Wireless Sensor Network for in situ Soil Moisture Monitoring

Jianing Fang

, Chuheng Hu

, Nour Smaoui

, Doug Carlson

4,∗

, Jayant Gupchup

5,∗

Razvan Musaloiu-E.

6,∗

, Chieh-Jan Mike Liang

7,∗

, Marcus Chang

8,∗

, Omprakash Gnawali

Tamas Budavari

, Andreas Terzis

6,∗

, Katalin Szlavecz

and Alexander S. Szalay

Department of Earth & Planetary Sciences, Johns Hopkins University, 3400 N Charles Street, Baltimore, U.S.A.

Department of Computer Science, Johns Hopkins University, 3400 N Charles Street, Baltimore, U.S.A.

Department of Computer Science, University of Houston, 4800 Calhoun Rd., Houston, Houston, U.S.A.

CEVA, Inc., 15245 Shady Grove Rd., Rockville, U.S.A.

Microsoft Corporation, One Microsoft Way, Redmond, U.S.A.

Independent Researcher, U.S.A.

Microsoft Research, No. 5 Danling Street, Beijing, China

Arm Ltd., 5707 Southwest Pkwy #100, Austin, U.S.A.

Department of Applied Mathematics & Statistics, Johns Hopkins University, 3400 N Charles Street, Baltimore, U.S.A.

Dept. of Computer Science and Physics & Astronomy, Johns Hopkins University, 3400 N Charles Street, Baltimore, U.S.A.

razvanm@cs.jhu.edu, liang.mike@microsoft.com, marcus.chang@arm.com, gnawali@gmail.com, budavari@jhu.edu,

aterzis@gmail.com, szlavecz@jhu.edu, szalay@jhu.edu

∗

Previous members of the team project. Work performed while at Johns Hopkins University.

Keywords:

Wireless Sensor Network, Soil Moisture, In-situ Environmental Monitoring.

Abstract:

We discuss the history and lessons learned from a series of deployments of environmental sensors measuring

soil parameters and CO

ﬂuxes over the last ﬁfteen years, in an outdoor environment. We present the hardware

and software architecture of our current Gen-3 system, and then discuss how we are simplifying the user facing

part of the software, to make it easier and friendlier for the environmental scientist to be in full control of the

system. Finally, we describe the current effort to build a large-scale Gen-4 sensing platform consisting of

hundreds of nodes to track the environmental parameters for urban green spaces in Baltimore, Maryland.

1 INTRODUCTION

Soil is a semi-aquatic habitat harboring a huge diver-

sity of terrestrial as well as aquatic organisms. The

amount and availability of soil water affect survival

and activity of organisms from bacteria to macrofauna

to plants directly, but also indirectly, by redistributing

nutrients or toxic substances. Consequently soil water

content is a main driver of belowground biological ac-

tivity. Pathways and rates of complex biogeochemical

processes can shift dramatically especially if the soil

dries out or becomes waterlogged. For instance the

same ecosystem can switch between being a methane

sink or source depending on ﬂuctuations of water ta-

ble or precipitation patterns.

Soil is inherently spatially heterogeneous in all

three dimensions and at many spatial scales. Young

and Crawford (2004) called soils “the most compli-

cated biomaterials on the planet” (Young and Craw-

ford, 2004). On ﬁeld scale high degree of patchiness

exists even in seemingly uniform landscapes (Robert-

son et al., 1994). Soil physical, chemical and bio-

logical characteristics vary even more in fragmented

landscapes. An extreme example of this is the urban-

suburban environment with highly modiﬁed surface

topography and soil conditions. Land use-land man-

agement decisions (construction, plant cover, irriga-

tion regimes) are often made on a parcel level to

the extent that it overrides the characteristics of the

underlying natural soil (Pouyat et al., 2010; Pickett

et al., 2011). To obtain spatial soil moisture data

with modeling approaches, such as using digital ele-

vation models to produce topographic moisture index,

proved to be challenging in such disturbed environ-

ments (Tenenbaum et al., 2006).

For soil ecology research, understanding soil

moisture conditions on a ﬁeld scale is paramount

Fang, J., Hu, C., Smaoui, N., Carlson, D., Gupchup, J., Musaloiu-E., R., Liang, C., Chang, M., Gnawali, O., Budavari, T., Terzis, A., Szlavecz, K. and Szalay, A.

Wireless Sensor Network for in situ Soil Moisture Monitoring.

DOI: 10.5220/0010261500250036

In Proceedings of the 10th International Conference on Sensor Networks (SENSORNETS 2021), pages 25-36

ISBN: 978-989-758-489-3

to identify biogeochemical activity ‘hot spots’ and

‘hot moments’, and to explain spatial distribution of

soil organisms. Spatially explicit approach to study

soil communities is relatively new, but increasingly

recognized as a signiﬁcant tool in explaining high

soil biodiversity (Ettema and Wardle, 2002; Chust

et al., 2003). One approach to collect in situ, spa-

tially explicit, continuous soil data is to use sensor

systems, more speciﬁcally wireless sensor networks

(WSN hereafter). This technology emerged in the

late 1990’s (Cerpa et al., 2001), and promised inex-

pensive, hands-free, low-cost, and low-impact envi-

ronmental data collection — an attractive alternative

to manual data logging — in addition to providing

considerably more data at ﬁner spatial and tempo-

ral granularities. Since originally complete off-the

shelf solutions were not available, and both the sen-

sors themselves and the related data systems needed

customization, we, a group of soil ecologists, physi-

cists and computer scientists started building our own

WSN to monitor the soil ecosystem.

Our project, dubbed LifeUnderYourFeet (LUYF)

developed three generations of hardware and soft-

ware, and deployed the systems in various habitats

including tropical and temperate forests, agricultural

ﬁelds and a desert. In this paper we discuss our ef-

forts and experiences in designing and deploying an

end-to-end system using wireless sensor networks to

monitor soil conditions.

