wavelength is determined by the state of the exit port (channels on appropriate exit link). If all the channels are occupied, the wavelength determines the buffer state if the buffering is used as the contention resolution. To reduce the buffer occupancy in order to decrease the packet loss probability of transit packets, the optical packet creation can be delayed until one of the appropriate egress channels becomes available. The delay is optional and can be introduced in the holder module adjacent to the aggregator module. The principle of buffering is the same as in the aggregator, but the number of buffers differs, as it equals the number of egress links in the holder case. Ingress data is forwarded to appropriate buffer by the module denoted as fwd. The transparency of the network is preserved as the possible input delays in buffers occur while the aggregated loads for optical packet are still in the electrical domain.

V. RESULTS The analyzed network consists of 5 nodes which

corresponding to cities included in Cost 266 reference networks [7]. The traffic was generated using the enhanced population-distance model [7]. The year 2004 traffic volumes over 4 wavelength 40 Gbps channel system were assumed. Generated traffic demands were scaled in a way that the highest link load corresponds to the chosen network load.

The aim was to determine the influences of the aggregation parameters on the packet loss probability and the packet loss probability decrease due to the used holder mechanism. A discrete event simulation with self-similar traffic generator based on the fractional Gaussian noise was used [SI. The used packet length probability density function corresponds to the one described in [9] in order to imitate the Internet traffic.

Figure 12 shows the influence of the aggregation to the mean PLR of all network demands. The thick line depicts the packet loss probability without the input aggregation implying that the IP packets were packed into optical packets as they arrive. With the increase of MPS from 3000 to 6000 octets the packet loss probability decreases. The value of 500 octets is considered just for comparison purposes as it implies the fragmentation of input packets in most cases. This result can be easily justified by taking into account that with the increase of optical packet lengths the OPS actually becomes a hybrid of optical burst switching, and finally more similar to circuit switching

517

which can be approximated with infinite packet size. Moreover the variability of the input self-similar traffic is reduced by traffic shaping introduced by aggregation.

1.OEt00 - 5 1.OE-01 -

% 1.OE-02 -

g 1.OE-03 -

1.OE-04 -

3 1.OE-05 - U 2 1.OE-06

-

VI

- +

1.OEt00 - r .z 1.OE-01- I P g 1.OE-02 - UI 3 1.OE-03 -

0

5 1 OE-04 -

-

L

2

: / i 1 0 E 0 7 1 1 7 ,z’, I , I I I I I I

0 5 055 06 0 6 5 07 075 08 065 0 9 095 1

Network Load [erl]

Figure 1 1. Influence of ingress aggregation on packet loss probability

The second group of results focuses on holder influence on the packet loss probability and the delay the mechanism introduces. Figure 13 depicts packet loss probability with and without a holder. The used buffers in the holder were ideal without capacity limitation.

[Worst (no holder)[

1 OE-05 0 1 0 2 0 3 0 4 05 0 6 07 O B 09

Network load [erl]

Figure 12. Influence of input hold to packet loss probability

The results have a good match on lower network loads, but differ on the loads above 0.5 erl. The worst case without holder has the linear increase of the packet loss probability, while the worst case with holder has almost negligible increase on loads above 0.6 erl. This is the consequence of almost all add traffic being buffered as the buffers are limitless.

The buffering introduces a considerable delay as depicted in figure 13.

1 E105 - ]Delay (Unlimited Holder)

1 Et04 -

\Rejected Packets1 1 E+03 -

1 Et02 - /

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Network load [erl]

Figure 13. Normalized delay time and number of rejected packets dependence on holder structure and network load

A delay time normalized to the duration of the IP packet with mean load has been shown in the case of a holder with unlim.ited buffer size, as well as with limited buffer size to 1000 means IP packet lengths. These results are accompanied by the number of rejected packets due to buffer overflow in the second case. With the load increase

the delay increases in the ideal case due to more buffered packets. In the second case a limited holder can buffer almost all packets up to 0.5 er1 of load, while the rejection increases rapidly with larger loads.

VI. CONCLUSION The IP to photonic layer interface has to be carefully

constructed as it can reduce the packet loss probabilities in the packet switched optical network and increase the offered quality of service. The optical network model along with the IPPoP and WDMPoP network nodes has been proposed. As a part of node and network performance evaluation the influence of the interface part of the WDMPoP node on the packet loss ratio has been analyzed. The interface has been modeled as the assembler module performing input aggregation of the IP packets in order to form optical packets and the input delay of aggregated data as a form of contention resolution. Used network and traffic data were based on the Cost 266 project, while the traffic source followed the fractional Gaussian noise model.

The results show that the increase of the maximum optical packet size and the input delay can reduce the packet loss ratio. These results are not to be taken as a general conclusion as they reflect the influence of those two mechanisms alone on the OPS based network performance. The increase of the packet size is limited by the switch performance and the delay line lengths in the buffers, while the holder introduces input delay which can be unacceptable for certain requests with quality of service guarantees.

ACKNOWLEDGMENT This work has been conducted with the support of the

Swiss Federal Office for Education and Science within the European Action Cost 266.

REFERENCES S. Yao, B. Mukhejer, and S. Dixit, “Advances in Photonic Packet Switching: An Overview”, IEEE Comm. Mug., Feb. 2000, pp. 84- 94. M. Lackovic and R. Inkret, “Network Design, Optimization and Simulation Tool Cosmos”, In Proc. of WAON (Zagreb, Croatia, June 13-14,2001), pp. 37-44. S. Yao , B. Mukhejee, S. J. B. Yoo, and S. Dixit, “A unified study of contention-resolution schemes in optical packet-switched networks,” IEEE J. of Lighrwuve Tech., Mar., 2003., pp. 612 - 683 F. Xue, S. Yao, B. Mukhejee and S. J. B. Yoo, “The performance improvement in optical packet-switched networks by traffic shaping of self-similar traffic”, In Proc. of OFC 2002, Mar. 2002. Ge, A.: Callegati, F.: Tamil, L.S., “On optical burst switching and self-similar traffic”, IEEE Comm. Letters, 4(3) , Mar. 2000, pp. 98 - 100 M. J. 0’ Mahony, D. Simeonidou, D. K. Hunter, A. Tzanakaki, “The application of optical packet switching in future communication networks”, IEEE Comm. Mug., Mar. 2001, pp.

R. Inkret, A. Kuchar and B. Mikac (editors), “Advanced Infrastructure for Photonic Networks”, extended final report of Cost 266 action, FER, Zagreb, Croatia, 2003. M. Lackovic, B. Mikac, V. Sinkovic, “Network Performance Evaluation by Means of Self Similar Traffic Model”, In Proc. of Mipro (Opatija, Croatia, May 19-23,2003), pp. 82-87. http://advanced.comms.agilent.com/insight/2OOl- OS/TestingTips/lMxdPktSzThroughput.pd

128-1 35.

518