Computers and Electronics in Agriculture 205 (2023) 107591

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

Original papers

Placement and drone flight path mapping of agricultural soil sensors using
machine learning
Payton Goodrich a ,∗, Omar Betancourt a , Ana Claudia Arias b , Tarek Zohdi a
a
University of California Berkeley, Department of Mechanical Engineering, Berkeley, 94720, California, USA
b
University of California Berkeley, Department of Electrical Engineering and Computer Science, Berkeley, 94720, California, USA

ARTICLE INFO ABSTRACT

Keywords: Accurate soil data, which can be collected with agricultural sensors spaced at the half-variogram range, is
Agricultural sensors crucial information for precision agriculture. Drones offer a unique advantage over other existing methods to
UAVs sample data from soil sensors because of the high density of sensors required to gather spatially granular data
Genetic algorithms
to capture soil variability. To determine the placement of sensors within an agricultural field, a novel sequential
Agent-based model
gap-reduction algorithm that minimizes sensor overlap was developed and used to assign sensor placement for
Geostatistics
four types of agricultural fields. A genetic algorithm optimized multi-agent flight paths for scanning sensors in
a simulated agricultural field using a robust agent-based model. The locations of soil sensors for a simulated
400 m center-pivot irrigation field were used to determine flight paths for swarms of 1-8 drones. Increasing the
number of drones in the swarm had a negligible effect on total energy expenditure but reduced the time to scan
the sensors from 19 min to less than three minutes. The proposed sequential gap reduction algorithm maximizes
the coverage of any arbitrarily-shaped agricultural field with fewer sensors than a grid-based distribution in
most cases. The proposed multi-agent flight path mapping can effectively and efficiently generate flight paths
for variable numbers of drones to scan all sensors. The framework proposed here can be utilized and expanded
for variable rate irrigation, precision application of biological control agents, and smart farming.

1. Introduction water consumption by 26.3% (Hedley and Yule, 2009). Meanwhile,

fixing nitrogen from the air to produce fertilizers is an extraordi-
Precision agriculture offers a pathway to increase crop yield while narily energy-intensive process (Gellings and Parmenter, 2016) and
reducing water consumption, carbon footprint, and chemicals leaching accounts for nearly 2% of the U.S.’s annual CO2 emissions (Ritchie
into groundwater. Precision agriculture is the practice of collecting and Roser, 2020). Crop plants recover only 30%–50% of nitrogen in
spatial and temporal data in an agricultural field to match the inputs fertilizers (Cassman et al., 2002), which means that over half of the
to the site-specific conditions (Diacono et al., 2013). While industrial nitrogen becomes a potential source of environmental pollution, such
agriculture seeks to maximize crop yield, there is also the consideration as groundwater contamination, eutrophication, acid rain, ammonia
of maintaining a healthy ecosystem. Fortunately, these are not compet- redeposition, and greenhouse gases (Ladha et al., 2005). Fortunately,
ing interests; numerous case studies have demonstrated that adopting precision agriculture practices have demonstrated an increase in ni-
precision agriculture techniques increases crop yield while lessening
trogen use efficiency (Ahrens et al., 2010; Gupta and Khosla, 2012),
detrimental environmental effects (Diacono et al., 2013; Ahrens et al.,
thereby reducing both the production volume of fertilizer and the
2010; Sela et al., 2018). Consider first the use of irrigation in agri-
amount that is polluted into the environment.
culture, which accounts for approximately 36.7% of the freshwater
Accurate soil data is crucial information for precision agricul-
consumption in the U.S. (Dieter, 2018), 65% in China (Dalin et al.,
ture (Robert et al., 1995). In particular, the moisture content (SU et al.,
2014), and 77% in New Zealand (Hedley and Yule, 2009). This is
2014) and the concentration of various chemical analytes (Robertson
in part because crops need a large amount of water to grow. For
example, high-production maize crops require 600,000 gallons of water et al., 1997; Shahandeh et al., 2005; Anthony et al., 2012) in soil
per acre per season—about the same volume as an Olympic swim- have a significant influence on crop health and yield. These proper-
ming pool (Comis, 2011). However, adopting precision agriculture ties vary considerably over short distances, which begs the question:
practices such as variable-rate irrigation methods has proven to reduce What spatial density does soil need to be sampled at to capture soil

∗ Corresponding author.
E-mail address: paytongoodrich@berkeley.edu (P. Goodrich).

https://doi.org/10.1016/j.compag.2022.107591
Received 21 June 2022; Received in revised form 16 December 2022; Accepted 21 December 2022
Available online 27 December 2022
0168-1699/© 2022 Elsevier B.V. All rights reserved.
P. Goodrich et al. Computers and Electronics in Agriculture 205 (2023) 107591