2 HISTORY OF OUR WSN

Generation 1. Since our project started in 2005 we

have gone through three generations of data collection

equipment. The ﬁrst wireless devices were manufac-

tured by CrossBow Inc. These MicaZ “mote” (Fig-

ure 1) contained processing, storage, and communi-

cation capabilities in a small form factor (matchstick

box size). Besides two simple on-board sensors for

temperature and light, ﬁve additional external ana-

log sensors could be connected simultaneously to the

mote via a separate fan-out board. In practice, this

board turned out to be very fragile and was one of

the reasons we abandoned this platform. The device’s

on board ﬂash was capable of storing thousands of

sensor measurements for later retrieval. However, the

MicaZ motes were quite expensive costing close to

$200 each at the time.

For the soil measurements, we ﬁrst built our own

temperature sensors by placing small thermistors into

plastic vials, about 1.75 in long, and ﬁlling them with

epoxy. While the construction was moderately labor

intensive, the material cost was less than $5 (Figure

2). For soil moisture we ﬁrst experimented with the

Watermark moisture sensor by Irrometer. The perfor-

mance of these sensors rapidly degraded with time,

thus we switched to the ECH

O EC-5 moisture sen-

sors by Decagon. We used two AA batteries to power

both the MicaZ mote and its sensors.

Figure 1: Our ﬁrst platform, based on the MicaZ in Ham-

mond boxes.

In order to protect the motes from the environ-

ment we build our own enclosures out of watertight

IP67-rated Hammond enclosures. While this enclo-

sure setup survived the “bathtub test”, i.e. remained

dry for extended periods under water in a bathtub, the

prolonged exposure to the changing weather condi-

tions along with the holes that we drilled for the sen-

sor cables, even though ample sealant was used, was

enough to compromise its water-tight feature.

Generation 2. In 2007 we switched to a new wire-

less device, the Tmote Sky from MoteIV, but kept

the custom-built temperature sensors and the Decagon

moisture sensors (Figure 2). The Tmote Sky is in

many ways quite similar to the MicaZ, but consider-

ably less expensive. This platform offered four times

the measurement ﬁdelity and twice as much space for

storing samples. We designed and built an expansion

board that allowed us to handle four external analog

sensors.

Figure 2: Our Gen-2 platform, based on the Telos mote.

Besides the external sensors, this mote also had

a built in digital humidity and temperature sensor by

Sensirion (as opposed to the analog temperature sen-

sor on the MicaZ) and two light sensors, one for vis-

ible light and the other one for photosynthetically ac-

tive radiation. However, as with the MicaZ, the light

sensors were quite sensitive, so under normal daylight

SENSORNETS 2021 - 10th International Conference on Sensor Networks

they were either entirely “off” or “on”.

For the enclosure we still used a Hammond box,

but in order to increase the range of the radios, we at-

tached an external antenna (5 dBi gain), directly con-

nected to the mote, but extending outside the box. The

hole for the antenna was sealed with silicone based

caulk. The expansion board was put in its own case

and ﬁlled with insulation foam after the moisture and

temperature sensors had been attached. We also up-

graded the power supply to a single D-sized Li-SoCl2

battery that increased the capacity by an order of mag-

nitude.

The new platform was much more stable and pro-

fessional, but not without its weak points. First, the

sensor connection was quite expensive. The connec-

tors and the cable for the expansion board cost al-

most $30. Furthermore, assembling the whole setup

was cumbersome, requiring considerable manual la-

bor. While there were almost no cases in which we

had standing water inside the enclosures, moisture

tended to accumulate over time, and we placed mois-

ture absorbing silicone beads inside each enclosure to

mitigate this problem.

Figure 3: Our Gen-3 platform, based on our PCB design.

Generation 3. In 2013 we have designed and started

deploying our third generation sensing platform, the

Breakfast Suite, consisting of the Bacon mote (Fig-

ure 3), and the Toast multiplexer module. These were

all designed in house and manufactured by an OEM

in Hungary. This new mote encapsulates our expe-

riences over the past seven years and is speciﬁcally

designed for soil monitoring at different scales. With

the same form factor as the Tmote Sky, the Bacon

mote has ﬁve times the processing power and eight

times the storage space. Moreover, its average power

consumption is an order of magnitude lower than the

Tmote sky and the manufacturing cost for thousands

of units was only $20, in 2012. To achieve large

scalability, both in terms of coverage area and sen-

sor density, we increased the communication range

of the Bacon mote with more efﬁcient radios and de-

signed a new expansion board, the Toast sensor board,

capable of reading eight external sensors at a time.

Multiple Toast boards can be attached to a single Ba-

con mote at a time thereby facilitating different sensor

densities and more ﬂexible sensor compositions. Fur-

thermore, we have added a much simpler multiplexer

board (miniToast) to the palette, which is code com-

patible with the Toast, but contains an on-board high

precision thermistor and a port for additional analog

sensor.

3 HISTORY OF DEPLOYMENTS

Roughly parallel to the evolution of hardware plat-

forms over the years, our deployment architecture

went through several changes as we identiﬁed weak

points and expanded our usage scenarios. At a

high level, we ﬁrst increased reliability by adding a

permanently-powered “basestation” collection mote

(with Internet access), then worked to maintain reli-

ability while decoupling network performance from

the presence of the basestation.

Figure 4: System architecture of Generation 2 wireless sen-

sor network. Many deployments of sensing nodes send data

through a Gateway computer to a web server. The raw

analog sensor measurements are assigned timestamps con-

verted to physical units, and screened for faults in the Stage-

database. This data is mapped to geographical locations and

the data is organized for retrieval in the science database.

Early Deployments. The early deployments (Fig-

ure 4) had no persistent infrastructure: motes logged

data internally and a ﬁeld visit was required to col-

lect their data over a wireless connection to a laptop.

While this approach was more convenient than using

wired data loggers, it was still a tedious manual pro-

cess. More importantly, due to the lack of a reliable

clock on the motes, periodic references to some exter-

nal time source were required to map measurements

to points in time. If a mote became unstable (due to

low battery, condensation, or software faults), it could

record data that lacked time references. This problem

was addressed by using the on-board light sensors to

estimate the date (by matching the estimated length-

of-day to a physical model of the Earth’s movement

around the Sun) and local noon (the midpoint between

Wireless Sensor Network for in situ Soil Moisture Monitoring

dawn and sunset). Even though the precision of this

approach was limited to a few minutes, it was still

adequate for the relatively low sensor sampling rate

we used (one sample every 5-10 minutes) (Gupchup

et al., 2009).

