660 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 1, NO.

4, OCTOBER 2002

An Application-Specific Protocol Architecture for

Wireless Microsensor Networks
Wendi B. Heinzelman, Member, IEEE, Anantha P. Chandrakasan, Senior Member, IEEE, and
Hari Balakrishnan, Member, IEEE

Abstract—Networking together hundreds or thousands of cheap In order to design good protocols for wireless microsensor
microsensor nodes allows users to accurately monitor a remote en- networks, it is important to understand the parameters that are
vironment by intelligently combining the data from the individual relevant to the sensor applications. While there are many ways
nodes. These networks require robust wireless communication pro-
tocols that are energy efficient and provide low latency. In this in which the properties of a sensor network protocol can be eval-
paper, we develop and analyze low-energy adaptive clustering hier- uated, we use the following metrics.
archy (LEACH), a protocol architecture for microsensor networks
that combines the ideas of energy-efficient cluster-based routing A. Ease of Deployment
and media access together with application-specific data aggrega-
tion to achieve good performance in terms of system lifetime, la- Sensor networks may contain hundreds or thousands of
tency, and application-perceived quality. LEACH includes a new, nodes, and they may need to be deployed in remote or dan-
distributed cluster formation technique that enables self-organiza- gerous environments, allowing users to extract information
tion of large numbers of nodes, algorithms for adapting clusters in ways that would not have been possible otherwise. This
and rotating cluster head positions to evenly distribute the energy requires that nodes be able to communicate with each other
load among all the nodes, and techniques to enable distributed
signal processing to save communication resources. Our results even in the absence of an established network infrastructure
show that LEACH can improve system lifetime by an order of mag- and predefined node locations.
nitude compared with general-purpose multihop approaches.
Index Terms—Data aggregation, protocol architecture, wireless
B. System Lifetime
microsensor networks. These networks should function for as long as possible. It may
be inconvenient or impossible to recharge node batteries. There-
I. INTRODUCTION fore, all aspects of the node, from the hardware to the protocols,
must be designed to be extremely energy efficient.

A DVANCES iN sensor technology, low-power electronics,

and low-power radio frequency (RF) design have enabled
the development of small, relatively inexpensive and low-power
C. Latency
Data from sensor networks are typically time sensitive, so it
sensors, called microsensors, that can be connected via a wire- is important to receive the data in a timely manner.
less network. These wireless microsensor networks represent a
new paradigm for extracting data from the environment and en- D. Quality
able the reliable monitoring of a variety of environments for ap-
The notion of “quality” in a microsensor network is very
plications that include surveillance, machine failure diagnosis,
different than in traditional wireless data networks. For sensor
and chemical/biological detection. An important challenge in
networks, the end user does not require all the data in the
the design of these networks is that two key resources—com-
network because 1) the data from neighboring nodes are highly
munication bandwidth and energy—are significantly more lim-
correlated, making the data redundant and 2) the end user
ited than in a tethered network environment. These constraints
cares about a higher-level description of events occurring in
require innovative design techniques to use the available band-
the environment being monitored. The quality of the network
width and energy efficiently.
is, therefore, based on the quality of the aggregate data set,
so protocols should be designed to optimize for the unique,
Manuscript received January 9, 2001; revised July 1, 2001 and August application- specific quality of a sensor network.
24, 2001; accepted August 24, 2001. The editor coordinating the review
of this paper and approving it for publication is M. Zorzi. The work of This paper builds on the work described in [11] by giving
W. B. Heinzelman was supported by a Kodak Fellowship. This work was a detailed description and analysis of low-energy adaptive
supported in part by the Defense Advanced Research Project Agency (DARPA) clustering hierarchy (LEACH), an application-specific protocol
Power Aware Computing/Communication Program and the U. S. Air Force
Research Laboratory, Air Force Materiel Command, under Agreement architecture for wireless microsensor networks. LEACH
F30602-00-2-0551. employs the following techniques to achieve the design goals
W. B. Heinzelman was with the Massachusetts Institute of Technology, Cam- stated: 1) randomized, adaptive, self-configuring cluster for-
bridge, MA 02139 USA. She is now with the Department of Electrical and Com-
puter Engineering, University of Rochester, Rochester, NY 14627-0126 USA mation; 2) localized control for data transfers; 3) low-energy
(e-mail: [email protected]). media access control (MAC); and 4) application-specific data
A. P. Chandrakasan and H. Balakrishnan are with the Massachusetts Insti- processing, such as data aggregation or compression. Simula-
tute of Technology, Cambridge, MA 02139 USA (e-mail: [email protected];
[email protected]). tion results show that LEACH is able to achieve the desired
Digital Object Identifier 10.1109/TWC.2002.804190 properties of sensor networks.
1536-1276/02$17.00 © 2002 IEEE
Authorized licensed use limited to: Indian Institute of Technology (Ropar). Downloaded on September 15,2024 at 18:16:55 UTC from IEEE Xplore. Restrictions apply.
HEINZELMAN et al.: AN APPLICATION-SPECIFIC PROTOCOL ARCHITECTURE FOR WIRELESS MICROSENSOR NETWORKS 661