variability? Half of the spatial range – referred to hereafter as the ‘half- coordinate, and its size corresponds to an area. Three methods of
variogram range’ (Bachmaier and Backes, 2008) – can be used as a expressing an agricultural field in a digital format are discussed.
‘‘rule-of-thumb’’ to account for the spatial dependency of agricultural For agricultural fields that a simple geometric shape can approxi-
measurements (Kerry et al., 2010). The variance of a measurand, 𝑧, as mate – such as a rectangular field or a center-pivot irrigation field –
a function of distance is empirically given by: expressing the farm digitally is trivial. For a rectangular-shaped field,
⃗ [ (
the space is discretized into a grid of uniform pixels with dimensions
∑
𝑁( ℎ) ) ( )]2
⃗ = 1⋅ 1 ⃗ − 𝑧 𝑥⃗𝑖 proportional to the length and width of the physical field. For a center-
𝛾̂ (ℎ) 𝑧 𝑥⃗𝑖 + ℎ (1)
2 𝑁(ℎ) ⃗ 𝑖=1 pivot irrigation field, the field is bounded by a square grid of uniform
( ) ( ) pixels. Then, each pixel in the grid is tested to determine if the pixel’s
⃗ are the measured values of
where 𝛾̂ is the variance, 𝑧 𝑥⃗𝑖 and 𝑧 𝑥⃗𝑖 + ℎ coordinates are equal to or less than the physical field radius. This
the measurand 𝑧 at 𝑁(ℎ) ⃗ pairs of comparisons separated by the vector technique is demonstrated in Fig. 1A for a rectangular-shaped field and
⃗ Numerous studies have determined the spatial range of various soil
ℎ. Fig. 1B for a center-pivot irrigation field.
properties in various soil conditions (Robertson et al., 1997; Shahandeh When the boundaries of the agricultural field are not regularly
et al., 2005; Anthony et al., 2012; Kerry et al., 2010; Longchamps shaped, the field is defined by a list of consecutive coordinate points
et al., 2015), which demonstrates the fact that the half-variogram that form an enclosed shape when piecewise connected by polynomial
range itself varies depending on the geographic location and sampling curves. A ray tracing algorithm is adopted to determine whether or not
method (Kerry et al., 2010). Recently, Longchamps & Khosla analyzed a pixel is inside or outside of this boundary (Kumar and Bangi, 2018).
and tabulated the spatial ranges of numerous soil properties reported Given an enclosed boundary and a point in space, if one were to draw
in literature (Longchamps and Khosla, 2017), which can be used as an infinite vector in any direction originating from that point, it will
informed estimates for spacing sensors when no other information intersect the boundary an odd-numbered amount of times if-and-only-
about the soil is known. if the point is within the enclosed space, which is shown in Fig. 1C.
Geostatistically representative soil data can be collected with agri- This holds for all points in space except for points on the boundary,
cultural sensors spaced at the half-variogram range (Kerry et al., 2010). which must be determined explicitly. This way, the coordinates of each
The data gathered from these sensors can inform management tech- pixel are used as a point to determine if a pixel is inside the boundary
niques such as variable-rate technologies, which adapt to the hetero- and append it to a list. A more detailed explanation of this method is
geneities of an agricultural field and thus enable site-specific manage- described in the Supplementary Material.
ment (Kanter et al., 2019; Evans, 2001; Evans et al., 2013; Lo et al., Finally, satellite or drone visible-spectra images of agricultural land
2017). For example, farmers could tailor their nitrogen and water man- are already stored in a digital, pixelized format. Such images and
agement to site-specific conditions (Finger et al., 2019), which would, datasets are widely available from Google Earth, NASA Earth Observa-
tory, or the USDA cropland data layer, to name a few. Computer vision
in turn, reduce nitric oxide emissions, increase yields, and reduce
techniques can differentiate the arable land on a field from obstructions
fertilizer use (Ahrens et al., 2010). Other researchers have investigated
(such as roads, buildings, trees, and ponds) and store those pixels in
the use of machine-learning algorithms as tools for decision-making
a list (TOMBE, 2020; Akbari et al., 2021). This process is shown in
in precision agriculture (Jourdan and de Weck, 2004; Chlingaryan
Fig. 1D.
et al., 2018; Akbarzadeh et al., 2014; Patrício and Rieder, 2018; Sun
In all cases, it is important to note the physical dimensions that
et al., 2022). However, to the best of our knowledge, no authors have
a single pixel represents. It should also be noted that because each
optimized agricultural soil sensor placement using the half-variogram
method requires discretization of the field, the results are approxima-
range to inform the placement of sensors.
tions whose accuracy increases proportionally to the number of pixels
Because many sensors need to be distributed across an agricul-
used.
tural field to acquire granular enough data to capture soil variability,
drones offer a unique advantage over other existing methods to sample
2.2. Optimized sensor placement
data from the sensors (Sørensen et al., 2017). With drones and drone
accessories becoming less expensive, using multiple drones to simulta-
The optimal layout of sensors in an agricultural field is achieved
neously map sensors has become an attractive route to efficiently gather
when, using the fewest number of sensors possible, all points in the
data (Tosato et al., 2019; Bertalan et al., 2022; Pereira et al., 2022).
field are statistically represented by the data collected by sensors in
Machine-learning algorithms are a promising approach for generating
that field. For a given sensor, the data collected from that sensor is
flight path maps due to their ability to solve highly non-convex prob-
statistically significant for all points within a radial distance equal to
lems rapidly and even operate in real-time as a digital twin (Zohdi,
the half-variogram range of that sensor (Kerry et al., 2010). Thus,
2021b,a, 2022).
if one considers an agricultural field as a two-dimensional collection
This work develops a sequential gap reduction (SGR) algorithm that of pixels described previously, one can model sensors as circles with
determines an optimal distribution of soil sensors across an agricultural a radius equal to the half-variogram range. Using this definition for
field. The SGR algorithm is used to solve optimal sensor distribution optimal sensor placement, the problem is similar to the circle pack-
using the half-variogram range as the basis in four field geometries: ing problem. Circle packing (or more broadly, ‘‘object packing’’) is
a circular field, a rectangular field, a field with both circular and a well-researched area in mathematics that has many practical ap-
rectangular features, and a field shape determined from an image. plications (Hifi and M’hallah, 2009). Object packing aims to fit as
Finally, a methodology for sampling the sensor measurements using many of some objects within a domain as possible without any overlap
UAV swarms is proposed and characterized. between the objects. There are several algorithms that aim to optimize
object packing, such as random sequential addition (Widom, 1966),
2. Theory the Metropolis algorithm (Metropolis et al., 1953), and various particle
growth schemes (Donev et al., 2005a,b). The limit of packing efficiency
2.1. Digital expression of an agricultural field for equal-size circles in two dimensions is about 91% for a hexagonal
grid (Steinhaus, 1999). While circle-packing nearly describes the model
Similar to how an agricultural field can be defined in the real problem, there is one major caveat: no physical justification prevents
world as a geographic area at a location, a digital representation – or the circles (sensors) from overlapping. This ‘soft boundary’ makes it
‘simulation’ – of an agricultural field can be defined as some number of possible to achieve 100% coverage of the domain by allowing overlap.
discrete pixels, where each pixel’s position corresponds to a geographic If the only objective was to maximize the effective areal coverage of

2
P. Goodrich et al. Computers and Electronics in Agriculture 205 (2023) 107591

depicting the evolution of sensor placement in an arbitrary field shape

is shown in Fig. 2B.

2.3. Data sampling

To make informed decisions about allocating resources for treating

the soil, one must sample or collect the soil sensor data periodi-
cally throughout the growing season. When deployed, soil sensors are
partially or completely buried under the soil using common farming
equipment such as post augers or portable excavators. Depending on
the design of the sensor, there are several ways that the data might be
communicated. One method would be through a hardwire connection,
such as a USB or Serial cable. An underground cable network would
interfere with regular farming activities, but the sensors could be
sampled by temporary connecting through a data port. However, this
would require a data-reading device to come into direct contact with
the sensor and a moderate level of precision to plug into the port.
A simpler data transfer solution is integrating a transceiver into the
sensor and transmitting the data wirelessly using RFID, Bluetooth Low
Energy (BLE), Wi-Fi, or similar technologies. In this case, the data-
reading device would only need to be within the transmission radius
of the sensor to sample the sensor data. Generally speaking, the cost
and complexity of the sensor increase dramatically as the transmission
radius increases. For the technology to be practical for agricultural field
Fig. 1. Methods of defining agricultural fields in a digital domain. (A) Rectangular
operations, the associated costs of the sensor need to be minimal, so
fields were defined by discretizing the field into a grid of uniform pixels proportional short-range transmission technologies are more feasible at scale (Deng
to the length and width of the physical field. (B) Center-pivot irrigation fields were et al., 2020).
defined by superimposing a uniform grid of pixels over the field and appending all There are several options one can choose to sample the sensor data.
pixels whose coordinates are within the radius of the pivot arm to a list of field
One option is using ground machinery (e.g. tractors) fitted with readers
pixels. (C) Irregularly shaped fields were defined by superimposing a uniform grid
over the field and appending all pixels whose rays pass through an odd number of to receive sensor data as it traverses above them. Tractors are well
boundary lines. In this case, pixel 𝛼 has a ray that passes through a single boundary suited to traverse through muddy or uneven ground, however, they can
line, while pixel 𝛽 has a ray that passes through five boundary lines. (D) Digital images become costly as the fuel economy of such ground machinery can be
of agricultural fields are in a digital domain by definition but require computer vision poor. For a typical Fendt 1050 tractor model, the diesel consumption
techniques to determine which pixels in the image correspond to arable land.
can be eight gallons per hour, and covering a large agricultural field
may take the majority of a day (Hoy, 2021). This is not to mention the
cost of a manned operator. A much more attractive option would be
the field, then one could simply distribute sensors next to one another to incorporate unmanned vehicles, such as unmanned aerial vehicles
without discretion. However, the monetary cost of sensors, sensor (UAVs) or unmanned ground vehicles (UGVs). Unmanned vehicles
operation, and sensor maintenance makes this approach unreasonable, typically have lower energy consumption than typical ground farm
which motivates the stated objective of maximizing field coverage with machinery and, as the name suggests, operate remotely, minimizing
the fewest sensors possible. labor and operational costs. There are important distinctions to make
The SGR algorithm was developed and applied to place sensors between deploying ground-based and air-based vehicles. UGVs can
within the field to minimize the overlap of each sensor’s coverage typically withstand a larger range of weather conditions than UAVs
radius. The general process is as follows: but are limited to driving along the field rows to avoid damaging the
crops. On the other hand, UAVs can traverse directly from sensor to
1. INITIALIZE: Select a random pixel in the field and place a sensor
sensor above the crop canopy. Because the energy consumption of small
there
UAVs and UGVs are similar, the ability to traverse freely throughout a
2. GENERATE: Select a random pixel in the field
field makes UAVs more energy-efficient than UGVs. Assuming the soil
3. TEST: Check to see if a sensor placed at that pixel would overlap
sensors have a wireless transceiver integrated, the UAVs can sample
with any other sensor already placed in the field. If not, append
data from the sensor from anywhere within the transmission radius,
that pixel to the list of placed sensors within the field.
meaning the UAVs’ positioning around the sensor does not have to
4. SCORE: Compute the score of the design.
be precise, and slight perturbations to the UAVs’ position caused by
5. ITERATE: If the score is below the threshold, iterate through the
weather will have a minor effect on the overall performance.
algorithm.
It is possible that a coordinated effort of multiple drones can further
The outcome of this algorithm is that sensors are placed throughout reduce the time and energy required to sample data from each sensor
the field such that sensors are placed in the largest gaps between in the field. Furthermore, it may be difficult for a single UAV to scan all
sensors. This is done by incrementing the acceptable distance between sensors due to battery life limitations. For these reasons, a flight path
sensors by a small amount, then making many attempts at placing optimization model was developed for single-drone and multi-drone
a sensor before repeating the process. The fitness of each placement swarms to investigate their trade-offs.
design was scored by the ratio of the number of field pixels within the
half-variogram range of a sensor to the total number of pixels in the 2.4. Flight path mapping
field. In other words, what percentage of the field area is within the
half-variogram range of one or more sensors? This process is repeated A robust agent-based model was developed that generates flight
until it is impossible to place a sensor outside the range of all other paths for each drone within a swarm to scan all sensors within a
sensors in the design, or until the field is completely covered. The simulated agriculture field. The simulations determine each drone’s
flowchart for this algorithm is shown in Fig. 2A, and a schematic aerial route for optimal flight path planning (Zohdi, 2019; Mueller