The ﬁrst deployment occurred on the Johns Hop-

kins University campus in a wooded area and lasted

for seven months. Subsequent deployment took place

in urban forests in Baltimore (Leakin park, Cub Hill),

deciduous forests in the Mid-Atlantic Coastal Plains

(Smithsonian Environmental Research Center, SERC,

Jug Bay), agricultural ﬁelds (USDA), tropical rain-

forests (Ecuador), and high altitude deserts (Atacama)

(Figure 5). The objectives of collecting soil moisture

and temperature data in these projects were diverse

and included measuring soil carbon cycling, monitor-

ing turtle overwintering behavior and egg incubation,

predicting soil denitriﬁcation activity and testing the

system at low barometric pressure and high UV expo-

sure.

The Cub Hill Project. The largest and longest de-

ployment was in an urban residential area, Called Cub

Hill. The Cub Hill site (39.412507 N,-76.520903 W)

is located in Parkville, Baltimore County, MD, 14 km

north from the center of Baltimore City. The site

has several ongoing studies related to the Baltimore

Ecosystem Study LTER (www.beslter.org) including

long-term anthropogenic effects focused on the atmo-

sphere, soils, hydrology, and vegetation (Pickett et al.,

2008; Pickett et al., 2011). A permanent urban CO

ﬂux tower operates at the juxtaposition of forest and

residential areas (Figure 6). To the North and West of

the tower is a poplar-oak-hickory stand with a canopy

height of 20-26 meters. To the South is a mix of

medium density residential areas made up of several

subdivisions built in the 70’s and 80’s. Detailed soil

and land use mapping has been carried out in the 1

footprint of the area (Ellis et al., 2006; Yesilonis

et al., 2016).

The majority of motes were deployed in the two

main land cover types, forest and grass, but a small

number of sensors were installed in planting beds,

covered with mulch or English Ivy. At each sam-

pling location soil temperature and soil moisture were

measured at 10 and 20 cm depths, thus at maximum

capacity a total of 106 soil moisture values were col-

lected at each sampling. Measurements were taken

every 10 minutes.

Maintenance ended in June 2011 and the project

terminated in 2012, at which point we brought all

sensors back to the lab. We were interested in the

performance of the soil moisture sensors after being

in the ﬁeld for several years. First, all sensors were

inspected for physical damage and categorized as fol-

lows: no visible damage, minor cracks in probe, and

major damage due to large cracks with internal wires

exposed. Laboratory tests were performed in a 10 gal

Sterilite storage bin using mineral soil from one of our

ﬁeld sites that had been previously dried and sieved

for homogeneity. The soil was gradually remoist-

ened and mixed to approximately 0.2 volumetric wa-

ter content (VWC). For testing, batches of eight sen-

sors were fully inserted into the soil and their mois-

ture readings were recorded every 1.5 sec (approx.)

for two minutes using custom developed data log-

gers (www.lifeunderyourfeet.org). Sensors that pro-

vided measurements that were noticeably different

than what others were reading often were tested sev-

eral times. Sensors were assessed based on how sim-

ilar their mean VWC was to the grand mean of all

sensor VWC means (excluding obviously bad values)

and how tight their sampling distribution was accord-

ing to their coefﬁcient of variation (CV). Sensors were

categorized as “good” if their mean VWC was within

0.03 VWC of the grand mean of all sensor means

(based on the 3% error rate speciﬁed by Decagon De-

vices for mineral soil) and their CV was less than

0.03. “Fair” sensors had a mean VWC within 0.03

VWC of the grand mean, but a CV greater than 0.03.

“Bad” sensors fell out of the allowable range for both

criteria.

The Cub Hill experiment in Baltimore City stud-

ied the urban/forest gradient, near an AmeriFlux

tower. It exempliﬁes our second generation of de-

ployments. These added a persistent basestation mote

(attached to an Internet-connected PC) that could au-

tomatically download data periodically and transfer it

to a database (Figure 4). These deployments also used

the Koala (Musaloiu-E et al., 2008) low power collec-

tion protocol to enable multi-hop downloads (where

the network cooperates to relay data for motes which

are out of the communication range of the basesta-

tion). While regular contacts with the basestation

helped to obviate the need for a reliable on-board

clock, there were still a few periods during which

power outages at the deployment site, heavy snow,

or general poor network conditions led to timestamp

losses.

In order to support deployments in remote regions

(lacking permanent power), we made a few modiﬁ-

cations to the Gen-2 software. We implemented a

system where motes exchange local time references

with each other which, when collected with the data,

can be used to map measurements to their correct

point in time, even when there is no basestation for

weeks (Gupchup et al., 2010). We added several GPS-

equipped motes to the network which served as reli-

able time references. Finally, we implemented data

SENSORNETS 2021 - 10th International Conference on Sensor Networks

Figure 5: Cumulative number of soil moisture samples col-

lected in the LifeUnderYourFeet project. Major hardware

changes are marked with vertical lines, while horizontal

lines show the start and end of deployments.

compression on the motes, so that even if conditions

prevented them from being accessed at all once de-

ployed, they would have enough space to buffer all

their measurements for the entire deployment dura-

tion (Carlson et al., 2010).

Figure 6: Land cover map of the Cub Hill site. Each sam-

pling location had one mote with two soil temperature and

two soil moisture sensors attached. Measurements were

taken at 10 and 20 cm depths.

Smithsonian Environmental Research Center. The

Gen-3 software decoupled the energy requirements

of individual motes from the overall network size

by segmenting deployments into semi-autonomous

patches. This multi-tiered network approach is a crit-

ical step towards enabling large-scale deployments.

Our largest Gen-3 deployment was at the Smithso-

nian Environmental Research Center in Edgewater,

Maryland. The system monitored air and soil tem-

perature and soil moisture in deciduous forest stands.