II. BACKGROUND not suited for microsensor networks. LEACH builds on this
work by creating a new ad-hoc cluster formation algorithm that
Since both device and battery technology have only recently
better suits microsensor network applications.
matured to the point where microsensor nodes are feasible, this
is a fairly new field of study. Researchers have begun discussing
not only the uses and challenges facing sensor networks [2], III. LEACH PROTOCOL ARCHITECTURE
[7], [20], but have also been developing preliminary ideas as to To meet the unique requirements of wireless microsensor
how these networks should function [4], [5], [13] as well as the networks, we developed LEACH, an application-specific
appropriate low-energy architecture for the sensor nodes them- protocol architecture [10], [11]. The application that typical
selves [6], [21]. microsensor networks support is the monitoring of a remote
There have been some application-specific protocols devel- environment. Since individual nodes’ data are often correlated
oped for microsensor networks. Clare et al. developed a time- in a microsensor network, the end user does not require all
divison multiple-access (TDMA) MAC protocol for low-energy the (redundant) data; rather, the end user needs a high-level
operation [5]. Using a TDMA approach saves energy by al- function of the data that describes the events occurring in the
lowing the nodes to remain in the sleep state, with radios pow- environment. Because the correlation is strongest between data
ered-down, for a long time. Intanagonwiwat et al. developed di- signals from nodes located close to each other, we chose to
rected diffusion, a protocol that employs a data-driven model to use a clustering infrastructure as the basis for LEACH. This
achieve low-energy routing [13]. allows all data from nodes within the cluster to be processed
Recently, there has been much work on “power-aware” locally, reducing the data set that needs to be transmitted to
routing protocols for wireless networks [19], [25]. In these the end user. In particular, data aggregation techniques can be
protocols, optimal routes are chosen based on the energy at used to combine several correlated data signals into a smaller
each node along the route. Routes that are longer, but which set of information that maintains the effective data (i.e., the
use nodes with more energy than the nodes along the shorter information content) of the original signals [9]. Therefore,
routes, are favored, helping avoid “hot spots” in the network. much less actual data needs to be transmitted from the cluster
In LEACH, we use randomized rotation of the cluster head to the base station (BS).
positions to achieve the same goal. For the development of LEACH, we made some assumptions
One method of choosing routes is to use “minimum transmis- about the sensor nodes and the underlying network model. For
sion energy” (MTE) routing [8], [24], where intermediate nodes the sensor nodes, we assume that all nodes can transmit with
are chosen such that the sum of squared distances (and, hence, enough power to reach the BS if needed, that the nodes can
the total transmit energy , assuming a power loss) is use power control to vary the amount of transmit power, and
minimized. Thus, for three nodes A, B, and C, node A would that each node has the computational power to support different
MAC protocols and perform signal processing functions. These
transmit to node C through node B if and only if
assumptions are reasonable due to technological advances in
(1) radio hardware and low-power computing. For the network, we
or . This approach ignores the energy dis- use a model where nodes always have data to send to the end
sipated in the radio to send and receive the data and, therefore, user and nodes located close to each other have correlated data.
may not actually produce the lowest energy routes. Although LEACH is optimized for this situation, it will continue
Another method of wireless communication is to use clus- to work if it were not true. In Section V, we discuss ways in
tering. In this case, nodes send their data to a central cluster which LEACH may be improved when these assumptions do
not hold.
head that forwards the data to get it closer to the desired recip-
In LEACH, the nodes organize themselves into local clusters,
ient. Clustering enables bandwidth reuse and can, thus, increase
with one node acting as the cluster head. All non-cluster head
system capacity. Using clustering enables better resource allo-
nodes transmit their data to the cluster head, while the cluster
cation and helps improve power control [14].
head node receives data from all the cluster members, performs
While conventional cluster-based networks rely on a fixed
signal processing functions on the data (e.g., data aggregation),
infrastructure, new research is focusing on ways to deploy clus- and transmits data to the remote BS. Therefore, being a cluster
tering architectures in an ad-hoc fashion [3], [15], [23]. Early head node is much more energy intensive than being a non-
work by Baker et al. developed a linked cluster architecture, cluster head node. If the cluster heads were chosen a priori and
where nodes are assigned to be either ordinary nodes, cluster fixed throughout the system lifetime, these nodes would quickly
head nodes, or gateways between different clusters [3]. The use up their limited energy. Once the cluster head runs out of en-
cluster heads act as local control centers, whereas the gate- ergy, it is no longer operational, and all the nodes that belong to
ways act as the backbone network, transporting data between the cluster lose communication ability. Thus, LEACH incorpo-
clusters. This enables robust networking with point-to-point rates randomized rotation of the high-energy cluster head posi-
connectivity. Another ad-hoc clustering protocol, the near term tion among the sensors to avoid draining the battery of any one
digital radio (NTDR), uses a clustering approach with a two-tier sensor in the network. In this way, the energy load of being a
hierarchical routing algorithm [23]. Nodes form local clusters, cluster head is evenly distributed among the nodes.
and intra-cluster data are sent directly from one node to the The operation of LEACH is divided into rounds. Each round
next, whereas inter-cluster data are routed through the cluster begins with a set-up phase when the clusters are organized, fol-
head nodes. This protocol enables point-to-point connectivity lowed by a steady-state phase when data are transfered from
and does not use low-energy routing or MAC; therefore, it is the nodes to the cluster head and on to the BS, as shown in
Authorized licensed use limited to: Indian Institute of Technology (Ropar). Downloaded on September 15,2024 at 18:16:55 UTC from IEEE Xplore. Restrictions apply.
662 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 1, NO. 4, OCTOBER 2002

This ensures that the energy at all nodes are approximately equal
to each other after every rounds. Using (3) and (4), the
expected number of cluster heads per round is
Fig. 1. Time line showing LEACH operation. Adaptive clusters are formed
during the set-up phase and data transfers occur during the steady-state phase.

Fig. 1. The following sections describe the cluster head selection

and distributed cluster formation algorithms and the steady-state
operation of LEACH.
(5)

A. Cluster Head Selection Algorithms In Section IV-B, we analytically determine the optimal
based on our energy dissipation models for computation and
LEACH forms clusters by using a distributed algorithm, communication.
where nodes make autonomous decisions without any central- This choice of probability for becoming a cluster head is
ized control. Our goal is to design a cluster formation algorithm based on the assumption that all nodes start with an equal
such that there are a certain number of clusters, , during each amount of energy, and that all nodes have data to send during
round. In addition, if nodes begin with equal energy, our goal is each frame. If nodes have different amounts of energy (or an
to try to evenly distribute the energy load among all the nodes event-driven model is used, whereby nodes only send data
in the network so that there are no overly-utilized nodes that when some event occurs in the environment), the nodes with
will run out of energy before the others. As being a cluster head more energy should be cluster heads more often than the nodes
node is much more energy intensive than being a non-cluster with less energy, to ensure that all nodes die at approximately
head node, this requires that each node take its turn as cluster the same time. This can be achieved by setting the probability
head. of becoming a cluster head as a function of a node’s energy
Each sensor elects itself to be a cluster head at the beginning level relative to the aggregate energy remaining in the network,
of round (which starts at time ) with probability . rather than purely as a function of the number of times the node
is chosen such that the expected number of cluster head has been cluster head, Thus
nodes for this round is . Thus, if there are nodes in the net-
work (6)

where is the current energy of node and

(2)