3
P. Goodrich et al. Computers and Electronics in Agriculture 205 (2023) 107591

Fig. 2. (A) Flowchart of the Sequential Gap Reduction algorithm used to place sensors in a digitized agricultural field. (B) Schematic depicting the action of the Sequential Gap
Reduction algorithm. (i) First, sensors are added to the field at randomly selected pixels within the field boundary such that there is no overlap between the sensors’ effective
coverage radius. The sensors are depicted as small black circles, and the sensor’s effective coverage area is depicted as a solid line encircling the sensor. (ii) Once it is no longer
possible to add any more sensors this way, the allowable minimum distance between sensors is decreased – represented by a dashed circle – while the effective coverage area of
the sensors remains the same. (iii) This process is repeated until all gaps have been filled and all pixels within the boundary are within the effective coverage radius of at least
one sensor.

et al., 2015). Each drone within the simulated framework, referred to where 𝑚𝑖 is the agent’s mass, 𝑎𝑖 is the agent’s acceleration, and 𝒏⋆ 𝑖 is
as an ‘agent’, has unique characteristics that determine how it interacts the agent’s propulsion vector.
with its surroundings, such as its environment and other drones. These The distance between an agent and other agents or sensors (here-
characteristic parameters take inspiration from the physics of molecular after referred to as ‘objects’) is imperative to calculate the propulsion’s
dynamics, where each agent is modeled as a point-mass particle that vector 𝒏⋆𝑖 . At each time step, the model calculates the Euclidean dis-
is attracted and repelled by other objects within the system (Zohdi, tance between objects, defined between agent 𝑖 with position 𝒓𝑖 and
2021b). The framework inputs are the field’s shape, the number of another object 𝑗 in the system at 𝑨𝑗 , as:
agents, and the sensor locations (targets). Depending on the field √
def
geometry and the locations of sensors within that field, the framework 𝑑𝑖𝑗 = ‖𝒓𝑖 − 𝑨𝑗 ‖ = (𝑟𝑖1 − 𝐴𝑗1 )2 + (𝑟𝑖2 − 𝐴𝑗2 )2 + (𝑟𝑖3 − 𝐴𝑗3 )2 . (4)
will output several suggestions of each drone’s trajectory. Agents follow
the simplified assumptions: The distance between an agent and other objects influences the
magnitude of the attraction or repulsion force it has towards that
• The effects of buoyancy, lift, drag, and gravity are of secondary particular object; Distant objects should have a weaker influence than
importance and may be neglected. near objects. Therefore, an exponentially decaying function is used to
• The agents may propel themselves in any direction in 3D space. calculate the interaction vector between an agent and an object:
• The agents may be idealized as point masses.
𝒏̂ 𝑖→𝑗 = (𝑤𝑎𝑡𝑡 𝑒−𝑐𝑎𝑡𝑡 𝑑𝑖𝑗 − 𝑤𝑟𝑒𝑝 𝑒−𝑐𝑟𝑒𝑝 𝑑𝑖𝑗 )𝒏𝑖→𝑗 . (5)
• The agents know the locations of all targets, obstacles, and other ⏟⏞⏞⏞⏞⏟⏞⏞⏞⏞⏟ ⏟⏞⏞⏞⏞⏟⏞⏞⏞⏞⏟
agents. 𝑎𝑡𝑡𝑟𝑎𝑐𝑡𝑖𝑜𝑛 𝑟𝑒𝑝𝑢𝑙𝑠𝑖𝑜𝑛

This framework is modeled in a fixed Cartesian basis in 𝒆1 , 𝒆2 , and where 𝑤𝑎𝑡𝑡 and 𝑐𝑎𝑡𝑡 are the weight and exponential decay coefficients of
𝒆3 where the position 𝒓, velocity 𝒗 and acceleration 𝒂 of a drone are the attraction term, and 𝑤𝑟𝑒𝑝 and 𝑐𝑟𝑒𝑝 are the weight and exponential
described as: decay coefficients of the repulsion term, respectively. 𝒏̂ 𝑖→𝑗 is calculated
for each type of object within the system, therefore agents and sensors
𝒓 = 𝑟1 𝒆1 + 𝑟2 𝒆2 + 𝑟3 𝒆3 ,
will have their own values associated with 𝑤𝑎𝑡𝑡 , 𝑐𝑎𝑡𝑡 , 𝑤𝑟𝑒𝑝 and 𝑐𝑟𝑒𝑝 . The
direction 𝒏𝑖→𝑗 is the unit normal vector in the direction of the object 𝑗
𝒗 = 𝒓̇ = 𝑟̇ 1 𝒆1 + 𝑟̇ 2 𝒆2 + 𝑟̇ 3 𝒆3 , (2)
relative to agent 𝑖 and is given by:
𝒂 = 𝒓̈ = 𝑟̈1 𝒆1 + 𝑟̈2 𝒆2 + 𝑟̈3 𝒆3 𝑨𝑗 − 𝒓𝑖
𝒏𝑖→𝑗 = . (6)
respectively. A schematic of this framework is shown in Fig. 3A. ‖𝑨𝑗 − 𝒓𝑖 ‖
The only force, 𝐹 , imposed on each agent, 𝑖, is the agent’s propul-
Generally, a larger attraction term relative to the repulsion term
sion, which is assumed to be of constant magnitude. Hence each agent’s
causes net propulsion towards sensor objects. The total interaction
equation of motion is described using Newton’s second law:
vector between agent 𝑖 and all objects of a particular type is the sum
𝑚𝑖 𝒂𝑖 = 𝑭 𝑝,𝑖 = 𝐹 𝒏⋆
𝑖 . (3) of all their interaction vectors, such as those shown in Fig. 3B. For

4
P. Goodrich et al. Computers and Electronics in Agriculture 205 (2023) 107591

If an agent is within a ‘crash radius’ of another agent, both will

be removed from the domain.
5. ITERATE: If there are unscanned sensors in the domain, iterate
through the algorithm.