For the Mid-Atlantic Region of North America, cli-

mate models predict increased amount of rainfall, dis-

tributed more unevenly throughout the year. Under

constructed rainout shelters we manipulated the fre-

quency and intensity of rainfall, and followed the fate

of carbon using isotopically enriched leaf litter. The

Figure 7: Daily data yield (as a fraction of expected data

yield) for the Cub Hill deployment. Starting point (July

2010) indicates a system-wide battery replacement. The

solid line indicates how much data was recovered as a frac-

tion of the entire 50-node network’s expected yield, while

the dashed line only considers the remaining operational

nodes.

pilot experiment lasted for ﬁve months. There were

two patches (old and young forest), and two manip-

ulations (wet and dry) in each stand. Each rainout

shelter was equipped with three sensor assemblies,

collecting data from 24 sensors. Another two were

deployed randomly in the forest to collect background

data. Over the course of ﬁve months we collected over

90,000 data points.

4 LESSONS LEARNED

The LifeUnderYourFeet project has collected over

400MB of raw sensor data, approximately 20% which

were soil moisture data (Figure 5). The Cub Hill de-

ployment itself collected 100,745,909 samples, out of

which 39,471,983 and 17,627,640 were environmen-

tal and speciﬁcally soil moisture data, respectively.

Below we present a selected set of these results, both

on the performance of the WSN and the sensors and

on soil conditions. Other aspects of the data collec-

tion and analysis are reported in (Musaloiu-E et al.,

2008; Gupchup et al., 2010; Savva et al., 2013).

Performance of the Wireless Sensor Network. Fig-

ure 7 shows how the number of operational nodes and

total data yield (timestamped and fault-free data) var-

ied over an 18-month maintenance-free period (Fig-

ure 7). All nodes were installed with fresh batteries

at the end of June 2010 and were not serviced again.

While the yield declined slowly as devices failed, the

remaining nodes continued to work well even as the

network degraded around them. It’s not until Novem-

ber of 2011 that so few nodes remain in operation that

it became difﬁcult to reach them and obtain time ref-

erences for them.

Figure 8 shows the distribution of maintenance in-

tervals on our system. While a handful of devices

Wireless Sensor Network for in situ Soil Moisture Monitoring

fail within the ﬁrst few months (likely due to physi-

cal/moisture damage), 90% of the nodes ran for more

than 6 months without maintenance, and 77% ran

for over a year without maintenance. With regular

twice-yearly site visits, such a deployment could be

expected to continuously deliver 90%+ yields.

Figure 8: Mote maintenance statistics of the Cub Hill

deployment following site-wide battery replacement on

06/24/10. Solid line is the cumulative distribution of the

node of maintenance intervals. The x axis indicates the

number of days from installation of a node to the last date

where 90% or more of its expected data volume was recov-

ered.

We evaluated mote failure according to the un-

derlying causes. Software defects caused 11% of

the mote failures. About half of the failures were

due to low batteries or moisture in their enclosures

(43% and 9%, respectively). The rest were marked as

“unknown”, because they showed neither low battery

nor high moisture before their disappearance. These

could be nodes with a catastrophic leak during a rain

event, nodes that froze and drained their battery in one

shot, nodes that suffered physical damage (such as a

falling tree or animal chewing), or nodes that failed

for no apparent reason. Given that the deployment

was in an urban environment, surprisingly, no motes

were stolen or vandalized.

Wireless Frequency Band. First and foremost, as we

have learned the hard way, the standard 2.4-2.5 GHz

band is substantially absorbed by water vapor, and on

days with a lot rain and humid air the range of the

2.5GHz radios have dropped substantially. So with

our Gen 3 hardware we have switched to the 900MHz

frequency band, and these problems have gone away.

Waterprooﬁng. Even though the assembled boxes

initially provided a perfect seal, the changing envi-

ronmental conditions naturally led to increasing mois-

ture. With the on-board moisture sensor, we were able

to monitor the conditions inside the boxes and have

seen the moisture steadily increasing. Temperature

inside the box became quite high, sometime reach-

ing 70

C, due to sun exposure (greenhouse effect).

This led to a substantial over-pressure inside, which

over time created small micro-cracks in the gasket and

the sealant, resulting in an equalization of the pres-

sure. Once the weather changed and suddenly cold

rain fell on the box, the temperature and pressure of

the box dropped, sucking in moist air. Thus, during

every cycle the moisture kept steadily increasing. Af-

ter about 6 months to a year this has caused the motes

to stop working. In the end, we have decided to use

semi-sealed acrylic tubes with an open bottom, which

are always in a temperature and humidity equilibrium

with the outside.

Power Usage. Battery consumption was in impor-

tant factor, and the typical analog sensors required a

current draw of about 10mA. In environmental sci-

ence the can afford a modest data collection rate, so

in practice we ended up using a 10ms/sensor sampling

interval, taken at every 20 minutes. Even with 8 sen-

sors this only amounts to about 100µA for an 9-sensor

mote, and about 25µA for a mote with 2 sensors.

Another lesson we learned that operating the

motes at 3.3V leaves too small margin for the battery

depletion. Due to temperature variations the voltages

of the LiPo batteries ﬂuctuate in excess of 100mV

during atypical daily cycle. The dropout voltage on

a typical voltage regulator does not leave much head-

room for operating at 3.3V, so we have gone to 3V op-

erating voltage and a very low dropout voltage chip,

giving us a much longer battery life.

Deployment Tradeoffs. We have also learned that

requiring a full peer-to-peer network with real-time

Internet connectivity so that we can assess the sen-

sors’ status at any moment makes practical deploy-

ments much too complex. While this was used for

our big campaigns, e.g. Cub Hill in Baltimore, many

other use cases require just a few sensors but spread

out in many small clusters over a wide area. The price

of sensors come into play as well. In the dense de-

ployment we had many CO

sensors, with a typical

cost of a few hundred to a few thousand dollars – it

was good to know that they were in place. The dense

deployment use case was big and expensive enough

that we have permanent internet gateway, with access

to line power and a cable connection.

The wide and shallow deployments typically re-

quire only a few temperature and moisture sensors,

placed at various depths, typically 5-10-15 cm. The

moisture sensors are typically around $100, and thus

deploying them on a dedicated mote with a mini-

Toast temperature sensor is a reasonable deployment

choice. In this setup we can operate the motes in a

data logger mode, waking up the whole network about

only once a month and using the mesh communica-

tion to incrementally download the new data. This

mode saves power, thus enabled us to use even smaller

batteries, making the whole assembly much smaller.