(7)
Ensuring that all nodes are cluster heads the same number of
times requires each node to be a cluster head once in Using these probabilities, the nodes with higher energy are more
rounds on average. If is the indicator function determining likely to become cluster heads than nodes with less energy. The
whether or not node has been a cluster head in the most recent expected number of cluster head nodes is1
( ) rounds (i.e., if node has been a
cluster head and one otherwise), then each node should choose
to become a cluster head at round with probability
(8)
Equation (6) can be approximated by (3) when the nodes begin
(3)
with equal energy [10].
To use the probabilities in (6), each node must have an es-
Therefore, only nodes that have not already been cluster heads timate of the total energy of all nodes in the network. This re-
recently, and which presumably have more energy available than quires a routing protocol that allows each node to determine the
nodes that have recently performed this energy-intensive func- total energy, whereas the probabilities in (3) enable each node to
tion, may become cluster heads at round . make completely autonomous decisions. One approach to avoid
The expected number of nodes that have not been cluster this might be to approximate the aggregate node energy by mul-
heads in the first rounds is . After rounds, all tiplying the average energy of the nodes in each cluster by .
nodes are expected to have been cluster head once, following Note that to compute the probabilities in (3) and (6) requires
which they are all eligible to perform this task in the next se- that each node knows the parameters and . In this paper,
quence of rounds. Since is one if node is eligible to be a we assume these parameters are programmed into the nodes a
cluster head at time and zero otherwise, the term priori. However, this approach does not work well in dynamic
represents the total number of nodes that are eligible to be a networks. As we show in Section IV-B, the optimal number of
cluster head at time and clusters is a function of the number of nodes distributed
throughout an region of space. Therefore, the nodes
(4) 1Note that if any node i has E > (E =k ), which occurs with a small but
nonzero probability, the expected number of cluster heads will be less than k .
Authorized licensed use limited to: Indian Institute of Technology (Ropar). Downloaded on September 15,2024 at 18:16:55 UTC from IEEE Xplore. Restrictions apply.
HEINZELMAN et al.: AN APPLICATION-SPECIFIC PROTOCOL ARCHITECTURE FOR WIRELESS MICROSENSOR NETWORKS 663

Fig. 2. Flowchart of the distributed cluster formation algorithm for LEACH.

only need to determine assuming there is a predefined spec-

ification for . To do this, nodes can send “hello” messages
to all neighbors within a predetermined number of hops (to ap-
proximate ). Each node can count the number of “hello” mes-
sages it receives—this is that node’s estimate for . The desired
number of clusters can then be determined based on these pa-
rameters. This approach allows LEACH to adapt to changing
networks at the cost of increased overhead.
Fig. 3. Dynamic cluster formation during two different rounds of LEACH. All
B. Cluster Formation Algorithm nodes marked with a given symbol belong to the same cluster, and the cluster
head nodes are marked with .
Once the nodes have elected themselves to be cluster heads
using the probabilities in (3) or (6), the cluster head nodes must
MAC protocol. This message is again a short message, con-
let all the other nodes in the network know that they have chosen
this role for the current round. To do this, each cluster head node sisting of the node’s ID and the cluster head’s ID.
broadcasts an advertisement message (ADV) using a nonpersis- The cluster heads in LEACH act as local control centers to
coordinate the data transmissions in their cluster. The cluster
tent carrier-sense multiple access (CSMA) MAC protocol [18].
head node sets up a TDMA schedule and transmits this schedule
This message is a small message containing the node’s ID and a
header that distinguishes this message as an announcement mes- to the nodes in the cluster. This ensures that there are no colli-
sage. Each non-cluster head node determines its cluster for this sions among data messages and also allows the radio compo-
nents of each non-cluster head node to be turned off at all times
round by choosing the cluster head that requires the minimum
except during their transmit time, thus reducing the energy con-
communication energy, based on the received signal strength
of the advertisement from each cluster head. Assuming sym- sumed by the individual sensors. After the TDMA schedule is
metric propagation channels for pure signal strength, the cluster known by all nodes in the cluster, the set-up phase is complete
and the steady-state operation (data transmission) can begin.
head advertisement heard with the largest signal strength is the
A flowchart of this distributed cluster formation algorithm is
cluster head that requires the minimum amount of transmit en-
ergy to communicate with. Note that typically this will be the shown in Fig. 2. Fig. 3 shows an example of the clusters formed
cluster head closest to the sensor, unless there is an obstacle im- during two different rounds of LEACH.
peding communication. In the case of ties, a random cluster head
is chosen. C. Steady-State Phase
After each node has decided to which cluster it belongs, it The steady-state operation is broken into frames, where nodes
must inform the cluster head node that it will be a member of the send their data to the cluster head at most once per frame during
cluster. Each node transmits a join-request message (Join-REQ) their allocated transmission slot. The duration of each slot in
back to the chosen cluster head using a nonpersistent CSMA which a node transmits data is constant, so the time to send a
Authorized licensed use limited to: Indian Institute of Technology (Ropar). Downloaded on September 15,2024 at 18:16:55 UTC from IEEE Xplore. Restrictions apply.
664 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 1, NO. 4, OCTOBER 2002