The magnitude and range of the interactive vectors between the

agents and the targets were varied to optimize the flight paths, and
the results were simulated. In this simulation, the number of mapped
sensors were maximized while flight time and the number of collisions
were minimized. Written as a cost function:
𝑁𝑠𝑒𝑛𝑠𝑜𝑟𝑠 − 𝑁𝑚𝑎𝑝𝑝𝑒𝑑 𝑡 𝑁
𝛱 = 𝑊1 + 𝑊2 𝑎𝑐𝑡𝑢𝑎𝑙 + 𝑊3 𝑐𝑟𝑎𝑠ℎ𝑒𝑑 (9)
𝑁𝑠𝑒𝑛𝑠𝑜𝑟𝑠 𝑡𝑚𝑎𝑥 𝑁𝑑𝑟𝑜𝑛𝑒𝑠
where 𝑁𝑠𝑒𝑛𝑠𝑜𝑟𝑠 is the total number of sensors in the field, 𝑁𝑚𝑎𝑝𝑝𝑒𝑑 is the
number of sensors that were mapped by the drones, 𝑡𝑎𝑐𝑡𝑢𝑎𝑙 is the time it
takes for the drones to map the sensors, 𝑡𝑚𝑎𝑥 is the maximum allowed
time for the mapping to take place, 𝑁𝑑𝑟𝑜𝑛𝑒𝑠 is the number of drones in
the swarm, and 𝑁𝑐𝑟𝑎𝑠ℎ𝑒𝑑 is the number of drones that crashed into one
another during the simulation. 𝑊1 , 𝑊2 , and 𝑊3 are the user-defined
weights given to each cost term.
A genetic algorithm was employed to minimize the cost function,
using the magnitude and range of the interaction vectors as the design
strings.

3. Computational model

3.1. Digital expression of an agricultural field

The two-dimensional problem for an agricultural field in a ge-

ographic coordinate system was encoded. First, the domain of the
problem is expressed as 𝐷, ̄ a 𝑁𝑑𝑝 × 2 vector of values where 𝑁𝑑𝑝 is
the number of pixels within the domain of consideration and the two
Fig. 3. (A) Schematic of a single-agent model in a Cartesian coordinate system. columns are the longitudinal and latitudinal coordinates, respectively.
(B) Schematic of the interaction vectors in a two-agent, three-objective system. The Similarly, the agricultural field is expressed as 𝐹̄ , a 𝑁𝑓 𝑝 × 2 vector of
instantaneous direction of drone acceleration will be in the net direction of the values where 𝑁𝑓 𝑝 is the number of pixels within the bounds of the field
attraction and repulsion vectors acting on the drone, which changes in magnitude and
direction as the drones move within the domain.
and the two columns are the longitudinal and latitudinal coordinates,
respectively.

instance, the total interaction vector given by all sensors on agent 𝑖 3.2. Optimized sensor placement
is:
𝑁𝑠𝑒𝑛𝑠𝑜𝑟𝑠
∑ Sensor optimization throughout a field is achieved using the SGR
𝑵 𝑠𝑖 = 𝒏̂ 𝑠𝑖→𝑗 (7) algorithm. The algorithm is as follows:
𝑗=1

where 𝑁𝑠𝑒𝑛𝑠𝑜𝑟𝑠 is the total number of sensors in the system. A similar 1. INITIALIZE: For a given vector of field pixels, 𝐹̄ , select a random
̄ a 𝑁𝑠𝑒𝑛𝑠𝑜𝑟𝑠 × 2 vector of sensor
pixel within it and append it to 𝑆,
calculation is used for finding the interaction vector influenced by
all other agents, 𝑵 𝑎𝑖 . A weighted sum of the total interaction forces coordinates, where 𝑁𝑠𝑒𝑛𝑠𝑜𝑟𝑠 is the number of sensors in the field.
of sensors and agents is normalized to give the final direction of 2. GENERATE: Select a random pixel, 𝑃 = (𝑝𝑥 , 𝑝𝑦 ), within the field
propulsion, i.e, 𝐹̄
3. TEST: For all 𝑆𝑖 , where 𝑆𝑖 = (𝑠𝑥 , 𝑠𝑦 ), if ‖𝑃 − 𝑆𝑖 ‖ ≥ 𝑅ℎ𝑣 , where
𝑊𝑠 𝑵 𝑠𝑖 + 𝑊𝑎 𝑵 𝑎𝑖
𝒏⋆
𝑖 = . (8) 𝑅ℎ𝑣 is the half-variogram range of the sensor, then append 𝑃 to
‖𝑊𝑠 𝑵 𝑠𝑖 + 𝑊𝑎 𝑵 𝑎𝑖 ‖
𝑆̄
Using this framework, the following algorithm determines the flight 4. SCORE: Calculate fitness as the ratio of 𝑁𝛤 to 𝑁𝑓 𝑝 , where 𝑁𝛤
paths: is the number of pixels within distance 𝑅 from any sensor pixel
𝑁
𝑆𝑖 : 𝛱(𝛬) = 𝑁 𝛤 .
1. INITIALIZE: Load in the sensor location data within the sim- 𝑓𝑝

ulated domain and position agents along one edge of the do- 5. ITERATE: If 𝛱 ≤ 𝑇 𝑂𝐿, loop to Step 2.
main. Assign weight and exponential decay coefficients for each
The algorithm’s speed can be further increased by incorporating a
object’s attraction and repulsion terms in Eq. (5).
spatial hash. Spatial hashing is further discussed in the Supplementary
2. OBSERVE: Each agent observes it’s surroundings to calculate
Material.
the attraction and repulsion terms and the optimal direction of
propulsion as described in Eqs. (3)–(8).
3. STEP: Each agent will apply its thrust and accelerate toward the 3.3. Flight path mapping
optimal direction.
4. TEST: If an agent is within the transmission range of a sensor, The algorithm for optimizing the system parameters for flight paths
the sensor is considered scanned and removed from the domain. for each drone is as follows:

5
P. Goodrich et al. Computers and Electronics in Agriculture 205 (2023) 107591

1. POPULATION GENERATION: For a given number of drones, 5. Results and discussion

𝑁𝑑𝑟𝑜𝑛𝑒𝑠 , randomly generate a population of 𝐺 genetic strings, 𝛬𝑖 ,
(i = 1,2, 3, . . . , G): The placement of sensors determined by the SGR algorithm is shown
def def
in Fig. 4A–D. In Fig. 4A, the placement of sensors in a circular field
𝛬𝑖 = {𝜆𝑖1 , 𝜆𝑖2 , 𝜆𝑖3 , 𝜆𝑖4 , … , 𝜆𝑖𝐺 } = {𝜆1 , 𝜆2 , 𝜆3 , … , 𝜆𝐺 }𝑖 with a 400 m radius – which is a common length for a center-pivot
irrigation arm (Evans, 2001) – was determined. Fig. 4B shows the opti-
def mal placement of sensors for a rectangular farm field with dimensions
𝛬𝑖 = {𝑤𝑎𝑡𝑡 , 𝑤𝑟𝑒𝑝 , 𝑐𝑎𝑡𝑡 , 𝑐𝑟𝑒𝑝 }𝑖
corresponding to the average small-family U.S. farm (MacDonald et al.,
2. PERFORMANCE EVALUATION: Compute fitness of each 2013). As of 2017, small family farms make up 89% of farms in the
string, 𝛱(𝛬𝑖 ), (i=1, . . . , G): U.S. (Hellerstein and Vilorio, 2019). Fig. 4C demonstrates that the SGR
algorithm could adapt and generate sensor placement for an arbitrary
def 𝑁𝑠𝑒𝑛𝑠𝑜𝑟𝑠 − 𝑁𝑚𝑎𝑝𝑝𝑒𝑑 𝑡𝑎𝑐𝑡𝑢𝑎𝑙 𝑁
𝛱(𝛬𝑖 ) = 𝑊1 + 𝑊2 + 𝑊3 𝑐𝑟𝑎𝑠ℎ𝑒𝑑 field shape defined by several boundary points. In Fig. 4D, the SGR
𝑁𝑠𝑒𝑛𝑠𝑜𝑟𝑠 𝑡𝑚𝑎𝑥 𝑁𝑑𝑟𝑜𝑛𝑒𝑠 algorithm distributed sensors in a field generated from a digital image.
The sensors were efficiently placed such that they cover all of the pixels
3. RANK: Rank each string based on their cost output 𝛱, where
of the image while ignoring the islands of non-field pixels, such as
Rank 1 is the best performing design string that produced the
those in the ‘a’ and ‘l’. In real-world applications, this image could
lowest cost and Rank G the worst performing string:
be captured from satellite imagery, such as those available on Google
𝛱(𝛬𝑖 , 𝑖 = 1, … , 𝐺) Earth, ArcGIS, or other publicly available data sets.
𝛱(𝛬1 ) ≤ 𝛱(𝛬2 ) ≤ ⋯ ≤ 𝛱(𝛬𝐺 ) A hexagonal-grid sensor distribution scheme is shown in Fig. 4E–
H. Hexagonal packing is the most efficient packing method known to
4. MATE: Mate design strings to produce offspring: date, and circles can be packed in two dimensions
√ to completely cover
an area when staggered by a distance of 𝑅 3 in the 𝑥-direction and 3𝑅 2
def
𝛬𝑖 = 𝛷(1) 𝛬𝑖 + (1 − 𝛷(1) )𝛬𝑖+1 in the 𝑦-direction.
The field coverage for varying numbers of sensors for the SGR and
where 0 ≤ 𝛷 ≤ 1 hexagonal-packing schemes are shown in Fig. 4I–L. The SGR algorithm
5. GENE ELIMINATION: Eliminate poorly performing genetic outperformed the efficient hexagonal-grid distribution scheme when
strings, keep top parents and generated offspring less than 95% of the field is covered for the circular and rectangular
6. POPULATION REGENERATION: Repeat the process with the field types. In the case of total field coverage, however, the hexagonal-
new gene pool and new random genetic strings packing method could cover the entire field with fewer sensors. This
is because the SGR algorithm prioritizes maximizing field coverage at
This process is repeated until the performance of a genetic string, each step, making it superior to other distribution schemes when the
𝛱(𝛬𝑖 ), falls below the tolerance limit, indicating that the cost function domain is sufficiently spacious. However, if the field shape is relatively
has been minimized. The minimization of the cost function 𝛱 is guar- simple and total coverage is needed, then the hexagonal-packing ap-
anteed to be monotone with increasing generations if the parent strings proach outperforms SGR because the entire field area can be covered
are retained, i.e., 𝛱(𝛬𝑜𝑝𝑡,𝐼 ) ≥ 𝛱(𝛬𝑜𝑝𝑡,𝐼+1 ), where 𝛬𝑜𝑝𝑡,𝐼+1 and 𝛬𝑜𝑝𝑡,𝐼 are with fewer sensors. Distributing sensors this way could have other
the best genetic strings from generations 𝐼 + 1 and 𝐼, respectively. If benefits during field operations. Sensor distribution and maintenance
one does not retain the parents in the algorithm above, it is possible would be simpler than the SGR placement approach, which requires
that inferior-performing offspring may replace superior parents. Thus, a calibrated GPS and field map for the sensors’ initial placement
top parents were kept for the subsequent generation. and subsequent maintenance events. As the complexity of the field
becomes high, however, the SGR algorithm becomes preferential to the
hexagonal-packing method. The number of sensors required to cover
4. Simulation
the user-defined boundary type field shown in Fig. 4C and G was the
same. For the field generated from an image in Fig. 4D and H, the
Digital expressions of four types of agricultural fields were gener- SGR algorithm outperformed the hexagonal packing scheme for any
ated to show the flexibility and range of the SGR sensor placement number of sensors, as shown in Fig. 4L, and required only 52 sensors to
method. First, a 50-ha circular field with a half-mile (400 m) radius completely cover the field compared to 59 for the hexagonal packing
was generated. Second, a rectangular-shaped 93-ha field was generated. method.
Third, a 22-ha field with straight and curved boundaries was gener- To validate that UAVs operating with optimized flight paths are
ated. Finally, a field from an image of pixels was generated. For all more energy efficient than a traditional sweeping method or a UGV, the
simulations, the pixel size was set to one square meter. paths for all three were simulated for a 400 m center pivot irrigation
Next, for each field type, optimal sensor placement was found using field with sensors distributed using the SGR algorithm, shown in Fig. 5.
the SGR algorithm with 40 m as the half-variogram range, which is a Fig. 5A shows the path for a UGV, which is constrained to traversing
conservative value for soil nitrate (Longchamps and Khosla, 2017; Cao through the concentric crop rows to avoid damaging the growing crop.
Row spacing varies, though it is typically around 0.5–0.75 m for corn
et al., 2012).
and generally less for smaller row crops. Since this is less than the
The optimized locations of soil sensors in a center-pivot irrigation
transmission radius of the sensors, the UGV does not have to traverse
field were then inserted into the agent-based model to determine
every row of the field. Instead, the effective row spacing was set to 4 m
flight paths for UAV drone swarms of varying sizes. The simulation
to minimize the number of concentric sweeps performed by the UGV,
parameters are shown in Table 1. In this simulation, the transmission resulting in a 59.6 km path. If the sensors were paired with a stronger
radius of the sensors was assumed to be 2 m, corresponding to the transceiver – such as a BLE or WiFi module – than the UGV would
transmission radius of an RFID tag, which has been shown recently in be able to skip more rows with each concentric sweep. However, the
other agricultural monitoring applications (Deng et al., 2020; Hardin cost of such modules is inhibitive, and an economic analysis motivating
et al., 2022). the use of RFID is presented in the Supplementary Materials. Fig. 5B
To test the hypothesis that a swarm of drones would outperform shows the flight path for a UAV scanning the sensors using a traditional
a single drone, the simulation was executed for swarm sizes of 1–8 sweeping method. Because of the short transmission radius, the UAV
drones. has to make tight turns to reach all of the sensors, making for a 63.9 km

6
P. Goodrich et al. Computers and Electronics in Agriculture 205 (2023) 107591

Table 1
Flight path simulation parameters.
Symbol Type Units Value Description
𝑅𝑓 𝑖𝑒𝑙𝑑 Scalar m 400 Radius of the center-pivot irrigation field
𝑅ℎ𝑣 Scalar m 40 Half-variogram range
𝑅𝑡𝑟𝑎𝑛𝑠𝑚𝑖𝑠𝑠𝑖𝑜𝑛 Scalar m 2 Sensor transmission radius
𝑁𝑑𝑟𝑜𝑛𝑒𝑠 Scalar None 1–8 Number of drones in the simulation
𝑊𝑎 Scalar None [1, 10] Search bounds of interaction weight for agents
𝑊𝑠 Scalar None [1, 10] Search bounds of interaction weight for sensors
𝑤−𝑎𝑡𝑡 , 𝑤+𝑎𝑡𝑡 Scalar None [0, 1] Search bounds for attraction term weight coefficient
− +
𝑐𝑎𝑡𝑡 , 𝑐𝑎𝑡𝑡 Scalar None [0, 1] Search bounds for attraction term decay coefficient
𝑤−𝑟𝑒𝑝 , 𝑤+𝑟𝑒𝑝 Scalar None [0, 1] Search bounds for repulsion term weight coefficient
− +
𝑐𝑟𝑒𝑝 , 𝑐𝑟𝑒𝑝 , Scalar None [0, 1] Search bounds for repulsion term decay coefficient
𝐹 Scalar N 100 Drone magnitude of propulsion force
𝑚 Scalar kg 5 Drone mass
𝑣𝑚𝑎𝑥 Scalar m/s 10 Drone max velocity
𝑅𝑐𝑟𝑎𝑠ℎ Scalar m 2 Crash radius
𝑊1 Scalar None 0.6 Cost function weight associated with mapping sensors
𝑊2 Scalar None 0.1 Cost function weight associated with mapping speed
𝑊3 Scalar None 0.3 Cost function weight associated with crashed drones
𝑇 𝑂𝐿 Scalar None 0.05 Cost tolerance for GA

Fig. 4. Sensor placement using the (A–D) SGR algorithm and (E–H) hexagonal grid distributions using a 40 m half-variogram range as the effective sensing radius for various
field types. The orange circular markers indicate the sensor coordinates, and the black ring around the sensor indicates the half-variogram range that the sensor covers. (A–D)
Sensor placement for a circular field, rectangular field, user-defined boundary, and fields generated from an image using the SGR algorithm. (E–H) Sensor placement for a circular
field, rectangular field, user-defined boundary, and fields generated from an image using a hexagonal grid spacing. (I–L) Plots of the effective field coverage for a given number of
sensors corresponding to the field types above. The black curve is the field coverage for a distribution using the SGR algorithm, while the red line is for the hexagonal distribution.