SENSORNETS 2021 - 10th International Conference on Sensor Networks

Software Infrastructure. It was also extremely im-

portant that we adopted the well-tested TinyOS plat-

form even when we moved to build our customized

Gen 3 hardware. This was the reason why we stayed

with the MSP430 based chip, but with the embedded

multi-band radio, enabling us to go to 900MHz.

5 SOIL MOISTURE RESULTS

At the end of the day, what matters is whether these

systems can provide insight into physical systems that

would otherwise be impractical to attain. The rich

dataset allows for a variety of analyses. For instance,

we can compare conditions in different land use-land

cover types using daily or monthly means and sea-

sonal trends (Figure 10). Interestingly, monthly mean

values in forest and grass were very similar, while val-

ues in the “other” category were consistently lower.

Figures 9 and 10 illustrate what makes WSNs so

well-suited for this domain. Even though the averages

are similar, the individual sampling locations varied

a great deal. Moreover, our dataset shows how soil

moisture responds to rainfall events at ﬁne tempo-

ral resolution, grouping locations by ground cover.

Not only can we see clear differences between the

cover types, we can capture the small-scale hetero-

geneity between sampling locations under the same

cover. More detailed analysis of these data and the

interactions on soil temperature and soil moisture are

discussed in (Savva et al., 2013). Spatial analysis al-

lows us to observe the soil response to extreme events

(Figure 10), to explore temporal stability of spatial

structure (Savva et al., 2013), and to interpret soil

fauna distribution (Szlavecz et al., 2011) in this res-

idential neighborhood. Soil moisture data, along with

continuous soil CO

concentration data collected by

the WSN can be used to build a model for soil CO

efﬂux in urban forest and grass (Chun et al., 2011).

Combined with the urban CO

ﬂux tower data this in-

formation is then used to build a coupled urban carbon

and water cycle model (John Hom, pers. comm.).

The primary objective of the Life Under Your Feet

project was to design and build an end-to-end wire-

less sensor network (WSN) for monitoring the soil

ecosystem and to test the system in a variety of en-

vironmental conditions. Similar efforts have been

reported (Cardell-Oliver et al., 2005; Sikka et al.,

2006; Ramanathan et al., 2009), or still ongoing

(http://soilscape.eecs.umich.edu/). To our knowledge,

at the time of the deployment, the Cub Hill project

was the largest and it was the longest experiment of its

size. In general, our deployments met their key goals:

they survived for long periods in the ﬁeld, while reli-

Figure 9: Local variation of soil moisture (VMC) at two

land cover types in the Cub Hill residential neighborhood

in Baltimore County, MD, between March and June 2009.

A subset of locations are shown. Each cell represents daily

average values at 10 cm depth. Daily precipitation values

are also shown.

Figure 10: Response of soil moisture to extreme precipita-

tion at Cub Hill. Spatial map shows the change in volumet-

ric soil moisture content between hours before rain and dur-

ing peak precipitation intensity.measurements were taken at

10 cm depths. A: regular rain event, B: Hurricane Irene.

ably collecting large volumes of data. The data can be

utilized for a variety of projects from building hydrol-

ogy models to identifying biological activity hotspots.

A wireless sensor network can be used for validation

of remotely sensed soil moisture data. As technology

advances, the system keeps evolving. However, fur-

ther development of hardware will always have to op-

timize among conﬂicting demands: increasing scale

of deployment, while keeping the system reliable, and

keeping the cost low. The Gen-3 platform and soft-

ware addresses many of the lessons we learned from

previous deployments, and the cost of one mote ($20)

is now negligible compared to the cost of the sensors

themselves. Ultimately the goal of such approach is

to gain insight to the spatio-temporally complex and

extremely species rich soil ecosystem.

Wireless Sensor Network for in situ Soil Moisture Monitoring

6 THE GEN-3 HARDWARE

The Bacon. The Breakfast Suite has several modu-

lar components. The main unit is the so called Bacon

board, which has the wireless transmitter. We decided

to use the CC430F5137IRGZ MCU, which has sev-

eral useful properties. It has the MSP430 core, im-

mediately providing compatibility with the existing

TinyOS stack. It has a multiband radio transmitter

supporting the 900Mhz band, it still uses the ZigBee

packet format on all frequencies, has more processing

power and eight times more memory than the original

MSP430 (Figure 11). Furthermore, it has ultra-low

power consumption. We have also placed a 64 Mbit

serial ﬂash memory on board (Numonyx M25P64-

VME6G) for buffering the acquired samples. We are

using a low dropout LDO providing 3V of stabilized

voltage (MCP1700T-3002E/TT, with a dropout volt-

age < 25mV at a 25mA current).

Figure 11: The CC430-based Bacon node, our Gen-3 radio

platform.

The board has options for an on-board 1/2 AA

3.7V LiPo battery as well as it has an external battery

connector. The boards have been also designed to ac-

cept a radio power ampliﬁer, and an external antenna,

but we have only populated about 100 boards with

that option. These motes can serve as longer range

relay nodes. We have an on-board analog switch turn-

ing the ﬂash power off when it is not required, saving

additional current in deep sleep mode. We have an ad-

ditional 10-pin JTAG connector, for deep debugging

of the system.

We have two sensors on board, one is a linear ther-

mistor (Microchip MCP9700AT-E/LT) running in 12-

bit, 0.1 degC sensitivity, the other is an Avago APDS-

9007 ambient light sensor. We are using two mi-

cro USB connectors, one to connect to a programmer

board, the other to connect to the multiplexers. The

board is also actively monitoring (and saving) the bat-

tery voltage.

The Toast. The Toast board is an intelligent multi-

plexer/interface for up to eight analog sensors. It is

based on a light-weight MSP430F235TRGCR MCU,

which takes commands from the Bacon node, turns

the sensors on for measurement for a predetermined

period (10ms), acquires an analog voltage measure-

ment for each enabled channel and uploads the data

to the Bacon for storage in ﬂash. The communication

between the Bacon and the Toast is using the I2C pro-

tocol. Several Toast modules can be daisy chained.