Therefore, there will be few overlapping transmissions and

little spreading of the data is actually needed to ensure a low
probability of collision.
Data is sent from the cluster head nodes to the BS using a
Fig. 4. Time line showing LEACH operation. Data transmissions are explicitly
scheduled to avoid collisions and increase the amount of time each non-cluster fixed spreading code and CSMA. When a cluster head has data
head node can remain in the sleep state. to send (at the end of its frame), it must sense the channel to
see if anyone else is transmitting using the BS spreading code.
frame of data depends on the number of nodes in the cluster. If so, the cluster head waits to transmit the data. Otherwise, the
Fig. 4 shows the time line for one round of LEACH. We as- cluster head sends the data using the BS spreading code.
sume that the nodes are all time synchronized and start the set-up Other channelization techniques, such as having each cluster
phase at the same time. This could be achieved, for example, by use a different frequency band (e.g., FDMA), are possible. How-
having the BS send out synchronization pulses to the nodes. ever, since the number of clusters in LEACH is not fixed, using
To reduce energy dissipation, each non-cluster head node DSSS ensures that all nodes will receive better communication
uses power control to set the amount of transmit power based channels when there are fewer clusters. It is much harder to dy-
on the received strength of the cluster head advertisement.2 namically assign frequency bands so that all the bandwidth is
Furthermore, the radio of each non-cluster head node is turned utilized in a fixed channelization scheme. Of course, the draw-
off until its allocated transmission time. Since we optimize our back of using DSSS is the need for tight timing synchronization,
design for the situation when all the nodes have data to send to which may necessitate extra communication between the cluster
the cluster head, using a TDMA schedule is an efficient use of head and the non-cluster head nodes.
bandwidth and represents a low-latency and energy-efficient
approach. D. LEACH-C: BS Cluster Formation
The cluster head must be awake to receive all the data from the While there are advantages to using LEACHs distributed
nodes in the cluster. Once the cluster head receives all the data, cluster formation algorithm, this protocol offers no guarantee
it performs data aggregation to enhance the common signal and about the placement and/or number of cluster head nodes.
reduce the uncorrelated noise among the signals. In our anal- Since the clusters are adaptive, obtaining a poor clustering
ysis, we assume perfect correlation such that all individual sig- set-up during a given round will not greatly affect overall
nals can be combined into a single representative signal. The performance. However, using a central control algorithm to
resultant data are sent from the cluster head to the BS. Since the form the clusters may produce better clusters by dispersing
BS may be far away and the data messages are large, this is a the cluster head nodes throughout the network. This is the
high-energy transmission. basis for LEACH-centralized (LEACH-C), a protocol that uses
The preceding discussion describes communication within a centralized clustering algorithm and the same steady-state
a cluster, where the MAC and routing protocols are designed protocol as LEACH.
to ensure low energy dissipation in the nodes and no collisions During the set-up phase of LEACH-C, each node sends in-
of data messages within a cluster. However, radio is inherently formation about its current location (possibly determined using
a broadcast medium. As such, transmission in one cluster will a GPS receiver) and energy level to the BS. In addition to de-
affect (and often degrade) communication in a nearby cluster. termining good clusters, the BS needs to ensure that the energy
To reduce inter-cluster interference, each cluster in LEACH load is evenly distributed among all the nodes. To do this, the BS
communicates using direct-sequence spread spectrum (DSSS). computes the average node energy, and whichever nodes have
Each cluster uses a unique spreading code; all the nodes in energy below this average cannot be cluster heads for the current
the cluster transmit their data to the cluster head using this round. Using the remaining nodes as possible cluster heads, the
spreading code and the cluster head filters all received energy BS finds clusters using the simulated annealing algorithm [16]
using this spreading code. This is known as transmitter-based to solve the NP-hard problem of finding optimal clusters [1].
code assignment [12], since all transmitters within the cluster This algorithm attempts to minimize the amount of energy for
use the same code. The first cluster head to advertise its position the non-cluster head nodes to transmit their data to the cluster
is assigned the first code on a predefined list, the second cluster head, by minimizing the total sum of squared distances between
head to advertise its position is assigned the second code, and all the non-cluster head nodes and the closest cluster head.4
so on.3 With enough spreading, neighboring clusters’ radio Once the cluster heads and associated clusters are found, the
signals will be filtered out as noise during decorrelation and not BS broadcasts a message that contains the cluster head ID for
corrupt the transmission from nodes in the cluster. To reduce each node. If a node’s cluster head ID matches its own ID, the
the possibility of interfering with nearby clusters and reduce its node is a cluster head; otherwise, the node determines its TDMA
own energy dissipation, each node adjusts its transmit power. slot for data transmission and goes to sleep until it is time to
2To ensure connectivity in a dynamic environment, the node can either set its transmit data. The steady-state phase of LEACH-C is identical
transmit power slightly greater than the minimum needed to reach the cluster to that of LEACH.
head, or the cluster head can send short feedback messages to each of the nodes
telling them to increase or decrease their transmitted power, as is done in cellular 4Communication energy often does not scale exactly with distance. How-
systems. ever, gathering information about the communication channel between all nodes
3If there are more clusters than spreading codes, some clusters will use the is impractical. Using distance calculated from the nodes’ GPS coordinates is,
same code, possibly causing data collisions if the clusters are located close to therefore, an approximation to the energy that will be required for communica-
each other. tion.
Authorized licensed use limited to: Indian Institute of Technology (Ropar). Downloaded on September 15,2024 at 18:16:55 UTC from IEEE Xplore. Restrictions apply.
HEINZELMAN et al.: AN APPLICATION-SPECIFIC PROTOCOL ARCHITECTURE FOR WIRELESS MICROSENSOR NETWORKS 665

Fig. 5. Radio energy dissipation model.