flight path, which is even longer than the UGV. Finally, the optimized calculated from the path lengths and the average power of the vehicles
flight path for a single UAV is shown in Fig. 5C. When the agent- while neglecting the impact of start-stop behavior and turning. The
based optimization model is applied, the total path length is reduced to total energy consumed to scan all of the sensors in the field was 165
11.4 km. While the agent-based method is more energy efficient than kWh for UGVs, 0.32 kWh for UAVs performing a traditional sweep, and
the traditional sweeping method, the comparison is more difficult to 0.06 kWh for UAVs with optimized flight paths.
make against the case of UGVs, which consume power at a different To determine the ideal number of drones to use for sampling
rate than UAVs. The energy consumption of the three scenarios was data from the distributed sensors, flight paths for swarms of n = 1–8

7
P. Goodrich et al. Computers and Electronics in Agriculture 205 (2023) 107591

Fig. 5. Comparison of different data sampling strategies in a 400 m radius center-pivot irrigation field. (A) Manned or unmanned ground vehicles are restricted to traversing crop
rows to prevent damaging the crops, which are radial in a center-pivot irrigation field. (B) A UAV performing a traditional sweep scan must execute tight sweeps to scan all the
sensors. (C) A UAV with an optimized flight path has a significantly shorter path than the other sampling options making it the most energy efficient.

Fig. 6. Flight paths of drones scanning sensors in a 400 m radius center-pivot irrigation field for swarms of (A) 1, (B) 2, (C) 3, (D) 4, (E) 5, (F) 6, (G) 7, and (H) 8 agents,
respectively.

drones were simulated. The optimized flight paths for the varying-sized
swarms are shown in Fig. 6, and the total flight path length and flight
time are shown in Fig. 7. Fig. 6A shows the flight path of a single agent.
Fig. 6B and C show the flight paths for two and three agent swarms,
respectively. In both cases, the agents frequently crossed each other’s
paths. As swarm size increases, such as for four agent swarms shown
in Fig. 6D, the agents crossed paths less, and each agent subdivided
the field into its own sections to scan the sensors within it, which is
generally more efficient and less prone to accidental crashes in the
event of interference or GPS malfunction. For swarms with five or
more sensors, as shown in Fig. 6E–H, more overlapping occurred as
agents began to compete over the same sensors. These results may
also depend on each drone’s initial starting position. The agents in this
example were placed at the edge and linearly spaced along the width
of the domain. However, one may choose to have the agents start along
the circumference of the field, somewhere within the field, or from
Fig. 7. Split-axis plot showing the total distance traveled (left) and the flight time to
a single point. A staggered start time would also vary the flight path read all sensors in a 400 m radius center-pivot irrigation field (right) for swarms of n
recommendations. = 1–8 agents. Each bar in the stacked bar plot shows the distance traveled for each
The benefits of employing a multi-agent swarm of drones are high- agent.
lighted in Fig. 7 where a different color indicates each agent’s total
energy consumed on the stacked bar chart. As the number of agents in-
creases, the time required to map all the sensors decreases. Meanwhile, path lengths decreased while the total flight path remained relatively
as the number of drones in the swarm increased, their individual flight consistent for all simulated swarm sizes. The distance traveled can be

8
P. Goodrich et al. Computers and Electronics in Agriculture 205 (2023) 107591

The flight paths for a four-agent drone swarm in a 400 m radius

center-pivot irrigation field are shown in Fig. 9. Starting from the edge
of the simulated domain, the agents traversed through the center-pivot
irrigation field and sampled the distributed sensors, as seen in Fig. 9A
and B. Initially, the drones seek out the nearest sensors as because
they have the strongest interaction force. However, as the flight paths
continued to develop (as in Fig. 9C–F) the number of unscanned sensors
dwindled, and the relative attraction strength of more distant sensors
grew. This continued until the last sensor was scanned, as shown in
Fig. 9G, and the drones returned to their initial positions.
Because the flight paths are optimized using a genetic algorithm
and the initialization is random, the outcome could theoretically be
different every time the simulation is executed, even when the same
parameters are used. Table 2, for example, shows the optimal param-
eters for the top three performing design strings when the four-drone
𝑎𝑔𝑒𝑛𝑡
condition was simulated. For parameters such as 𝑐𝑟𝑒𝑝 , which is the
exponential decay coefficient for the repulsion term emitted by the
agent, the ‘optimal value’ is practically the same for each design string.
Fig. 8. Box-whisker plot showing the variance of energy consumption per sensor 𝑠𝑒𝑛𝑠𝑜𝑟 , the value varies by
scanned. Low variance indicates the amount of energy spent per target is evenly
Meanwhile, for other parameters, such as 𝑐𝑎𝑡𝑡
distributed across all agents. A larger box size corresponds to a larger variance. about one-third of the search bounds across the design strings. While
When five or more agents are simulated, there is a significant imbalance in average it may seem like this indicates that the output is unstable, when the
energy expenditure per sensor scanned across the drones. This was primarily due to flight paths of the top-performing design strings are simulated, as
a phenomenon where multiple agents ‘race’ to scan the same sensor and develop a
shown in Fig. 10, it becomes evident that only minor differences exist
‘leader–follower’ behavior, rendering ‘followers’ completely ineffective.
between the resulting flight paths. This is because genetic algorithms
excel at solving highly multivariate problems quickly. Furthermore, the
similarities between the top three flight paths indicate the robustness
correlated with the amount of fuel or battery power a drone consumes,
of the solution.
meaning that a swarm of drones uses a comparable amount of energy
as a single drone, but performs the scan much faster.
To further analyze the trade-offs in swarm size, the drones’ energy 6. Conclusions
expenditure per sensor scanned as a function of swarm size was exam-
ined, as shown in Fig. 8. As the number of agents increased, the total For any agricultural field geometry, the proposed SGR algorithm
average energy consumption per sensor scanned did not significantly maximizes the coverage of an agricultural field with fewer sensors than
change. This agrees with Fig. 7, which shows that the total flight path a grid-based distribution in most cases. The proposed multi-agent flight
to scan all of the sensors in the field also did not significantly change path mapping can effectively and efficiently generate flight paths for
when the number of drones was increased. However, the variance of variable numbers of drones to scan all sensors. Thus, a crop grower
the energy consumption per sensor scanned increased as the number of could use the algorithms described in this paper to determine the opti-
drones increased. When inspecting the simulations, it was found that mal placement of soil sensors inside their own field, the ideal number
this was primarily caused by a scenario where several agents would of drones to sample data from the sensors, and their flight paths. This
‘race’ to scan the same sensor. Then, once that sensor was scanned framework can be adopted for variable-rate irrigation applications to
and removed from the domain, the same agents would often continue maximize crop yield and minimize environmental pollution.
to race one another for subsequent sensors. This ‘leader–follower’ be- The aforementioned framework was used to determine the optimal
havior is wasteful as the agents consume energy, thus creating more placement of sensors for a 400 m radius center-pivot irrigation field
conflict between agents instead of collaboration. Four was the optimal and simulated the flight paths of drone swarms of varying sizes. Four
number of drones for this scenario because it was the fastest scan that drones provided the best balance between energy efficiency and flight
did not demonstrate ‘leader–follower’ behavior. time for the simulated scenario. Using multiple drones also significantly

Fig. 9. Sequences of flight paths in a 400 m radius center-pivot irrigation field for a four-agent swarm. (A) Initial configuration of the agents and sensors. (B–G) Development of
each agent’s flight path throughout simulated time.