Figure 12: The CC430-based Toast and miniToast analog

interface/multiplexer boards.

The miniToast. The miniToast is a compact, sim-

pliﬁed version of the Toast for simple mote con-

ﬁgurations, when only a single analog sensor is

needed. It includes two additional internal channels,

one is a high precision analog thermistor (US Sen-

sor PS103J2), the other is a measurement of the sup-

ply voltage to the sensors. The miniToast module is

placed in epoxy so that it can be buried underground,

together with the additional external sensor.

The USB Programmer. This board enables the pro-

gramming of the Bacon and Toast boards from a PC

host. It contains an FTDI232R interface chip, and

an ADG715T 3.3V voltage regulator. For program-

ming the Bacon and Toast combination, we need to

ﬁrst connect a Bacon board to the programmer, and

then the Toast. One can update the ﬁrmware on each

boards, as well as add id numbers (barcodes) to both

devices as well as for the individual sensors. These

ID numbers are than contained in the various status

messages and data records.

Figure 13: The USB programmer board, with and without

its 3D-printed enclosure.

SENSORNETS 2021 - 10th International Conference on Sensor Networks

7 THE GEN-3 SOFTWARE STACK

Concurrent Transmission w. Forwarder Selection.

One major goal of the Gen-3 platform is to to avoid

the inherent limitations in selecting data transmission

routes over unreliable hardware and ﬁckle link dy-

namics. The Koala data collection system(Musaloiu-

E et al., 2008) uses a central source-routing mecha-

nism where the base station decides the route of each

download by analyzing the probes sent from nodes

in the network. However, when the batteries of a

few nodes in the network are depleted, the base sta-

tion often fails to ﬁnd a reliable route, since a node

on a low voltage may still send wake-up probes and

be included in the forwarder route, even if it could

no longer reliably transmit data. Consequently, fre-

quent download retries cause high duty cycles that

drain the batteries even faster. One possible solution

to the problem is to use detailed node status infor-

mation (e.g. battery voltage, humidity in the enclo-

sure) in the route selection process, yet obtaining de-

tailed link quality information demand signiﬁcant en-

ergy and computation costs, placing a high burden on

resource-constrained motes.

In Gen-3 network, we take an alternative method

by leveraging non-destructive interference to create

redundant simultaneous delivery paths to multiply the

probability of reaching the end node over an unreli-

able network. Under this approach, we use simple hop

counts to identify a subset of network that roughly lies

between the source and destination as the set of poten-

tial forwarders. By using the precisely-timed radio on

board bacon motes to schedule transmissions to occur

simultaneously, we can send a data packet over this

set of nodes with very low risk of loss due to interfer-

ence. This protocol, which we call concurrent trans-

mission with forwarder selection (CXFS), requires

neither the overhead of calculating a single optimal

path between the the source and the sink nor the costs

of indiscriminately forwarding the message across the

entire network as in a simple ﬂooding method. This

balance between simplicity and selectivity helps the

sensor network attain both high throughout and low

energy consumption.

Multi-tiered Network Hierarchy. Studies in soil

ecology sometimes require comparing distinct habi-

tats over a relatively large region while simultane-

ously capturing the heterogeneity within each habitat.

This conﬁguration often leads to the need of setting

up multiple sensors in small clusters. To account for

the spatial patchiness of the deployment, the entire de-

ployment is divided into one or more patches, where

each patch contains a router mote equipped with long

range transceivers and a cluster of leaves motes. A

router independently collects data from the leaves in

its patch, and a basestation periodically downloads

data from all routers. This multi-tiered structure

avoids the need to install multiple relay motes be-

tween patches and allows energy-efﬁcient sampling

over heterogeneous landscape.

Figure 14: Multi-tiered network segments.

Data Pipeline and Meta Data Management. With

a hardware interface to analog sensors and a reli-

able networking structure in place, the next step is

to build an end-to-end data pipeline from sensors to

database. On each mote, we divide the ﬂash stor-

age into a setting storage and a log records storage.

The former is used to store customized conﬁgurations

such as sampling interval and unique id, whereas the

later is used by both the Bacon and Toast modules to

record samples collected. An auto-push component

keeps track of data recorded in the ﬂash, pushes the

data from the ﬂash to the network stack when needed,

and handles recover requests for missing data based

on speciﬁc cookie and length parameters. If cases

where log record packets are forwarded from leaf

motes to a router mote, they will be temporarily stored

in the router’s log with a TUNNELED MESSAGE

record appended. During a download, the base sta-

tion will ﬁrst attempt to receive all information from

the router’s log, and it will retrieve log records from

the leaf motes only if the information is absent from

the router.

Figure 15: Interactions between major components of the

software.

Wireless Sensor Network for in situ Soil Moisture Monitoring

A Python program on the PC will process each

log record packet when they arrive from the base sta-

tion. The program ﬁrst stores a safe copy of the raw

packet in a binary table, and then parses the packet

to extract individual records. Each record is iden-

tiﬁed by its type and dispatched to a corresponding

decoder. The decoder then decodes the record to an

ASCII string, extracts the ﬁelds from the string, and

inserts the ﬁelds into a matching table in the database.

The above description highlights the ﬂow of data

from the sensor to the database. Yet for the system to

collect scientiﬁcally meaningful data, it is also critical

to systematically track deployment metadata such as

the type of the sensor, the channel assignment, and the

pairing between bacon and toast boards. In the third

generation software, metadata tracking is enforced at

installation stage by a dedicated labeler program. The

manufacturer ID, barcode ID, and sensor assignments

are stored in both a local database and the setting stor-

age volumes of the Bacon and Toast boards. The pro-

gram will not generate data reports unless the meta-

data associated with the deployment is present.

Figure 16: Data ﬂow in PC.

8 THE GEN-4 SOFTWARE

PLATFORM

Integrated GUI. The completion of the Gen-3 plat-

form enabled a wide range of environmental moni-

toring opportunities. After the development, the sys-

tem has been test-deployed in SERC and yielded

promising scientiﬁc data. Despite the technical suc-

cess, our ﬁeld tests revealed several stability issues

that required a consistent involvement of technicians,

and the platform’s operational complexity hindered

its wide adoption by soil scientists. The current in-

terface only allows a user to download all the data

from a patch of sensors, yet a user may be interested

in retrieving data from a speciﬁc mote to obtain data

with higher granularity. Furthermore, the Gen-3 sys-

tem does not perform quality-control measures on the

data collected or alert the user of any potential issues,

so it might take an end-user several months to real-

ize the issue while the critical window of research is

wasted.