IV. ANALYSIS AND SIMULATION OF LEACH A. Experiment Setup

For our experiments, we used a 100-node network where
For even moderately-sized networks with tens of nodes, it
nodes were randomly distributed between ( , ) and
is extremely difficult to analytically model the interactions
( , ) with the BS at location ( , ).
between all the nodes. Therefore, we used the network sim-
The bandwidth of the channel was set to 1 Mb/s, each data
ulator ns [17] to evaluate LEACH and compare it to other
message was 500 bytes long, and the packet header for each
protocols. We compare LEACH to LEACH-C, MTE routing,
type of packet was 25 bytes long.
and static clustering in terms of system lifetime, energy dis-
We assume a simple model for the radio hardware energy dis-
sipation, amount of data transfer, and latency. The code for
sipation where the transmitter dissipates energy to run the radio
our experiments can be found at http://www-mtl.mit.edu/re-
electronics and the power amplifier, and the receiver dissipates
search/icsystems/uamps/cadtools.
energy to run the radio electronics, as shown in Fig. 5. For the
For MTE routing, each node runs a start-up routine to deter-
experiments described here, both the free space ( power loss)
mine its next-hop neighbor, defined to be the closest node that
and the multipath fading ( power loss) channel models were
is in the direction of the BS. We assume that each node knows
used, depending on the distance between the transmitter and re-
the location of all nodes in the network, to simplify the set-up of
ceiver5 [22]. Power control can be used to invert this loss by ap-
MTE routes. In general, some sort of initialization phase would
propriately setting the power amplifier—if the distance is less
be needed where this information is disseminated through the
than a threshold , the free space (fs) model is used; otherwise,
network. Data packets are passed along via next-hop neigh-
the multipath (mp) model is used. Thus, to transmit an -bit mes-
bors until they reach the BS. As there is no central control in
sage a distance , the radio expends
MTE routing, it is difficult to set up fixed MAC protocols (e.g.,
TDMA), so each node uses CSMA to listen to the channel be-
fore transmitting data. If the channel is busy, the node backs
off; otherwise, the node transmits its data to the next-hop node. (9)
As nodes run out of energy, the routes are recomputed to en-
and to receive this message, the radio expends:
sure connectivity with the BS. We do not account for the en-
ergy requirements or delay for such updates in our simulations. (10)
Each node transmits its own data once every seconds,
where is set to minimize congestion but ensure efficient The electronics energy, , depends on factors such as the
digital coding, modulation, filtering, and spreading of the signal,
use of the channel bandwidth. If is too small, nodes end
whereas the amplifier energy, or , depends on the
up sending their own data before the previous set of data was
able to reach the BS. Large queues will build up, there will be distance to the receiver and the acceptable bit-error rate. For the
many collisions of data, and nothing will be transmitted to the experiments described in this paper, the communication energy
parameters are set as: nJ bit, pJ bit m ,
BS. If is too large, the channel is idle when it could be
and pJ bit m . Using our previous experimental
used for data transmission. We set based on , the total
number of nodes in the network, the average number of hops to results [26], the energy for data aggregation is set as
get a message to the BS, and the time it takes a message to tra- nJ bit signal.
verse a single hop. B. Optimum Number of Clusters
For static clustering, nodes are organized into clusters ini-
tially by the BS using the same method as in LEACH-C to en- In LEACH, the cluster formation algorithm was created to en-
sure that good clusters are formed. These clusters and cluster sure that the expected number of clusters per round is , a system
heads remain fixed throughout the lifetime of the network. As in parameter. We can analytically determine the optimal value of
LEACH and LEACH-C, nodes transmit their data to the cluster in LEACH using the computation and communication energy
head node during each frame of data transfer (using TDMA and models. Assume that there are nodes distributed uniformly
a DSSS spreading code to ensure minimal inter-cluster interfer- in an region. If there are clusters, there are on av-
ence), and the cluster head aggregates the data and sends the erage nodes per cluster (one cluster head and
resultant data to the BS. When the cluster head node’s energy non-cluster head nodes). Each cluster head dissipates energy
is depleted, the nodes in the cluster lose communication ability 5Note that this is a simplified model; in general, radio wave propagation is
with the BS and are essentially “dead.” highly variable and difficult to model.
Authorized licensed use limited to: Indian Institute of Technology (Ropar). Downloaded on September 15,2024 at 18:16:55 UTC from IEEE Xplore. Restrictions apply.
666 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 1, NO. 4, OCTOBER 2002

receiving signals from the nodes, aggregating the signals, and

transmitting the aggregate signal to the BS. Since the BS is far
from the nodes, presumably the energy dissipation follows the
multipath model ( power loss). Therefore, the energy dissi-
pated in the cluster head node during a single frame is

(11)
where is the number of bits in each data message, is
the distance from the cluster head node to the BS, and we have
assumed perfect data aggregation.
Each non-cluster head node only needs to transmit its data to
the cluster head once during a frame. Presumably the distance
to the cluster head is small, so the energy dissipation follows the
Friss free-space model ( power loss). Thus, the energy used
in each non-cluster head node is
(12) Fig. 6. Average energy dissipated per round in LEACH as the number of
clusters is varied between 1 and 11. This graph shows that LEACH is most
where is the distance from the node to the cluster head. energy efficient when there are between 3 and 5 clusters in the 100-node
The area occupied by each cluster is approximately . In network, as predicted by the analysis.
general, this is an arbitrary-shaped region with a node distri-
bution . The expected squared distance from the nodes
For our experiments, nodes, m,
to the cluster head (assumed to be at the center of mass of the
pJ, pJ, and 75 m 185 m, so we
cluster) is given by
expect the optimum number of clusters to be .
These analytical results were verified using simulations on
a 100-node network where we varied the number of clusters
between 1 and 11 and ran LEACH for 1000 simulated sec-
(13)
onds. Note that for these simulations, the nodes were placed
randomly throughout the 100 m 100 m area and we made
If we assume this area is a circle with radius
no restrictions on the distance between the nodes and their
and is constant for and , (13) simplifies to
cluster heads (e.g., ) or between the nodes and the BS
(e.g., ). Even though we made these assumptions for the
(14) analysis, Fig. 6, which shows the average energy dissipated per
round as a function of the number of clusters, shows that the
If the density of nodes is uniform throughout the cluster area, simulation agrees well with the analysis. This graph shows that
then and the optimum number of clusters is around 3–5 for the 100-node
network. When there is only one cluster, the non-cluster head
(15) nodes often have to transmit data very far to reach the cluster
Therefore, in this case head node, draining their energy, and when there are more than
five clusters, there is not as much local data aggregation being
(16) performed. For the rest of the experiments, we set to five.

The energy dissipated in a cluster during the frame is C. Energy Gains

In these experiments, each node begins with only 2 J of energy
and an unlimited amount of data to send to the BS. Each node
(17) uses the probabilities in (3) to determine its cluster head status at
and the total energy for the frame is the beginning of each round, and each round lasts for 20 s.6 We
tracked the rate at which the data packets are transfered to the
BS and the amount of energy required to get the data to the BS.
When the nodes use up their limited energy during the course
of the simulation, they can no longer transmit or receive data.
(18)
For these simulations, energy is consumed whenever a node
We can find the optimum number of clusters by setting the transmits or receives data or performs data aggregation. Using
derivative of with respect to to zero spread-spectrum increases the number of bits transmitted,
6The time for a round was chosen so that on average each node has enough

(19) energy to act as cluster head once and non-cluster head several times throughout
the simulation lifetime [10].
Authorized licensed use limited to: Indian Institute of Technology (Ropar). Downloaded on September 15,2024 at 18:16:55 UTC from IEEE Xplore. Restrictions apply.
HEINZELMAN et al.: AN APPLICATION-SPECIFIC PROTOCOL ARCHITECTURE FOR WIRELESS MICROSENSOR NETWORKS 667