9
P. Goodrich et al. Computers and Electronics in Agriculture 205 (2023) 107591

Table 2
Optimal parameters for the top three performing design strings for flight paths of four drones.
𝑊𝑠 𝑊𝑎 𝑤𝑠𝑒𝑛𝑠𝑜𝑟
𝑎𝑡𝑡
𝑤𝑠𝑒𝑛𝑠𝑜𝑟
𝑟𝑒𝑝
𝑠𝑒𝑛𝑠𝑜𝑟
𝑐𝑎𝑡𝑡 𝑠𝑒𝑛𝑠𝑜𝑟
𝑐𝑟𝑒𝑝 𝑤𝑎𝑔𝑒𝑛𝑡
𝑎𝑡𝑡 𝑤𝑎𝑔𝑒𝑛𝑡
𝑟𝑒𝑝
𝑎𝑔𝑒𝑛𝑡
𝑐𝑎𝑡𝑡 𝑎𝑔𝑒𝑛𝑡
𝑐𝑟𝑒𝑝
2.38897 6.60861 0.38679 0.34980 0.41784 0.73018 0.28894 0.48590 0.69733 0.32812
2.38633 3.69574 0.36072 0.32646 0.35214 0.78264 0.29032 0.48498 0.72134 0.33702
2.42084 6.98402 0.39193 0.54399 0.69210 0.67569 0.27516 0.48624 0.68220 0.32287

Fig. 10. Flight paths for the top three performing design strings for a four-agent swarm in a 400 m radius center-pivot irrigation field.

decreased the amount of the total time it takes to complete the sensor Acknowledgments
scanning task within the field compared to a single drone.
The drone flight path mapping shown here does not account for The drone photography of the farm field shown in Fig. 1D is
topological features or weather conditions. Further extensions of this from the publicly available data set from senseFly. This and simi-
framework include adding variations in terrain height and additional lar datasets can be accessed at https://www.sensefly.com/education/
outside forces such as wind drag synced with weather data, to obtain datasets/. This work was partially supported by the National Science
a more accurate framework for digital twin capabilities. Foundation Graduate Research Fellowships, USA under Grant No. DGE
Agricultural drone flight path mapping applications are not limited 1752814, and the Advanced Research Projects Agency - Energy, USA
to reading sensors. Recently, drones in precision agriculture have taken award DE-AR0001013. This work was also partially supported by AFRI
a more physical role, such as aerial application of fluids, solids, and Competitive, USA Grant no. 2020-67021-32855/project accession no.
biological control agents. In 2015, the Federal Aviation Administration 1024262 from the USDA National Institute of Food and Agriculture,
approved the first crop-dusting drone capable of spraying pesticides USA. This grant is being administered through AIFS: the AI Institute
with tanks weighing more than 55 lbs. This approval allowed the rapid for Next Generation Food Systems. https://aifs.ucdavis.edu.
deployment of drones to spray soil amendments and seeds to precise
areas indicated by the soil-embedded sensors, reducing the amount of Appendix A. Supplementary data
irrigation runoff and seeds dispensed in unwanted areas. The flight path
model could also be modified for tasks such as orchard harvesting, Supplementary material related to this article can be found online
where the drone is programmed to collect fruits or nuts from trees and at https://doi.org/10.1016/j.compag.2022.107591.
deposit them into boxes.
References

CRediT authorship contribution statement Ahrens, T., Lobell, D., Ortiz-Monasterio, J., Li, Y., Matson, P., 2010. Narrowing the
agronomic yield gap with improved nitrogen use efficiency: a modeling approach.
Payton Goodrich: Generated the code and figures for the ‘‘digital Ecol. Appl. 20 (1), 91–100.
Akbari, Y., Almaadeed, N., Al-maadeed, S., Elharrouss, O., 2021. Applications, databases
expression of an agricultural field’’ and ‘‘optimized sensor placement’’ and open computer vision research from drone videos and images: a survey. Artif.
sections, Generated the figures for the ‘‘Results and Discussion’’ section Intell. Rev. 54 (5), 3887–3938.
and wrote the manuscript. Omar Betancourt: Generated the code and Akbarzadeh, V., Lévesque, J.-C., Gagné, C., Parizeau, M., 2014. Efficient sensor
figures for the ‘‘flight path mapping’’ sections, Generated the figures for placement optimization using gradient descent and probabilistic coverage. Sensors
14 (8), 15525–15552.
the ‘‘Results and Discussion’’ section and wrote the manuscript.
Anthony, P., Malzer, G., Zhang, M., Sparrow, S., 2012. Soil nitrogen and phosphorus
behavior in a long-term fertilization experiment. Agron. J. 104 (5), 1223–1237.
Declaration of competing interest Bachmaier, M., Backes, M., 2008. Variogram or semivariogram? Understanding the
variances in a variogram. Precis. Agric. 9 (3), 173–175.
Bertalan, L., Holb, I., Pataki, A., Négyesi, G., Szabó, G., Kupásné Szalóki, A., Szabó, S.,
The authors declare that they have no known competing finan- 2022. UAV-based multispectral and thermal cameras to predict soil water content –
cial interests or personal relationships that could have appeared to A machine learning approach. Comput. Electron. Agric. 200, 107262. http://dx.doi.
org/10.1016/j.compag.2022.107262, URL https://www.sciencedirect.com/science/
influence the work reported in this paper.
article/pii/S0168169922005750.
Cao, Q., Cui, Z., Chen, X., Khosla, R., Dao, T.H., Miao, Y., 2012. Quantifying spatial
Data availability variability of indigenous nitrogen supply for precision nitrogen management in
small scale farming. Precis. Agric. 13 (1), 45–61.
Cassman, K.G., Dobermann, A., Walters, D.T., 2002. Agroecosystems, nitrogen-use
Data will be made available on request. efficiency, and nitrogen management. AMBIO: J. Human Environ. 31 (2), 132–140.