Figure 17: Prototype design of the generation-4 software

platform. The new software will provide a convenient inter-

face for users to interact with the sensor network.

These limitations have prompted us to redesign

the high-level, user facing components, and to cre-

ate an integrated software platform to streamline de-

vice management, ﬁeld deployment, quality control

and data reporting. A recent Python-based web-app

framework (Dash) provides us the ability to imple-

ment such a platform with comparative ease. Through

a web-based graphic interface, the user will be ﬁrst

guided through the process of labelling motes and

sensors, setting radio channels and sampling inter-

vals, and managing deployment-wide metadata (Fig-

ure 17). When in ﬁeld, the user can manage the de-

ployment, edit the metadata for each unit, or add or

remove a device if needed. Through an interactive

portal, the user can download data from the entire de-

ployment, a patch within the deployment, or a speciﬁc

mote as desired. Progress bars and a color grid on

the screen will display the status of download. When

a download completes, the status of each mote, in-

cluding battery voltage and percentage of samples re-

trieved, will be reported. The platform will then run

a quality control algorithm to check the database for

missing, duplicate, or anomalous data, and alert the

user if the problem is not recoverable by automatic

correction. This platform will also provide customiz-

able data visualization and report functions, allowing

a user to generate spreadsheets and graphics with only

a few clicks “in situ”. Finally, the platform will also

provide optional user authentication and data upload

SENSORNETS 2021 - 10th International Conference on Sensor Networks

function, allowing a user to share data while control-

ling access.

Data Hosting and Analytics. Our data will be hosted

on a scalable collaborative data analytics platform,

which integrates about 6PB of ﬁle storage, several

PB of databases, automatically generates user logs,

and provides about 100 virtual machines for interac-

tive and batch computations. All code, database ta-

bles and data products are shareable at different gran-

ularities (users, groups, world). The system today is

supporting more than 70 different projects in a variety

of science domains, from astronomy to turbulence.

The analyses can also use server-side Jupyter/iPython

notebooks. These are preconﬁgured with database

access, thus users can run their SQL queries out of

Python. The Jupyter environment also enables Matlab

and R. All major machine learning toolkits are avail-

able on the system, including several V100-based

GPU servers. Most of the sensor data collected so

far has already been moved to the SciServer, to alle-

viate the need for maintaining a separate data system

for our project. We will use this system for all future

data hosting and analysis.

9 SUMMARY

Soil moisture is a key driver of many soil physical and

biogeochemical processes, and an important compo-

nent of the water cycle. In the US several platforms

exist to monitor soil moisture, but a collective effort

to establish a National Soil Monitoring Network has

emerged only recently. Currently neither the exist-

ing networks nor the SMAP satellite provide data on

soil moisture in urban environments. Due to the enor-

mous heterogeneity of the urban landscape, the diver-

sity of urban soils from remnants of naturally devel-

oping soils to engineered substrates, and various man-

agement practices, physico-chemical properties can-

not be simply derived or modeled from natural soil

forming factors and thus have to be measured directly.

An ideal urban soil monitoring network would

consist of many small sensor clusters, and should de-

ployed on all major land use types. Lawns parks are

a permanent feature in the urban-suburban landscape,

but additional land uses, vacant lots, community gar-

dens, remnant patches of regional biome, and other

unmanaged areas should also be monitored. Wireless

data collection allows less disturbance and less intru-

sion of the sites, which are often privately owned.

In Baltimore, in our next major deployment

project, we are planning to monitor over 200 sites,

called “parcels”. A parcel is a unit of investigation

and it is managed by a single entity: a company, orga-

nization, or an individual. The plan is to deploy a set

of soil temperature and moisture sensors at different

depths, and in several patches within the parcel, for

example lawns and planting beds.

There is an increasing need for this type of in-

formation not only in the scientiﬁc community but

in various sectors of the community. For instance,

cities are relying on, and even by policy requiring, al-

ternative stormwater management technologies, from

green roofs, to living walls, bioswales, and bioreten-

tion systems. Yet, there is little opportunity for de-

sign and engineering professionals to evaluate their

post-construction performance in order to improve fu-

ture projects and revise sustainable design guidelines.

Sensor technology would enable designers, engineers,

and scientists to collect the data required to meet the

needs of urban populations by improving the develop-

ment high-performance landscapes and green infras-

tructure projects that are multi-functional in protect-

ing environmental quality, while providing social and

economic beneﬁts in urban centers.

Our Gen-3 platform can be ideally used for the

wide area deployment described above. The main

challenge for such wide area deployments is to make

the whole conﬁguration, deployment and ongoing

data collection process much more streamlined and

simple, and enable a variety of users, including soil

scientists, urban ecologists, practitioners, as well as

community scientists to be in full control of the whole

experiment. In the same spirit, the analysis of the data

must be made much simpler by using an integrated

collaborative analysis environment, which integrates

databases, ﬁle stores and Jupyter notebooks.

In this paper we presented the history of develop-

ing several generations of inexpensive sensors to mea-

sure soil properties. Our architecture evolved a lot

over the years and now we have an inexpensive and

robust hardware/software architecture that can collect

data at an extremely low power consumption, and can

be used in deployments such as the one described

above.

ACKNOWLEDGEMENTS

The Life Under Your Feet project started out with

seed grants from Microsoft Research and the Seaver

Institute. We are most grateful to Jim Gray who was

instrumental in getting this project off the ground,

and to Dan Fay and Tony Hey for their continuing

support and encouragement. Later this project was

partially supported by several grants from the Na-

tional Science Foundation (NSF IDBR-0754782 and

NSF DEB-0423476, NSF-ERC EEC-0540832). The

Wireless Sensor Network for in situ Soil Moisture Monitoring

Gordon and Betty Moore Foundation sponsored the

development of the Breakfast Suite. Undergradu-

ates Josh Cogan, Julia Klofas and Justin Silverman

helped building and testing the WSN. Mike Liang

contributed to the software development; Jordan Rad-

dick, Taesung Kim and Luis Grimaldo were instru-

mental in developing Grazor. Thanks are due to the

Maryland Department of Natural Resources for al-

lowing us to use their site as testbed, and to John Hom

of the US Forest Service for hosting the gateway com-

puter in his lab at Cub Hill.