(a) (b)
Fig. 7. Data for the limited energy simulations, where each node begins with 2 J of energy. (a) Total amount of data received at the BS over time. (b) Total amount
of data received at the BS per given amount of energy. These graphs show that LEACH distributes an order of magnitude more data per unit energy than MTE
routing, LEACH-C delivers 40% more data per unit energy than LEACH, and static clustering does not perform well when the nodes have limited energy.

thereby increasing the amount of energy dissipated in the traverses several hops. In the other protocols, each message
electronics of the radio. We do not assume any static energy is transmitted over a single hop, to the cluster head, where
dissipation nor do we assume energy is consumed during data aggregation occurs. The aggregate signals are sent to the
carrier-sense operations; hence, the results here do not account BS, greatly reducing the amount of data transmitted. Fig. 7(b)
for the potential energy benefits of using TDMA in LEACH shows the total data received at the BS for a given amount of
compared with CSMA in MTE. energy. This graph shows that LEACH and LEACH-C deliver
Although quality is an application-specific and data-de- the most data per unit energy, achieving both energy and latency
pendent quantity, one application-independent method of efficiency. A routing protocol such as MTE does not enable
determining quality is to measure the amount of data (number local computation to reduce the amount of data that needs to
of actual data signals or number of data signals represented be transmitted to the BS.
by an aggregate signal) received at the BS. The more data Fig. 7 shows that LEACH is not as efficient as LEACH-C
the BS receives, the more accurate its view of the remote (LEACH-C delivers about 40% more data per unit energy than
environment will be. If all the nodes within a cluster are sensing LEACH). This is because the BS has global knowledge of the
the same event, the actual and effective data will contain the location and energy of all the nodes in the network, so it can
same information, and there is no loss in quality by sending produce better clusters that require less energy for data trans-
effective or aggregate data. If, on the other hand, the nodes mission. In addition, the BS formation algorithm ensures that
are seeing different events, the cluster head will pick out the there are clusters during each round of operation. As
strongest event (strongest signal within the signals of the cluster there are only 100 nodes in the simulation, even though the ex-
members) and send that as the data from the cluster. In this pected number of clusters per round is in LEACH, each
case, there will be a loss in quality by aggregating signals into round does not always have five clusters.
a single representative signal. As with radio wave propagation, Fig. 8(a) shows the total number of nodes that remain alive
it is difficult to quantify signal propagation as it depends on over the simulation time. While nodes remain alive for a long
factors such as the nature of the signal, the path between the time in MTE, this is because a much smaller amount of data
source and the sensor, and the sensitivity of the sensors. If the has been transmitted to the BS. If we plot the total number of
distance between nodes within a cluster is small compared with nodes that remain alive per amount of data received at the BS
the distance from which events can be sensed, or if the distance [Fig. 8(b)], we see that nodes in LEACH can deliver ten times
between events occurring in the environment is large, there is a more effective data than MTE for the same number of node
high probability that the nodes will be sensing the same event. deaths. There are two reasons that MTE requires more energy
For our experiments, we assume that all nodes in a cluster sense to send data to the BS (hence, causing more node deaths for
the same events. the same amount of data delivery): collisions and lack of data
Fig. 7 shows the total number of data signals (actual for MTE aggregation. Because MTE does not have any centralized con-
and effective for LEACH, LEACH-C, and static clustering) trol over when nodes transmit and receive packets, collisions
received at the BS over time and the total data received at increase the amount of energy required to send each successful
the BS for a given amount of energy. Fig. 7(a) shows that message. Furthermore, each message in MTE must traverse
LEACH sends much more data to the BS in the simulation time approximately hops to get to the BS,7 whereas
than MTE routing. The reason MTE requires so much time 7The analysis for finding the average number of hops is similar to the analysis
to send data from the nodes to the BS is that each message for finding E [d ] in (13).

Authorized licensed use limited to: Indian Institute of Technology (Ropar). Downloaded on September 15,2024 at 18:16:55 UTC from IEEE Xplore. Restrictions apply.
668 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 1, NO. 4, OCTOBER 2002

(a) (b)
Fig. 8. Data for the limited energy simulations, where each node begins with 2 J of energy. (a) Number of nodes alive over time. (b) Number of nodes alive per
amount of data sent to the BS. LEACH can deliver ten times the amount of effective data to the BS as MTE routing for the same number of node deaths. The benefit
of rotating cluster heads in LEACH is clearly seen by comparing the number of nodes alive in LEACH and static clustering.