10
P. Goodrich et al. Computers and Electronics in Agriculture 205 (2023) 107591

Chlingaryan, A., Sukkarieh, S., Whelan, B., 2018. Machine learning approaches for Longchamps, L., Khosla, R., 2017. Precision maize cultivation techniques. In: Achieving
crop yield prediction and nitrogen status estimation in precision agriculture: A Sustainable Cultivation of Maize Volume 2. Burleigh Dodds Science Publishing, pp.
review. Comput. Electron. Agric. (ISSN: 0168-1699) 151, 61–69. http://dx.doi. 127–157.
org/10.1016/j.compag.2018.05.012, URL https://www.sciencedirect.com/science/ Longchamps, L., Khosla, R., Reich, R., Gui, D., 2015. Spatial and temporal variability
article/pii/S0168169917314710. of soil water content in leveled fields. Soil Sci. Am. J. 79 (5), 1446–1454.
Comis, D., 2011. Growing crops and saving water in the west. Agric. Res. 59 (7), 12–14. MacDonald, J.M., Korb, P., Hoppe, R.A., 2013. Farm Size and the Organization of US
Dalin, C., Hanasaki, N., Qiu, H., Mauzerall, D.L., Rodriguez-Iturbe, I., 2014. Water Crop Farming. Tech. rep.
resources transfers through Chinese interprovincial and foreign food trade. Proc. Metropolis, N., Rosenbluth, A.W., Rosenbluth, M.N., Teller, A.H., Teller, E., 1953.
Natl. Acad. Sci. 111 (27), 9774–9779. Equation of state calculations by fast computing machines. J. Chem. Phys. 21 (6),
Deng, F., Zuo, P., Wen, K., Wu, X., 2020. Novel soil environment monitoring 1087–1092.
system based on RFID sensor and LoRa. Comput. Electron. Agric. 169, 105169. Mueller, M.W., Hehn, M., D’Andrea, R., 2015. A computationally efficient motion
http://dx.doi.org/10.1016/j.compag.2019.105169, URL https://www.sciencedirect. primitive for quadrocopter trajectory generation. IEEE Trans. Robot. 31 (6),
com/science/article/pii/S0168169919319222. 1294–1310.
Diacono, M., Rubino, P., Montemurro, F., 2013. Precision nitrogen management of Patrício, D.I., Rieder, R., 2018. Computer vision and artificial intelligence in precision
wheat. A review. Agron. Sustain. Dev. 33 (1), 219–241. agriculture for grain crops: A systematic review. Comput. Electron. Agric. 153,
Dieter, C.A., 2018. Water Availability and Use Science Program: Estimated Use of Water 69–81.
in the United States in 2015. Geological Survey. Pereira, F.S., de Lima, J., Freitas, R., Dos Reis, A., Amaral, L., Figueiredo, G.,
Donev, A., Torquato, S., Stillinger, F.H., 2005a. Neighbor list collision-driven molecular Lamparelli, R., Magalhães, P., 2022. Nitrogen variability assessment of pasture
dynamics simulation for nonspherical hard particles. I. Algorithmic details. J. fields under an integrated crop-livestock system using UAV, PlanetScope, and
Comput. Phys. 202 (2), 737–764. Sentinel-2 data. Comput. Electron. Agric. 193, 106645. http://dx.doi.org/10.
Donev, A., Torquato, S., Stillinger, F.H., 2005b. Neighbor list collision-driven molecular 1016/j.compag.2021.106645, URL https://www.sciencedirect.com/science/article/
dynamics simulation for nonspherical hard particles.: Ii. applications to ellipses and pii/S0168169921006621.
ellipsoids. J. Comput. Phys. 202 (2), 765–793. Ritchie, H., Roser, M., 2020. 𝐶𝑂2 and Greenhouse Gas Emissions. Our World in Data,
Evans, R.G., 2001. Center pivot irrigation. In: Agricultural Systems Research Unit, Vol. https://ourworldindata.org/co2-and-other-greenhouse-gas-emissions.
1500. North Plain Agric. Res. Laboratory. USDN-Agric. Res. Service. Robert, P.C., Rust, R.H., Larson, W.E., 1995. Proceedings of site-specific management
Evans, R.G., LaRue, J., Stone, K.C., King, B.A., 2013. Adoption of site-specific variable for agricultural systems: second international conference, march 27-30, 1994 thun-
rate sprinkler irrigation systems. Irrigation Sci. 31 (4), 871–887. derbird hotel, 2201 east 78th st., minneapolis, MN: conducted by the department
Finger, R., Swinton, S.M., El Benni, N., Walter, A., 2019. Precision farming at the nexus of soil science and minnesota extension service, university of minnesota. In: 2nd
of agricultural production and the environment. International Conference on Site-Specific Management for Agricultural Systems.,
Gellings, C.W., Parmenter, K.E., 2016. Energy efficiency in fertilizer production and use. Minneapolis, Minn.(USA), 1994. American Society of Agronomy.
In: Gellings, C.W. (Ed.), Efficient Use and Conservation of Energy. In: Encyclopedia Robertson, G.P., Klingensmith, K.M., Klug, M.J., Paul, E.A., Crum, J.R., Ellis, B.G., 1997.
of Life Support Systems, pp. 123–136. Soil resources, microbial activity, and primary production across an agricultural
Gupta, M., Khosla, R., 2012. Precision nitrogen management and global nitrogen ecosystem. Ecol. Appl. 7 (1), 158–170.
use efficiency. In: Proceedings of the 11th International Conference on Precision Sela, S., Woodbury, P., Van Es, H., 2018. Dynamic model-based N management reduces
Agriculture. Indianapolis, Ind, USA. surplus nitrogen and improves the environmental performance of corn production.
Hardin, R.G., Barnes, E.M., Delhom, C.D., Wanjura, J.D., Ward, J.K., 2022. Internet of Environ. Res. Lett. 13 (5), 054010.
things: Cotton harvesting and processing. Comput. Electron. Agric. 202, 107294. Shahandeh, H., Wright, A., Hons, F., Lascano, R., 2005. Spatial and temporal variation
http://dx.doi.org/10.1016/j.compag.2022.107294, URL https://www.sciencedirect. of soil nitrogen parameters related to soil texture and corn yield. Agron. J. 97 (3),
com/science/article/pii/S0168169922006068. 772–782.
Hedley, C.B., Yule, I.J., 2009. Soil water status mapping and two variable-rate irrigation Sørensen, L.Y., Jacobsen, L.T., Hansen, J.P., 2017. Low cost and flexible UAV
scenarios. Precis. Agric. 10 (4), 342–355. deployment of sensors. Sensors 17 (1), 154.
Hellerstein, D., Vilorio, D., 2019. Agricultural Resources and Environmental Indicators, Steinhaus, H., 1999. Mathematical Snapshots. Courier Corporation.
2019. Tech. rep. SU, S.L., Singh, D., Baghini, M.S., 2014. A critical review of soil moisture measurement.
Hifi, M., M’hallah, R., 2009. A literature review on circle and sphere packing problems: Measurement 54, 92–105.
Models and methodologies. Adv. Oper. Res. 2009. Sun, X., Fang, W., Gao, C., Fu, L., Majeed, Y., Liu, X., Gao, F., Yang, R., Li, R., 2022.
Hoy, R., 2021. Nebraska summary S1150A for Fendt 1050. Institute of Agriculture and Remote estimation of grafted apple tree trunk diameter in modern orchard with
Natural Resources University of Nebraska–Lincoln, URL https://digitalcommons.unl. RGB and point cloud based on SOLOv2. Comput. Electron. Agric. 199, 107209.
edu/tractormuseumlit/3481/. http://dx.doi.org/10.1016/j.compag.2022.107209, URL https://www.sciencedirect.
Jourdan, D.B., de Weck, O.L., 2004. Layout optimization for a wireless sensor network com/science/article/pii/S0168169922005245.
using a multi-objective genetic algorithm. In: 2004 IEEE 59th Vehicular Technology TOMBE, R., 2020. Computer vision for smart farming and sustainable agriculture. In:
Conference. VTC 2004-Spring (IEEE Cat. No. 04CH37514), Vol. 5. IEEE, pp. 2020 IST-Africa Conference (IST-Africa). IEEE, pp. 1–8.
2466–2470. Tosato, P., Facinelli, D., Prada, M., Gemma, L., Rossi, M., Brunelli, D., 2019. An
Kanter, D.R., Bell, A.R., McDermid, S.S., 2019. Precision agriculture for Small- autonomous swarm of drones for industrial gas sensing applications. In: 2019 IEEE
holder Nitrogen Management. One Earth 1 (3), 281–284. http://dx.doi.org/10. 20th International Symposium on" a World of Wireless, Mobile and Multimedia
1016/j.oneear.2019.10.015, URL https://www.sciencedirect.com/science/article/ Networks"(WoWMoM). IEEE, pp. 1–6.
pii/S2590332219301290. Widom, B., 1966. Random sequential addition of hard spheres to a volume. J. Chem.
Kerry, R., Oliver, M., Frogbrook, Z., 2010. Sampling in precision agriculture. In: Phys. 44 (10), 3888–3894.
Geostatistical Applications for Precision Agriculture. Springer, pp. 35–63. Zohdi, T., 2019. The Game of Drones: rapid agent-based machine-learning models for
Kumar, G.N., Bangi, M., 2018. An extension to winding number and point-in-polygon multi-UAV path planning. Comput. Mech. 65, 217–228.
algorithm. IFAC-PapersOnLine 51 (1), 548–553. Zohdi, T., 2021a. A digital-twin and machine-learning framework for the design of
Ladha, J.K., Pathak, H., Krupnik, T.J., Six, J., van Kessel, C., 2005. Efficiency of multiobjective agrophotovoltaic solar farms. Comput. Mech. 68 (2), 357–370.
fertilizer nitrogen in cereal production: retrospects and prospects. Adv. Agron. 87, Zohdi, T., 2021b. A digital-twin and machine-learning framework for ventilation system
85–156. optimization for capturing infectious disease respiratory emissions. Arch. Comput.
Lo, T.H., Heeren, D.M., Mateos, L., Luck, J.D., Martin, D.L., Miller, K.A., Barker, J.B., Methods Eng. 28 (6), 4317–4329.
Shaver, T.M., 2017. Field characterization of field capacity and root zone available Zohdi, T., 2022. A digital-twin and machine-learning framework for precise heat and
water capacity for variable rate irrigation. Appl. Eng. Agric. 33 (4), 559–572. energy management of data-centers. Comput. Mech. 1–16.

11