REFERENCES

Cardell-Oliver, R., Kranz, M., Smettem, K., and Mayer, K.

(2005). A reactive soil moisture sensor network: De-

sign and ﬁeld evaluation. International journal of dis-

tributed sensor networks, 1(2):149–162.

Carlson, D., Gupchup, J., Fatland, R., and Terzis, A. (2010).

K2: a system for campaign deployments of wire-

less sensor networks. In International Workshop on

Real-world Wireless Sensor Networks, pages 1–12.

Springer.

Cerpa, A., Elson, J., Estrin, D., Girod, L., Hamilton, M.,

and Zhao, J. (2001). Habitat monitoring: Application

driver for wireless communications technology. ACM

SIGCOMM Computer Communication Review, 31(2

supplement):20–41.

Chun, J., Szlavecz, K., Ferrer, D., Bernard, M., Pitz, S.,

Hom, J., and Zaitchik, B. (2011). Estimation of soil

co2 efﬂuxes from suburban forest and lawn using con-

tinuous measurements of co2 proﬁles in soils and a

process-based model. AGUFM, 2011:B11A–0462.

Chust, G., Pretus, J., Ducrot, D., Bedos, A., and Deharveng,

L. (2003). Response of soil fauna to landscape hetero-

geneity: determining optimal scales for biodiversity

modeling. Conservation Biology, 17(6):1712–1723.

Ellis, E. C., Wang, H., Xiao, H. S., Peng, K., Liu, X. P.,

Li, S. C., Ouyang, H., Cheng, X., and Yang, L. Z.

(2006). Measuring long-term ecological changes in

densely populated landscapes using current and his-

torical high resolution imagery. Remote Sensing of

Environment, 100(4):457–473.

Ettema, C. H. and Wardle, D. A. (2002). Spatial soil ecol-

ogy. Trends in ecology & evolution, 17(4):177–183.

Gupchup, J., Carlson, D., Mus

aloiu-e, R., Szalay, A., and

Terzis, A. (2010). Phoenix: An epidemic approach

to time reconstruction. In European Conference on

Wireless Sensor Networks, pages 17–32. Springer.

Gupchup, J., Mus

aloiu-e, R., Szalay, A., and Terzis, A.

(2009). Sundial: Using sunlight to reconstruct global

timestamps. In European Conference on Wireless Sen-

sor Networks, pages 183–198. Springer.

Musaloiu-E, R., Liang, C.-J. M., and Terzis, A. (2008).

Koala: Ultra-low power data retrieval in wireless sen-

sor networks. In 2008 International Conference on

Information Processing in Sensor Networks (IPSN

2008), pages 421–432. IEEE.

Pickett, S. T., Cadenasso, M. L., Grove, J. M., Boone, C. G.,

Groffman, P. M., Irwin, E., Kaushal, S. S., Marshall,

V., McGrath, B. P., Nilon, C. H., et al. (2011). Ur-

ban ecological systems: Scientiﬁc foundations and a

decade of progress. Journal of environmental man-

agement, 92(3):331–362.

Pickett, S. T., Cadenasso, M. L., Grove, J. M., Groffman,

P. M., Band, L. E., Boone, C. G., Burch, W. R., Grim-

mond, C. S. B., Hom, J., Jenkins, J. C., et al. (2008).

Beyond urban legends: an emerging framework of ur-

ban ecology, as illustrated by the baltimore ecosystem

study. BioScience, 58(2):139–150.

Pouyat, R. V., Szlavecz, K., Yesilonis, I. D., Groffman,

P. M., and Schwarz, K. (2010). Chemical, physical,

and biological characteristics of urban soils. Urban

ecosystem ecology, 55:119–152.

Ramanathan, N., Schoellhammer, T., Kohler, E., White-

house, K., Harmon, T., and Estrin, D. (2009). Suelo:

human-assisted sensing for exploratory soil monitor-

ing studies. In Proceedings of the 7th ACM Confer-

ence on Embedded Networked Sensor Systems, pages

197–210.

Robertson, G. P., Gross, K. L., Caldwell, M., and Pearcy,

R. (1994). Assessing the heterogeneity of below-

ground resources: quantifying pattern and scale. Ex-

ploitation of environmental heterogeneity by plants:

ecophysiological processes above-and belowground,

pages 237–253.

Savva, Y., Szlavecz, K., Carlson, D., Gupchup, J., Szalay,

A., and Terzis, A. (2013). Spatial patterns of soil

moisture under forest and grass land cover in a subur-

ban area, in maryland, usa. Geoderma, 192:202–210.

Sikka, P., Corke, P., Valencia, P., Crossman, C., Swain, D.,

and Bishop-Hurley, G. (2006). Wireless adhoc sensor

and actuator networks on the farm. In Proceedings of

the 5th international conference on Information pro-

cessing in sensor networks, pages 492–499.

Szlavecz, K., Warren, P., and Pickett, S. (2011). Biodiver-

sity on the urban landscape. In Human Population,

pages 75–101. Springer.

Tenenbaum, D., Band, L., Kenworthy, S., and Tague, C.

(2006). Analysis of soil moisture patterns in forested

and suburban catchments in baltimore, maryland, us-

ing high-resolution photogrammetric and lidar digital

elevation datasets. Hydrological Processes: An Inter-

national Journal, 20(2):219–240.

Yesilonis, I. D., Pouyat, R., Russell-Anelli, J., and Powell,

E. (2016). The effects of landscape cover on surface

soils in a low density residential neighborhood in bal-

timore, maryland. Urban ecosystems, 19(1):115–129.

Young, I. M. and Crawford, J. W. (2004). Interactions and

self-organization in the soil-microbe complex. Sci-

ence, 304(5677):1634–1637.

SENSORNETS 2021 - 10th International Conference on Sensor Networks