each message in LEACH need only traverse one hop due to that assigns the cluster head role to another node in the cluster
data aggregation at the cluster head. Of course, this assumes when the current cluster head node dies. Adapting the clusters
perfect aggregation—the advantages of using MTE become depending on which nodes are cluster heads for a particular
greater when this assumption is relaxed. round (as in LEACH) is advantageous because it ensures that
Fig. 8 shows why static clustering performs poorly (as seen nodes communicate with the cluster head node that requires
in the results from Fig. 7)—the cluster head nodes die quickly, the lowest amount of transmit power. In addition to reducing
ending the lifetime of all nodes belonging to those clusters. energy dissipation, this ensures minimum inter-cluster in-
Therefore, rotating the cluster head position enables LEACH to terference. If, on the other hand, the clusters were fixed and
achieve a longer lifetime than static clustering. only the cluster head nodes were rotated, a node may have to
use a large amount of power to communicate with its cluster
head when there is another cluster’s cluster head close by.
V. DISCUSSION Therefore, using fixed clusters and rotating cluster head nodes
While LEACH appears to be a promising protocol, there are within the cluster may require more transmit power from the
some areas for improvement to make the protocol more widely nodes, increasing non-cluster head node energy dissipation and
applicable. In the current implementation of LEACH, we as- increasing inter-cluster interference. However, the advantage
sume sensors always transmit data to the cluster head during of fixed clusters is that once the clusters are formed, there is no
their allocated TDMA slot (or, for MTE routing, each sec- set-up overhead at the beginning of each round. Depending on
onds). To save energy, nodes may only need to transmit data the cost of forming adaptive clusters, an approach where the
after they detect some interesting event. In this case, we may clusters are formed once and fixed and the cluster head position
need to rethink the intra-cluster communication scheme to make rotates among the nodes in the cluster may be more energy
sure that we efficiently utilize bandwidth when not all nodes efficient than LEACH.
communicate to the cluster head all the time. Finally, we showed that using data aggregation reduces en-
Another assumption we have made is that all nodes are within ergy dissipation and latency in data transfer compared with an
communication range of each other and the BS. This assump- approach like MTE that cannot take advantage of local data
tion limits the scalability of the protocol but can be relaxed by correlation. However, if there is no correlation among the data
using collision-avoidance techniques during the set-up phase to (and, hence, the cluster head cannot compress the data from the
reduce collisions in ADV and Join-REQ messages and using a cluster members), a multihop approach like MTE will outper-
hierarchical or multihop routing approach to get data from the form LEACH. In Section III-A we discussed an approach to de-
cluster head nodes to the BS. The cluster heads could form a termine the number of nodes in an approximately
multihop backbone whereby data are transmitted among cluster
region of space around each node. We can set appropriately
heads until they reach the BS. Alternatively, LEACH can evolve
so that there is a high probability that all the sensors within the
into a hierarchical protocol by forming “super clusters” out of
the cluster head nodes and having a “super-cluster head” that area have correlated data. Using the determined value
processes the data from all the cluster head nodes in the super for , each node can find the approximate optimal value (19)
cluster. These changes will make LEACH suitable for a wider and compute the appropriate probability that it should become a
range of wireless microsensor networks. cluster head during the next round. Using this approach we can
As our results have clearly shown the advantage of rotating assure that, with high probability, the clusters have correlated
the cluster head position among all the nodes, it would be data and the protocol can scale to a large number of nodes and
interesting to compare LEACH to a fixed clustering protocol a large network area, and can handle dynamic nodes.
Authorized licensed use limited to: Indian Institute of Technology (Ropar). Downloaded on September 15,2024 at 18:16:55 UTC from IEEE Xplore. Restrictions apply.
HEINZELMAN et al.: AN APPLICATION-SPECIFIC PROTOCOL ARCHITECTURE FOR WIRELESS MICROSENSOR NETWORKS 669

VI. CONCLUSION [20] G. Pottie, “Wireless sensor networks,” in Proc. Information Theory
Workshop, San Diego, CA, June 1998, pp. 139–140.
When designing protocol architectures for wireless mi- [21] G. Pottie and W. Kaiser, “Wireless integrated network sensors,”
crosensor networks, it is important to consider the function Commun. ACM, vol. 43, pp. 51–58, May 2000.
[22] T. Rappaport, Wireless Communications: Principles & Prac-
of the application, the need for ease of deployment, and the tice. Englewood Cliffs, NJ: Prentice-Hall, 1996.
severe energy constraints of the nodes. These features led us [23] R. Ruppe, S. Griswald, P. Walsh, and R. Martin, “Near term digital radio
to design LEACH, a protocol architecture where computation (NTDR) system,” in Proc. MILCOM, vol. 3, Monterey, CA, Nov. 1997,
pp. 1282–1287.
is performed locally to reduce the amount of transmitted [24] T. Shepard, “A channel access scheme for large dense packet radio
data, network configuration and operation is done using local networks,” in Proc. ACM SIGCOMM, Stanford, CA, Aug. 1996, pp.
control, and media access control (MAC) and routing protocols 219–230.
[25] S. Singh, M. Woo, and C. Raghavendra, “Power-aware routing in mobile
enable low-energy networking. Results from our experiments ad hoc networks,” in Proc. 4th Annual ACM/IEEE Int. Conf. Mobile
show that LEACH provides the high performance needed under Computing Networking (MobiCom), Oct. 1998.
the tight constraints of the wireless channel. [26] A. Wang, W. Heinzelman, and A. Chandrakasan, “Energy-scalable
protocols for battery-operated microsensor networks,” Proc. 1999 IEEE
Workshop Signal Processing Systems (SiPS ’99), pp. 483–492, Oct.
1999.
ACKNOWLEDGMENT
The authors would like to thank the reviewers for their de-
tailed comments on earlier versions of this paper.
Wendi B. Heinzelman (S’95–M’01) received the
B.S. degree in electrical engineering from Cornell
REFERENCES University, Ithaca, NY, in 1995, and the M.S.
and Ph.D. degrees in electrical engineering and
[1] P. Agarwal and C. Procopiuc, “Exact and approximation algorithms for computer science from the Massachusetts Institute
clustering,” in Proc. 9th Annu. ACM-SIAM Symp. Discrete Algorithms, of Technology (MIT), Cambridge, in 1997 and 2000,
Baltimore, MD, Jan. 1999, pp. 658–667. respectively.
[2] J. Agre and L. Clare, “An integrated architecture for cooperative sensing Since January 2001, she has been an Assistant Pro-
networks,” IEEE Computer, vol. 33, pp. 106–108, May 2000. fessor in the Department of Electrical and Computer
[3] D. Baker, A. Ephremides, and J. Flynn, “The design and simulation of Engineering, University of Rochester, Rochester, NY.
a mobile radio network with distributed control,” IEEE J. Select. Areas Her current research interests lie in the area of low-
Commun., vol. SAC–2, pp. 226–237, Jan. 1984.
power ad-hoc wireless protocol architectures, scalable sensor networks, and
[4] A. Chandrakasan, R. Amirtharajah, S.-H. Cho, J. Goodman, G. Kon-
multimedia communication.
duri, J. Kulik, W. Rabiner, and A. Wang, “Design considerations for
distributed microsensor systems,” in Proc. IEEE Custom Integrated Cir- Dr. Heinzelman is an Elected Member of the Design and Implementation of
cuits Conf. (CICC), San Diego, CA, May 1999, pp. 279–286. Signal Processing Systems (DISPS) Technical Committee of the IEEE Signal
[5] L. Clare, G. Pottie, and J. Agre, “Self-organizing distributed sensor net- Processing Society and a member of Sigma Xi and the ACM.
works,” in Proc. SPIE Conf. Unattended Ground Sensor Technologies
and Applications, vol. 3713, Orlando, FL, Apr. 1999, pp. 229–237.
[6] M. Dong, K. Yung, and W. Kaiser, “Low power signal processing ar-
chitectures for network microsensors,” in Proc. Int. Symp. Low Power
Electronics and Design, Monterey, CA, Aug. 1997, pp. 173–177. Anantha P. Chandrakasan (S’87–M’95–SM’01)
[7] D. Estrin, R. Govindan, J. Heidemann, and S. Kumar, “Next century received the B.S, M.S. and the Ph.D. degrees in
challenges: Scalable coordination in sensor networks,” in Proc. 5th electrical engineering and computer sciences from
Annual ACM Int. Confe. Mobile Computing Networking (MobiCom), the University of California, Berkeley, in 1989,
Seattle, WA, Aug. 1999, pp. 263–270. 1990, and 1994, respectively.
[8] M. Ettus, “System capacity, latency, and power consumption in mul- Since September 1994, he has been with the
tihop-routed SS-CDMA wireless networks,” in Proc. Radio and Wire- Massachusetts Institute of Technology (MIT),
less Conf. (RAWCON), Colorado Springs, CO, Aug. 1998, pp. 55–58. Cambridge, and is currently an Associate Professor
[9] D. Hall, Mathematical Techniques in Multisensor Data Fu- of electrical engineering and computer science. He
sion. Boston, MA: Artech House, 1992.
held the Analog Devices Career Development Chair
[10] W. Heinzelman, “Application-specific protocol architectures for wire-
less networks,” Ph.D. dissertstion, Mass. Inst. Technol., Cambridge, from 1994 to 1997. He is a co-author of Low Power
2000. Digital CMOS Design (Norwell, MA: Kluwer) and a co-editor of Low Power
[11] W. Heinzelman, A. Chandrakasan, and H. Balakrishnan, “Energy-ef- CMOS Design and Design of High-Performance Microprocessor Circuits
ficient routing protocols for wireless microsensor networks,” in Proc. (Piscataway, NJ: IEEE Press). His research interests include the ultra low
33rd Hawaii Int. Conf. System Sciences (HICSS), Maui, HI, Jan. 2000. power implementation of custom and programmable digital signal processors,
[12] L. Hu, “Distributed code assignments for CDMA packet radio net- distributed wireless sensors, multimedia devices, emerging technologies, and
works,” IEEE/ACM Trans. Networking, vol. 1, pp. 668–677, Dec. 1993. CAD tools for VLSI.
[13] C. Intanagonwiwat, R. Govindan, and D. Estrin, “Directed diffusion: A Dr. Chandrakasan received the NSF Career Development Award in 1995, the
scalable and robust communication paradigm for sensor networks,” in IBM Faculty Development Award in 1995, and the National Semiconductor Fac-
Proc. Fourth Annu. ACM Int. Conf. Mobile Computing and Networking ulty Development Award in 1996 and 1997, respectively. He has received several
(MobiCom), Boston, MA, Aug. 2000, pp. 56–67. best paper awards, including the 1993 IEEE Communications Society’s Best
[14] T. Kwon and M. Gerla, “Clutering with power control,” in Proc. Tutorial Paper Award, the IEEE Electron Devices Society’s 1997 Paul Rappa-
MILCOM, vol. 2, Atlantic City, NJ, Nov. 1999. port Award for the Best Paper in an EDS publication during 1997, and the 1999
[15] C. Lin and M. Gerla, “Adaptive clustering for mobile wireless net- Design Automation Conference Design Contest Award. He has served on the
works,” IEEE J. Select. Areas Commun., vol. 15, pp. 1265–1275, Sept. technical program committee of various conferences, including ISSCC, VLSI
1997. Circuits Symposium, DAC, and ISLPED. He has served as a Technical Program
[16] T. Murata and H. Ishibuchi, “Performance evaluation of genetic algo- Co-Chair for the 1997 International Symposium on Low-Power Electronics and
rithms for flowshop scheduling problems,” Proc. 1st IEEE Conf. Evolu- Design (ISLPED), VLSI Design ’98, and the 1998 IEEE Workshop on Signal
tionary Computation, vol. 2, pp. 812–817, June 1994.
Processing Systems, and as a General Co-Chair of the 1998 ISLPED. He was an
[17] UCB/LBNL/VINT Network Simulator – ns (2000). [Online]. Available:
http://www.isi.edu/vint/ nsnam/ Associate Editor for the IEEE JOURNAL ON SOLID-STATE CIRCUITS from 1998
[18] K. Pahlavan and A. Levesque, Wireless Information Networks. New to 2001. He served as an Elected Member of the Design and Implementation
York: Wiley, 1995. of Signal Processing Systems (DISPS) Technical Committee of the Signal Pro-
[19] S. Park and M. Srivastava, “Power aware routing in sensor networks cessing Society. He was the Signal Processing Subcommittee Chair for ISSCC
using dynamic source routing,” ACM MONET Special Issue on Energy 1999 through 2001 and the Program Vice-Chair for ISSCC 2002. He is the Tech-
Conserving Protocols in Wireless Networks, 1999. nical Program Chair for ISSCC 2003.

Authorized licensed use limited to: Indian Institute of Technology (Ropar). Downloaded on September 15,2024 at 18:16:55 UTC from IEEE Xplore. Restrictions apply.
670 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 1, NO. 4, OCTOBER 2002

Hari Balakrishnan (S’95–M’98) received the

B.Tech. degree from the Indian Institute of Tech-
nology, Madras, in 1993 and the Ph.D. degree in
computer science from the University of California,
Berkeley, in 1998.
He is an Assistant Professor in the Department of
Electrical Engineering and Computer Science and
a Member of the Laboratory for Computer Science
(LCS) at Massachusetts Institute of Technology
(MIT), Cambridge. He leads the Networks and Mo-
bile Systems group at LCS, exploring research issues
in wireless networking, mobile systems, network protocols and architecture,
and pervasive computing.
Dr. Balakrishnan is the recipient of a National Science Foundation CAREER
Award (2000), an IBM University Partnership Prize (1999-2001), the ACM
Doctoral Dissertation Award for his work on reliable data transport over wire-
less networks (1998), and Best Paper Awards at the ACM MOBICOM (1995
and 2000), IEEE HotOS (2001), and USENIX (1995) technical conferences.
He holds the KDD Career Development Chair in Communications Technology
at MIT, and was awarded the MIT School of Engineering Ruth and Joel Spira
Award for Distinguished Teaching in 2001.

Authorized licensed use limited to: Indian Institute of Technology (Ropar). Downloaded on September 15,2024 at 18:16:55 UTC from IEEE Xplore. Restrictions apply.